diff --git "a/vectors/vector_index/docstore.json" "b/vectors/vector_index/docstore.json"
new file mode 100644--- /dev/null
+++ "b/vectors/vector_index/docstore.json"
@@ -0,0 +1 @@
+{"docstore/data": {"3431a00b-b2c3-4aa3-9597-d5c79a81f9c5": {"__data__": {"id_": "3431a00b-b2c3-4aa3-9597-d5c79a81f9c5", "embedding": null, "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6d6fd0f4-5a6d-47a0-b304-59417f3cadb6", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ca4e49fe86b8da8044f500a128481893279d656bbf3dece51a9aa5073341cc0d", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "135d7103-a9ad-43cd-b4e3-619fe02e901f", "node_type": "1", "metadata": {}, "hash": "b1e0c8d8656573e5e76dba4f8ffd3e7a1f33f5ce4b2cbe96dabd46c44ad82bd4", "class_name": "RelatedNodeInfo"}}, "text": "Provided proper attribution is provided, Google hereby grants permission to\nreproduce the tables and figures in this paper solely for use in journalistic or\nscholarly works.\nAttention Is All You Need\nAshish Vaswani\u2217\nGoogle Brain\navaswani@google.comNoam Shazeer\u2217\nGoogle Brain\nnoam@google.comNiki Parmar\u2217\nGoogle Research\nnikip@google.comJakob Uszkoreit\u2217\nGoogle Research\nusz@google.com\nLlion Jones\u2217\nGoogle Research\nllion@google.comAidan N. Gomez\u2217 \u2020\nUniversity of Toronto\naidan@cs.toronto.edu\u0141ukasz Kaiser\u2217\nGoogle Brain\nlukaszkaiser@google.com\nIllia Polosukhin\u2217 \u2021\nillia.polosukhin@gmail.com\nAbstract\nThe dominant sequence transduction models are based on complex recurrent or\nconvolutional neural networks that include an encoder and a decoder. The best\nperforming models also connect the encoder and decoder through an attention\nmechanism. We propose a new simple network architecture, the Transformer,\nbased solely on attention mechanisms, dispensing with recurrence and convolutions\nentirely. Experiments on two machine translation tasks show these models to\nbe superior in quality while being more parallelizable and requiring significantly\nless time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-\nto-German translation task, improving over the existing best results, including\nensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task,\nour model establishes a new single-model state-of-the-art BLEU score of 41.8 after\ntraining for 3.5 days on eight GPUs, a small fraction of the training costs of the\nbest models from the literature. We show that the Transformer generalizes well to\nother tasks by applying it successfully to English constituency parsing both with\nlarge and limited training data.\n\u2217Equal contribution. ", "start_char_idx": 0, "end_char_idx": 1758, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "135d7103-a9ad-43cd-b4e3-619fe02e901f": {"__data__": {"id_": "135d7103-a9ad-43cd-b4e3-619fe02e901f", "embedding": null, "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6d6fd0f4-5a6d-47a0-b304-59417f3cadb6", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ca4e49fe86b8da8044f500a128481893279d656bbf3dece51a9aa5073341cc0d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "3431a00b-b2c3-4aa3-9597-d5c79a81f9c5", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e5fad06d02d35abea195f627bac19ad00bca2e1036c8f8810f3b4ee531ca14c7", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "7fa60ea2-8f3d-403a-91ab-ca2409c0a30f", "node_type": "1", "metadata": {}, "hash": "14e592bbcb900896dff54bddde810f729c4223c17986db11fccf3512bd7d2cea", "class_name": "RelatedNodeInfo"}}, "text": "Listing order is random. Jakob proposed replacing RNNs with self-attention and started\nthe effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and\nhas been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head\nattention and the parameter-free position representation and became the other person involved in nearly every\ndetail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and\ntensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and\nefficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and\nimplementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating\nour research.\n\u2020Work performed while at Google Brain.\n", "start_char_idx": 1758, "end_char_idx": 2682, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "7fa60ea2-8f3d-403a-91ab-ca2409c0a30f": {"__data__": {"id_": "7fa60ea2-8f3d-403a-91ab-ca2409c0a30f", "embedding": null, "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6d6fd0f4-5a6d-47a0-b304-59417f3cadb6", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ca4e49fe86b8da8044f500a128481893279d656bbf3dece51a9aa5073341cc0d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "135d7103-a9ad-43cd-b4e3-619fe02e901f", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ff03134021863208563480fbef84ac81faeb0087f4be4f89a65e000553e9e08e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "3ea43674-021d-4a53-bf50-8b7ba146a699", "node_type": "1", "metadata": {}, "hash": "b76f5cbda6706f433a99870c215cd0693b5fc8fc5b1c5374bbcab748006a73f1", "class_name": "RelatedNodeInfo"}}, "text": "\u2021Work performed while at Google Research.\n31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.arXiv:1706.03762v7  [cs.CL]  2 Aug 2023", "start_char_idx": 2682, "end_char_idx": 2853, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "3ea43674-021d-4a53-bf50-8b7ba146a699": {"__data__": {"id_": "3ea43674-021d-4a53-bf50-8b7ba146a699", "embedding": null, "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "49b3abf2a4ad66160f00161741a1b57d04eb861fd7af8e1d38f6ccd91900278d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "7fa60ea2-8f3d-403a-91ab-ca2409c0a30f", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c622a92b6bc9e23792425b98d86dd29232b783cd3e5c099cd76330a5790730db", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "40ceaf90-165a-4ba9-91cb-e441a1b06a33", "node_type": "1", "metadata": {}, "hash": "39db70cbd822139819b32529ad79ccc5614a06bd7a95830e5a5911e54ac039f2", "class_name": "RelatedNodeInfo"}}, "text": "1 Introduction\nRecurrent neural networks, long short-term memory [ 13] and gated recurrent [ 7] neural networks\nin particular, have been firmly established as state of the art approaches in sequence modeling and\ntransduction problems such as language modeling and machine translation [ 35,2,5]. Numerous\nefforts have since continued to push the boundaries of recurrent language models and encoder-decoder\narchitectures [38, 24, 15].\nRecurrent models typically factor computation along the symbol positions of the input and output\nsequences. Aligning the positions to steps in computation time, they generate a sequence of hidden\nstates ht, as a function of the previous hidden state ht\u22121and the input for position t. This inherently\nsequential nature precludes parallelization within training examples, which becomes critical at longer\nsequence lengths, as memory constraints limit batching across examples. Recent work has achieved\nsignificant improvements in computational efficiency through factorization tricks [ 21] and conditional\ncomputation [ 32], while also improving model performance in case of the latter. The fundamental\nconstraint of sequential computation, however, remains.\nAttention mechanisms have become an integral part of compelling sequence modeling and transduc-\ntion models in various tasks, allowing modeling of dependencies without regard to their distance in\nthe input or output sequences [ 2,19]. In all but a few cases [ 27], however, such attention mechanisms\nare used in conjunction with a recurrent network.\nIn this work we propose the Transformer, a model architecture eschewing recurrence and instead\nrelying entirely on an attention mechanism to draw global dependencies between input and output.\nThe Transformer allows for significantly more parallelization and can reach a new state of the art in\ntranslation quality after being trained for as little as twelve hours on eight P100 GPUs.\n2 Background\nThe goal of reducing sequential computation also forms the foundation of the Extended Neural GPU\n[16], ByteNet [ 18] and ConvS2S [ 9], all of which use convolutional neural networks as basic building\nblock, computing hidden representations in parallel for all input and output positions. In these models,\nthe number of operations required to relate signals from two arbitrary input or output positions grows\nin the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. ", "start_char_idx": 0, "end_char_idx": 2434, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "40ceaf90-165a-4ba9-91cb-e441a1b06a33": {"__data__": {"id_": "40ceaf90-165a-4ba9-91cb-e441a1b06a33", "embedding": null, "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "49b3abf2a4ad66160f00161741a1b57d04eb861fd7af8e1d38f6ccd91900278d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "3ea43674-021d-4a53-bf50-8b7ba146a699", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "2461968fbfdb377e396e4fcd7628cdff5e9e23bfaf2994f850fef85b09157907", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "4e00cc5e-34a8-4d28-96cd-521a8613082e", "node_type": "1", "metadata": {}, "hash": "3d5478f5af12a0260f05b9e221b5b229d887414b80201489bee91af10d773eee", "class_name": "RelatedNodeInfo"}}, "text": "This makes\nit more difficult to learn dependencies between distant positions [ 12]. ", "start_char_idx": 2434, "end_char_idx": 2518, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "4e00cc5e-34a8-4d28-96cd-521a8613082e": {"__data__": {"id_": "4e00cc5e-34a8-4d28-96cd-521a8613082e", "embedding": null, "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "49b3abf2a4ad66160f00161741a1b57d04eb861fd7af8e1d38f6ccd91900278d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "40ceaf90-165a-4ba9-91cb-e441a1b06a33", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7de9d0ce7f1f67f6746d3dd0073dcfce83e23a2d541737cfb5a5e72c7056933e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "72dfb8e6-9f71-4211-a72d-5ec77bd4b2fe", "node_type": "1", "metadata": {}, "hash": "316ec4d731eae17aca2a5ad669da2323a90e5dda811f87fadc8ba763f44ba0e9", "class_name": "RelatedNodeInfo"}}, "text": "In the Transformer this is\nreduced to a constant number of operations, albeit at the cost of reduced effective resolution due\nto averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as\ndescribed in section 3.2.\nSelf-attention, sometimes called intra-attention is an attention mechanism relating different positions\nof a single sequence in order to compute a representation of the sequence. Self-attention has been\nused successfully in a variety of tasks including reading comprehension, abstractive summarization,\ntextual entailment and learning task-independent sentence representations [4, 27, 28, 22].\nEnd-to-end memory networks are based on a recurrent attention mechanism instead of sequence-\naligned recurrence and have been shown to perform well on simple-language question answering and\nlanguage modeling tasks [34].\nTo the best of our knowledge, however, the Transformer is the first transduction model relying\nentirely on self-attention to compute representations of its input and output without using sequence-\naligned RNNs or convolution. In the following sections, we will describe the Transformer, motivate\nself-attention and discuss its advantages over models such as [17, 18] and [9].\n3 Model Architecture\nMost competitive neural sequence transduction models have an encoder-decoder structure [ 5,2,35].\nHere, the encoder maps an input sequence of symbol representations (x1, ..., x n)to a sequence\nof continuous representations z= (z1, ..., z n). ", "start_char_idx": 2518, "end_char_idx": 4019, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "72dfb8e6-9f71-4211-a72d-5ec77bd4b2fe": {"__data__": {"id_": "72dfb8e6-9f71-4211-a72d-5ec77bd4b2fe", "embedding": null, "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "49b3abf2a4ad66160f00161741a1b57d04eb861fd7af8e1d38f6ccd91900278d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "4e00cc5e-34a8-4d28-96cd-521a8613082e", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ee9d9ba0f9c4dee86ba934c329cbd5c2561d61f5da47753803c99c6e53e586b4", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "bb6aa732-aa8c-440f-86bb-de07bd602ea1", "node_type": "1", "metadata": {}, "hash": "ce1a545e91283af06ea36a8e0aa3ac75d56f32791ea80eb61dedb56180ed81d3", "class_name": "RelatedNodeInfo"}}, "text": "Given z, the decoder then generates an output\nsequence (y1, ..., y m)of symbols one element at a time. At each step the model is auto-regressive\n[10], consuming the previously generated symbols as additional input when generating the next.\n2", "start_char_idx": 4019, "end_char_idx": 4260, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "bb6aa732-aa8c-440f-86bb-de07bd602ea1": {"__data__": {"id_": "bb6aa732-aa8c-440f-86bb-de07bd602ea1", "embedding": null, "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "10071188-42e8-4929-98dd-642ee4131169", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "43ae829564edc9b5ed17c359ae6edcaa482d4371370a9e7539468351fc9292ff", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "72dfb8e6-9f71-4211-a72d-5ec77bd4b2fe", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8f3f71fcecd60992e39150337420806a2a1342fa9d895e95c3c1d6319d861979", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "536fb5fb-63e5-4170-80b6-739234a9149f", "node_type": "1", "metadata": {}, "hash": "adced3075ee07d80dd5a064953b3e6345338a7788e2ed5df75314c4c19302de4", "class_name": "RelatedNodeInfo"}}, "text": "Figure 1: The Transformer - model architecture.\nThe Transformer follows this overall architecture using stacked self-attention and point-wise, fully\nconnected layers for both the encoder and decoder, shown in the left and right halves of Figure 1,\nrespectively.\n3.1 Encoder and Decoder Stacks\nEncoder: The encoder is composed of a stack of N= 6 identical layers. Each layer has two\nsub-layers. ", "start_char_idx": 0, "end_char_idx": 394, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "536fb5fb-63e5-4170-80b6-739234a9149f": {"__data__": {"id_": "536fb5fb-63e5-4170-80b6-739234a9149f", "embedding": null, "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "10071188-42e8-4929-98dd-642ee4131169", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "43ae829564edc9b5ed17c359ae6edcaa482d4371370a9e7539468351fc9292ff", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "bb6aa732-aa8c-440f-86bb-de07bd602ea1", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ec844fda5ac16742fcb4be7508ed70c02440962916828d5582d53939c797fd65", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "fce102f2-322f-4e80-8130-96ff394dd00d", "node_type": "1", "metadata": {}, "hash": "bc306d0d8071010116dc3b5eb47c080c7a19eb91730296df10a30301cbbf6289", "class_name": "RelatedNodeInfo"}}, "text": "The first is a multi-head self-attention mechanism, and the second is a simple, position-\nwise fully connected feed-forward network. We employ a residual connection [ 11] around each of\nthe two sub-layers, followed by layer normalization [ 1]. That is, the output of each sub-layer is\nLayerNorm( x+ Sublayer( x)), where Sublayer( x)is the function implemented by the sub-layer\nitself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding\nlayers, produce outputs of dimension dmodel = 512 .\nDecoder: The decoder is also composed of a stack of N= 6identical layers. In addition to the two\nsub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head\nattention over the output of the encoder stack. Similar to the encoder, we employ residual connections\naround each of the sub-layers, followed by layer normalization. We also modify the self-attention\nsub-layer in the decoder stack to prevent positions from attending to subsequent positions. This\nmasking, combined with fact that the output embeddings are offset by one position, ensures that the\npredictions for position ican depend only on the known outputs at positions less than i.\n3.2 Attention\nAn attention function can be described as mapping a query and a set of key-value pairs to an output,\nwhere the query, keys, values, and output are all vectors. ", "start_char_idx": 394, "end_char_idx": 1784, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "fce102f2-322f-4e80-8130-96ff394dd00d": {"__data__": {"id_": "fce102f2-322f-4e80-8130-96ff394dd00d", "embedding": null, "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "10071188-42e8-4929-98dd-642ee4131169", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "43ae829564edc9b5ed17c359ae6edcaa482d4371370a9e7539468351fc9292ff", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "536fb5fb-63e5-4170-80b6-739234a9149f", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "2e731d358bfcfda1c02958e2f95c5290561829d2a32e82ac769b3dda94ffaf86", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "637186e0-6b70-47af-9395-a5f23c9922c7", "node_type": "1", "metadata": {}, "hash": "a6b53683e0a2b96be0935e39d0574511e5955296c192a44644e7b2becae4d936", "class_name": "RelatedNodeInfo"}}, "text": "The output is computed as a weighted sum\n3", "start_char_idx": 1784, "end_char_idx": 1826, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "637186e0-6b70-47af-9395-a5f23c9922c7": {"__data__": {"id_": "637186e0-6b70-47af-9395-a5f23c9922c7", "embedding": null, "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "de79a486-751a-414d-b21e-078ab065ed82", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6d7eafcd20aa872de73d70cafddd26517476b3a8947a3be7c4b572699123d453", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "fce102f2-322f-4e80-8130-96ff394dd00d", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1ae967901d22e2c27573427e0a60288b2a0d6e75e5a20f39487c84d8f6faa5db", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "422f9866-8f04-463f-8d50-eab9a4886128", "node_type": "1", "metadata": {}, "hash": "d3b4b9f322caa182e14014ca418b27538a6229a94e5bc613b44004202ff454fc", "class_name": "RelatedNodeInfo"}}, "text": "Scaled Dot-Product Attention\n Multi-Head Attention\nFigure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several\nattention layers running in parallel.\nof the values, where the weight assigned to each value is computed by a compatibility function of the\nquery with the corresponding key.\n3.2.1 Scaled Dot-Product Attention\nWe call our particular attention \"Scaled Dot-Product Attention\" (Figure 2). The input consists of\nqueries and keys of dimension dk, and values of dimension dv. We compute the dot products of the\nquery with all keys, divide each by\u221adk, and apply a softmax function to obtain the weights on the\nvalues.\nIn practice, we compute the attention function on a set of queries simultaneously, packed together\ninto a matrix Q. The keys and values are also packed together into matrices KandV. We compute\nthe matrix of outputs as:\nAttention( Q, K, V ) = softmax(QKT\n\u221adk)V (1)\nThe two most commonly used attention functions are additive attention [ 2], and dot-product (multi-\nplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor\nof1\u221adk. ", "start_char_idx": 0, "end_char_idx": 1134, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "422f9866-8f04-463f-8d50-eab9a4886128": {"__data__": {"id_": "422f9866-8f04-463f-8d50-eab9a4886128", "embedding": null, "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "de79a486-751a-414d-b21e-078ab065ed82", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6d7eafcd20aa872de73d70cafddd26517476b3a8947a3be7c4b572699123d453", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "637186e0-6b70-47af-9395-a5f23c9922c7", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "25d2f70ef16cdd181bcf39f32b2fa03c410a8ea0d81c8a4e7350866c7d4e3f0a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "8687e9ed-235b-4278-a496-45395cddaf37", "node_type": "1", "metadata": {}, "hash": "7e96d9c89ff8d9bc7bbaeb8f6a828ef8a4aa67f18210fd16213798b1a60972f9", "class_name": "RelatedNodeInfo"}}, "text": "Additive attention computes the compatibility function using a feed-forward network with\na single hidden layer. While the two are similar in theoretical complexity, dot-product attention is\nmuch faster and more space-efficient in practice, since it can be implemented using highly optimized\nmatrix multiplication code.\nWhile for small values of dkthe two mechanisms perform similarly, additive attention outperforms\ndot product attention without scaling for larger values of dk[3]. We suspect that for large values of\ndk, the dot products grow large in magnitude, pushing the softmax function into regions where it has\nextremely small gradients4. To counteract this effect, we scale the dot products by1\u221adk.\n3.2.2 Multi-Head Attention\nInstead of performing a single attention function with dmodel-dimensional keys, values and queries,\nwe found it beneficial to linearly project the queries, keys and values htimes with different, learned\nlinear projections to dk,dkanddvdimensions, respectively. On each of these projected versions of\nqueries, keys and values we then perform the attention function in parallel, yielding dv-dimensional\n4To illustrate why the dot products get large, assume that the components of qandkare independent random\nvariables with mean 0and variance 1. Then their dot product, q\u00b7k=Pdk\ni=1qiki, has mean 0and variance dk.\n", "start_char_idx": 1134, "end_char_idx": 2480, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "8687e9ed-235b-4278-a496-45395cddaf37": {"__data__": {"id_": "8687e9ed-235b-4278-a496-45395cddaf37", "embedding": null, "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "de79a486-751a-414d-b21e-078ab065ed82", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6d7eafcd20aa872de73d70cafddd26517476b3a8947a3be7c4b572699123d453", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "422f9866-8f04-463f-8d50-eab9a4886128", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6d8d07e874fb92b09011ea27bcd9ef831e9cabce9634f81d8862de406b9ab9ed", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "6524bd60-c48e-4539-b3e1-78075b31bac0", "node_type": "1", "metadata": {}, "hash": "21ff6be24feb44f64834efc2217a726e625aaed495dbe2203cb6c7c174bfdea1", "class_name": "RelatedNodeInfo"}}, "text": "4", "start_char_idx": 1778, "end_char_idx": 1779, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "6524bd60-c48e-4539-b3e1-78075b31bac0": {"__data__": {"id_": "6524bd60-c48e-4539-b3e1-78075b31bac0", "embedding": null, "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "1006cf3f-32fc-4c2e-b31a-f519be43fd1b", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c814a65337d0e89ec59292c4b363d8a7cecf44e9b5155e17e5b71decb6e52b0c", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "8687e9ed-235b-4278-a496-45395cddaf37", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "26988312de96dce95a0094d5e86a130f91e5f69e4d6a720c5a581491f23ed818", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "412d7574-4fb3-4985-8745-7494188dd8e6", "node_type": "1", "metadata": {}, "hash": "c33d6baeaa634ded1a452943784e6ac9010ae35bb2cd35e9795434c6b2172e08", "class_name": "RelatedNodeInfo"}}, "text": "output values. These are concatenated and once again projected, resulting in the final values, as\ndepicted in Figure 2.\nMulti-head attention allows the model to jointly attend to information from different representation\nsubspaces at different positions. With a single attention head, averaging inhibits this.\nMultiHead( Q, K, V ) = Concat(head 1, ...,head h)WO\nwhere head i= Attention( QWQ\ni, KWK\ni, V WV\ni)\nWhere the projections are parameter matrices WQ\ni\u2208Rdmodel\u00d7dk,WK\ni\u2208Rdmodel\u00d7dk,WV\ni\u2208Rdmodel\u00d7dv\nandWO\u2208Rhdv\u00d7dmodel.\nIn this work we employ h= 8 parallel attention layers, or heads. ", "start_char_idx": 0, "end_char_idx": 586, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "412d7574-4fb3-4985-8745-7494188dd8e6": {"__data__": {"id_": "412d7574-4fb3-4985-8745-7494188dd8e6", "embedding": null, "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "1006cf3f-32fc-4c2e-b31a-f519be43fd1b", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c814a65337d0e89ec59292c4b363d8a7cecf44e9b5155e17e5b71decb6e52b0c", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "6524bd60-c48e-4539-b3e1-78075b31bac0", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "23f9f422947d199bca0707b35985de63fb61e9b64a4622365a90a1cb8d4f5cbe", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "ff341210-7431-4b1b-a4d2-0b144b35d985", "node_type": "1", "metadata": {}, "hash": "ae61ebb23bddaa99d04fdda048329f2ab389e9b5a4e9d1f98e168d0eb697a090", "class_name": "RelatedNodeInfo"}}, "text": "For each of these we use\ndk=dv=dmodel/h= 64 . Due to the reduced dimension of each head, the total computational cost\nis similar to that of single-head attention with full dimensionality.\n3.2.3 Applications of Attention in our Model\nThe Transformer uses multi-head attention in three different ways:\n\u2022In \"encoder-decoder attention\" layers, the queries come from the previous decoder layer,\nand the memory keys and values come from the output of the encoder. This allows every\nposition in the decoder to attend over all positions in the input sequence. ", "start_char_idx": 586, "end_char_idx": 1138, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "ff341210-7431-4b1b-a4d2-0b144b35d985": {"__data__": {"id_": "ff341210-7431-4b1b-a4d2-0b144b35d985", "embedding": null, "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "1006cf3f-32fc-4c2e-b31a-f519be43fd1b", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c814a65337d0e89ec59292c4b363d8a7cecf44e9b5155e17e5b71decb6e52b0c", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "412d7574-4fb3-4985-8745-7494188dd8e6", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6f7ef2d54429da3b4bfe849e20e66927288c8f200a426a593a4ce40fced9bb5e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "abdc2077-fd4a-4184-bd93-07b2088061de", "node_type": "1", "metadata": {}, "hash": "7cbbe3cedcd85f13e8877f623107b685e449ab6affec17847118f51cab2d9384", "class_name": "RelatedNodeInfo"}}, "text": "This mimics the\ntypical encoder-decoder attention mechanisms in sequence-to-sequence models such as\n[38, 2, 9].\n\u2022The encoder contains self-attention layers. In a self-attention layer all of the keys, values\nand queries come from the same place, in this case, the output of the previous layer in the\nencoder. Each position in the encoder can attend to all positions in the previous layer of the\nencoder.\n\u2022Similarly, self-attention layers in the decoder allow each position in the decoder to attend to\nall positions in the decoder up to and including that position. We need to prevent leftward\ninformation flow in the decoder to preserve the auto-regressive property. We implement this\ninside of scaled dot-product attention by masking out (setting to \u2212\u221e) all values in the input\nof the softmax which correspond to illegal connections. See Figure 2.\n3.3 Position-wise Feed-Forward Networks\nIn addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully\nconnected feed-forward network, which is applied to each position separately and identically. This\nconsists of two linear transformations with a ReLU activation in between.\nFFN( x) = max(0 , xW 1+b1)W2+b2 (2)\nWhile the linear transformations are the same across different positions, they use different parameters\nfrom layer to layer. Another way of describing this is as two convolutions with kernel size 1.\n", "start_char_idx": 1138, "end_char_idx": 2534, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "abdc2077-fd4a-4184-bd93-07b2088061de": {"__data__": {"id_": "abdc2077-fd4a-4184-bd93-07b2088061de", "embedding": null, "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "1006cf3f-32fc-4c2e-b31a-f519be43fd1b", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c814a65337d0e89ec59292c4b363d8a7cecf44e9b5155e17e5b71decb6e52b0c", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "ff341210-7431-4b1b-a4d2-0b144b35d985", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "9df99719393b4ec72ccf855808242c10151dcb245937fcda88f7e8456ec7e8fb", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "67d23967-8054-4c86-9a3f-213604c7f8d6", "node_type": "1", "metadata": {}, "hash": "d38477a528396f5cc112bcb61b7c9befc0bba3596c07341c247686f85296a7e7", "class_name": "RelatedNodeInfo"}}, "text": "The dimensionality of input and output is dmodel = 512 , and the inner-layer has dimensionality\ndff= 2048 .\n3.4 Embeddings and Softmax\nSimilarly to other sequence transduction models, we use learned embeddings to convert the input\ntokens and output tokens to vectors of dimension dmodel. We also use the usual learned linear transfor-\nmation and softmax function to convert the decoder output to predicted next-token probabilities. In\nour model, we share the same weight matrix between the two embedding layers and the pre-softmax\nlinear transformation, similar to [ 30]. In the embedding layers, we multiply those weights by\u221admodel.\n5", "start_char_idx": 2534, "end_char_idx": 3169, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "67d23967-8054-4c86-9a3f-213604c7f8d6": {"__data__": {"id_": "67d23967-8054-4c86-9a3f-213604c7f8d6", "embedding": null, "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "e95894c0-aec1-4614-b98f-d141b63efd5c", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f618674d2e12b9eac8d947f3c475ce6dadc2b597e3418b93ade5726a62dc0a14", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "abdc2077-fd4a-4184-bd93-07b2088061de", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ed5894d3e4cc2756a80d2170b70d9eb799ba0dd9bd09c0986a226bfedba5444a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "67607d1c-0394-46f4-8a3b-6bb6cff1bfe1", "node_type": "1", "metadata": {}, "hash": "4e63986289fc951fa89b559329ed84e62002d2262af2efe5b4418ffa4a933ccc", "class_name": "RelatedNodeInfo"}}, "text": "Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations\nfor different layer types. nis the sequence length, dis the representation dimension, kis the kernel\nsize of convolutions and rthe size of the neighborhood in restricted self-attention.\nLayer Type Complexity per Layer Sequential Maximum Path Length\nOperations\nSelf-Attention O(n2\u00b7d) O(1) O(1)\nRecurrent O(n\u00b7d2) O(n) O(n)\nConvolutional O(k\u00b7n\u00b7d2) O(1) O(logk(n))\nSelf-Attention (restricted) O(r\u00b7n\u00b7d) O(1) O(n/r)\n3.5 Positional Encoding\nSince our model contains no recurrence and no convolution, in order for the model to make use of the\norder of the sequence, we must inject some information about the relative or absolute position of the\ntokens in the sequence. To this end, we add \"positional encodings\" to the input embeddings at the\nbottoms of the encoder and decoder stacks. ", "start_char_idx": 0, "end_char_idx": 874, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "67607d1c-0394-46f4-8a3b-6bb6cff1bfe1": {"__data__": {"id_": "67607d1c-0394-46f4-8a3b-6bb6cff1bfe1", "embedding": null, "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "e95894c0-aec1-4614-b98f-d141b63efd5c", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f618674d2e12b9eac8d947f3c475ce6dadc2b597e3418b93ade5726a62dc0a14", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "67d23967-8054-4c86-9a3f-213604c7f8d6", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "4149d0733a2077929fc0fc7832725ed2c7850610d31f1fa2e6770bea04f69a4d", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "84010686-fc60-4ee0-b307-f08a0e19cfe6", "node_type": "1", "metadata": {}, "hash": "529205e3a6d8b748eba53794c5d10346f0dc59338979ad554e1a9f3eefb693ee", "class_name": "RelatedNodeInfo"}}, "text": "The positional encodings have the same dimension dmodel\nas the embeddings, so that the two can be summed. There are many choices of positional encodings,\nlearned and fixed [9].\nIn this work, we use sine and cosine functions of different frequencies:\nPE(pos,2i)=sin(pos/100002i/d model)\nPE(pos,2i+1)=cos(pos/100002i/d model)\nwhere posis the position and iis the dimension. That is, each dimension of the positional encoding\ncorresponds to a sinusoid. The wavelengths form a geometric progression from 2\u03c0to10000 \u00b72\u03c0. We\nchose this function because we hypothesized it would allow the model to easily learn to attend by\nrelative positions, since for any fixed offset k,PEpos+kcan be represented as a linear function of\nPEpos.\nWe also experimented with using learned positional embeddings [ 9] instead, and found that the two\nversions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version\nbecause it may allow the model to extrapolate to sequence lengths longer than the ones encountered\nduring training.\n4 Why Self-Attention\nIn this section we compare various aspects of self-attention layers to the recurrent and convolu-\ntional layers commonly used for mapping one variable-length sequence of symbol representations\n(x1, ..., x n)to another sequence of equal length (z1, ..., z n), with xi, zi\u2208Rd, such as a hidden\nlayer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we\nconsider three desiderata.\n", "start_char_idx": 874, "end_char_idx": 2350, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "84010686-fc60-4ee0-b307-f08a0e19cfe6": {"__data__": {"id_": "84010686-fc60-4ee0-b307-f08a0e19cfe6", "embedding": null, "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "e95894c0-aec1-4614-b98f-d141b63efd5c", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f618674d2e12b9eac8d947f3c475ce6dadc2b597e3418b93ade5726a62dc0a14", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "67607d1c-0394-46f4-8a3b-6bb6cff1bfe1", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "087f13d1ed39ef41492b3d272f5c1b21b4b783865a4247e17bc158debda912f5", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "bbfd17cd-9afc-4fae-9279-52965ee289d1", "node_type": "1", "metadata": {}, "hash": "2296b0274cd92b7809a3629bdd9cbfddde0e5225d8eca903f310f8d0ee302d9a", "class_name": "RelatedNodeInfo"}}, "text": "One is the total computational complexity per layer. ", "start_char_idx": 2350, "end_char_idx": 2403, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "bbfd17cd-9afc-4fae-9279-52965ee289d1": {"__data__": {"id_": "bbfd17cd-9afc-4fae-9279-52965ee289d1", "embedding": null, "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "e95894c0-aec1-4614-b98f-d141b63efd5c", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f618674d2e12b9eac8d947f3c475ce6dadc2b597e3418b93ade5726a62dc0a14", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "84010686-fc60-4ee0-b307-f08a0e19cfe6", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ef3f6df3b02956ee87ad5c023687353efda09b51ec3d0682d81ad9e3c0af0e23", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "0c7f8f31-d6ab-4f30-803d-7a0056b15963", "node_type": "1", "metadata": {}, "hash": "c2571f9c9bd39ffc7f852bd73b31b821838adcf8b8eb6c1b2b019f0eb1d62b8e", "class_name": "RelatedNodeInfo"}}, "text": "Another is the amount of computation that can\nbe parallelized, as measured by the minimum number of sequential operations required.\nThe third is the path length between long-range dependencies in the network. Learning long-range\ndependencies is a key challenge in many sequence transduction tasks. One key factor affecting the\nability to learn such dependencies is the length of the paths forward and backward signals have to\ntraverse in the network. The shorter these paths between any combination of positions in the input\nand output sequences, the easier it is to learn long-range dependencies [ 12]. Hence we also compare\nthe maximum path length between any two input and output positions in networks composed of the\ndifferent layer types.\nAs noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially\nexecuted operations, whereas a recurrent layer requires O(n)sequential operations. In terms of\ncomputational complexity, self-attention layers are faster than recurrent layers when the sequence\n6", "start_char_idx": 2403, "end_char_idx": 3448, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "0c7f8f31-d6ab-4f30-803d-7a0056b15963": {"__data__": {"id_": "0c7f8f31-d6ab-4f30-803d-7a0056b15963", "embedding": null, "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ab05337f-6fcc-408d-8685-7654ec1f476d", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3d32a5c8f07d78a73fefaf20f8136d569775b557772bd46a2ce2d90c3365c26b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "bbfd17cd-9afc-4fae-9279-52965ee289d1", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "930cc5f06b3219abf8c66d02b61b945a23f8268250497902ad05ed77a71d848f", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "07227a15-44fc-4950-9f7d-7c1461364365", "node_type": "1", "metadata": {}, "hash": "7b2fa15fa24f3d13a562b21e7da4ac31a1dc40e480c702d86266667f3d7087d5", "class_name": "RelatedNodeInfo"}}, "text": "length nis smaller than the representation dimensionality d, which is most often the case with\nsentence representations used by state-of-the-art models in machine translations, such as word-piece\n[38] and byte-pair [ 31] representations. To improve computational performance for tasks involving\nvery long sequences, self-attention could be restricted to considering only a neighborhood of size rin\nthe input sequence centered around the respective output position. This would increase the maximum\npath length to O(n/r). ", "start_char_idx": 0, "end_char_idx": 520, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "07227a15-44fc-4950-9f7d-7c1461364365": {"__data__": {"id_": "07227a15-44fc-4950-9f7d-7c1461364365", "embedding": null, "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ab05337f-6fcc-408d-8685-7654ec1f476d", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3d32a5c8f07d78a73fefaf20f8136d569775b557772bd46a2ce2d90c3365c26b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "0c7f8f31-d6ab-4f30-803d-7a0056b15963", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "58db0b38b6326b1bf61137c8dbe62d75f8d11c21e85243ec183c9bda8fb05027", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "6c2e74fb-f80a-4074-97a1-2f78727f1fc7", "node_type": "1", "metadata": {}, "hash": "e995007d39659f8ed1b131ab550becf6731d17bd6470483e2afe60328761ef1d", "class_name": "RelatedNodeInfo"}}, "text": "We plan to investigate this approach further in future work.\nA single convolutional layer with kernel width k < n does not connect all pairs of input and output\npositions. Doing so requires a stack of O(n/k)convolutional layers in the case of contiguous kernels,\norO(logk(n))in the case of dilated convolutions [ 18], increasing the length of the longest paths\nbetween any two positions in the network. Convolutional layers are generally more expensive than\nrecurrent layers, by a factor of k. Separable convolutions [ 6], however, decrease the complexity\nconsiderably, to O(k\u00b7n\u00b7d+n\u00b7d2). Even with k=n, however, the complexity of a separable\nconvolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer,\nthe approach we take in our model.\nAs side benefit, self-attention could yield more interpretable models. ", "start_char_idx": 520, "end_char_idx": 1371, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "6c2e74fb-f80a-4074-97a1-2f78727f1fc7": {"__data__": {"id_": "6c2e74fb-f80a-4074-97a1-2f78727f1fc7", "embedding": null, "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ab05337f-6fcc-408d-8685-7654ec1f476d", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3d32a5c8f07d78a73fefaf20f8136d569775b557772bd46a2ce2d90c3365c26b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "07227a15-44fc-4950-9f7d-7c1461364365", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f48d0ff5ec08366539996b772ac04c071acf8af22b72c11e1d8f8931409f4bcc", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "612fbc2e-e4ab-4d6e-b3d6-af6dc250aaa8", "node_type": "1", "metadata": {}, "hash": "08b98bd856e3b568167819c21f23dfe2d27afd29dc5197004a61d3446a50b002", "class_name": "RelatedNodeInfo"}}, "text": "We inspect attention distributions\nfrom our models and present and discuss examples in the appendix. Not only do individual attention\nheads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic\nand semantic structure of the sentences.\n5 Training\nThis section describes the training regime for our models.\n5.1 Training Data and Batching\nWe trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million\nsentence pairs. Sentences were encoded using byte-pair encoding [ 3], which has a shared source-\ntarget vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT\n2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece\nvocabulary [ 38]. Sentence pairs were batched together by approximate sequence length. Each training\nbatch contained a set of sentence pairs containing approximately 25000 source tokens and 25000\ntarget tokens.\n5.2 Hardware and Schedule\nWe trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using\nthe hyperparameters described throughout the paper, each training step took about 0.4 seconds. We\ntrained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the\nbottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps\n(3.5 days).\n5.3 Optimizer\nWe used the Adam optimizer [ 20] with \u03b21= 0.9,\u03b22= 0.98and\u03f5= 10\u22129. We varied the learning\nrate over the course of training, according to the formula:\nlrate =d\u22120.5\nmodel\u00b7min(step_num\u22120.5, step _num\u00b7warmup _steps\u22121.5) (3)\nThis corresponds to increasing the learning rate linearly for the first warmup _steps training steps,\nand decreasing it thereafter proportionally to the inverse square root of the step number. We used\nwarmup _steps = 4000 .\n", "start_char_idx": 1371, "end_char_idx": 3228, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "612fbc2e-e4ab-4d6e-b3d6-af6dc250aaa8": {"__data__": {"id_": "612fbc2e-e4ab-4d6e-b3d6-af6dc250aaa8", "embedding": null, "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ab05337f-6fcc-408d-8685-7654ec1f476d", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3d32a5c8f07d78a73fefaf20f8136d569775b557772bd46a2ce2d90c3365c26b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "6c2e74fb-f80a-4074-97a1-2f78727f1fc7", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "18393b3aa39c791d80bb634cbf5b3e31f34a2417b1d83fc705b84c50c0e7d67b", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "826d10fd-380f-47e8-b970-51cd5c3e9270", "node_type": "1", "metadata": {}, "hash": "2ee52b486177a249fde2286c059b570f769f98a23868517e0103c77d53fb230a", "class_name": "RelatedNodeInfo"}}, "text": "5.4 Regularization\nWe employ three types of regularization during training:\n7", "start_char_idx": 3228, "end_char_idx": 3305, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "826d10fd-380f-47e8-b970-51cd5c3e9270": {"__data__": {"id_": "826d10fd-380f-47e8-b970-51cd5c3e9270", "embedding": null, "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8df67ec6-1e0f-4188-9687-0323556948cb", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "104f868c9ec7bb6f6157b5bbd4bd5478e9edf52dbafbc104103eb9f004b9a858", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "612fbc2e-e4ab-4d6e-b3d6-af6dc250aaa8", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "eb179bdea2450b5906d42af007dd524a30ad1d4232843531eed033060a74a62f", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "b3bfb5b0-4f15-45b3-970b-ea71328a74cf", "node_type": "1", "metadata": {}, "hash": "60789b4d682bfe6a26431e4dbe92bb206917ff4c763c2a94bd513ecdfe8967e9", "class_name": "RelatedNodeInfo"}}, "text": "Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the\nEnglish-to-German and English-to-French newstest2014 tests at a fraction of the training cost.\nModelBLEU Training Cost (FLOPs)\nEN-DE EN-FR EN-DE EN-FR\nByteNet [18] 23.75\nDeep-Att + PosUnk [39] 39.2 1.0\u00b71020\nGNMT + RL [38] 24.6 39.92 2.3\u00b710191.4\u00b71020\nConvS2S [9] 25.16 40.46 9.6\u00b710181.5\u00b71020\nMoE [32] 26.03 40.56 2.0\u00b710191.2\u00b71020\nDeep-Att + PosUnk Ensemble [39] 40.4 8.0\u00b71020\nGNMT + RL Ensemble [38] 26.30 41.16 1.8\u00b710201.1\u00b71021\nConvS2S Ensemble [9] 26.36 41.29 7.7\u00b710191.2\u00b71021\nTransformer (base model) 27.3 38.1 3.3\u00b71018\nTransformer (big) 28.4 41.8 2.3\u00b71019\nResidual Dropout We apply dropout [ 33] to the output of each sub-layer, before it is added to the\nsub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the\npositional encodings in both the encoder and decoder stacks. ", "start_char_idx": 0, "end_char_idx": 917, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "b3bfb5b0-4f15-45b3-970b-ea71328a74cf": {"__data__": {"id_": "b3bfb5b0-4f15-45b3-970b-ea71328a74cf", "embedding": null, "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8df67ec6-1e0f-4188-9687-0323556948cb", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "104f868c9ec7bb6f6157b5bbd4bd5478e9edf52dbafbc104103eb9f004b9a858", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "826d10fd-380f-47e8-b970-51cd5c3e9270", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f8a87f87a229ea433cafaddb9a7b882283a1255e295a8e9f20c489580811c1b8", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "42d035cb-f399-442d-b7da-260217dd8a0c", "node_type": "1", "metadata": {}, "hash": "cc6ddb6cf1dd1fc9fd3b3083c970a010d117eee847efa7092e4894f37d0c123e", "class_name": "RelatedNodeInfo"}}, "text": "For the base model, we use a rate of\nPdrop= 0.1.\nLabel Smoothing During training, we employed label smoothing of value \u03f5ls= 0.1[36]. This\nhurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.\n6 Results\n6.1 Machine Translation\nOn the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big)\nin Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0\nBLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is\nlisted in the bottom line of Table 3. ", "start_char_idx": 917, "end_char_idx": 1515, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "42d035cb-f399-442d-b7da-260217dd8a0c": {"__data__": {"id_": "42d035cb-f399-442d-b7da-260217dd8a0c", "embedding": null, "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8df67ec6-1e0f-4188-9687-0323556948cb", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "104f868c9ec7bb6f6157b5bbd4bd5478e9edf52dbafbc104103eb9f004b9a858", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "b3bfb5b0-4f15-45b3-970b-ea71328a74cf", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "847192d4d9919bf8dc8014956a35a1777b97e1b9ac70ad634785cfbb7705755a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "aef4eca4-aecd-4f06-99ee-3c35d2427597", "node_type": "1", "metadata": {}, "hash": "51d2c03cc287c61e329de8841e02495d7c4f75c096bd196f124b7cdf49004162", "class_name": "RelatedNodeInfo"}}, "text": "Training took 3.5days on 8P100 GPUs. Even our base model\nsurpasses all previously published models and ensembles, at a fraction of the training cost of any of\nthe competitive models.\nOn the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0,\noutperforming all of the previously published single models, at less than 1/4the training cost of the\nprevious state-of-the-art model. The Transformer (big) model trained for English-to-French used\ndropout rate Pdrop= 0.1, instead of 0.3.\nFor the base models, we used a single model obtained by averaging the last 5 checkpoints, which\nwere written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We\nused beam search with a beam size of 4and length penalty \u03b1= 0.6[38]. These hyperparameters\nwere chosen after experimentation on the development set. We set the maximum output length during\ninference to input length + 50, but terminate early when possible [38].\nTable 2 summarizes our results and compares our translation quality and training costs to other model\narchitectures from the literature. We estimate the number of floating point operations used to train a\nmodel by multiplying the training time, the number of GPUs used, and an estimate of the sustained\nsingle-precision floating-point capacity of each GPU5.\n6.2 Model Variations\nTo evaluate the importance of different components of the Transformer, we varied our base model\nin different ways, measuring the change in performance on English-to-German translation on the\n5We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.\n8", "start_char_idx": 1515, "end_char_idx": 3149, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "aef4eca4-aecd-4f06-99ee-3c35d2427597": {"__data__": {"id_": "aef4eca4-aecd-4f06-99ee-3c35d2427597", "embedding": null, "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8e0c2ca4-0215-412a-ba2e-f675e6a861af", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3555b87155aec9aa8a75b706c876df3cf0558b251144393e5e38cc83b87f3d3b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "42d035cb-f399-442d-b7da-260217dd8a0c", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b4f25e74ebd0475b910b654d5a93c49c99526a91afd42bb12bd4b92824d630b6", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "9c91e9f4-91a7-403b-9a4e-6f517810bcad", "node_type": "1", "metadata": {}, "hash": "f3277f5fec6ed95385f3dbd440c02404609c1387a95b86c76a32467d8ee3a95c", "class_name": "RelatedNodeInfo"}}, "text": "Table 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base\nmodel. ", "start_char_idx": 0, "end_char_idx": 111, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "9c91e9f4-91a7-403b-9a4e-6f517810bcad": {"__data__": {"id_": "9c91e9f4-91a7-403b-9a4e-6f517810bcad", "embedding": null, "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8e0c2ca4-0215-412a-ba2e-f675e6a861af", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3555b87155aec9aa8a75b706c876df3cf0558b251144393e5e38cc83b87f3d3b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "aef4eca4-aecd-4f06-99ee-3c35d2427597", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a0f9febd7e187836acd4a8a2f32e8e1c8dcf61714ac8c19e23a90cc151f71418", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "739d85c4-e25b-49ae-a049-96c25851a609", "node_type": "1", "metadata": {}, "hash": "77c42a81387029e301b36817bcbfc5f0b5b9f6f996b7a8449cf67da0914770c2", "class_name": "RelatedNodeInfo"}}, "text": "All metrics are on the English-to-German translation development set, newstest2013. Listed\nperplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to\nper-word perplexities.\nN d model dff h d k dvPdrop \u03f5lstrain PPL BLEU params\nsteps (dev) (dev) \u00d7106\nbase 6 512 2048 8 64 64 0.1 0.1 100K 4.92 25.8 65\n(A)1 512 512 5.29 24.9\n4 128 128 5.00 25.5\n16 32 32 4.91 25.8\n32 16 16 5.01 25.4\n(B)16 5.16 25.1 58\n32 5.01 25.4 60\n(C)2 6.11 23.7 36\n4 5.19 25.3 50\n8 4.88 25.5 80\n256 32 32 5.75 24.5 28\n1024 128 128 4.66 26.0 168\n1024 5.12 25.4 53\n4096 4.75 26.2 90\n(D)0.0 5.77 24.6\n0.2 4.95 25.5\n0.0 4.67 25.3\n0.2 5.47 25.7\n(E) positional embedding instead of sinusoids 4.92 25.7\nbig 6 1024 4096 16 0.3 300K 4.33 26.4 213\ndevelopment set, newstest2013. We used beam search as described in the previous section, but no\ncheckpoint averaging. ", "start_char_idx": 111, "end_char_idx": 975, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "739d85c4-e25b-49ae-a049-96c25851a609": {"__data__": {"id_": "739d85c4-e25b-49ae-a049-96c25851a609", "embedding": null, "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8e0c2ca4-0215-412a-ba2e-f675e6a861af", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3555b87155aec9aa8a75b706c876df3cf0558b251144393e5e38cc83b87f3d3b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "9c91e9f4-91a7-403b-9a4e-6f517810bcad", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fd1f6f3268954ed574af42f2a7adca2cb23a5b323cc99293d58a8a7f2e51f71f", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "d463eec0-8dbb-4558-b9e9-5b3ff0ed9137", "node_type": "1", "metadata": {}, "hash": "0a4812b67002733db4519b06b1d218113d2dfef8b1bc223dd2721ebd5185cade", "class_name": "RelatedNodeInfo"}}, "text": "We present these results in Table 3.\nIn Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions,\nkeeping the amount of computation constant, as described in Section 3.2.2. While single-head\nattention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.\nIn Table 3 rows (B), we observe that reducing the attention key size dkhurts model quality. This\nsuggests that determining compatibility is not easy and that a more sophisticated compatibility\nfunction than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected,\nbigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our\nsinusoidal positional encoding with learned positional embeddings [ 9], and observe nearly identical\nresults to the base model.\n6.3 English Constituency Parsing\nTo evaluate if the Transformer can generalize to other tasks we performed experiments on English\nconstituency parsing. This task presents specific challenges: the output is subject to strong structural\nconstraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence\nmodels have not been able to attain state-of-the-art results in small-data regimes [37].\nWe trained a 4-layer transformer with dmodel = 1024 on the Wall Street Journal (WSJ) portion of the\nPenn Treebank [ 25], about 40K training sentences. We also trained it in a semi-supervised setting,\nusing the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences\n[37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens\nfor the semi-supervised setting.\nWe performed only a small number of experiments to select the dropout, both attention and residual\n(section 5.4), learning rates and beam size on the Section 22 development set, all other parameters\nremained unchanged from the English-to-German base translation model. During inference, we\n9", "start_char_idx": 975, "end_char_idx": 2969, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "d463eec0-8dbb-4558-b9e9-5b3ff0ed9137": {"__data__": {"id_": "d463eec0-8dbb-4558-b9e9-5b3ff0ed9137", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "739d85c4-e25b-49ae-a049-96c25851a609", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ffdd464141d42dff9893063dd9743953ef246c6f3caedc88f3a461ac8d2ada65", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "1f60dd32-a23b-4678-880c-cee36938e787", "node_type": "1", "metadata": {}, "hash": "2c01788355b4081bae742c8ff8e18d40d08d89aa8dcc02f0954bb8e243417195", "class_name": "RelatedNodeInfo"}}, "text": "Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23\nof WSJ)\nParser Training WSJ 23 F1\nVinyals & Kaiser el al. (2014) [37] WSJ only, discriminative 88.3\nPetrov et al. ", "start_char_idx": 0, "end_char_idx": 215, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "1f60dd32-a23b-4678-880c-cee36938e787": {"__data__": {"id_": "1f60dd32-a23b-4678-880c-cee36938e787", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "d463eec0-8dbb-4558-b9e9-5b3ff0ed9137", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "2452e5b3e5ff62d94096818af5a1d3620d61489948275f630fc54745090bf3a3", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "680e94b6-62fb-4536-a7ff-32dc41f99c89", "node_type": "1", "metadata": {}, "hash": "b38cd8e544f44a26f93d4cb275a25febb200b6e4b330287b6e7435b29a0e3823", "class_name": "RelatedNodeInfo"}}, "text": "(2006) [29] WSJ only, discriminative 90.4\nZhu et al. (2013) [40] WSJ only, discriminative 90.4\nDyer et al. (2016) [8] WSJ only, discriminative 91.7\nTransformer (4 layers) WSJ only, discriminative 91.3\nZhu et al. (2013) [40] semi-supervised 91.3\nHuang & Harper (2009) [14] semi-supervised 91.3\nMcClosky et al. (2006) [26] semi-supervised 92.1\nVinyals & Kaiser el al. (2014) [37] semi-supervised 92.1\nTransformer (4 layers) semi-supervised 92.7\nLuong et al. (2015) [23] multi-task 93.0\nDyer et al. (2016) [8] generative 93.3\nincreased the maximum output length to input length + 300. We used a beam size of 21and\u03b1= 0.3\nfor both WSJ only and the semi-supervised setting.\nOur results in Table 4 show that despite the lack of task-specific tuning our model performs sur-\nprisingly well, yielding better results than all previously reported models with the exception of the\nRecurrent Neural Network Grammar [8].\nIn contrast to RNN sequence-to-sequence models [ 37], the Transformer outperforms the Berkeley-\nParser [29] even when training only on the WSJ training set of 40K sentences.\n7 Conclusion\nIn this work, we presented the Transformer, the first sequence transduction model based entirely on\nattention, replacing the recurrent layers most commonly used in encoder-decoder architectures with\nmulti-headed self-attention.\nFor translation tasks, the Transformer can be trained significantly faster than architectures based\non recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014\nEnglish-to-French translation tasks, we achieve a new state of the art. In the former task our best\nmodel outperforms even all previously reported ensembles.\nWe are excited about the future of attention-based models and plan to apply them to other tasks. We\nplan to extend the Transformer to problems involving input and output modalities other than text and\nto investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs\nsuch as images, audio and video. Making generation less sequential is another research goals of ours.\nThe code we used to train and evaluate our models is available at https://github.com/\ntensorflow/tensor2tensor .\nAcknowledgements We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful\ncomments, corrections and inspiration.\n", "start_char_idx": 215, "end_char_idx": 2526, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "680e94b6-62fb-4536-a7ff-32dc41f99c89": {"__data__": {"id_": "680e94b6-62fb-4536-a7ff-32dc41f99c89", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "1f60dd32-a23b-4678-880c-cee36938e787", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ca90a964e379767202d4fd498fdb92350f9e2a87fda59e4392e11f70fd3ae227", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "3a6477bc-b068-4c75-90c1-c406affa9259", "node_type": "1", "metadata": {}, "hash": "40cc5a74c9b96205b43ae910c6c5ca3ea7f192127db83336ff242a9c8cfa00ad", "class_name": "RelatedNodeInfo"}}, "text": "References\n[1]Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint\narXiv:1607.06450 , 2016.\n[2]Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly\nlearning to align and translate. CoRR , abs/1409.0473, 2014.\n", "start_char_idx": 2526, "end_char_idx": 2810, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "3a6477bc-b068-4c75-90c1-c406affa9259": {"__data__": {"id_": "3a6477bc-b068-4c75-90c1-c406affa9259", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "680e94b6-62fb-4536-a7ff-32dc41f99c89", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7ebcbdca9beb0f2b2055033447e5820ee04856ee3b6f8376a7a43470e27568fc", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "51d4b860-054a-4489-b47a-281b1b4f11ad", "node_type": "1", "metadata": {}, "hash": "8c3a33e9ff98dc22faaa89324e411b4fe7680f578975442fffb7fee2e3bc6f4b", "class_name": "RelatedNodeInfo"}}, "text": "[3]Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V . ", "start_char_idx": 2810, "end_char_idx": 2870, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "51d4b860-054a-4489-b47a-281b1b4f11ad": {"__data__": {"id_": "51d4b860-054a-4489-b47a-281b1b4f11ad", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "3a6477bc-b068-4c75-90c1-c406affa9259", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "0be6230d8a0e1ec43cfcd886930a0beb9ed97ac6da6889b42bc879831d7bf862", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "27a20239-3af4-4aac-841c-c53dbde13a14", "node_type": "1", "metadata": {}, "hash": "5b8dbb6c47ab8f1b8f392bbd43ea822198828f36d1e67f37933a787830641e9b", "class_name": "RelatedNodeInfo"}}, "text": "Le. Massive exploration of neural\nmachine translation architectures. CoRR , abs/1703.03906, 2017.\n[4]Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine\nreading. arXiv preprint arXiv:1601.06733 , 2016.\n10", "start_char_idx": 2870, "end_char_idx": 3111, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "27a20239-3af4-4aac-841c-c53dbde13a14": {"__data__": {"id_": "27a20239-3af4-4aac-841c-c53dbde13a14", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "51d4b860-054a-4489-b47a-281b1b4f11ad", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "5e8ee706174d8b81333912e397fce755fc14b64f18580c768109292b8da150a4", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "d5d45e29-a0d7-4bb4-9cc7-c8d088f4c963", "node_type": "1", "metadata": {}, "hash": "e2bad031038c5330dddc478435ba8dcda240aa676cae3d3ce500b99125d0e81c", "class_name": "RelatedNodeInfo"}}, "text": "[5]Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk,\nand Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical\nmachine translation. CoRR , abs/1406.1078, 2014.\n", "start_char_idx": 0, "end_char_idx": 231, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "d5d45e29-a0d7-4bb4-9cc7-c8d088f4c963": {"__data__": {"id_": "d5d45e29-a0d7-4bb4-9cc7-c8d088f4c963", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "27a20239-3af4-4aac-841c-c53dbde13a14", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "990904d016b4432da73521124cde5423e69795eb577199c4f45c7108326be663", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "ad73e473-afa4-41aa-a795-c7a9b1082fa9", "node_type": "1", "metadata": {}, "hash": "75189f3288d37fc892233605f7ce71569e48d667a719fb43102c7b4bdb35f23b", "class_name": "RelatedNodeInfo"}}, "text": "[6]Francois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv\npreprint arXiv:1610.02357 , 2016.\n", "start_char_idx": 231, "end_char_idx": 355, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "ad73e473-afa4-41aa-a795-c7a9b1082fa9": {"__data__": {"id_": "ad73e473-afa4-41aa-a795-c7a9b1082fa9", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "d5d45e29-a0d7-4bb4-9cc7-c8d088f4c963", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "4fced396588f621fccc8cfcd794eeb4a6de05086890061e25ca1653793c73c56", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "de4d13e0-dfe7-4887-a0f6-e087cbcda550", "node_type": "1", "metadata": {}, "hash": "385615682866e774914e0762c469f5f7821153687d1af64ea13912819fdbfcf4", "class_name": "RelatedNodeInfo"}}, "text": "[7]Junyoung Chung, \u00c7aglar G\u00fcl\u00e7ehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation\nof gated recurrent neural networks on sequence modeling. CoRR , abs/1412.3555, 2014.\n[8]Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural\nnetwork grammars. In Proc. of NAACL , 2016.\n[9]Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolu-\ntional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2 , 2017.\n[10] Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint\narXiv:1308.0850 , 2013.\n[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im-\nage recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern\nRecognition , pages 770\u2013778, 2016.\n[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and J\u00fcrgen Schmidhuber. Gradient flow in\nrecurrent nets: the difficulty of learning long-term dependencies, 2001.\n[13] Sepp Hochreiter and J\u00fcrgen Schmidhuber. Long short-term memory. Neural computation ,\n9(8):1735\u20131780, 1997.\n[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations\nacross languages. In Proceedings of the 2009 Conference on Empirical Methods in Natural\nLanguage Processing , pages 832\u2013841. ACL, August 2009.\n[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring\nthe limits of language modeling. arXiv preprint arXiv:1602.02410 , 2016.\n", "start_char_idx": 355, "end_char_idx": 1837, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "de4d13e0-dfe7-4887-a0f6-e087cbcda550": {"__data__": {"id_": "de4d13e0-dfe7-4887-a0f6-e087cbcda550", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "ad73e473-afa4-41aa-a795-c7a9b1082fa9", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8711b9214dabab460c0c1db37828ef546489b3e37782eb8fce220a74cd5e5f81", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "6e5ad201-a5fe-4090-bfbb-17ef7fcb9bb8", "node_type": "1", "metadata": {}, "hash": "33de0027a55eeb67b971f5bb1fb80e804d230ce4ada148fcb121699ba8d27a7e", "class_name": "RelatedNodeInfo"}}, "text": "[16] \u0141ukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural\nInformation Processing Systems, (NIPS) , 2016.\n[17] \u0141ukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference\non Learning Representations (ICLR) , 2016.\n[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Ko-\nray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2 ,\n2017.\n", "start_char_idx": 1837, "end_char_idx": 2317, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "6e5ad201-a5fe-4090-bfbb-17ef7fcb9bb8": {"__data__": {"id_": "6e5ad201-a5fe-4090-bfbb-17ef7fcb9bb8", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "de4d13e0-dfe7-4887-a0f6-e087cbcda550", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a0e76f2b888e24855cd06dd7b08cfa2c499637b1331c61e5dfaad481bc9472e9", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "b4bdc4ea-ca2a-496d-b377-238c8447c25b", "node_type": "1", "metadata": {}, "hash": "26d7e191978a11862ca9e0645cf86ecf622b0002daf5fbeb7d7708666b7102cd", "class_name": "RelatedNodeInfo"}}, "text": "[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks.\nInInternational Conference on Learning Representations , 2017.\n[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR , 2015.\n", "start_char_idx": 2317, "end_char_idx": 2570, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "b4bdc4ea-ca2a-496d-b377-238c8447c25b": {"__data__": {"id_": "b4bdc4ea-ca2a-496d-b377-238c8447c25b", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "6e5ad201-a5fe-4090-bfbb-17ef7fcb9bb8", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "bc6c2e7b082588db7c8384a6a7b4eea48521d1809219b69c4422420906b566c6", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "6533451f-b299-4d94-ab0e-f50c4bb006cf", "node_type": "1", "metadata": {}, "hash": "56bd80b1bdf82527d61cd0d3f0bffe0c7c018356aa4616e8f0a9916690dfbab7", "class_name": "RelatedNodeInfo"}}, "text": "[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint\narXiv:1703.10722 , 2017.\n", "start_char_idx": 2570, "end_char_idx": 2692, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "6533451f-b299-4d94-ab0e-f50c4bb006cf": {"__data__": {"id_": "6533451f-b299-4d94-ab0e-f50c4bb006cf", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "b4bdc4ea-ca2a-496d-b377-238c8447c25b", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e4d727ce734f9ddfd9f0e20b4820ac060878d95d4915de9f724dbc88a3017055", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "cdf35e7c-fe91-4b76-83b7-a59e88b47277", "node_type": "1", "metadata": {}, "hash": "8d77459b8be2b95ef185499fabf082d83096113e57464684ceec6de056d85888", "class_name": "RelatedNodeInfo"}}, "text": "[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen\nZhou, and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint\narXiv:1703.03130 , 2017.\n[23] Minh-Thang Luong, Quoc V . Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task\nsequence to sequence learning. arXiv preprint arXiv:1511.06114 , 2015.\n[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-\nbased neural machine translation. arXiv preprint arXiv:1508.04025 , 2015.\n", "start_char_idx": 2692, "end_char_idx": 3227, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "cdf35e7c-fe91-4b76-83b7-a59e88b47277": {"__data__": {"id_": "cdf35e7c-fe91-4b76-83b7-a59e88b47277", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "6533451f-b299-4d94-ab0e-f50c4bb006cf", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "d3738f25a0cd8c4d20d3085de3e53869ef7e448baed6654c781e2885379d9496", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "91571f0f-6bcf-4172-870b-8a9ff8c2f42e", "node_type": "1", "metadata": {}, "hash": "84203037f8a0c53aaeb71352e2689230828a14154526d8423435388e249f5d75", "class_name": "RelatedNodeInfo"}}, "text": "11", "start_char_idx": 947, "end_char_idx": 949, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "91571f0f-6bcf-4172-870b-8a9ff8c2f42e": {"__data__": {"id_": "91571f0f-6bcf-4172-870b-8a9ff8c2f42e", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "cdf35e7c-fe91-4b76-83b7-a59e88b47277", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "9e6f66859ddfec72e70fb13359e1f388130a6c023a5c82cf503736651dc1f7c7", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "fdb3ba2e-2595-4e34-b4ff-55ddb6157c27", "node_type": "1", "metadata": {}, "hash": "fc2346583d85860528ee82bc01f5a0c13c812330901c2f3f1b8de4b4d49fe525", "class_name": "RelatedNodeInfo"}}, "text": "[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated\ncorpus of english: The penn treebank. Computational linguistics , 19(2):313\u2013330, 1993.\n[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In\nProceedings of the Human Language Technology Conference of the NAACL, Main Conference ,\npages 152\u2013159. ACL, June 2006.\n", "start_char_idx": 0, "end_char_idx": 401, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "fdb3ba2e-2595-4e34-b4ff-55ddb6157c27": {"__data__": {"id_": "fdb3ba2e-2595-4e34-b4ff-55ddb6157c27", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "91571f0f-6bcf-4172-870b-8a9ff8c2f42e", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8a443340e87718cb142ec1aa7a8da55f95242662bb1cf660d0b6f3da84dd5e02", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "d36501b8-45cd-4893-8869-e0deadd4d8f7", "node_type": "1", "metadata": {}, "hash": "524e63f2027f2bdaec83b5d9375718140fb4d1b15ee36987ed4cf363a3aafbd2", "class_name": "RelatedNodeInfo"}}, "text": "[27] Ankur Parikh, Oscar T\u00e4ckstr\u00f6m, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention\nmodel. In Empirical Methods in Natural Language Processing , 2016.\n[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive\nsummarization. arXiv preprint arXiv:1705.04304 , 2017.\n[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact,\nand interpretable tree annotation. In Proceedings of the 21st International Conference on\nComputational Linguistics and 44th Annual Meeting of the ACL , pages 433\u2013440. ACL, July\n2006.\n[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv\npreprint arXiv:1608.05859 , 2016.\n[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words\nwith subword units. arXiv preprint arXiv:1508.07909 , 2015.\n[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton,\nand Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts\nlayer. arXiv preprint arXiv:1701.06538 , 2017.\n", "start_char_idx": 401, "end_char_idx": 1502, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "d36501b8-45cd-4893-8869-e0deadd4d8f7": {"__data__": {"id_": "d36501b8-45cd-4893-8869-e0deadd4d8f7", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "fdb3ba2e-2595-4e34-b4ff-55ddb6157c27", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "cb2a3128ee8cd9093dd904f00de96c9f1420b1eef8ae224a6d32107045e1dfa6", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "df80c204-cf8e-47e7-a08e-ef0f828244ac", "node_type": "1", "metadata": {}, "hash": "3bccf818cb61eb7052c2a89b9fae81422dd21a412f5fff0a0c9d69eea5650e49", "class_name": "RelatedNodeInfo"}}, "text": "[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi-\nnov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine\nLearning Research , 15(1):1929\u20131958, 2014.\n", "start_char_idx": 1502, "end_char_idx": 1735, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "df80c204-cf8e-47e7-a08e-ef0f828244ac": {"__data__": {"id_": "df80c204-cf8e-47e7-a08e-ef0f828244ac", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "d36501b8-45cd-4893-8869-e0deadd4d8f7", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7b4cd38ff6e38b2dabbd1edd476a34157af122470c247af2dd9d4affce2bef75", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "0e66ad56-e2bc-41d3-9693-771e1655bd78", "node_type": "1", "metadata": {}, "hash": "42f93162cc74963f7dcc86c2334f13671856c452eebd670250bb1e7afd830a07", "class_name": "RelatedNodeInfo"}}, "text": "[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory\nnetworks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors,\nAdvances in Neural Information Processing Systems 28 , pages 2440\u20132448. Curran Associates,\nInc., 2015.\n[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural\nnetworks. In Advances in Neural Information Processing Systems , pages 3104\u20133112, 2014.\n[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna.\nRethinking the inception architecture for computer vision. CoRR , abs/1512.00567, 2015.\n", "start_char_idx": 1735, "end_char_idx": 2380, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "0e66ad56-e2bc-41d3-9693-771e1655bd78": {"__data__": {"id_": "0e66ad56-e2bc-41d3-9693-771e1655bd78", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "df80c204-cf8e-47e7-a08e-ef0f828244ac", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c044cdc18704ffff9d448325add32008d4dcd520fa42070a35cdd83e64df20d5", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "fe56f7b0-b0aa-4bab-b28c-c8eb52d2dab1", "node_type": "1", "metadata": {}, "hash": "8d68d3f6744e1b97e31e9af6a78f4a7184143d70a4d8d4a132a2cf3671f401bb", "class_name": "RelatedNodeInfo"}}, "text": "[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In\nAdvances in Neural Information Processing Systems , 2015.\n[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang\nMacherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google\u2019s neural machine\ntranslation system: Bridging the gap between human and machine translation. arXiv preprint\narXiv:1609.08144 , 2016.\n[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with\nfast-forward connections for neural machine translation. CoRR , abs/1606.04199, 2016.\n", "start_char_idx": 2380, "end_char_idx": 2994, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "fe56f7b0-b0aa-4bab-b28c-c8eb52d2dab1": {"__data__": {"id_": "fe56f7b0-b0aa-4bab-b28c-c8eb52d2dab1", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "0e66ad56-e2bc-41d3-9693-771e1655bd78", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "73b3e9c645e02dc8dfc1bc6e3be0cdfceaadbc390d6a4535f9a1a66dab3f750c", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "357c6c9e-81fe-40af-ac73-8af5a4a0797c", "node_type": "1", "metadata": {}, "hash": "a5db1cde4cdbcdaa1ca213a650e7168ef85409b04f996221bf72423f6d52b8f4", "class_name": "RelatedNodeInfo"}}, "text": "[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate\nshift-reduce constituent parsing. In Proceedings of the 51st Annual Meeting of the ACL (Volume\n1: Long Papers) , pages 434\u2013443. ", "start_char_idx": 2994, "end_char_idx": 3209, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "357c6c9e-81fe-40af-ac73-8af5a4a0797c": {"__data__": {"id_": "357c6c9e-81fe-40af-ac73-8af5a4a0797c", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "fe56f7b0-b0aa-4bab-b28c-c8eb52d2dab1", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "26f318ce06dcf322c1b7c7ba4e474ada94075c50979ee14b6336fb46c88464d1", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "f6bd0eec-8952-4d79-b677-d7589a0c9025", "node_type": "1", "metadata": {}, "hash": "1c18b295c041c2822394737be0afb1c51f9258ddc315596ebdddc4bc593a1d8c", "class_name": "RelatedNodeInfo"}}, "text": "ACL, August 2013.\n12", "start_char_idx": 3209, "end_char_idx": 3229, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "f6bd0eec-8952-4d79-b677-d7589a0c9025": {"__data__": {"id_": "f6bd0eec-8952-4d79-b677-d7589a0c9025", "embedding": null, "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "21c4166d-6661-46b2-955e-0e7a112b0a80", "node_type": "4", "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "83f916af064b4eef81dc72612f14a9e99b432e342bad25bb3ee9f350fa0caae4", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "357c6c9e-81fe-40af-ac73-8af5a4a0797c", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "39b99350971078f7fc3d0418b95292c0a299be0c6cc4df484ac63d18ad9c48db", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "79a1ef13-5aca-4612-8f62-bfcf74a5f248", "node_type": "1", "metadata": {}, "hash": "e90d26ea6082900350e55977cebfd27272137fcae3ff1643187482884be70731", "class_name": "RelatedNodeInfo"}}, "text": "Attention Visualizations\nInput-Input Layer5\nIt\nis\nin\nthis\nspirit\nthat\na\nmajority\nof\nAmerican\ngovernments\nhave\npassed\nnew\nlaws\nsince\n2009\nmaking\nthe\nregistration\nor\nvoting\nprocess\nmore\ndifficult\n.\n<EOS>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\nIt\nis\nin\nthis\nspirit\nthat\na\nmajority\nof\nAmerican\ngovernments\nhave\npassed\nnew\nlaws\nsince\n2009\nmaking\nthe\nregistration\nor\nvoting\nprocess\nmore\ndifficult\n.\n<EOS>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\nFigure 3: An example of the attention mechanism following long-distance dependencies in the\nencoder self-attention in layer 5 of 6. Many of the attention heads attend to a distant dependency of\nthe verb \u2018making\u2019, completing the phrase \u2018making...more difficult\u2019. ", "start_char_idx": 0, "end_char_idx": 694, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "79a1ef13-5aca-4612-8f62-bfcf74a5f248": {"__data__": {"id_": "79a1ef13-5aca-4612-8f62-bfcf74a5f248", "embedding": null, "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "21c4166d-6661-46b2-955e-0e7a112b0a80", "node_type": "4", "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "83f916af064b4eef81dc72612f14a9e99b432e342bad25bb3ee9f350fa0caae4", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "f6bd0eec-8952-4d79-b677-d7589a0c9025", "node_type": "1", "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "0f66dd1fa94e42224747a720646f13aa6361d47db9f333ac9611fef2be76bbc1", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "84327c1d-50ac-4c2c-bca4-773e888100f1", "node_type": "1", "metadata": {}, "hash": "82178916cb1f647546c46f05856544259e8a5253367f062c9d726d3f2bd111ee", "class_name": "RelatedNodeInfo"}}, "text": "Attentions here shown only for\nthe word \u2018making\u2019. Different colors represent different heads. Best viewed in color.\n13", "start_char_idx": 694, "end_char_idx": 812, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "84327c1d-50ac-4c2c-bca4-773e888100f1": {"__data__": {"id_": "84327c1d-50ac-4c2c-bca4-773e888100f1", "embedding": null, "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "d167b925-01c2-44fa-aad1-97716bc024bc", "node_type": "4", "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1aee818b1ec4a25980c87581f500a4cd556a90dc8e0a52ad4f6aae08ff93e6ef", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "79a1ef13-5aca-4612-8f62-bfcf74a5f248", "node_type": "1", "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7da9e349a032cba09bcb86f8bf8a29ffa22831d0745372e1fb2ffa7716163ea3", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "2f750689-4461-4756-8fde-01726308eaf3", "node_type": "1", "metadata": {}, "hash": "5258a15c954ee82b06d1aae6640117e9336cbedce629d4663d56af2389879f66", "class_name": "RelatedNodeInfo"}}, "text": "Input-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>Figure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top:\nFull attentions for head 5. ", "start_char_idx": 0, "end_char_idx": 675, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "2f750689-4461-4756-8fde-01726308eaf3": {"__data__": {"id_": "2f750689-4461-4756-8fde-01726308eaf3", "embedding": null, "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "d167b925-01c2-44fa-aad1-97716bc024bc", "node_type": "4", "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1aee818b1ec4a25980c87581f500a4cd556a90dc8e0a52ad4f6aae08ff93e6ef", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "84327c1d-50ac-4c2c-bca4-773e888100f1", "node_type": "1", "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "abdba424cd7be782708c822c2e7f151604871a6898738214cb86b325d9430ff3", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "16dc7d44-1a9e-4ed7-8959-6798f8529cac", "node_type": "1", "metadata": {}, "hash": "74e3304cb376cacaafb53934946d5a9be2ae44e4f093003f9bc27f8cfcdca178", "class_name": "RelatedNodeInfo"}}, "text": "Bottom: Isolated attentions from just the word \u2018its\u2019 for attention heads 5\nand 6. Note that the attentions are very sharp for this word.\n14", "start_char_idx": 675, "end_char_idx": 814, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "16dc7d44-1a9e-4ed7-8959-6798f8529cac": {"__data__": {"id_": "16dc7d44-1a9e-4ed7-8959-6798f8529cac", "embedding": null, "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "4bc641df-b009-48a6-8ba9-9d001e5879b7", "node_type": "4", "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ced7a073f0f0f4013feee9dcf052485cd29c25a4c52712cfe6f08c0d17107802", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "2f750689-4461-4756-8fde-01726308eaf3", "node_type": "1", "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c60f4996f8dcb81dd4154fa77aa567c50325ae00304d6361c33aee7aef1b34a6", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "f8bb10ed-3dd9-4d4d-9c27-d66565142fbd", "node_type": "1", "metadata": {}, "hash": "23c3dcefc794b12005578dcccfd379b5caefc4110796fdf1bf866a478d202c07", "class_name": "RelatedNodeInfo"}}, "text": "Input-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>Figure 5: Many of the attention heads exhibit behaviour that seems related to the structure of the\nsentence. We give two such examples above, from two different heads from the encoder self-attention\nat layer 5 of 6. The heads clearly learned to perform different tasks.\n", "start_char_idx": 0, "end_char_idx": 815, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "f8bb10ed-3dd9-4d4d-9c27-d66565142fbd": {"__data__": {"id_": "f8bb10ed-3dd9-4d4d-9c27-d66565142fbd", "embedding": null, "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "4bc641df-b009-48a6-8ba9-9d001e5879b7", "node_type": "4", "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ced7a073f0f0f4013feee9dcf052485cd29c25a4c52712cfe6f08c0d17107802", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "16dc7d44-1a9e-4ed7-8959-6798f8529cac", "node_type": "1", "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "638e638c4f36246e4e4551570c51e672d237533837ff9d3c7328588ef0741f8d", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "4ae62500-f4b7-4576-bfd1-dd7a206dd0b7", "node_type": "1", "metadata": {}, "hash": "b7814948fe2591d15a7b4784cd803a14e89e5221ee0e366d18579150aec8421e", "class_name": "RelatedNodeInfo"}}, "text": "15", "start_char_idx": 815, "end_char_idx": 817, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "4ae62500-f4b7-4576-bfd1-dd7a206dd0b7": {"__data__": {"id_": "4ae62500-f4b7-4576-bfd1-dd7a206dd0b7", "embedding": null, "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8c7447f6-6a91-435c-80ce-786595a1941b", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8404a369a6397095d669ba03727495bc99aecf425e3487b4f5ea280878e4e7cb", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "f8bb10ed-3dd9-4d4d-9c27-d66565142fbd", "node_type": "1", "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ffe3659172531da0d7313fc185d6509a6a6b811247ac5d61833440c86e1d0b99", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "03db1e24-2be5-417d-aca0-fd45ec65c9df", "node_type": "1", "metadata": {}, "hash": "03e45538219a26d0a020a34ddbbd105d505b4e2e04f3087452374e6a1c4897cc", "class_name": "RelatedNodeInfo"}}, "text": "arXiv:1502.03167v3  [cs.LG]  2 Mar 2015BatchNormalization: AcceleratingDeepNetworkTrainingb y\nReducingInternalCovariateShift\nSergey Ioffe\nGoogleInc., sioffe@google.comChristianSzegedy\nGoogleInc., szegedy@google.com\nAbstract\nTrainingDeepNeuralNetworksiscomplicatedbythefact\nthat the distributionofeach layer\u2019sinputschangesduring\ntraining, as the parametersof the previouslayers change.\nThis slows down the training by requiringlower learning\nratesandcarefulparameterinitialization,andmakesitno -\ntoriously hard to train models with saturating nonlineari-\nties. We refer to this phenomenon as internal covariate\nshift, and address the problem by normalizing layer in-\nputs. Ourmethoddrawsitsstrengthfrommakingnormal-\nizationapartofthemodelarchitectureandperformingthe\nnormalization for each training mini-batch . ", "start_char_idx": 0, "end_char_idx": 811, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "03db1e24-2be5-417d-aca0-fd45ec65c9df": {"__data__": {"id_": "03db1e24-2be5-417d-aca0-fd45ec65c9df", "embedding": null, "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8c7447f6-6a91-435c-80ce-786595a1941b", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8404a369a6397095d669ba03727495bc99aecf425e3487b4f5ea280878e4e7cb", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "4ae62500-f4b7-4576-bfd1-dd7a206dd0b7", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "854e1173728be619baa11d2a14d9a0da66822261ba6501caedeb38a116ae1359", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "ce5cdd06-2fdf-47b3-a43a-47e144522450", "node_type": "1", "metadata": {}, "hash": "24a6fde2a01135f3376938974c8109a3e758ad045efb2553ba1a74d29722b068", "class_name": "RelatedNodeInfo"}}, "text": "Batch Nor-\nmalizationallowsustousemuchhigherlearningratesand\nbe less careful about initialization. It also acts as a regu-\nlarizer, in some cases eliminating the need for Dropout.\nApplied to a state-of-the-art image classi\ufb01cation model,\nBatch Normalizationachievesthe same accuracy with 14\ntimes fewer training steps, and beats the original model\nby a signi\ufb01cant margin. Using an ensemble of batch-\nnormalizednetworks,weimproveuponthebestpublished\nresult on ImageNet classi\ufb01cation: reaching 4.9% top-5\nvalidation error (and 4.8% test error), exceeding the ac-\ncuracyofhumanraters.\n1 Introduction\nDeep learning has dramatically advanced the state of the\nart in vision, speech, and many other areas. Stochas-\ntic gradient descent (SGD) has proved to be an effec-\ntive way of training deep networks, and SGD variants\nsuch as momentum (Sutskeveret al., 2013) and Adagrad\n(Duchiet al.,2011)havebeenusedtoachievestate ofthe\nart performance. SGD optimizes the parameters \u0398of the\nnetwork,soasto minimizetheloss\n\u0398 = argmin\n\u03981\nNN\u2211\ni=1\u2113(xi,\u0398)\nwherex1...Nisthetrainingdataset. With SGD,thetrain-\ningproceedsinsteps,andateachstepweconsidera mini-\nbatchx1...mofsizem. The mini-batchis usedtoapprox-\nimate the gradient of the loss functionwith respect to the\nparameters,bycomputing\n1\nm\u2202\u2113(xi,\u0398)\n\u2202\u0398.Usingmini-batchesofexamples,asopposedtooneexam-\npleatatime,ishelpfulinseveralways. First,thegradient\nofthelossoveramini-batchisanestimateofthegradient\noverthetrainingset, whose qualityimprovesas thebatch\nsize increases. Second, computation over a batch can be\nmuch more ef\ufb01cient than mcomputations for individual\nexamples, due to the parallelism afforded by the modern\ncomputingplatforms.\nWhile stochastic gradient is simple and effective, it\nrequires careful tuning of the model hyper-parameters,\nspeci\ufb01callythelearningrateusedinoptimization,aswell\nas the initial values for the model parameters. The train-\ningiscomplicatedbythefactthattheinputstoeachlayer\nareaffectedbytheparametersofallprecedinglayers\u2013so\nthat small changes to the network parameters amplify as\nthenetworkbecomesdeeper.\n", "start_char_idx": 811, "end_char_idx": 2883, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ce5cdd06-2fdf-47b3-a43a-47e144522450": {"__data__": {"id_": "ce5cdd06-2fdf-47b3-a43a-47e144522450", "embedding": null, "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8c7447f6-6a91-435c-80ce-786595a1941b", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8404a369a6397095d669ba03727495bc99aecf425e3487b4f5ea280878e4e7cb", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "03db1e24-2be5-417d-aca0-fd45ec65c9df", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f112420a13a842d1bffbdde4d53bca9b21072e7c99ec41fb4b01dfac76d7cf16", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "5337277a-d5ed-4ff1-b72e-4e76d6d7b65e", "node_type": "1", "metadata": {}, "hash": "43642c4de7cee84b209a4ef2d90ec0a2f5c4703813095f3ee495f96857ed7985", "class_name": "RelatedNodeInfo"}}, "text": "The change in the distributions of layers\u2019 inputs\npresents a problem because the layers need to continu-\nously adapt to the new distribution. When the input dis-\ntributiontoalearningsystemchanges,itissaidtoexperi-\nencecovariateshift (Shimodaira, 2000). This is typically\nhandled via domain adaptation (Jiang, 2008). However,\nthe notion of covariate shift can be extended beyond the\nlearningsystemasawhole,toapplytoitsparts,suchasa\nsub-networkora layer. Considera networkcomputing\n\u2113=F2(F1(u,\u03981),\u03982)\nwhereF1andF2are arbitrary transformations, and the\nparameters \u03981,\u03982are to be learned so as to minimize\nthe loss\u2113. Learning \u03982can be viewed as if the inputs\nx =F1(u,\u03981)arefedintothesub-network\n\u2113=F2(x,\u03982).\nForexample,agradientdescentstep\n\u03982\u2190\u03982\u2212\u03b1\nmm\u2211\ni=1\u2202F2(xi,\u03982)\n\u2202\u03982\n(forbatchsize mandlearningrate \u03b1)isexactlyequivalent\nto that for a stand-alone network F2with input x. There-\nfore, the input distribution properties that make training\nmore ef\ufb01cient \u2013 such as having the same distribution be-\ntween the training and test data \u2013 apply to training the\nsub-network as well. As such it is advantageous for the\ndistributionof xtoremain\ufb01xedovertime. ", "start_char_idx": 2883, "end_char_idx": 4024, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "5337277a-d5ed-4ff1-b72e-4e76d6d7b65e": {"__data__": {"id_": "5337277a-d5ed-4ff1-b72e-4e76d6d7b65e", "embedding": null, "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "8c7447f6-6a91-435c-80ce-786595a1941b", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8404a369a6397095d669ba03727495bc99aecf425e3487b4f5ea280878e4e7cb", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "ce5cdd06-2fdf-47b3-a43a-47e144522450", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fb990e5b8378b925ee40845ae6e5c5674c2ebf9234e1b737f8cb3965a8d28924", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "1dae63a4-7407-4fec-911b-2befb2937f12", "node_type": "1", "metadata": {}, "hash": "a4b9b4f6ac14a11cabe470ed6f76ea84e643a5ebc568b0b65a431c970f53b64f", "class_name": "RelatedNodeInfo"}}, "text": "Then, \u03982does\n1", "start_char_idx": 4024, "end_char_idx": 4038, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "1dae63a4-7407-4fec-911b-2befb2937f12": {"__data__": {"id_": "1dae63a4-7407-4fec-911b-2befb2937f12", "embedding": null, "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7c1efa7b-b862-45f1-8f56-75407df01893", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "d0389bc180e2dc620e770f616e39ce98e691e832d1f375fa9583b4a88b1a5cfc", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "5337277a-d5ed-4ff1-b72e-4e76d6d7b65e", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3178c3ccad23d13d83701bd4fd1a779c64ad3ee09c190028be13ad4f6026b3a8", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "3dd4b47e-81ae-41a1-9efb-277f7ee3c064", "node_type": "1", "metadata": {}, "hash": "3e70962e41696ee61ab9ed72cd1a361178557be9542aae4b4bf9b6a74dded0f8", "class_name": "RelatedNodeInfo"}}, "text": "not have to readjust to compensate for the change in the\ndistributionof x.\nFixed distribution of inputs to a sub-network would\nhavepositiveconsequencesforthelayers outsidethesub-\nnetwork,as well. Consider a layer with a sigmoid activa-\ntion function z =g(Wu+b)whereuis the layer input,\nthe weight matrix Wand bias vector bare the layer pa-\nrameters to be learned, and g(x) =1\n1+exp(\u2212x). ", "start_char_idx": 0, "end_char_idx": 387, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "3dd4b47e-81ae-41a1-9efb-277f7ee3c064": {"__data__": {"id_": "3dd4b47e-81ae-41a1-9efb-277f7ee3c064", "embedding": null, "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7c1efa7b-b862-45f1-8f56-75407df01893", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "d0389bc180e2dc620e770f616e39ce98e691e832d1f375fa9583b4a88b1a5cfc", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "1dae63a4-7407-4fec-911b-2befb2937f12", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "0f85c02e4238f7eee72e00b3b4d3915fbcac0812d45b87fe943f6f4faa40fc72", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "7181f2d0-c9a5-45a7-86cb-3ee42f1833c0", "node_type": "1", "metadata": {}, "hash": "3d3d4fca784d732e481b42554a61deb4cc097479f6917c58d798aa388d544390", "class_name": "RelatedNodeInfo"}}, "text": "As|x|\nincreases, g\u2032(x)tends to zero. ", "start_char_idx": 387, "end_char_idx": 424, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "7181f2d0-c9a5-45a7-86cb-3ee42f1833c0": {"__data__": {"id_": "7181f2d0-c9a5-45a7-86cb-3ee42f1833c0", "embedding": null, "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7c1efa7b-b862-45f1-8f56-75407df01893", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "d0389bc180e2dc620e770f616e39ce98e691e832d1f375fa9583b4a88b1a5cfc", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "3dd4b47e-81ae-41a1-9efb-277f7ee3c064", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7b856afe11149f92c9cdeb662455d2e1537d3a7b0b851df90019d56ec9783d35", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "0f5f3713-6d91-42ee-a1d0-6f6d439ea009", "node_type": "1", "metadata": {}, "hash": "9b9b8c600099e5efa5bf2995552f837dac61b3a72c73a02908e64747265ee6ce", "class_name": "RelatedNodeInfo"}}, "text": "This means that for all di-\nmensionsof x =Wu+bexceptthosewithsmallabsolute\nvalues,thegradient\ufb02owingdownto uwillvanishandthe\nmodel will train slowly. However, since xis affected by\nW,band the parameters of all the layers below, changes\ntothoseparametersduringtrainingwilllikelymovemany\ndimensions of xinto the saturated regime of the nonlin-\nearity and slow down the convergence. ", "start_char_idx": 424, "end_char_idx": 803, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "0f5f3713-6d91-42ee-a1d0-6f6d439ea009": {"__data__": {"id_": "0f5f3713-6d91-42ee-a1d0-6f6d439ea009", "embedding": null, "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7c1efa7b-b862-45f1-8f56-75407df01893", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "d0389bc180e2dc620e770f616e39ce98e691e832d1f375fa9583b4a88b1a5cfc", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "7181f2d0-c9a5-45a7-86cb-3ee42f1833c0", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "050d25b181f9347622d6be2e67d5acc2f749ece9eb7e540841ac987c17e2c91a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "e339daf1-51f9-4679-a79b-40564c5ccf0b", "node_type": "1", "metadata": {}, "hash": "462e1418787b5e1e5cd3311fc307e96f92e2fb2f7dd7b1310a0791b2e96ea635", "class_name": "RelatedNodeInfo"}}, "text": "This effect is\nampli\ufb01ed as the network depth increases. In practice,\nthe saturation problem and the resulting vanishing gradi-\nentsareusuallyaddressedbyusingRecti\ufb01edLinearUnits\n(Nair&Hinton, 2010) ReLU(x) = max( x,0), careful\ninitialization (Bengio&Glorot, 2010; Saxeet al., 2013),\nand small learning rates. If, however, we could ensure\nthat the distribution of nonlinearity inputs remains more\nstable as the network trains, then the optimizer would be\nless likely to get stuck in the saturated regime, and the\ntrainingwouldaccelerate.\nWe refer to the change in the distributions of internal\nnodes of a deep network, in the course of training, as In-\nternal Covariate Shift . Eliminating it offers a promise of\nfaster training. We propose a new mechanism, which we\ncallBatch Normalization , that takes a step towards re-\nducing internal covariate shift, and in doing so dramati-\ncally accelerates the training of deep neural nets. It ac-\ncomplishes this via a normalization step that \ufb01xes the\nmeansandvariancesoflayerinputs. BatchNormalization\nalso has a bene\ufb01cial effect on the gradient \ufb02ow through\nthe network, by reducing the dependence of gradients\non the scale of the parameters or of their initial values.\nThis allows us to use much higher learning rates with-\nout the risk of divergence. Furthermore, batch normal-\nization regularizes the model and reduces the need for\nDropout(Srivastavaet al., 2014). Finally, Batch Normal-\nization makes it possible to use saturating nonlinearities\nby preventingthe network from getting stuck in the satu-\nratedmodes.\nIn Sec. 4.2, we apply Batch Normalization to the best-\nperforming ImageNet classi\ufb01cation network, and show\nthat we can match its performance using only 7% of the\ntraining steps, and can further exceed its accuracy by a\nsubstantial margin. Using an ensemble of such networks\ntrained with Batch Normalization, we achieve the top-5\nerror rate that improves upon the best known results on\nImageNetclassi\ufb01cation.2 Towards Reducing Internal\nCovariateShift\nWe de\ufb01ne Internal Covariate Shift as the change in the\ndistribution of network activations due to the change in\nnetworkparametersduringtraining. Toimprovethetrain-\ning, we seek to reduce the internal covariate shift. ", "start_char_idx": 803, "end_char_idx": 3031, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "e339daf1-51f9-4679-a79b-40564c5ccf0b": {"__data__": {"id_": "e339daf1-51f9-4679-a79b-40564c5ccf0b", "embedding": null, "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7c1efa7b-b862-45f1-8f56-75407df01893", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "d0389bc180e2dc620e770f616e39ce98e691e832d1f375fa9583b4a88b1a5cfc", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "0f5f3713-6d91-42ee-a1d0-6f6d439ea009", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "2c03e6fdfaa7612e38e9c3530588951a3928a7b29384e995cfbc73e040516232", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "e7c5b44d-5ae0-4cc6-9d00-dee9cb2a8814", "node_type": "1", "metadata": {}, "hash": "2a319f2ae1b7be008286f8fd2f07b70d1fd7867fe3b45d0b1b49f967fd05600b", "class_name": "RelatedNodeInfo"}}, "text": "By\n\ufb01xingthe distributionof the layer inputs xas the training\nprogresses,weexpecttoimprovethetrainingspeed. Ithas\nbeen long known (LeCunetal., 1998b; Wiesler &Ney,\n2011) that the network training convergesfaster if its in-\nputsarewhitened\u2013i.e.,linearlytransformedtohavezero\nmeansandunitvariances,anddecorrelated. Aseachlayer\nobservestheinputsproducedbythelayersbelow,itwould\nbe advantageousto achieve the same whiteningof the in-\nputsof each layer. By whitening the inputsto each layer,\nwe would take a step towards achieving the \ufb01xed distri-\nbutions of inputs that would remove the ill effects of the\ninternalcovariateshift.\nWe couldconsiderwhiteningactivationsat everytrain-\ning step or at some interval, either by modifying the\nnetwork directly or by changing the parameters of the\noptimization algorithm to depend on the network ac-\ntivation values (Wiesleret al., 2014; Raikoetal., 2012;\nPoveyet al., 2014; Desjardins&Kavukcuoglu). How-\never, if these modi\ufb01cations are interspersed with the op-\ntimization steps, then the gradient descent step may at-\ntempt to update the parameters in a way that requires\nthe normalization to be updated, which reduces the ef-\nfect of the gradient step. For example, consider a layer\nwith the input uthat addsthe learned bias b, and normal-\nizes the result by subtracting the mean of the activation\ncomputed over the training data: \u02c6x=x\u2212E[x]where\nx=u+b,X={x1...N}is the set of values of xover\nthe training set, and E [x] =1\nN\u2211N\ni=1xi. If a gradient\ndescent step ignores the dependence of E [x]onb, then it\nwill update b\u2190b+ \u2206b, where\u2206b\u221d\u2212\u2202\u2113/\u2202\u02c6x. Then\nu+ (b+ \u2206b)\u2212E[u+ (b+ \u2206b)] =u+b\u2212E[u+b].\nThus, the combination of the update to band subsequent\nchange in normalization led to no change in the output\nof the layer nor, consequently, the loss. As the training\ncontinues, bwill grow inde\ufb01nitely while the loss remains\n\ufb01xed. Thisproblemcangetworseifthenormalizationnot\nonly centers but also scales the activations. We have ob-\nserved this empirically in initial experiments, where the\nmodel blows up when the normalization parameters are\ncomputedoutsidethe gradientdescentstep.\nThe issue with the above approach is that the gradient\ndescent optimization does not take into account the fact\nthat the normalization takes place. To address this issue,\nwe would like to ensure that, for any parameter values,\nthe network alwaysproducesactivationswith the desired\ndistribution. Doing so would allow the gradient of the\nloss with respect to the model parameters to account for\nthe normalization, and for its dependence on the model\nparameters \u0398. Let again xbe a layer input, treated as a\n2", "start_char_idx": 3031, "end_char_idx": 5645, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "e7c5b44d-5ae0-4cc6-9d00-dee9cb2a8814": {"__data__": {"id_": "e7c5b44d-5ae0-4cc6-9d00-dee9cb2a8814", "embedding": null, "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "495b8b80-518b-4889-9658-6901e0229bc0", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b5513447a5e5ed0ff2404ba31a47ad4e748c8662478e8a2f355dd1ee39229480", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "e339daf1-51f9-4679-a79b-40564c5ccf0b", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c0dd363079f7880ac69a36b1905de189031268f45d010ff3872722418c10a53e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "285134f3-afad-4ca8-a9a3-d4d13b2438c1", "node_type": "1", "metadata": {}, "hash": "02c9ae43cb38f4ccbcf23a7e5eddb27e39dcc2d07694b307cc43cb64d6ae1334", "class_name": "RelatedNodeInfo"}}, "text": "vector, andXbe the set of these inputs over the training\ndataset. Thenormalizationcanthenbewrittenasatrans-\nformation\n\u02c6x =Norm(x,X)\nwhich depends not only on the given training example x\nbut on all examples X\u2013 each of which depends on \u0398if\nxis generatedby anotherlayer. For backpropagation,we\nwouldneedtocomputetheJacobians\n\u2202Norm(x,X)\n\u2202xand\u2202Norm(x,X)\n\u2202X;\nignoring the latter term would lead to the explosion de-\nscribedabove. Withinthisframework,whiteningthelayer\ninputs is expensive, as it requires computing the covari-\nance matrix Cov [x] =Ex\u2208X[xxT]\u2212E[x]E[x]Tand its\ninverse square root, to produce the whitened activations\nCov[x]\u22121/2(x\u2212E[x]), as well as the derivatives of these\ntransformsforbackpropagation.Thismotivatesustoseek\nan alternative that performs input normalization in a way\nthat is differentiable and does not require the analysis of\ntheentiretrainingset aftereveryparameterupdate.\nSome of the previous approaches (e.g.\n", "start_char_idx": 0, "end_char_idx": 937, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "285134f3-afad-4ca8-a9a3-d4d13b2438c1": {"__data__": {"id_": "285134f3-afad-4ca8-a9a3-d4d13b2438c1", "embedding": null, "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "495b8b80-518b-4889-9658-6901e0229bc0", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b5513447a5e5ed0ff2404ba31a47ad4e748c8662478e8a2f355dd1ee39229480", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "e7c5b44d-5ae0-4cc6-9d00-dee9cb2a8814", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e635cc1759213b8b444e6fd44991b0c21a24965f0ef7546c05f02f3131d9ee70", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "f4519fee-459f-4522-b07d-82747620a572", "node_type": "1", "metadata": {}, "hash": "08e6f12eac00128ce65a1528bf1990cd0ee1c8cf2216bbe81cb7205535b1643f", "class_name": "RelatedNodeInfo"}}, "text": "(Lyu&Simoncelli, 2008)) use statistics computed\nover a single training example, or, in the case of image\nnetworks, over differentfeature maps at a given location.\nHowever, this changes the representation ability of a\nnetwork by discarding the absolute scale of activations.\nWe want to a preservethe informationin the network,by\nnormalizing the activations in a training example relative\ntothe statisticsoftheentiretrainingdata.\n3 Normalization via Mini-Batch\nStatistics\nSince the full whitening of each layer\u2019s inputs is costly\nand not everywhere differentiable, we make two neces-\nsary simpli\ufb01cations. The \ufb01rst is that instead of whitening\nthe features in layer inputs and outputs jointly, we will\nnormalizeeachscalarfeatureindependently,bymakingit\nhave the mean of zero and the variance of 1. For a layer\nwithd-dimensionalinput x = (x(1)...x(d)),wewillnor-\nmalizeeachdimension\n\u02c6x(k)=x(k)\u2212E[x(k)]\u221a\nVar[x(k)]\nwheretheexpectationandvariancearecomputedoverthe\ntrainingdataset. Asshownin(LeCunetal.,1998b),such\nnormalizationspeedsupconvergence,evenwhenthefea-\nturesarenotdecorrelated.\nNotethatsimplynormalizingeachinputofalayermay\nchange what the layer can represent. For instance, nor-\nmalizing the inputsof a sigmoid wouldconstrain them to\nthe linear regime of the nonlinearity. To address this, we\nmakesurethat thetransformationinsertedin thenetwork\ncan represent the identity transform . To accomplish this,weintroduce,foreachactivation x(k),apairofparameters\n\u03b3(k),\u03b2(k),whichscale andshift thenormalizedvalue:\ny(k)=\u03b3(k)\u02c6x(k)+\u03b2(k).\nThese parameters are learned along with the original\nmodel parameters, and restore the representation power\nofthenetwork. Indeed,bysetting \u03b3(k)=\u221a\nVar[x(k)]and\n\u03b2(k)=E[x(k)], we couldrecoverthe originalactivations,\nifthatwerethe optimalthingto do.\nInthebatchsettingwhereeachtrainingstepisbasedon\ntheentire trainingset, we woulduse the wholeset to nor-\nmalize activations. ", "start_char_idx": 937, "end_char_idx": 2839, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f4519fee-459f-4522-b07d-82747620a572": {"__data__": {"id_": "f4519fee-459f-4522-b07d-82747620a572", "embedding": null, "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "495b8b80-518b-4889-9658-6901e0229bc0", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b5513447a5e5ed0ff2404ba31a47ad4e748c8662478e8a2f355dd1ee39229480", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "285134f3-afad-4ca8-a9a3-d4d13b2438c1", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "4f472d203cf887d9754200c958ebdff874b6839cfb4e60123d4c1a5617b6cd23", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "ccc697df-34bf-4228-a9c7-5834f08cd529", "node_type": "1", "metadata": {}, "hash": "a12b1781d99102e4089ad75bacb7d618b96125716f8f3813fc4d8343c71264b7", "class_name": "RelatedNodeInfo"}}, "text": "However,this is impracticalwhen us-\ning stochastic optimization. Therefore, we make the sec-\nondsimpli\ufb01cation: since we use mini-batchesin stochas-\ntic gradient training, each mini-batch produces estimates\nofthemeanandvariance ofeachactivation. Thisway,the\nstatistics used for normalization can fully participate in\nthe gradient backpropagation. Note that the use of mini-\nbatchesis enabledbycomputationof per-dimensionvari-\nances rather than joint covariances; in the joint case, reg-\nularizationwouldbe requiredsince the mini-batchsize is\nlikely to be smaller than the number of activations being\nwhitened,resultinginsingularcovariancematrices.\nConsider a mini-batch Bof sizem. Since the normal-\nization is applied to each activation independently, let us\nfocusonaparticularactivation x(k)andomitkforclarity.\nWe havemvaluesofthisactivationinthemini-batch,\nB={x1...m}.\nLetthenormalizedvaluesbe \u02c6x1...m,andtheirlineartrans-\nformationsbe y1...m. We referto thetransform\nBN\u03b3,\u03b2:x1...m\u2192y1...m\nas theBatch Normalizing Transform . We present the BN\nTransforminAlgorithm1. Inthealgorithm, \u01ebisaconstant\naddedtothemini-batchvariancefornumericalstability.\nInput:Valuesof xovera mini-batch: B={x1...m};\nParametersto belearned: \u03b3,\u03b2\nOutput:{yi=BN\u03b3,\u03b2(xi)}\n\u00b5B\u21901\nmm\u2211\ni=1xi // mini-batchmean\n\u03c32\nB\u21901\nmm\u2211\ni=1(xi\u2212\u00b5B)2// mini-batchvariance\n\u02c6xi\u2190xi\u2212\u00b5B\u221a\n\u03c32\nB+\u01eb// normalize\nyi\u2190\u03b3\u02c6xi+\u03b2\u2261BN\u03b3,\u03b2(xi) // scale andshift\nAlgorithm 1: Batch Normalizing Transform, applied to\nactivation xoveramini-batch.\nTheBNtransformcanbeaddedtoanetworktomanip-\nulate any activation. ", "start_char_idx": 2839, "end_char_idx": 4373, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ccc697df-34bf-4228-a9c7-5834f08cd529": {"__data__": {"id_": "ccc697df-34bf-4228-a9c7-5834f08cd529", "embedding": null, "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "495b8b80-518b-4889-9658-6901e0229bc0", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b5513447a5e5ed0ff2404ba31a47ad4e748c8662478e8a2f355dd1ee39229480", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "f4519fee-459f-4522-b07d-82747620a572", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "efb449c09a7384b0bc47ed7d0f48350c3fadcc92d4d4f85565f6f62347dfe23e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "2eadfedb-30e8-41e5-8a28-756700b37955", "node_type": "1", "metadata": {}, "hash": "51a1cd252d1760a5cdbb75d24a5a4d6e6666f782263c90b9a2207c1c705b270b", "class_name": "RelatedNodeInfo"}}, "text": "In the notation y=BN\u03b3,\u03b2(x), we\n3", "start_char_idx": 4373, "end_char_idx": 4405, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "2eadfedb-30e8-41e5-8a28-756700b37955": {"__data__": {"id_": "2eadfedb-30e8-41e5-8a28-756700b37955", "embedding": null, "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6ad9b48a-2a55-40b1-9808-07cd8e439c05", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1be2d9a83122331f8962df7ed2a776b99969a41d85319d0d3365f913d7440a3a", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "ccc697df-34bf-4228-a9c7-5834f08cd529", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "80de42688b76a29cabb2b2b7194c9405c5e85541b3938c88f8fdc08d20b19bef", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "6bc1dfa7-bb04-4d5f-ae1e-90f9723a74a7", "node_type": "1", "metadata": {}, "hash": "71eec491ba260c3146252d98a99502bc6a13dddae1798b29d5cc25876aeef1af", "class_name": "RelatedNodeInfo"}}, "text": "indicate that the parameters \u03b3and\u03b2are to be learned,\nbut it should be noted that the BN transform does not\nindependently process the activation in each training ex-\nample. Rather, BN \u03b3,\u03b2(x)depends both on the training\nexampleand the other examples in the mini-batch . The\nscaled and shifted values yare passed to other network\nlayers. ", "start_char_idx": 0, "end_char_idx": 335, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "6bc1dfa7-bb04-4d5f-ae1e-90f9723a74a7": {"__data__": {"id_": "6bc1dfa7-bb04-4d5f-ae1e-90f9723a74a7", "embedding": null, "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6ad9b48a-2a55-40b1-9808-07cd8e439c05", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1be2d9a83122331f8962df7ed2a776b99969a41d85319d0d3365f913d7440a3a", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "2eadfedb-30e8-41e5-8a28-756700b37955", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "692bca53190ce8ef17128416e23cb12445935daa8c9eabae4be5b8832fe1b1c0", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "1c0272bc-94c8-41e3-8fe7-a93957197edd", "node_type": "1", "metadata": {}, "hash": "1e04210fa6f9e606104ba18959a79de25ebfc229e51242d8256b6d19297f5142", "class_name": "RelatedNodeInfo"}}, "text": "The normalized activations \u02c6xare internal to our\ntransformation, but their presence is crucial. The distri-\nbutions of values of any \u02c6xhas the expected value of 0\nand the variance of 1, as long as the elements of each\nmini-batch are sampled from the same distribution, and\nif we neglect \u01eb. This can be seen by observing that\u2211m\ni=1\u02c6xi= 0and1\nm\u2211m\ni=1\u02c6x2\ni= 1, and taking expec-\ntations. Eachnormalizedactivation \u02c6x(k)canbeviewedas\nan input to a sub-network composed of the linear trans-\nformy(k)=\u03b3(k)\u02c6x(k)+\u03b2(k), followed by the other pro-\ncessing doneby the originalnetwork. Thesesub-network\ninputs all have \ufb01xed means and variances, and although\nthe jointdistributionofthese normalized \u02c6x(k)canchange\nover the course of training, we expect that the introduc-\ntion of normalized inputs accelerates the training of the\nsub-networkand,consequently,thenetworkasawhole.\nDuring training we need to backpropagate the gradi-\nent of loss \u2113through this transformation,as well as com-\npute the gradients with respect to the parameters of the\nBN transform. We use chainrule,as follows(beforesim-\npli\ufb01cation):\n\u2202\u2113\n\u2202\u02c6xi=\u2202\u2113\n\u2202yi\u00b7\u03b3\n\u2202\u2113\n\u2202\u03c32\nB=\u2211m\ni=1\u2202\u2113\n\u2202\u02c6xi\u00b7(xi\u2212\u00b5B)\u00b7\u22121\n2(\u03c32\nB+\u01eb)\u22123/2\n\u2202\u2113\n\u2202\u00b5B=(\u2211m\ni=1\u2202\u2113\n\u2202\u02c6xi\u00b7\u22121\u221a\n\u03c32\nB+\u01eb)\n+\u2202\u2113\n\u2202\u03c32\nB\u00b7\u2211m\ni=1\u22122(xi\u2212\u00b5B)\nm\n\u2202\u2113\n\u2202xi=\u2202\u2113\n\u2202\u02c6xi\u00b71\u221a\n\u03c32\nB+\u01eb+\u2202\u2113\n\u2202\u03c32\nB\u00b72(xi\u2212\u00b5B)\nm+\u2202\u2113\n\u2202\u00b5B\u00b71\nm\n\u2202\u2113\n\u2202\u03b3=\u2211m\ni=1\u2202\u2113\n\u2202yi\u00b7\u02c6xi\n\u2202\u2113\n\u2202\u03b2=\u2211m\ni=1\u2202\u2113\n\u2202yi\nThus,BNtransformisadifferentiabletransformationthat\nintroduces normalized activations into the network. This\nensures that as the model is training, layers can continue\nlearningoninputdistributionsthatexhibitlessinternal co-\nvariate shift, thus accelerating the training. ", "start_char_idx": 335, "end_char_idx": 1927, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "1c0272bc-94c8-41e3-8fe7-a93957197edd": {"__data__": {"id_": "1c0272bc-94c8-41e3-8fe7-a93957197edd", "embedding": null, "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6ad9b48a-2a55-40b1-9808-07cd8e439c05", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1be2d9a83122331f8962df7ed2a776b99969a41d85319d0d3365f913d7440a3a", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "6bc1dfa7-bb04-4d5f-ae1e-90f9723a74a7", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e60ee4ab698375af92d3103104908c062b4d9c1219e48a8b14e0342feac2dc3e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "23cc386c-71cf-4b1a-bdcd-d2fd8be427cc", "node_type": "1", "metadata": {}, "hash": "c67e5c62bac7fa443c0a201cd0f1465136dbbd0d6781a0b8f16f056896bcb960", "class_name": "RelatedNodeInfo"}}, "text": "Furthermor e,\nthe learned af\ufb01ne transform applied to these normalized\nactivationsallowstheBNtransformtorepresenttheiden-\ntity transformationandpreservesthenetworkcapacity.\n3.1 Training and Inference with Batch-\nNormalizedNetworks\nToBatch-Normalize anetwork,wespecifyasubsetofac-\ntivations and insert the BN transform for each of them,\naccording to Alg. 1. ", "start_char_idx": 1927, "end_char_idx": 2283, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "23cc386c-71cf-4b1a-bdcd-d2fd8be427cc": {"__data__": {"id_": "23cc386c-71cf-4b1a-bdcd-d2fd8be427cc", "embedding": null, "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6ad9b48a-2a55-40b1-9808-07cd8e439c05", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1be2d9a83122331f8962df7ed2a776b99969a41d85319d0d3365f913d7440a3a", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "1c0272bc-94c8-41e3-8fe7-a93957197edd", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7f62a1e26d1b342046d0f61f84de2ccddb69c779de57200bdf914b84b062fd6c", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "1223e2d0-f3f1-45f1-9962-24e1d2a3139c", "node_type": "1", "metadata": {}, "hash": "d9cc012c95db1910e4785984d538c57029f1343884de7cf8444f2ba046ba110c", "class_name": "RelatedNodeInfo"}}, "text": "Any layer that previously received\nxas the input, now receives BN (x). A model employing\nBatch Normalization can be trained using batch gradient\ndescent,orStochasticGradientDescentwithamini-batch\nsizem >1, or with any of its variants such as Adagrad(Duchiet al.,2011). Thenormalizationofactivationsthat\ndependsonthemini-batchallowsef\ufb01cienttraining,but is\nneithernecessarynordesirableduringinference;wewant\nthe output to depend only on the input, deterministically.\nFor this, once the network has been trained, we use the\nnormalization\n\u02c6x=x\u2212E[x]\u221a\nVar[x]+\u01eb\nusing the population, rather than mini-batch, statistics.\nNeglecting \u01eb, these normalized activations have the same\nmean0 and variance1 as duringtraining. We use the un-\nbiased variance estimate Var [x] =m\nm\u22121\u00b7EB[\u03c32\nB], where\ntheexpectationisovertrainingmini-batchesofsize mand\n\u03c32\nBaretheirsamplevariances. Usingmovingaveragesin-\nstead, we can track the accuracy of a model as it trains.\nSincethemeansandvariancesare\ufb01xedduringinference,\nthe normalization is simply a linear transform applied to\neachactivation. Itmayfurtherbecomposedwiththescal-\ning by\u03b3and shift by \u03b2, to yield a single linear transform\nthat replacesBN (x). Algorithm 2 summarizesthe proce-\ndurefortrainingbatch-normalizednetworks.\nInput:NetworkNwith trainableparameters \u0398;\nsubsetofactivations {x(k)}K\nk=1\nOutput: Batch-normalizednetworkforinference, Ninf\nBN\n1:Ntr\nBN\u2190N// TrainingBN network\n2:fork= 1...Kdo\n3:Add transformation y(k)=BN\u03b3(k),\u03b2(k)(x(k))to\nNtr\nBN(Alg.1)\n4:Modify each layer in Ntr\nBNwith input x(k)to take\ny(k)instead\n5:end for\n6:TrainNtr\nBNto optimize the parameters \u0398\u222a\n{\u03b3(k),\u03b2(k)}K\nk=1\n7:Ninf\nBN\u2190Ntr\nBN// InferenceBN networkwithfrozen\n// parameters\n8:fork= 1...Kdo\n9:// Forclarity, x\u2261x(k),\u03b3\u2261\u03b3(k),\u00b5B\u2261\u00b5(k)\nB, etc.\n10:Process multiple training mini-batches B, each of\nsizem,andaverageoverthem:\nE[x]\u2190EB[\u00b5B]\nVar[x]\u2190m\nm\u22121EB[\u03c32\nB]\n11:InNinf\nBN, replace the transform y=BN\u03b3,\u03b2(x)with\ny=\u03b3\u221a\nVar[x]+\u01eb\u00b7x+(\n\u03b2\u2212\u03b3E[x]\u221a\nVar[x]+\u01eb)\n12:end for\nAlgorithm2: Traininga Batch-NormalizedNetwork\n3.2 Batch-Normalized Convolutional Net-\nworks\nBatch Normalization can be applied to any set of acti-\nvations in the network. Here, we focus on transforms\n4", "start_char_idx": 2283, "end_char_idx": 4443, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "1223e2d0-f3f1-45f1-9962-24e1d2a3139c": {"__data__": {"id_": "1223e2d0-f3f1-45f1-9962-24e1d2a3139c", "embedding": null, "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "501f9969-eca6-4e92-9080-b03e896be613", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c950dd2e15cab928c213c532a52debc6dc332b8149f6a264c891c209974514fe", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "23cc386c-71cf-4b1a-bdcd-d2fd8be427cc", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "63beb233a676f65a336090ad756d1aadd9007a32cc59922d110161a79dae3d09", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "92b507ee-ce43-4dd8-b017-7dcbdf070435", "node_type": "1", "metadata": {}, "hash": "eab8e509f0317ccb0828163cb554e9835e3f207e46b2290fda93893e16f6a402", "class_name": "RelatedNodeInfo"}}, "text": "that consist of an af\ufb01ne transformation followed by an\nelement-wisenonlinearity:\nz =g(Wu+b)\nwhereWandbare learned parametersof the model, and\ng(\u00b7)isthenonlinearitysuchassigmoidorReLU.Thisfor-\nmulation covers both fully-connected and convolutional\nlayers. We add the BN transform immediately before the\nnonlinearity,bynormalizing x =Wu+b. Wecouldhave\nalso normalized the layer inputs u, but since uis likely\nthe output of another nonlinearity, the shape of its distri-\nbutionislikelytochangeduringtraining,andconstrainin g\nits \ufb01rst and second moments would not eliminate the co-\nvariate shift. In contrast, Wu + bis more likely to have\na symmetric,non-sparsedistribution,that is \u201cmoreGaus-\nsian\u201d(Hyv\u00a8 arinen&Oja,2000);normalizingitislikelyto\nproduceactivationswithastable distribution.\nNotethat,sincewenormalize Wu+b,thebiasbcanbe\nignoredsinceitseffectwillbecanceledbythesubsequent\nmeansubtraction(theroleofthebiasissubsumedby \u03b2in\nAlg.1). Thus, z =g(Wu+b)is replacedwith\nz =g(BN(Wu))\nwhere the BN transformis applied independentlyto each\ndimension of x =Wu, with a separate pair of learned\nparameters \u03b3(k),\u03b2(k)perdimension.\nForconvolutionallayers,we additionallywant the nor-\nmalization to obey the convolutional property \u2013 so that\ndifferent elements of the same feature map, at different\nlocations, are normalized in the same way. ", "start_char_idx": 0, "end_char_idx": 1331, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "92b507ee-ce43-4dd8-b017-7dcbdf070435": {"__data__": {"id_": "92b507ee-ce43-4dd8-b017-7dcbdf070435", "embedding": null, "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "501f9969-eca6-4e92-9080-b03e896be613", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c950dd2e15cab928c213c532a52debc6dc332b8149f6a264c891c209974514fe", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "1223e2d0-f3f1-45f1-9962-24e1d2a3139c", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f3388435b0e94eae6a26de8ab56b586ab5cf73364bc23877542c4435763ff135", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "02afc34b-32f4-4c24-b536-349dd39356e3", "node_type": "1", "metadata": {}, "hash": "75ed1fe0302d1dad0b1af945b3c346230616a2abd208b94b6fe20dcd72842b85", "class_name": "RelatedNodeInfo"}}, "text": "To achieve\nthis, we jointly normalize all the activations in a mini-\nbatch, overall locations. In Alg. 1, we let Bbe the set of\nall values in a feature map across both the elements of a\nmini-batch and spatial locations \u2013 so for a mini-batch of\nsizemand feature maps of size p\u00d7q, we use the effec-\ntive mini-batch of size m\u2032=|B|=m\u00b7pq. We learn a\npair of parameters \u03b3(k)and\u03b2(k)per feature map, rather\nthan per activation. ", "start_char_idx": 1331, "end_char_idx": 1751, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "02afc34b-32f4-4c24-b536-349dd39356e3": {"__data__": {"id_": "02afc34b-32f4-4c24-b536-349dd39356e3", "embedding": null, "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "501f9969-eca6-4e92-9080-b03e896be613", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c950dd2e15cab928c213c532a52debc6dc332b8149f6a264c891c209974514fe", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "92b507ee-ce43-4dd8-b017-7dcbdf070435", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "53f59ef1f5e610b5e4f0b15dba8440677d52d2b910c550957109b3b818164709", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "65332716-e35c-496a-b0bc-e6a4ed0cc983", "node_type": "1", "metadata": {}, "hash": "1424829f5cc767e4385fac93a659e49391d70f4b6b1019107f1609fad0711131", "class_name": "RelatedNodeInfo"}}, "text": "Alg. 2 is modi\ufb01ed similarly, so that\nduringinferencetheBNtransformappliesthesamelinear\ntransformationtoeachactivationina givenfeaturemap.\n3.3 Batch Normalization enables higher\nlearning rates\nIn traditional deep networks, too-high learning rate may\nresult in the gradients that explode or vanish, as well as\ngetting stuck in poor local minima. Batch Normaliza-\ntion helps address these issues. By normalizing activa-\ntions throughout the network, it prevents small changes\nto the parameters from amplifying into larger and subop-\ntimal changes in activations in gradients; for instance, it\nprevents the training from getting stuck in the saturated\nregimesofnonlinearities.\nBatchNormalizationalsomakestrainingmoreresilient\ntotheparameterscale. Normally,largelearningratesmay\nincreasethescaleoflayerparameters,whichthenamplifythegradientduringbackpropagationandleadtothemodel\nexplosion. However, with Batch Normalization, back-\npropagation through a layer is unaffected by the scale of\nitsparameters. Indeed,fora scalar a,\nBN(Wu) =BN((aW)u)\nandwe canshowthat\n\u2202BN((aW)u)\n\u2202u=\u2202BN(Wu)\n\u2202u\n\u2202BN((aW)u)\n\u2202(aW)=1\na\u00b7\u2202BN(Wu)\n\u2202W\nThe scale does not affect the layer Jacobian nor, con-\nsequently, the gradient propagation. Moreover, larger\nweights lead to smallergradients, and Batch Normaliza-\ntionwill stabilize theparametergrowth.\nWe further conjecture that Batch Normalization may\nleadthelayerJacobianstohavesingularvaluescloseto1,\nwhich is known to be bene\ufb01cial for training (Saxeet al.,\n2013). Consider two consecutive layers with normalized\ninputs, and the transformation between these normalized\nvectors:\u02c6z =F(\u02c6x). Ifweassumethat \u02c6xand\u02c6zareGaussian\nanduncorrelated,andthat F(\u02c6x)\u2248J\u02c6xisalineartransfor-\nmationforthe givenmodelparameters,thenboth \u02c6xand\u02c6z\nhave unit covariances, and I=Cov[\u02c6z] =JCov[\u02c6x]JT=\nJJT. Thus,JJT=I, and so all singular values of J\nare equal to 1, which preserves the gradient magnitudes\nduring backpropagation. In reality, the transformation is\nnotlinear,andthenormalizedvaluesarenotguaranteedto\nbe Gaussian nor independent, but we nevertheless expect\nBatch Normalization to help make gradient propagation\nbetter behaved. ", "start_char_idx": 1751, "end_char_idx": 3884, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "65332716-e35c-496a-b0bc-e6a4ed0cc983": {"__data__": {"id_": "65332716-e35c-496a-b0bc-e6a4ed0cc983", "embedding": null, "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "501f9969-eca6-4e92-9080-b03e896be613", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c950dd2e15cab928c213c532a52debc6dc332b8149f6a264c891c209974514fe", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "02afc34b-32f4-4c24-b536-349dd39356e3", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "4625d48a162c194f959462b3cdd39131c9ec62082453cbd70eb6d44456f37301", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "8c2c7fe2-0b52-4c53-81be-65d29a7cb679", "node_type": "1", "metadata": {}, "hash": "49e4f40b4a409c481e09b33f49464b8f96b8f844b1449b9216fb40000e0b4a7d", "class_name": "RelatedNodeInfo"}}, "text": "The precise effect of Batch Normaliza-\ntion on gradient propagation remains an area of further\nstudy.\n3.4 Batch Normalization regularizes the\nmodel\nWhen training with Batch Normalization, a training ex-\nample is seen in conjunction with other examples in the\nmini-batch, and the training network no longer produc-\ning deterministic values for a given training example. In\nour experiments,we foundthis effect to be advantageous\nto the generalization of the network. Whereas Dropout\n(Srivastavaet al., 2014) is typically used to reduce over-\n\ufb01tting,inabatch-normalizednetworkwefoundthatitcan\nbeeitherremovedorreducedinstrength.\n4 Experiments\n4.1 Activationsovertime\nTo verify the effects of internal covariate shift on train-\ning, and the ability of Batch Normalization to combat it,\nweconsideredtheproblemofpredictingthedigitclasson\ntheMNISTdataset(LeCunetal.,1998a). Weusedavery\nsimple network, with a 28x28binary image as input, and\n5", "start_char_idx": 3884, "end_char_idx": 4819, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "8c2c7fe2-0b52-4c53-81be-65d29a7cb679": {"__data__": {"id_": "8c2c7fe2-0b52-4c53-81be-65d29a7cb679", "embedding": null, "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "defb123a-181a-439f-9b2a-c8baddeb016e", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "542d2ebc712807be19d69dc83b63126983d3aaaba4a523b2bb078429a62e8588", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "65332716-e35c-496a-b0bc-e6a4ed0cc983", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "517bf1cba2d2b86e1d2be7ec59c2f325c0bd949a649cdc5e7ec09bf76fe75057", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "15d23f05-58ab-4ae9-92c7-9ea57abb2ae3", "node_type": "1", "metadata": {}, "hash": "3a5d0f5d7769a2f8d1b9164d0147473fee304408ceb9f773f86179be022f5bb6", "class_name": "RelatedNodeInfo"}}, "text": "10K20K30K40K50K0.70.80.91\n  \nWithout BN\nWith BN\n\u2212202\n\u2212202\n(a) (b)WithoutBN (c)With BN\nFigure 1: (a) The test accuracy of the MNIST network\ntrained with and without Batch Normalization, vs. the\nnumber of training steps. ", "start_char_idx": 0, "end_char_idx": 219, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "15d23f05-58ab-4ae9-92c7-9ea57abb2ae3": {"__data__": {"id_": "15d23f05-58ab-4ae9-92c7-9ea57abb2ae3", "embedding": null, "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "defb123a-181a-439f-9b2a-c8baddeb016e", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "542d2ebc712807be19d69dc83b63126983d3aaaba4a523b2bb078429a62e8588", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "8c2c7fe2-0b52-4c53-81be-65d29a7cb679", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b1afd56513692b317ef9d86b9ca4217f0a6a5fb1e7b233805ed2b61f00aecf83", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "e01a3c34-8381-4a3d-aad8-4d3de10bd55f", "node_type": "1", "metadata": {}, "hash": "458752a7aedc2adc2391aad1f3abdc822167c75c9f570d18abd33a292cd99363", "class_name": "RelatedNodeInfo"}}, "text": "Batch Normalization helps the\nnetwork train faster and achieve higher accuracy. (b,\nc)The evolution of input distributions to a typical sig-\nmoid,overthecourseoftraining,shownas {15,50,85}th\npercentiles. Batch Normalization makes the distribution\nmorestableandreducestheinternalcovariateshift.\n3fully-connectedhiddenlayerswith100activationseach.\n", "start_char_idx": 219, "end_char_idx": 565, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e01a3c34-8381-4a3d-aad8-4d3de10bd55f": {"__data__": {"id_": "e01a3c34-8381-4a3d-aad8-4d3de10bd55f", "embedding": null, "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "defb123a-181a-439f-9b2a-c8baddeb016e", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "542d2ebc712807be19d69dc83b63126983d3aaaba4a523b2bb078429a62e8588", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "15d23f05-58ab-4ae9-92c7-9ea57abb2ae3", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fe85b2dec32804668809c04e83c6101883c7e74867fb62ca340cc951bd96d445", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "8040ef1b-6d72-49bd-a743-013c777a8bc0", "node_type": "1", "metadata": {}, "hash": "602b089b439125d68fb0bf3d59024701d1037031cbd3f9de336a04857b3daa20", "class_name": "RelatedNodeInfo"}}, "text": "Eachhiddenlayercomputes y =g(Wu+b)withsigmoid\nnonlinearity, and the weights Winitialized to small ran-\ndom Gaussian values. The last hidden layer is followed\nby a fully-connected layer with 10 activations (one per\nclass) and cross-entropyloss. We trained the network for\n50000steps, with 60 examplespermini-batch. We added\nBatchNormalizationtoeachhiddenlayerofthenetwork,\nas in Sec. ", "start_char_idx": 565, "end_char_idx": 948, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "8040ef1b-6d72-49bd-a743-013c777a8bc0": {"__data__": {"id_": "8040ef1b-6d72-49bd-a743-013c777a8bc0", "embedding": null, "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "defb123a-181a-439f-9b2a-c8baddeb016e", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "542d2ebc712807be19d69dc83b63126983d3aaaba4a523b2bb078429a62e8588", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "e01a3c34-8381-4a3d-aad8-4d3de10bd55f", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "05aaf00d208a1e422e5acdfd67c245231de26c7b5319e94c561f252054141c72", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "d0f23a06-1048-4e37-9e50-ca0035756ec6", "node_type": "1", "metadata": {}, "hash": "4da86f490ee500cb6ced12240dc6760c94f200f5928afb819b0c8ce55d7fe253", "class_name": "RelatedNodeInfo"}}, "text": "3.1. We were interested in the comparison be-\ntweenthebaselineandbatch-normalizednetworks,rather\nthanachievingthestateoftheartperformanceonMNIST\n(whichthe describedarchitecturedoesnot).\nFigure 1(a) shows the fraction of correct predictions\nby the two networks on held-out test data, as training\nprogresses. The batch-normalized network enjoys the\nhigher test accuracy. To investigate why, we studied in-\nputs to the sigmoid, in the original network Nand batch-\nnormalizednetwork Ntr\nBN(Alg.2)overthecourseoftrain-\ning. InFig.1(b,c)weshow,foronetypicalactivationfrom\nthe last hidden layer of each network, how its distribu-\ntion evolves. The distributions in the original network\nchange signi\ufb01cantly over time, both in their mean and\nthe variance, which complicates the training of the sub-\nsequent layers. In contrast, the distributions in the batch -\nnormalizednetworkaremuchmorestableastrainingpro-\ngresses,whichaidsthe training.\n4.2 ImageNetclassi\ufb01cation\nWe applied Batch Normalization to a new variant of the\nInception network (Szegedyetal., 2014), trained on the\nImageNet classi\ufb01cation task (Russakovskyet al., 2014).\nThe network has a large number of convolutional and\npooling layers, with a softmax layer to predict the image\nclass, out of 1000 possibilities. Convolutional layers use\nReLU asthenonlinearity. Themaindifferenceto thenet-\nwork described in (Szegedyet al., 2014) is that the 5\u00d75\nconvolutionallayersare replacedby two consecutivelay-\ners of3\u00d73convolutionswith up to 128\ufb01lters. The net-\nwork contains 13.6\u00b7106parameters, and, other than the\ntop softmax layer, has no fully-connected layers. Moredetails are given in the Appendix. ", "start_char_idx": 948, "end_char_idx": 2597, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "d0f23a06-1048-4e37-9e50-ca0035756ec6": {"__data__": {"id_": "d0f23a06-1048-4e37-9e50-ca0035756ec6", "embedding": null, "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "defb123a-181a-439f-9b2a-c8baddeb016e", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "542d2ebc712807be19d69dc83b63126983d3aaaba4a523b2bb078429a62e8588", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "8040ef1b-6d72-49bd-a743-013c777a8bc0", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "4fbc6ad5867b4893855fbcca008106baa59d145d72a240cb8329be7aaa6e0afc", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "282e536f-fe88-4f7a-a179-754220179b24", "node_type": "1", "metadata": {}, "hash": "bc73cc9e4b5cfebb8b59890dd65e5ac788c0dfded5516fb90fe76640a80ecd31", "class_name": "RelatedNodeInfo"}}, "text": "We refer to this model\nasInception intherestofthetext. Themodelwastrained\nusing a version of Stochastic Gradient Descent with mo-\nmentum(Sutskeveretal.,2013),usingthemini-batchsize\nof32. Thetrainingwasperformedusingalarge-scale,dis-\ntributed architecture (similar to (Deanet al., 2012)). All\nnetworksare evaluatedastrainingprogressesbycomput-\ning the validation accuracy @1, i.e. the probability of\npredicting the correct label out of 1000 possibilities, on\naheld-outset,usinga singlecropperimage.\nInourexperiments,weevaluatedseveralmodi\ufb01cations\nofInceptionwithBatchNormalization. Inallcases,Batch\nNormalizationwasappliedtotheinputofeachnonlinear-\nity, in a convolutional way, as described in section 3.2,\nwhilekeepingtherestofthearchitectureconstant.\n4.2.1 AcceleratingBN Networks\nSimplyaddingBatchNormalizationtoanetworkdoesnot\ntake full advantage of our method. To do so, we further\nchanged the network and its training parameters, as fol-\nlows:\nIncrease learning rate. In a batch-normalized model,\nwe have been able to achieve a training speedup from\nhigherlearningrates,with noill sideeffects(Sec.3.3).\nRemoveDropout. ", "start_char_idx": 2597, "end_char_idx": 3720, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "282e536f-fe88-4f7a-a179-754220179b24": {"__data__": {"id_": "282e536f-fe88-4f7a-a179-754220179b24", "embedding": null, "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "defb123a-181a-439f-9b2a-c8baddeb016e", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "542d2ebc712807be19d69dc83b63126983d3aaaba4a523b2bb078429a62e8588", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "d0f23a06-1048-4e37-9e50-ca0035756ec6", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "337a4eb6f4d8a41034a11178e27d97f95fb25d446ec9bdfed3c48cbe28449728", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "b5e4fffe-eede-4a33-ade6-23fcded34c28", "node_type": "1", "metadata": {}, "hash": "d85aa62eb365c6e3757dd0c89a1b33b1d87f7a5e670ec94d991f6ce47e9aea88", "class_name": "RelatedNodeInfo"}}, "text": "As describedin Sec. 3.4, Batch Nor-\nmalizationful\ufb01llssomeofthesamegoalsasDropout. Re-\nmoving Dropout from Modi\ufb01ed BN-Inception speeds up\ntraining,withoutincreasingover\ufb01tting.\nReduce the L2weight regularization. While in Incep-\ntion anL2loss on the model parameters controls over\ufb01t-\nting, in Modi\ufb01ed BN-Inception the weight of this loss is\nreduced by a factor of 5. We \ufb01nd that this improves the\naccuracyontheheld-outvalidationdata.\nAccelerate the learning rate decay. In training Incep-\ntion, learning rate was decayed exponentially. Because\nour network trains faster than Inception, we lower the\nlearningrate 6timesfaster.\nRemove Local Response Normalization While Incep-\ntion and other networks (Srivastavaet al., 2014) bene\ufb01t\nfrom it, we found that with Batch Normalization it is not\nnecessary.\nShuf\ufb02etrainingexamplesmorethoroughly. Weenabled\nwithin-shardshuf\ufb02ingofthetrainingdata,whichprevents\nthesameexamplesfromalwaysappearinginamini-batch\ntogether. This led to about 1% improvements in the val-\nidation accuracy, which is consistent with the view of\nBatch Normalization as a regularizer (Sec. 3.4): the ran-\ndomization inherent in our method should be most bene-\n\ufb01cialwhenitaffectsanexampledifferentlyeachtimeitis\nseen.\nReduce the photometric distortions. Because batch-\nnormalized networks train faster and observe each train-\ningexamplefewertimes,weletthetrainerfocusonmore\n\u201creal\u201dimagesbydistortingthemless.\n6", "start_char_idx": 3720, "end_char_idx": 5138, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "b5e4fffe-eede-4a33-ade6-23fcded34c28": {"__data__": {"id_": "b5e4fffe-eede-4a33-ade6-23fcded34c28", "embedding": null, "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e95d1ba8ae1a665d0b6cbee394fbbebf268bc8762e4c7b7ea1d167470376f3a7", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "282e536f-fe88-4f7a-a179-754220179b24", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "68cad5e0fa0b9cb32544333f236c6e2c818681cecc1e163dfa3b271d43b3259e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "e2ca30f9-a9f6-409f-8018-03870a276b86", "node_type": "1", "metadata": {}, "hash": "574ba339e1f3fd6abb7ce7e1afc8d42ad5ea786ad903ab602b7c7e0751a9a947", "class_name": "RelatedNodeInfo"}}, "text": "5M 10M 15M 20M 25M 30M0.40.50.60.70.8\nInception\nBN\u2212Baseline\nBN\u2212x5\nBN\u2212x30\nBN\u2212x5\u2212Sigmoid\nSteps to match Inception\nFigure 2: Single crop validation accuracy of Inception\nand its batch-normalized variants, vs. the number of\ntrainingsteps.Model Stepsto72.2% Maxaccuracy\nInception 31.0\u00b710672.2%\nBN-Baseline 13.3\u00b710672.7%\nBN-x5 2.1\u00b710673.0%\nBN-x30 2.7\u00b710674.8%\nBN-x5-Sigmoid 69.8%\nFigure 3: For Inception and the batch-normalized\nvariants, the number of training steps required to\nreach the maximum accuracy of Inception(72.2%),\nand the maximum accuracy achieved by the net-\nwork.\n4.2.2 Single-NetworkClassi\ufb01cation\nWe evaluated the following networks, all trained on the\nLSVRC2012 training data, and tested on the validation\ndata:\nInception : the network described at the beginning of\nSection4.2,trainedwiththeinitiallearningrateof0.001 5.\nBN-Baseline : Same as Inception with Batch Normal-\nizationbeforeeachnonlinearity.\n", "start_char_idx": 0, "end_char_idx": 915, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "e2ca30f9-a9f6-409f-8018-03870a276b86": {"__data__": {"id_": "e2ca30f9-a9f6-409f-8018-03870a276b86", "embedding": null, "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e95d1ba8ae1a665d0b6cbee394fbbebf268bc8762e4c7b7ea1d167470376f3a7", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "b5e4fffe-eede-4a33-ade6-23fcded34c28", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ce4e3dcb8c15078726ada683b9ea412d68c56d23e0dde61984a972a2aad5950a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "8a29b80e-3241-4249-8c79-3058f85811ed", "node_type": "1", "metadata": {}, "hash": "733b84ad93f189676c91c6ab54d95bd021a754f13e87ec365712f0066bcf40f6", "class_name": "RelatedNodeInfo"}}, "text": "BN-x5: Inception with Batch Normalization and the\nmodi\ufb01cations in Sec. 4.2.1. The initial learning rate was\nincreased by a factor of 5, to 0.0075. The same learning\nrateincreasewithoriginalInceptioncausedthemodelpa-\nrameterstoreachmachinein\ufb01nity.\nBN-x30: LikeBN-x5, but with the initial learning rate\n0.045(30timesthatofInception).\nBN-x5-Sigmoid : LikeBN-x5, but with sigmoid non-\nlinearityg(t) =1\n1+exp(\u2212x)instead of ReLU. We also at-\ntempted to train the original Inception with sigmoid, but\nthemodelremainedat theaccuracyequivalenttochance.\n", "start_char_idx": 915, "end_char_idx": 1459, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "8a29b80e-3241-4249-8c79-3058f85811ed": {"__data__": {"id_": "8a29b80e-3241-4249-8c79-3058f85811ed", "embedding": null, "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e95d1ba8ae1a665d0b6cbee394fbbebf268bc8762e4c7b7ea1d167470376f3a7", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "e2ca30f9-a9f6-409f-8018-03870a276b86", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "5b5b2b6282b576441f9e3a864ae84fe24d90d51014a640d99901bd6d1862208f", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "2069f9a4-c773-4744-a410-67dfc05b1dc5", "node_type": "1", "metadata": {}, "hash": "576a63b104701fc4ac9c967d6e7d47f3f4b9e06a07dc78eb9a6fc0e2912d0005", "class_name": "RelatedNodeInfo"}}, "text": "In Figure 2, we show the validation accuracy of the\nnetworks, as a function of the number of training steps.\nInception reached the accuracy of 72.2% after 31\u00b7106\ntraining steps. The Figure 3 shows, for each network,\nthe number of training steps required to reach the same\n72.2%accuracy,aswellasthemaximumvalidationaccu-\nracy reached by the network and the number of steps to\nreachit.\nBy onlyusingBatch Normalization( BN-Baseline ),we\nmatchtheaccuracyofInceptioninlessthanhalfthenum-\nber of training steps. By applying the modi\ufb01cations in\nSec. 4.2.1, we signi\ufb01cantly increase the training speed of\nthe network. BN-x5needs 14 times fewer steps than In-\nception to reach the 72.2% accuracy. Interestingly, in-\ncreasing the learning rate further ( BN-x30) causes the\nmodel to train somewhat slowerinitially, but allows it to\nreachahigher\ufb01nalaccuracy. Itreaches74.8%after 6\u00b7106\nsteps, i.e. 5 times fewer steps than required by Inception\ntoreach72.2%.\nWe also veri\ufb01ed that the reduction in internal covari-\nate shift allows deep networks with Batch Normalizationto be trained when sigmoid is used as the nonlinearity,\ndespite the well-known dif\ufb01culty of training such net-\nworks. Indeed, BN-x5-Sigmoid achieves the accuracy of\n69.8%. WithoutBatchNormalization,Inceptionwithsig-\nmoidneverachievesbetterthan 1/1000accuracy.\n4.2.3 Ensemble Classi\ufb01cation\nThe current reported best results on the ImageNet Large\nScale Visual RecognitionCompetitionare reachedby the\nDeep Image ensemble of traditional models (Wuet al.,\n2015) and the ensemble model of (Heet al., 2015). The\nlatterreportsthetop-5errorof4.94%,asevaluatedbythe\nILSVRCserver. ", "start_char_idx": 1459, "end_char_idx": 3085, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "2069f9a4-c773-4744-a410-67dfc05b1dc5": {"__data__": {"id_": "2069f9a4-c773-4744-a410-67dfc05b1dc5", "embedding": null, "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e95d1ba8ae1a665d0b6cbee394fbbebf268bc8762e4c7b7ea1d167470376f3a7", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "8a29b80e-3241-4249-8c79-3058f85811ed", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "2a3332e07c8b4384a5e09cd9baf05824754cd6f744efc084d523e8c344218d00", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "612ad73d-a6a1-434e-a9c1-cb0f768657ed", "node_type": "1", "metadata": {}, "hash": "a29f3132fd637a7d040bffd17450d239529d87b5a5a80ede9ad9cbe57a35ceef", "class_name": "RelatedNodeInfo"}}, "text": "Herewereportatop-5validationerrorof\n4.9%, and test error of 4.82% (according to the ILSVRC\nserver). This improvesupon the previousbest result, and\nexceedstheestimatedaccuracyofhumanratersaccording\nto(Russakovskyet al.,2014).\nForourensemble,weused6networks. Eachwasbased\nonBN-x30,modi\ufb01edviasomeofthefollowing: increased\ninitial weights in the convolutionallayers; using Dropout\n(with the Dropout probability of 5% or 10%, vs. 40%\nfor the original Inception); and using non-convolutional,\nper-activation Batch Normalization with last hidden lay-\ners of the model. Each network achieved its maximum\naccuracyafter about 6\u00b7106training steps. The ensemble\nprediction was based on the arithmetic average of class\nprobabilities predicted by the constituent networks. The\ndetailsofensembleandmulticropinferencearesimilarto\n(Szegedyet al., 2014).\n", "start_char_idx": 3085, "end_char_idx": 3922, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "612ad73d-a6a1-434e-a9c1-cb0f768657ed": {"__data__": {"id_": "612ad73d-a6a1-434e-a9c1-cb0f768657ed", "embedding": null, "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e95d1ba8ae1a665d0b6cbee394fbbebf268bc8762e4c7b7ea1d167470376f3a7", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "2069f9a4-c773-4744-a410-67dfc05b1dc5", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "448e25e6e5abb25955acdd85ff93893027a1d5d37ebf100a352d14e512b395d9", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "1bced689-bd10-4d07-8faf-c77ead5b12ca", "node_type": "1", "metadata": {}, "hash": "e760ce7c3f41593d1864491e0869ef6d73a6718496ba601554b5333746e2b1ff", "class_name": "RelatedNodeInfo"}}, "text": "We demonstrate in Fig. 4 that batch normalization al-\nlowsusto set new state-of-the-artby a healthymarginon\ntheImageNetclassi\ufb01cationchallengebenchmarks.\n5 Conclusion\nWe have presented a novel mechanism for dramatically\naccelerating the training of deep networks. It is based on\nthe premise that covariate shift, which is known to com-\nplicate the trainingof machine learning systems, also ap-\n7", "start_char_idx": 3922, "end_char_idx": 4316, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "1bced689-bd10-4d07-8faf-c77ead5b12ca": {"__data__": {"id_": "1bced689-bd10-4d07-8faf-c77ead5b12ca", "embedding": null, "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "55f2872f-47ad-45b9-92b9-21548eaf276e", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a7a42a1d4edb917dc90745159960a923acd823566bab41f79286e7c9b4ac43a6", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "612ad73d-a6a1-434e-a9c1-cb0f768657ed", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "4e66e924330425b7075d15666249089f298e839f4728f88c8f1d692a92255ab8", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "afe7667f-e82d-42fb-8fde-0cb1954cf5d1", "node_type": "1", "metadata": {}, "hash": "83db96928cb53380282debefaeb0e57263c80f6aa0aebee246d4b60f3fa47d52", "class_name": "RelatedNodeInfo"}}, "text": "Model Resolution Crops Models Top-1error Top-5error\nGoogLeNetensemble 224 144 7 - 6.67%\nDeepImagelow-res 256 - 1 - 7.96%\nDeepImagehigh-res 512 - 1 24.88 7.42%\nDeepImageensemble variable - - - 5.98%\nBN-Inceptionsinglecrop 224 1 1 25.2% 7.82%\nBN-Inceptionmulticrop 224 144 1 21.99% 5.82%\nBN-Inceptionensemble 224 144 6 20.1% 4.9%*\nFigure 4: Batch-Normalized Inception comparison with previous stat e of the art on the provided validation set com-\nprising50000images. *BN-Inceptionensemblehasreached4 .82%top-5erroronthe100000imagesofthetestsetof\ntheImageNetasreportedbythe test server.\n", "start_char_idx": 0, "end_char_idx": 584, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "afe7667f-e82d-42fb-8fde-0cb1954cf5d1": {"__data__": {"id_": "afe7667f-e82d-42fb-8fde-0cb1954cf5d1", "embedding": null, "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "55f2872f-47ad-45b9-92b9-21548eaf276e", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a7a42a1d4edb917dc90745159960a923acd823566bab41f79286e7c9b4ac43a6", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "1bced689-bd10-4d07-8faf-c77ead5b12ca", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6e9d6dc7a3209b90923ecae233a451fdbe1a6e75b97027ba106c1e45befd0459", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "41e711b7-ffd9-4a80-9d2f-3b24107c2763", "node_type": "1", "metadata": {}, "hash": "601d011e79a408390222a28f55bea9c2f126ad5dfbbe04c2e6a0fc4dc41be1a0", "class_name": "RelatedNodeInfo"}}, "text": "plies to sub-networks and layers, and removing it from\ninternal activations of the network may aid in training.\nOur proposed method draws its power from normalizing\nactivations, and from incorporating this normalization in\nthe network architecture itself. This ensures that the nor-\nmalization is appropriately handled by any optimization\nmethod that is being used to train the network. To en-\nable stochastic optimization methods commonly used in\ndeep network training, we perform the normalization for\neachmini-batch,andbackpropagatethegradientsthrough\nthe normalization parameters. Batch Normalization adds\nonly two extra parameters per activation, and in doing so\npreserves the representation ability of the network. We\npresentedanalgorithmforconstructing,training,andper -\nforming inference with batch-normalized networks. The\nresulting networks can be trained with saturating nonlin-\nearities, are more tolerant to increased training rates, an d\noftendonotrequireDropoutforregularization.\nMerely adding Batch Normalization to a state-of-the-\nartimageclassi\ufb01cationmodelyieldsasubstantialspeedup\nin training. By further increasing the learning rates, re-\nmoving Dropout, and applying other modi\ufb01cations af-\nforded by Batch Normalization, we reach the previous\nstate of the art with onlya small fractionof trainingsteps\n\u2013andthenbeatthestateoftheartinsingle-networkimage\nclassi\ufb01cation. Furthermore, by combining multiple mod-\nels trained with Batch Normalization, we perform better\nthanthebestknownsystemonImageNet,byasigni\ufb01cant\nmargin.\nInterestingly, our method bears similarity to the stan-\ndardization layer of (G\u00a8 ulc \u00b8ehre& Bengio, 2013), though\nthe two methodsstem from very differentgoals, and per-\nform different tasks. The goal of Batch Normalization\nis to achieve a stable distribution of activation values\nthroughout training, and in our experiments we apply it\nbefore the nonlinearity since that is where matching the\n\ufb01rst and second moments is more likely to result in a\nstable distribution. On the contrary,(G\u00a8 ulc \u00b8ehre&Bengio ,\n2013) apply the standardizationlayer to the outputof the\nnonlinearity, which results in sparser activations. In our\nlarge-scaleimage classi\ufb01cation experiments,we havenot\nobservedthenonlinearity inputstobesparse,neitherwith\nnor without Batch Normalization. Other notable differ-entiating characteristics of Batch Normalization include\nthe learned scale and shift that allow the BN transform\nto representidentity (the standardizationlayer did not re -\nquirethissinceitwasfollowedbythelearnedlineartrans-\nform that, conceptually, absorbs the necessary scale and\nshift), handling of convolutional layers, deterministic i n-\nferencethatdoesnotdependonthemini-batch,andbatch-\nnormalizingeachconvolutionallayerin thenetwork.\nIn this work, we have not explored the full range of\npossibilitiesthatBatchNormalizationpotentiallyenabl es.\nOur future work includes applications of our method to\nRecurrent Neural Networks (Pascanuet al., 2013), where\ntheinternalcovariateshiftandthevanishingorexploding\ngradients may be especially severe, and which would al-\nlowustomorethoroughlytestthehypothesisthatnormal-\nizationimprovesgradientpropagation(Sec.3.3). Weplan\ntoinvestigatewhetherBatchNormalizationcanhelpwith\ndomain adaptation, in its traditional sense \u2013 i.e. whether\nthe normalization performed by the network would al-\nlow it to more easily generalize to new data distribu-\ntions,perhapswithjustarecomputationofthepopulation\nmeansandvariances(Alg.2). Finally,webelievethatfur-\nthertheoreticalanalysisofthealgorithmwouldallowstil l\nmoreimprovementsandapplications.\nReferences\nBengio, Yoshua and Glorot, Xavier. Understanding the\ndif\ufb01cultyoftrainingdeepfeedforwardneuralnetworks.\nInProceedings of AISTATS 2010 , volume 9, pp. ", "start_char_idx": 584, "end_char_idx": 4344, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "41e711b7-ffd9-4a80-9d2f-3b24107c2763": {"__data__": {"id_": "41e711b7-ffd9-4a80-9d2f-3b24107c2763", "embedding": null, "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "55f2872f-47ad-45b9-92b9-21548eaf276e", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a7a42a1d4edb917dc90745159960a923acd823566bab41f79286e7c9b4ac43a6", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "afe7667f-e82d-42fb-8fde-0cb1954cf5d1", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a8f4dda62982dac0fd4f0e9647019b3e17cdc6a880e77ac15e1ea873efafea64", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "9afdc6f5-b962-4b05-9964-28346b31319d", "node_type": "1", "metadata": {}, "hash": "98e1d81e288128e9003eb2863b6892dc118876e13b108e95e95af7369da78b5f", "class_name": "RelatedNodeInfo"}}, "text": "249\u2013\n256,May2010.\nDean,Jeffrey,Corrado,GregS.,Monga,Rajat,Chen,Kai,\nDevin,Matthieu,Le,QuocV., Mao,MarkZ.,Ranzato,\nMarc\u2019Aurelio,Senior,Andrew,Tucker,Paul,Yang,Ke,\nand Ng, Andrew Y. Large scale distributed deep net-\nworks. In NIPS,2012.\n", "start_char_idx": 4344, "end_char_idx": 4579, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "9afdc6f5-b962-4b05-9964-28346b31319d": {"__data__": {"id_": "9afdc6f5-b962-4b05-9964-28346b31319d", "embedding": null, "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "55f2872f-47ad-45b9-92b9-21548eaf276e", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a7a42a1d4edb917dc90745159960a923acd823566bab41f79286e7c9b4ac43a6", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "41e711b7-ffd9-4a80-9d2f-3b24107c2763", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3a90b5c06bcb8aa84e7a8e1fe848e63012a968fe47a11b799f462e3df1e56eb7", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "356b6eab-f375-4784-a9a9-b346e1d248f9", "node_type": "1", "metadata": {}, "hash": "a6df667172dab19187e639df99386071a4507cad216aa812eed538f2c290bc47", "class_name": "RelatedNodeInfo"}}, "text": "Desjardins, Guillaume and Kavukcuoglu,Koray. Natural\nneuralnetworks. (unpublished).\n", "start_char_idx": 4579, "end_char_idx": 4663, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "356b6eab-f375-4784-a9a9-b346e1d248f9": {"__data__": {"id_": "356b6eab-f375-4784-a9a9-b346e1d248f9", "embedding": null, "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "55f2872f-47ad-45b9-92b9-21548eaf276e", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a7a42a1d4edb917dc90745159960a923acd823566bab41f79286e7c9b4ac43a6", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "9afdc6f5-b962-4b05-9964-28346b31319d", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "29e1bb31e456e1fd8fa3330301c5ba66e57efa8e218c991ff57daabc4d8edb64", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "5fffd5af-e267-451e-bdef-ce3e373ee153", "node_type": "1", "metadata": {}, "hash": "75948197614b60201027c6ababcfc412cb02c84cfcd3f25c309ba5e23667fb20", "class_name": "RelatedNodeInfo"}}, "text": "Duchi, John, Hazan, Elad, and Singer, Yoram. Adaptive\nsubgradientmethodsfor onlinelearning and stochastic\n8", "start_char_idx": 4663, "end_char_idx": 4770, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "5fffd5af-e267-451e-bdef-ce3e373ee153": {"__data__": {"id_": "5fffd5af-e267-451e-bdef-ce3e373ee153", "embedding": null, "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b3e7c8aaf2c6c876c71fcab51900e6ecdf9d776f1c4c9385ec34aadacd8b0e40", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "356b6eab-f375-4784-a9a9-b346e1d248f9", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "5a21f50f8a04ce6f841393d90b70bbe9b4b5fd3ee85c723b0b6959a547e17eaf", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "6ef9eca0-2c8a-4e3e-85d4-d052726b1d8f", "node_type": "1", "metadata": {}, "hash": "a8cd34f0409a57272616af1520ccc4962b6954b516d1e34fca8a9975a6f04773", "class_name": "RelatedNodeInfo"}}, "text": "optimization. J.Mach.Learn.Res. ,12:2121\u20132159,July\n2011. ISSN1532-4435.\nG\u00a8 ulc \u00b8ehre, C \u00b8aglar and Bengio, Yoshua. Knowledge mat-\nters: Importanceof prior informationfor optimization.\nCoRR,abs/1301.4083,2013.\nHe, K., Zhang, X., Ren, S., and Sun, J. Delving Deep\ninto Recti\ufb01ers: Surpassing Human-Level Performance\non ImageNet Classi\ufb01cation. ArXiv e-prints , February\n2015.\nHyv\u00a8 arinen, A. and Oja, E. Independentcomponent anal-\nysis: Algorithms and applications. Neural Netw. , 13\n(4-5):411\u2013430,May2000.\nJiang, Jing. A literature survey on domain adaptation of\nstatistical classi\ufb01ers, 2008.\nLeCun, Y., Bottou, L., Bengio, Y., and Haffner, P.\nGradient-based learning applied to document recog-\nnition.Proceedings of the IEEE , 86(11):2278\u20132324,\nNovember1998a.\nLeCun, Y., Bottou, L., Orr, G., and Muller, K. Ef\ufb01cient\nbackprop. InOrr,G.andK.,Muller(eds.), NeuralNet-\nworks: Tricks ofthetrade .Springer,1998b.\nLyu, S and Simoncelli, E P. Nonlinear image representa-\ntion using divisive normalization. In Proc. Computer\nVision and Pattern Recognition , pp. 1\u20138. IEEE Com-\nputer Society, Jun 23-28 2008. doi: 10.1109/CVPR.\n2008.4587821.\nNair,VinodandHinton,GeoffreyE. Recti\ufb01edlinearunits\nimprove restricted boltzmann machines. In ICML, pp.\n", "start_char_idx": 0, "end_char_idx": 1233, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "6ef9eca0-2c8a-4e3e-85d4-d052726b1d8f": {"__data__": {"id_": "6ef9eca0-2c8a-4e3e-85d4-d052726b1d8f", "embedding": null, "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b3e7c8aaf2c6c876c71fcab51900e6ecdf9d776f1c4c9385ec34aadacd8b0e40", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "5fffd5af-e267-451e-bdef-ce3e373ee153", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "0125b97faa3a21500503bcd9c08f5156102f9490b91608d41ad74c4ecfa6297a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "124d2351-91ef-48a0-b5b0-eb5a7e3b0a78", "node_type": "1", "metadata": {}, "hash": "d608315dd7f2d9a7e852fdd6e9f966b7771ff7f74a79c470049bb7ced92214bb", "class_name": "RelatedNodeInfo"}}, "text": "807\u2013814.Omnipress,2010.\nPascanu, Razvan, Mikolov, Tomas, and Bengio, Yoshua.\nOn the dif\ufb01culty of training recurrent neural networks.\nInProceedingsofthe30thInternationalConferenceon\nMachineLearning,ICML 2013,Atlanta,GA,USA,16-\n21June2013 ,pp.1310\u20131318,2013.\nPovey, Daniel, Zhang, Xiaohui, and Khudanpur, San-\njeev. Parallel training of deep neural networks with\nnatural gradient and parameter averaging. CoRR,\nabs/1410.7455,2014.\nRaiko, Tapani, Valpola, Harri, and LeCun, Yann. Deep\nlearning made easier by linear transformations in per-\nceptrons. In International Conference on Arti\ufb01cial In-\ntelligenceandStatistics(AISTATS) ,pp.924\u2013932,2012.\nRussakovsky,Olga,Deng,Jia,Su,Hao,Krause,Jonathan,\nSatheesh, Sanjeev, Ma, Sean, Huang, Zhiheng, Karpa-\nthy, Andrej, Khosla, Aditya, Bernstein, Michael, Berg,\nAlexander C., and Fei-Fei, Li. ImageNet Large Scale\nVisual RecognitionChallenge,2014.Saxe, Andrew M., McClelland, James L., and Ganguli,\nSurya. Exact solutions to the nonlinear dynamics\nof learning in deep linear neural networks. CoRR,\nabs/1312.6120,2013.\n", "start_char_idx": 1233, "end_char_idx": 2289, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "124d2351-91ef-48a0-b5b0-eb5a7e3b0a78": {"__data__": {"id_": "124d2351-91ef-48a0-b5b0-eb5a7e3b0a78", "embedding": null, "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b3e7c8aaf2c6c876c71fcab51900e6ecdf9d776f1c4c9385ec34aadacd8b0e40", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "6ef9eca0-2c8a-4e3e-85d4-d052726b1d8f", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "779d1528dcec97358a3fa5f23046f233b6c261856ef6d1fcd1a4e6e839476553", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "1b37c06a-7824-45fd-8f61-d101b5a9df19", "node_type": "1", "metadata": {}, "hash": "0fc7d06bb03bec5f721713a1033bf61bae39d146652e28d5b37e0cadf643db55", "class_name": "RelatedNodeInfo"}}, "text": "Shimodaira, Hidetoshi. Improving predictive inference\nunder covariate shift by weighting the log-likelihood\nfunction. JournalofStatisticalPlanningandInference ,\n90(2):227\u2013244,October2000.\n", "start_char_idx": 2289, "end_char_idx": 2477, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "1b37c06a-7824-45fd-8f61-d101b5a9df19": {"__data__": {"id_": "1b37c06a-7824-45fd-8f61-d101b5a9df19", "embedding": null, "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b3e7c8aaf2c6c876c71fcab51900e6ecdf9d776f1c4c9385ec34aadacd8b0e40", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "124d2351-91ef-48a0-b5b0-eb5a7e3b0a78", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "476c9198f0b338b489bbf71a984c347f5981ca9fa093c928b3ed320d89df1211", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "8b1fa614-496b-43af-993f-498d707efb50", "node_type": "1", "metadata": {}, "hash": "ef0e7da658364cd5c3f38d0f2ec91f91638b9575fdaaa209bf7a96413e90d400", "class_name": "RelatedNodeInfo"}}, "text": "Srivastava, Nitish, Hinton, Geoffrey, Krizhevsky, Alex,\nSutskever, Ilya, and Salakhutdinov, Ruslan. Dropout:\nA simple way to preventneural networksfrom over\ufb01t-\nting.J. Mach. ", "start_char_idx": 2477, "end_char_idx": 2651, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "8b1fa614-496b-43af-993f-498d707efb50": {"__data__": {"id_": "8b1fa614-496b-43af-993f-498d707efb50", "embedding": null, "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b3e7c8aaf2c6c876c71fcab51900e6ecdf9d776f1c4c9385ec34aadacd8b0e40", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "1b37c06a-7824-45fd-8f61-d101b5a9df19", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fb215b41abeba0959faad13b87990e002b2b0ff18c6a11f58746800e049f8e61", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "7124a746-c0f8-4291-88f5-da1376271eb5", "node_type": "1", "metadata": {}, "hash": "f340b94c73c985e6d458758ecfde1833de3ddd6e0202b3b7aba161af2bf3b033", "class_name": "RelatedNodeInfo"}}, "text": "Learn. ", "start_char_idx": 2651, "end_char_idx": 2658, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "7124a746-c0f8-4291-88f5-da1376271eb5": {"__data__": {"id_": "7124a746-c0f8-4291-88f5-da1376271eb5", "embedding": null, "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b3e7c8aaf2c6c876c71fcab51900e6ecdf9d776f1c4c9385ec34aadacd8b0e40", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "8b1fa614-496b-43af-993f-498d707efb50", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "16ad24743741abee31f0dc2cc3941345ed47e0809edc75e056b78cd3cc5c6274", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "5979ee85-867e-4749-b395-408a98e32df4", "node_type": "1", "metadata": {}, "hash": "56fc7c7a1e3cb94290fed23535119b19daa02223853b6d567c415fa30940406d", "class_name": "RelatedNodeInfo"}}, "text": "Res. ", "start_char_idx": 27, "end_char_idx": 32, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "5979ee85-867e-4749-b395-408a98e32df4": {"__data__": {"id_": "5979ee85-867e-4749-b395-408a98e32df4", "embedding": null, "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b3e7c8aaf2c6c876c71fcab51900e6ecdf9d776f1c4c9385ec34aadacd8b0e40", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "7124a746-c0f8-4291-88f5-da1376271eb5", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "baa95f8666ed8ec5fdc1d41cc7ecb8fff97570b650fc04273f0e88af51e425ce", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "fcc49761-9190-43c7-b273-101c413fb492", "node_type": "1", "metadata": {}, "hash": "7e6189ad7574e60476a01c04448da769573cc2141201fb40965bbe8e38488852", "class_name": "RelatedNodeInfo"}}, "text": ", 15(1):1929\u20131958, January\n2014.\nSutskever, Ilya, Martens, James, Dahl, George E., and\nHinton, Geoffrey E. On the importance of initial-\nization and momentum in deep learning. In ICML\n(3), volume 28 of JMLR Proceedings , pp. 1139\u20131147.\nJMLR.org,2013.\nSzegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet,\nPierre, Reed, Scott, Anguelov, Dragomir, Erhan, Du-\nmitru, Vanhoucke, Vincent, and Rabinovich, An-\ndrew. Going deeper with convolutions. CoRR,\nabs/1409.4842,2014.\nWiesler, Simon and Ney, Hermann. A convergenceanal-\nysis of log-lineartraining. In Shawe-Taylor,J., Zemel,\nR.S.,Bartlett,P.,Pereira,F.C.N.,andWeinberger,K.Q.\n", "start_char_idx": 2663, "end_char_idx": 3290, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "fcc49761-9190-43c7-b273-101c413fb492": {"__data__": {"id_": "fcc49761-9190-43c7-b273-101c413fb492", "embedding": null, "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b3e7c8aaf2c6c876c71fcab51900e6ecdf9d776f1c4c9385ec34aadacd8b0e40", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "5979ee85-867e-4749-b395-408a98e32df4", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b1cc529152403413241e213ac736e951831bbdc314ee74ef7a806438f8950711", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "97c2d55b-945c-4c01-9dc0-86bd520257d0", "node_type": "1", "metadata": {}, "hash": "af0084e44ed3001688a07de3781019d8d256c29c4f23121f112c47b91d7712fd", "class_name": "RelatedNodeInfo"}}, "text": "(eds.),AdvancesinNeuralInformationProcessingSys-\ntems24,pp.657\u2013665,Granada,Spain,December2011.\nWiesler, Simon, Richard, Alexander, Schl\u00a8 uter, Ralf, and\nNey, Hermann. Mean-normalized stochastic gradient\nfor large-scale deep learning. In IEEE International\nConference on Acoustics, Speech, and Signal Process-\ning,pp.180\u2013184,Florence,Italy,May2014.\nWu, Ren, Yan, Shengen, Shan, Yi, Dang, Qingqing, and\nSun,Gang. Deepimage: Scalingupimagerecognition,\n2015.\nAppendix\nVariantofthe Inception Model Used\nFigure 5 documents the changes that were performed\ncompared to the architecture with respect to the\nGoogleNet archictecture. For the interpretation of this\ntable, please consult (Szegedyetal., 2014). The notable\narchitecture changes compared to the GoogLeNet model\ninclude:\n\u2022The 5\u00d75 convolutional layers are replaced by two\nconsecutive 3\u00d73 convolutional layers. This in-\ncreases the maximum depth of the network by 9\n9", "start_char_idx": 3290, "end_char_idx": 4206, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "97c2d55b-945c-4c01-9dc0-86bd520257d0": {"__data__": {"id_": "97c2d55b-945c-4c01-9dc0-86bd520257d0", "embedding": null, "metadata": {"page_label": "10", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "4a57759d-5ee3-4c11-8861-9bcad409f797", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7140d4bd4f57c2d093b60730facef063c99d44b721e69daa9818794391a0264e", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "fcc49761-9190-43c7-b273-101c413fb492", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c70af929ec42a31913065dfe5a801e860551412546424915ac47c385af840428", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "6a2b5ec3-b6df-4951-9797-fac4f6881808", "node_type": "1", "metadata": {}, "hash": "b2c5afc883f9e72b4015a2cc5286e0038f120902fd5ae7ddcb077b4be01ee24b", "class_name": "RelatedNodeInfo"}}, "text": "weight layers. Also it increases the number of pa-\nrameters by 25% and the computational cost is in-\ncreasedbyabout30%.\n\u2022The number 28\u00d728 inception modules is increased\nfrom2to 3.\n\u2022Inside the modules, sometimes average, sometimes\nmaximum-poolingis employed. This is indicated in\ntheentriescorrespondingtothepoolinglayersofthe\ntable.\n\u2022There are no across the board pooling layers be-\ntween any two Inception modules, but stride-2 con-\nvolution/pooling layers are employed before the \ufb01l-\nterconcatenationin themodules3c,4e.\nOur model employed separable convolution with depth\nmultiplier 8on the \ufb01rst convolutionallayer. This reduces\nthecomputationalcost while increasingthememorycon-\nsumptionat trainingtime.\n", "start_char_idx": 0, "end_char_idx": 707, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "6a2b5ec3-b6df-4951-9797-fac4f6881808": {"__data__": {"id_": "6a2b5ec3-b6df-4951-9797-fac4f6881808", "embedding": null, "metadata": {"page_label": "10", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "4a57759d-5ee3-4c11-8861-9bcad409f797", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7140d4bd4f57c2d093b60730facef063c99d44b721e69daa9818794391a0264e", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "97c2d55b-945c-4c01-9dc0-86bd520257d0", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "d3eb79886d6a4e5e44a9ea6f8a6e1e2834fb824b5af37a63ee29f7d6f7ce1401", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "38ac2daf-0a08-4e26-b28c-cc5f9a69f395", "node_type": "1", "metadata": {}, "hash": "c56a357b6ce32b1105014af797153af4355ad7339a3d621953e7bd24e250c785", "class_name": "RelatedNodeInfo"}}, "text": "10", "start_char_idx": 707, "end_char_idx": 709, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "38ac2daf-0a08-4e26-b28c-cc5f9a69f395": {"__data__": {"id_": "38ac2daf-0a08-4e26-b28c-cc5f9a69f395", "embedding": null, "metadata": {"page_label": "11", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "597b8534-da28-40e6-af8b-49dc5f661cad", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "70d4592261cdda5ef9025cb2d6f6e59d3b5b84174cde2260476d5360b6c97f35", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "6a2b5ec3-b6df-4951-9797-fac4f6881808", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3ac213c6f19275b96c8066f3305c620b9e7136bf65feacd99ffa4784b9fe1d49", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "b7c64a3c-4735-4672-815f-7cf7cd07c193", "node_type": "1", "metadata": {}, "hash": "0d4418137fc6355cd500426905e7de7140bf20d9f8a6fb91f4970a82aced69be", "class_name": "RelatedNodeInfo"}}, "text": "typepatch size/\nstrideoutput\nsizedepth #1\u00d71#3\u00d73\nreduce#3\u00d73double#3\u00d73\nreducedouble\n#3\u00d73Pool+proj\nconvolution* 7\u00d77/2112\u00d7112\u00d7641\nmaxpool 3\u00d73/256\u00d756\u00d764 0\nconvolution 3\u00d73/156\u00d756\u00d7192 1 64 192\nmaxpool 3\u00d73/228\u00d728\u00d7192 0\ninception (3a) 28\u00d728\u00d7256 364 64 64 64 96 avg+ 32\ninception (3b) 28\u00d728\u00d7320 364 64 96 64 96 avg+ 64\ninception (3c) stride 2 28\u00d728\u00d7576 3 0128 160 64 96max +pass through\ninception (4a) 14\u00d714\u00d7576 3224 64 96 96 128 avg+ 128\ninception (4b) 14\u00d714\u00d7576 3192 96 128 96 128 avg+ 128\ninception (4c) 14\u00d714\u00d7576 3160 128 160 128 160 avg+ 128\ninception (4d) 14\u00d714\u00d7576 396 128 192 160 192 avg+ 128\ninception (4e) stride 2 14\u00d714\u00d71024 3 0128 192 192 256max +pass through\ninception (5a) 7\u00d77\u00d71024 3352 192 320 160 224 avg+ 128\ninception (5b) 7\u00d77\u00d71024 3352 192 320 192 224 max+ 128\navgpool 7\u00d77/11\u00d71\u00d71024 0\nFigure5: Inceptionarchitecture\n11", "start_char_idx": 0, "end_char_idx": 827, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "b7c64a3c-4735-4672-815f-7cf7cd07c193": {"__data__": {"id_": "b7c64a3c-4735-4672-815f-7cf7cd07c193", "embedding": null, "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7bcf88e4-6827-4be9-988b-55bb1b49d948", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ca4e49fe86b8da8044f500a128481893279d656bbf3dece51a9aa5073341cc0d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "38ac2daf-0a08-4e26-b28c-cc5f9a69f395", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "70d4592261cdda5ef9025cb2d6f6e59d3b5b84174cde2260476d5360b6c97f35", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "d5f956f2-7566-4ce2-b90c-9b1972c8a919", "node_type": "1", "metadata": {}, "hash": "b1e0c8d8656573e5e76dba4f8ffd3e7a1f33f5ce4b2cbe96dabd46c44ad82bd4", "class_name": "RelatedNodeInfo"}}, "text": "Provided proper attribution is provided, Google hereby grants permission to\nreproduce the tables and figures in this paper solely for use in journalistic or\nscholarly works.\nAttention Is All You Need\nAshish Vaswani\u2217\nGoogle Brain\navaswani@google.comNoam Shazeer\u2217\nGoogle Brain\nnoam@google.comNiki Parmar\u2217\nGoogle Research\nnikip@google.comJakob Uszkoreit\u2217\nGoogle Research\nusz@google.com\nLlion Jones\u2217\nGoogle Research\nllion@google.comAidan N. Gomez\u2217 \u2020\nUniversity of Toronto\naidan@cs.toronto.edu\u0141ukasz Kaiser\u2217\nGoogle Brain\nlukaszkaiser@google.com\nIllia Polosukhin\u2217 \u2021\nillia.polosukhin@gmail.com\nAbstract\nThe dominant sequence transduction models are based on complex recurrent or\nconvolutional neural networks that include an encoder and a decoder. The best\nperforming models also connect the encoder and decoder through an attention\nmechanism. We propose a new simple network architecture, the Transformer,\nbased solely on attention mechanisms, dispensing with recurrence and convolutions\nentirely. Experiments on two machine translation tasks show these models to\nbe superior in quality while being more parallelizable and requiring significantly\nless time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-\nto-German translation task, improving over the existing best results, including\nensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task,\nour model establishes a new single-model state-of-the-art BLEU score of 41.8 after\ntraining for 3.5 days on eight GPUs, a small fraction of the training costs of the\nbest models from the literature. We show that the Transformer generalizes well to\nother tasks by applying it successfully to English constituency parsing both with\nlarge and limited training data.\n\u2217Equal contribution. ", "start_char_idx": 0, "end_char_idx": 1758, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "d5f956f2-7566-4ce2-b90c-9b1972c8a919": {"__data__": {"id_": "d5f956f2-7566-4ce2-b90c-9b1972c8a919", "embedding": null, "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7bcf88e4-6827-4be9-988b-55bb1b49d948", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ca4e49fe86b8da8044f500a128481893279d656bbf3dece51a9aa5073341cc0d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "b7c64a3c-4735-4672-815f-7cf7cd07c193", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e5fad06d02d35abea195f627bac19ad00bca2e1036c8f8810f3b4ee531ca14c7", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "2741d814-901a-40c1-b959-552157129a9d", "node_type": "1", "metadata": {}, "hash": "14e592bbcb900896dff54bddde810f729c4223c17986db11fccf3512bd7d2cea", "class_name": "RelatedNodeInfo"}}, "text": "Listing order is random. Jakob proposed replacing RNNs with self-attention and started\nthe effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and\nhas been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head\nattention and the parameter-free position representation and became the other person involved in nearly every\ndetail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and\ntensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and\nefficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and\nimplementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating\nour research.\n\u2020Work performed while at Google Brain.\n", "start_char_idx": 1758, "end_char_idx": 2682, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "2741d814-901a-40c1-b959-552157129a9d": {"__data__": {"id_": "2741d814-901a-40c1-b959-552157129a9d", "embedding": null, "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7bcf88e4-6827-4be9-988b-55bb1b49d948", "node_type": "4", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ca4e49fe86b8da8044f500a128481893279d656bbf3dece51a9aa5073341cc0d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "d5f956f2-7566-4ce2-b90c-9b1972c8a919", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ff03134021863208563480fbef84ac81faeb0087f4be4f89a65e000553e9e08e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "2923d6a1-d583-496a-81ed-0ab01dd2aaaf", "node_type": "1", "metadata": {}, "hash": "b76f5cbda6706f433a99870c215cd0693b5fc8fc5b1c5374bbcab748006a73f1", "class_name": "RelatedNodeInfo"}}, "text": "\u2021Work performed while at Google Research.\n31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.arXiv:1706.03762v7  [cs.CL]  2 Aug 2023", "start_char_idx": 2682, "end_char_idx": 2853, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "2923d6a1-d583-496a-81ed-0ab01dd2aaaf": {"__data__": {"id_": "2923d6a1-d583-496a-81ed-0ab01dd2aaaf", "embedding": null, "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "0ec692ee-efde-4036-bc5b-51ba9d6eca6c", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "49b3abf2a4ad66160f00161741a1b57d04eb861fd7af8e1d38f6ccd91900278d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "2741d814-901a-40c1-b959-552157129a9d", "node_type": "1", "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c622a92b6bc9e23792425b98d86dd29232b783cd3e5c099cd76330a5790730db", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "bfa3ce6d-c588-4f78-90f0-f0245cc7e6ba", "node_type": "1", "metadata": {}, "hash": "39db70cbd822139819b32529ad79ccc5614a06bd7a95830e5a5911e54ac039f2", "class_name": "RelatedNodeInfo"}}, "text": "1 Introduction\nRecurrent neural networks, long short-term memory [ 13] and gated recurrent [ 7] neural networks\nin particular, have been firmly established as state of the art approaches in sequence modeling and\ntransduction problems such as language modeling and machine translation [ 35,2,5]. Numerous\nefforts have since continued to push the boundaries of recurrent language models and encoder-decoder\narchitectures [38, 24, 15].\nRecurrent models typically factor computation along the symbol positions of the input and output\nsequences. Aligning the positions to steps in computation time, they generate a sequence of hidden\nstates ht, as a function of the previous hidden state ht\u22121and the input for position t. This inherently\nsequential nature precludes parallelization within training examples, which becomes critical at longer\nsequence lengths, as memory constraints limit batching across examples. Recent work has achieved\nsignificant improvements in computational efficiency through factorization tricks [ 21] and conditional\ncomputation [ 32], while also improving model performance in case of the latter. The fundamental\nconstraint of sequential computation, however, remains.\nAttention mechanisms have become an integral part of compelling sequence modeling and transduc-\ntion models in various tasks, allowing modeling of dependencies without regard to their distance in\nthe input or output sequences [ 2,19]. In all but a few cases [ 27], however, such attention mechanisms\nare used in conjunction with a recurrent network.\nIn this work we propose the Transformer, a model architecture eschewing recurrence and instead\nrelying entirely on an attention mechanism to draw global dependencies between input and output.\nThe Transformer allows for significantly more parallelization and can reach a new state of the art in\ntranslation quality after being trained for as little as twelve hours on eight P100 GPUs.\n2 Background\nThe goal of reducing sequential computation also forms the foundation of the Extended Neural GPU\n[16], ByteNet [ 18] and ConvS2S [ 9], all of which use convolutional neural networks as basic building\nblock, computing hidden representations in parallel for all input and output positions. In these models,\nthe number of operations required to relate signals from two arbitrary input or output positions grows\nin the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. ", "start_char_idx": 0, "end_char_idx": 2434, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "bfa3ce6d-c588-4f78-90f0-f0245cc7e6ba": {"__data__": {"id_": "bfa3ce6d-c588-4f78-90f0-f0245cc7e6ba", "embedding": null, "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "0ec692ee-efde-4036-bc5b-51ba9d6eca6c", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "49b3abf2a4ad66160f00161741a1b57d04eb861fd7af8e1d38f6ccd91900278d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "2923d6a1-d583-496a-81ed-0ab01dd2aaaf", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "2461968fbfdb377e396e4fcd7628cdff5e9e23bfaf2994f850fef85b09157907", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "300e8947-8fa3-417d-9a79-9ae7c242f4ef", "node_type": "1", "metadata": {}, "hash": "3d5478f5af12a0260f05b9e221b5b229d887414b80201489bee91af10d773eee", "class_name": "RelatedNodeInfo"}}, "text": "This makes\nit more difficult to learn dependencies between distant positions [ 12]. ", "start_char_idx": 2434, "end_char_idx": 2518, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "300e8947-8fa3-417d-9a79-9ae7c242f4ef": {"__data__": {"id_": "300e8947-8fa3-417d-9a79-9ae7c242f4ef", "embedding": null, "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "0ec692ee-efde-4036-bc5b-51ba9d6eca6c", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "49b3abf2a4ad66160f00161741a1b57d04eb861fd7af8e1d38f6ccd91900278d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "bfa3ce6d-c588-4f78-90f0-f0245cc7e6ba", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7de9d0ce7f1f67f6746d3dd0073dcfce83e23a2d541737cfb5a5e72c7056933e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "93964b05-4478-455c-87d2-d75c334e4c09", "node_type": "1", "metadata": {}, "hash": "316ec4d731eae17aca2a5ad669da2323a90e5dda811f87fadc8ba763f44ba0e9", "class_name": "RelatedNodeInfo"}}, "text": "In the Transformer this is\nreduced to a constant number of operations, albeit at the cost of reduced effective resolution due\nto averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as\ndescribed in section 3.2.\nSelf-attention, sometimes called intra-attention is an attention mechanism relating different positions\nof a single sequence in order to compute a representation of the sequence. Self-attention has been\nused successfully in a variety of tasks including reading comprehension, abstractive summarization,\ntextual entailment and learning task-independent sentence representations [4, 27, 28, 22].\nEnd-to-end memory networks are based on a recurrent attention mechanism instead of sequence-\naligned recurrence and have been shown to perform well on simple-language question answering and\nlanguage modeling tasks [34].\nTo the best of our knowledge, however, the Transformer is the first transduction model relying\nentirely on self-attention to compute representations of its input and output without using sequence-\naligned RNNs or convolution. In the following sections, we will describe the Transformer, motivate\nself-attention and discuss its advantages over models such as [17, 18] and [9].\n3 Model Architecture\nMost competitive neural sequence transduction models have an encoder-decoder structure [ 5,2,35].\nHere, the encoder maps an input sequence of symbol representations (x1, ..., x n)to a sequence\nof continuous representations z= (z1, ..., z n). ", "start_char_idx": 2518, "end_char_idx": 4019, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "93964b05-4478-455c-87d2-d75c334e4c09": {"__data__": {"id_": "93964b05-4478-455c-87d2-d75c334e4c09", "embedding": null, "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "0ec692ee-efde-4036-bc5b-51ba9d6eca6c", "node_type": "4", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "49b3abf2a4ad66160f00161741a1b57d04eb861fd7af8e1d38f6ccd91900278d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "300e8947-8fa3-417d-9a79-9ae7c242f4ef", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ee9d9ba0f9c4dee86ba934c329cbd5c2561d61f5da47753803c99c6e53e586b4", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "d5bcec10-5d92-4d3b-93e1-07345863d62e", "node_type": "1", "metadata": {}, "hash": "ce1a545e91283af06ea36a8e0aa3ac75d56f32791ea80eb61dedb56180ed81d3", "class_name": "RelatedNodeInfo"}}, "text": "Given z, the decoder then generates an output\nsequence (y1, ..., y m)of symbols one element at a time. At each step the model is auto-regressive\n[10], consuming the previously generated symbols as additional input when generating the next.\n2", "start_char_idx": 4019, "end_char_idx": 4260, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "d5bcec10-5d92-4d3b-93e1-07345863d62e": {"__data__": {"id_": "d5bcec10-5d92-4d3b-93e1-07345863d62e", "embedding": null, "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "c8a07910-0192-481a-bcf1-1a58d71a8f38", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "43ae829564edc9b5ed17c359ae6edcaa482d4371370a9e7539468351fc9292ff", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "93964b05-4478-455c-87d2-d75c334e4c09", "node_type": "1", "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8f3f71fcecd60992e39150337420806a2a1342fa9d895e95c3c1d6319d861979", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "dc04e657-c9d4-4ae4-9608-d506dcdba62d", "node_type": "1", "metadata": {}, "hash": "adced3075ee07d80dd5a064953b3e6345338a7788e2ed5df75314c4c19302de4", "class_name": "RelatedNodeInfo"}}, "text": "Figure 1: The Transformer - model architecture.\nThe Transformer follows this overall architecture using stacked self-attention and point-wise, fully\nconnected layers for both the encoder and decoder, shown in the left and right halves of Figure 1,\nrespectively.\n3.1 Encoder and Decoder Stacks\nEncoder: The encoder is composed of a stack of N= 6 identical layers. Each layer has two\nsub-layers. ", "start_char_idx": 0, "end_char_idx": 394, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "dc04e657-c9d4-4ae4-9608-d506dcdba62d": {"__data__": {"id_": "dc04e657-c9d4-4ae4-9608-d506dcdba62d", "embedding": null, "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "c8a07910-0192-481a-bcf1-1a58d71a8f38", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "43ae829564edc9b5ed17c359ae6edcaa482d4371370a9e7539468351fc9292ff", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "d5bcec10-5d92-4d3b-93e1-07345863d62e", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ec844fda5ac16742fcb4be7508ed70c02440962916828d5582d53939c797fd65", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "019fd1ce-c068-408b-8d2d-58d3a5d48b96", "node_type": "1", "metadata": {}, "hash": "bc306d0d8071010116dc3b5eb47c080c7a19eb91730296df10a30301cbbf6289", "class_name": "RelatedNodeInfo"}}, "text": "The first is a multi-head self-attention mechanism, and the second is a simple, position-\nwise fully connected feed-forward network. We employ a residual connection [ 11] around each of\nthe two sub-layers, followed by layer normalization [ 1]. That is, the output of each sub-layer is\nLayerNorm( x+ Sublayer( x)), where Sublayer( x)is the function implemented by the sub-layer\nitself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding\nlayers, produce outputs of dimension dmodel = 512 .\nDecoder: The decoder is also composed of a stack of N= 6identical layers. In addition to the two\nsub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head\nattention over the output of the encoder stack. Similar to the encoder, we employ residual connections\naround each of the sub-layers, followed by layer normalization. We also modify the self-attention\nsub-layer in the decoder stack to prevent positions from attending to subsequent positions. This\nmasking, combined with fact that the output embeddings are offset by one position, ensures that the\npredictions for position ican depend only on the known outputs at positions less than i.\n3.2 Attention\nAn attention function can be described as mapping a query and a set of key-value pairs to an output,\nwhere the query, keys, values, and output are all vectors. ", "start_char_idx": 394, "end_char_idx": 1784, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "019fd1ce-c068-408b-8d2d-58d3a5d48b96": {"__data__": {"id_": "019fd1ce-c068-408b-8d2d-58d3a5d48b96", "embedding": null, "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "c8a07910-0192-481a-bcf1-1a58d71a8f38", "node_type": "4", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "43ae829564edc9b5ed17c359ae6edcaa482d4371370a9e7539468351fc9292ff", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "dc04e657-c9d4-4ae4-9608-d506dcdba62d", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "2e731d358bfcfda1c02958e2f95c5290561829d2a32e82ac769b3dda94ffaf86", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "c304a4ca-c2b7-4211-9bc8-2a5108b28e07", "node_type": "1", "metadata": {}, "hash": "a6b53683e0a2b96be0935e39d0574511e5955296c192a44644e7b2becae4d936", "class_name": "RelatedNodeInfo"}}, "text": "The output is computed as a weighted sum\n3", "start_char_idx": 1784, "end_char_idx": 1826, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "c304a4ca-c2b7-4211-9bc8-2a5108b28e07": {"__data__": {"id_": "c304a4ca-c2b7-4211-9bc8-2a5108b28e07", "embedding": null, "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "cd926aa7-9c47-4517-abf3-097ddfe907f2", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6d7eafcd20aa872de73d70cafddd26517476b3a8947a3be7c4b572699123d453", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "019fd1ce-c068-408b-8d2d-58d3a5d48b96", "node_type": "1", "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1ae967901d22e2c27573427e0a60288b2a0d6e75e5a20f39487c84d8f6faa5db", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "bdf8130c-d544-420b-9005-33872a3f74c1", "node_type": "1", "metadata": {}, "hash": "d3b4b9f322caa182e14014ca418b27538a6229a94e5bc613b44004202ff454fc", "class_name": "RelatedNodeInfo"}}, "text": "Scaled Dot-Product Attention\n Multi-Head Attention\nFigure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several\nattention layers running in parallel.\nof the values, where the weight assigned to each value is computed by a compatibility function of the\nquery with the corresponding key.\n3.2.1 Scaled Dot-Product Attention\nWe call our particular attention \"Scaled Dot-Product Attention\" (Figure 2). The input consists of\nqueries and keys of dimension dk, and values of dimension dv. We compute the dot products of the\nquery with all keys, divide each by\u221adk, and apply a softmax function to obtain the weights on the\nvalues.\nIn practice, we compute the attention function on a set of queries simultaneously, packed together\ninto a matrix Q. The keys and values are also packed together into matrices KandV. We compute\nthe matrix of outputs as:\nAttention( Q, K, V ) = softmax(QKT\n\u221adk)V (1)\nThe two most commonly used attention functions are additive attention [ 2], and dot-product (multi-\nplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor\nof1\u221adk. ", "start_char_idx": 0, "end_char_idx": 1134, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "bdf8130c-d544-420b-9005-33872a3f74c1": {"__data__": {"id_": "bdf8130c-d544-420b-9005-33872a3f74c1", "embedding": null, "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "cd926aa7-9c47-4517-abf3-097ddfe907f2", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6d7eafcd20aa872de73d70cafddd26517476b3a8947a3be7c4b572699123d453", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "c304a4ca-c2b7-4211-9bc8-2a5108b28e07", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "25d2f70ef16cdd181bcf39f32b2fa03c410a8ea0d81c8a4e7350866c7d4e3f0a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "ab54a858-bc18-4ee5-b02f-1f549c149532", "node_type": "1", "metadata": {}, "hash": "7e96d9c89ff8d9bc7bbaeb8f6a828ef8a4aa67f18210fd16213798b1a60972f9", "class_name": "RelatedNodeInfo"}}, "text": "Additive attention computes the compatibility function using a feed-forward network with\na single hidden layer. While the two are similar in theoretical complexity, dot-product attention is\nmuch faster and more space-efficient in practice, since it can be implemented using highly optimized\nmatrix multiplication code.\nWhile for small values of dkthe two mechanisms perform similarly, additive attention outperforms\ndot product attention without scaling for larger values of dk[3]. We suspect that for large values of\ndk, the dot products grow large in magnitude, pushing the softmax function into regions where it has\nextremely small gradients4. To counteract this effect, we scale the dot products by1\u221adk.\n3.2.2 Multi-Head Attention\nInstead of performing a single attention function with dmodel-dimensional keys, values and queries,\nwe found it beneficial to linearly project the queries, keys and values htimes with different, learned\nlinear projections to dk,dkanddvdimensions, respectively. On each of these projected versions of\nqueries, keys and values we then perform the attention function in parallel, yielding dv-dimensional\n4To illustrate why the dot products get large, assume that the components of qandkare independent random\nvariables with mean 0and variance 1. Then their dot product, q\u00b7k=Pdk\ni=1qiki, has mean 0and variance dk.\n", "start_char_idx": 1134, "end_char_idx": 2480, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "ab54a858-bc18-4ee5-b02f-1f549c149532": {"__data__": {"id_": "ab54a858-bc18-4ee5-b02f-1f549c149532", "embedding": null, "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "cd926aa7-9c47-4517-abf3-097ddfe907f2", "node_type": "4", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6d7eafcd20aa872de73d70cafddd26517476b3a8947a3be7c4b572699123d453", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "bdf8130c-d544-420b-9005-33872a3f74c1", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6d8d07e874fb92b09011ea27bcd9ef831e9cabce9634f81d8862de406b9ab9ed", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "b334377e-b057-43e0-b76f-17f19781fc1d", "node_type": "1", "metadata": {}, "hash": "21ff6be24feb44f64834efc2217a726e625aaed495dbe2203cb6c7c174bfdea1", "class_name": "RelatedNodeInfo"}}, "text": "4", "start_char_idx": 1778, "end_char_idx": 1779, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "b334377e-b057-43e0-b76f-17f19781fc1d": {"__data__": {"id_": "b334377e-b057-43e0-b76f-17f19781fc1d", "embedding": null, "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "e1e2ea31-60e3-4801-805b-c4df3bc294dd", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c814a65337d0e89ec59292c4b363d8a7cecf44e9b5155e17e5b71decb6e52b0c", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "ab54a858-bc18-4ee5-b02f-1f549c149532", "node_type": "1", "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "26988312de96dce95a0094d5e86a130f91e5f69e4d6a720c5a581491f23ed818", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "b3784fbe-100b-49fc-b7f7-1d16b9296a1e", "node_type": "1", "metadata": {}, "hash": "c33d6baeaa634ded1a452943784e6ac9010ae35bb2cd35e9795434c6b2172e08", "class_name": "RelatedNodeInfo"}}, "text": "output values. These are concatenated and once again projected, resulting in the final values, as\ndepicted in Figure 2.\nMulti-head attention allows the model to jointly attend to information from different representation\nsubspaces at different positions. With a single attention head, averaging inhibits this.\nMultiHead( Q, K, V ) = Concat(head 1, ...,head h)WO\nwhere head i= Attention( QWQ\ni, KWK\ni, V WV\ni)\nWhere the projections are parameter matrices WQ\ni\u2208Rdmodel\u00d7dk,WK\ni\u2208Rdmodel\u00d7dk,WV\ni\u2208Rdmodel\u00d7dv\nandWO\u2208Rhdv\u00d7dmodel.\nIn this work we employ h= 8 parallel attention layers, or heads. ", "start_char_idx": 0, "end_char_idx": 586, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "b3784fbe-100b-49fc-b7f7-1d16b9296a1e": {"__data__": {"id_": "b3784fbe-100b-49fc-b7f7-1d16b9296a1e", "embedding": null, "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "e1e2ea31-60e3-4801-805b-c4df3bc294dd", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c814a65337d0e89ec59292c4b363d8a7cecf44e9b5155e17e5b71decb6e52b0c", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "b334377e-b057-43e0-b76f-17f19781fc1d", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "23f9f422947d199bca0707b35985de63fb61e9b64a4622365a90a1cb8d4f5cbe", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "897256a3-8c19-4b9b-8f51-204cff114902", "node_type": "1", "metadata": {}, "hash": "ae61ebb23bddaa99d04fdda048329f2ab389e9b5a4e9d1f98e168d0eb697a090", "class_name": "RelatedNodeInfo"}}, "text": "For each of these we use\ndk=dv=dmodel/h= 64 . Due to the reduced dimension of each head, the total computational cost\nis similar to that of single-head attention with full dimensionality.\n3.2.3 Applications of Attention in our Model\nThe Transformer uses multi-head attention in three different ways:\n\u2022In \"encoder-decoder attention\" layers, the queries come from the previous decoder layer,\nand the memory keys and values come from the output of the encoder. This allows every\nposition in the decoder to attend over all positions in the input sequence. ", "start_char_idx": 586, "end_char_idx": 1138, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "897256a3-8c19-4b9b-8f51-204cff114902": {"__data__": {"id_": "897256a3-8c19-4b9b-8f51-204cff114902", "embedding": null, "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "e1e2ea31-60e3-4801-805b-c4df3bc294dd", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c814a65337d0e89ec59292c4b363d8a7cecf44e9b5155e17e5b71decb6e52b0c", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "b3784fbe-100b-49fc-b7f7-1d16b9296a1e", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "6f7ef2d54429da3b4bfe849e20e66927288c8f200a426a593a4ce40fced9bb5e", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "989eb0dd-6927-47e5-91e3-d6f522a4bd2a", "node_type": "1", "metadata": {}, "hash": "7cbbe3cedcd85f13e8877f623107b685e449ab6affec17847118f51cab2d9384", "class_name": "RelatedNodeInfo"}}, "text": "This mimics the\ntypical encoder-decoder attention mechanisms in sequence-to-sequence models such as\n[38, 2, 9].\n\u2022The encoder contains self-attention layers. In a self-attention layer all of the keys, values\nand queries come from the same place, in this case, the output of the previous layer in the\nencoder. Each position in the encoder can attend to all positions in the previous layer of the\nencoder.\n\u2022Similarly, self-attention layers in the decoder allow each position in the decoder to attend to\nall positions in the decoder up to and including that position. We need to prevent leftward\ninformation flow in the decoder to preserve the auto-regressive property. We implement this\ninside of scaled dot-product attention by masking out (setting to \u2212\u221e) all values in the input\nof the softmax which correspond to illegal connections. See Figure 2.\n3.3 Position-wise Feed-Forward Networks\nIn addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully\nconnected feed-forward network, which is applied to each position separately and identically. This\nconsists of two linear transformations with a ReLU activation in between.\nFFN( x) = max(0 , xW 1+b1)W2+b2 (2)\nWhile the linear transformations are the same across different positions, they use different parameters\nfrom layer to layer. Another way of describing this is as two convolutions with kernel size 1.\n", "start_char_idx": 1138, "end_char_idx": 2534, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "989eb0dd-6927-47e5-91e3-d6f522a4bd2a": {"__data__": {"id_": "989eb0dd-6927-47e5-91e3-d6f522a4bd2a", "embedding": null, "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "e1e2ea31-60e3-4801-805b-c4df3bc294dd", "node_type": "4", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c814a65337d0e89ec59292c4b363d8a7cecf44e9b5155e17e5b71decb6e52b0c", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "897256a3-8c19-4b9b-8f51-204cff114902", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "9df99719393b4ec72ccf855808242c10151dcb245937fcda88f7e8456ec7e8fb", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "38465ff0-f5e8-43e3-a581-ff76dc2cea52", "node_type": "1", "metadata": {}, "hash": "d38477a528396f5cc112bcb61b7c9befc0bba3596c07341c247686f85296a7e7", "class_name": "RelatedNodeInfo"}}, "text": "The dimensionality of input and output is dmodel = 512 , and the inner-layer has dimensionality\ndff= 2048 .\n3.4 Embeddings and Softmax\nSimilarly to other sequence transduction models, we use learned embeddings to convert the input\ntokens and output tokens to vectors of dimension dmodel. We also use the usual learned linear transfor-\nmation and softmax function to convert the decoder output to predicted next-token probabilities. In\nour model, we share the same weight matrix between the two embedding layers and the pre-softmax\nlinear transformation, similar to [ 30]. In the embedding layers, we multiply those weights by\u221admodel.\n5", "start_char_idx": 2534, "end_char_idx": 3169, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "38465ff0-f5e8-43e3-a581-ff76dc2cea52": {"__data__": {"id_": "38465ff0-f5e8-43e3-a581-ff76dc2cea52", "embedding": null, "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f618674d2e12b9eac8d947f3c475ce6dadc2b597e3418b93ade5726a62dc0a14", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "989eb0dd-6927-47e5-91e3-d6f522a4bd2a", "node_type": "1", "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ed5894d3e4cc2756a80d2170b70d9eb799ba0dd9bd09c0986a226bfedba5444a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "c1fac594-7f52-4e5f-abbd-a8b9a2385ea1", "node_type": "1", "metadata": {}, "hash": "4e63986289fc951fa89b559329ed84e62002d2262af2efe5b4418ffa4a933ccc", "class_name": "RelatedNodeInfo"}}, "text": "Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations\nfor different layer types. nis the sequence length, dis the representation dimension, kis the kernel\nsize of convolutions and rthe size of the neighborhood in restricted self-attention.\nLayer Type Complexity per Layer Sequential Maximum Path Length\nOperations\nSelf-Attention O(n2\u00b7d) O(1) O(1)\nRecurrent O(n\u00b7d2) O(n) O(n)\nConvolutional O(k\u00b7n\u00b7d2) O(1) O(logk(n))\nSelf-Attention (restricted) O(r\u00b7n\u00b7d) O(1) O(n/r)\n3.5 Positional Encoding\nSince our model contains no recurrence and no convolution, in order for the model to make use of the\norder of the sequence, we must inject some information about the relative or absolute position of the\ntokens in the sequence. To this end, we add \"positional encodings\" to the input embeddings at the\nbottoms of the encoder and decoder stacks. ", "start_char_idx": 0, "end_char_idx": 874, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "c1fac594-7f52-4e5f-abbd-a8b9a2385ea1": {"__data__": {"id_": "c1fac594-7f52-4e5f-abbd-a8b9a2385ea1", "embedding": null, "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f618674d2e12b9eac8d947f3c475ce6dadc2b597e3418b93ade5726a62dc0a14", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "38465ff0-f5e8-43e3-a581-ff76dc2cea52", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "4149d0733a2077929fc0fc7832725ed2c7850610d31f1fa2e6770bea04f69a4d", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "abe19b59-2b09-4ccc-ad63-74dc10dd0989", "node_type": "1", "metadata": {}, "hash": "529205e3a6d8b748eba53794c5d10346f0dc59338979ad554e1a9f3eefb693ee", "class_name": "RelatedNodeInfo"}}, "text": "The positional encodings have the same dimension dmodel\nas the embeddings, so that the two can be summed. There are many choices of positional encodings,\nlearned and fixed [9].\nIn this work, we use sine and cosine functions of different frequencies:\nPE(pos,2i)=sin(pos/100002i/d model)\nPE(pos,2i+1)=cos(pos/100002i/d model)\nwhere posis the position and iis the dimension. That is, each dimension of the positional encoding\ncorresponds to a sinusoid. The wavelengths form a geometric progression from 2\u03c0to10000 \u00b72\u03c0. We\nchose this function because we hypothesized it would allow the model to easily learn to attend by\nrelative positions, since for any fixed offset k,PEpos+kcan be represented as a linear function of\nPEpos.\nWe also experimented with using learned positional embeddings [ 9] instead, and found that the two\nversions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version\nbecause it may allow the model to extrapolate to sequence lengths longer than the ones encountered\nduring training.\n4 Why Self-Attention\nIn this section we compare various aspects of self-attention layers to the recurrent and convolu-\ntional layers commonly used for mapping one variable-length sequence of symbol representations\n(x1, ..., x n)to another sequence of equal length (z1, ..., z n), with xi, zi\u2208Rd, such as a hidden\nlayer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we\nconsider three desiderata.\n", "start_char_idx": 874, "end_char_idx": 2350, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "abe19b59-2b09-4ccc-ad63-74dc10dd0989": {"__data__": {"id_": "abe19b59-2b09-4ccc-ad63-74dc10dd0989", "embedding": null, "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f618674d2e12b9eac8d947f3c475ce6dadc2b597e3418b93ade5726a62dc0a14", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "c1fac594-7f52-4e5f-abbd-a8b9a2385ea1", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "087f13d1ed39ef41492b3d272f5c1b21b4b783865a4247e17bc158debda912f5", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "700226de-87f3-4ff6-8023-2e460ab3599d", "node_type": "1", "metadata": {}, "hash": "2296b0274cd92b7809a3629bdd9cbfddde0e5225d8eca903f310f8d0ee302d9a", "class_name": "RelatedNodeInfo"}}, "text": "One is the total computational complexity per layer. ", "start_char_idx": 2350, "end_char_idx": 2403, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "700226de-87f3-4ff6-8023-2e460ab3599d": {"__data__": {"id_": "700226de-87f3-4ff6-8023-2e460ab3599d", "embedding": null, "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5", "node_type": "4", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f618674d2e12b9eac8d947f3c475ce6dadc2b597e3418b93ade5726a62dc0a14", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "abe19b59-2b09-4ccc-ad63-74dc10dd0989", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ef3f6df3b02956ee87ad5c023687353efda09b51ec3d0682d81ad9e3c0af0e23", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "5d14e9e0-9083-4dae-aa0f-a3b3ad856ae2", "node_type": "1", "metadata": {}, "hash": "c2571f9c9bd39ffc7f852bd73b31b821838adcf8b8eb6c1b2b019f0eb1d62b8e", "class_name": "RelatedNodeInfo"}}, "text": "Another is the amount of computation that can\nbe parallelized, as measured by the minimum number of sequential operations required.\nThe third is the path length between long-range dependencies in the network. Learning long-range\ndependencies is a key challenge in many sequence transduction tasks. One key factor affecting the\nability to learn such dependencies is the length of the paths forward and backward signals have to\ntraverse in the network. The shorter these paths between any combination of positions in the input\nand output sequences, the easier it is to learn long-range dependencies [ 12]. Hence we also compare\nthe maximum path length between any two input and output positions in networks composed of the\ndifferent layer types.\nAs noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially\nexecuted operations, whereas a recurrent layer requires O(n)sequential operations. In terms of\ncomputational complexity, self-attention layers are faster than recurrent layers when the sequence\n6", "start_char_idx": 2403, "end_char_idx": 3448, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "5d14e9e0-9083-4dae-aa0f-a3b3ad856ae2": {"__data__": {"id_": "5d14e9e0-9083-4dae-aa0f-a3b3ad856ae2", "embedding": null, "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3d32a5c8f07d78a73fefaf20f8136d569775b557772bd46a2ce2d90c3365c26b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "700226de-87f3-4ff6-8023-2e460ab3599d", "node_type": "1", "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "930cc5f06b3219abf8c66d02b61b945a23f8268250497902ad05ed77a71d848f", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "0716be40-0d7e-4a1a-9cf0-9f14437b82f9", "node_type": "1", "metadata": {}, "hash": "7b2fa15fa24f3d13a562b21e7da4ac31a1dc40e480c702d86266667f3d7087d5", "class_name": "RelatedNodeInfo"}}, "text": "length nis smaller than the representation dimensionality d, which is most often the case with\nsentence representations used by state-of-the-art models in machine translations, such as word-piece\n[38] and byte-pair [ 31] representations. To improve computational performance for tasks involving\nvery long sequences, self-attention could be restricted to considering only a neighborhood of size rin\nthe input sequence centered around the respective output position. This would increase the maximum\npath length to O(n/r). ", "start_char_idx": 0, "end_char_idx": 520, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "0716be40-0d7e-4a1a-9cf0-9f14437b82f9": {"__data__": {"id_": "0716be40-0d7e-4a1a-9cf0-9f14437b82f9", "embedding": null, "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3d32a5c8f07d78a73fefaf20f8136d569775b557772bd46a2ce2d90c3365c26b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "5d14e9e0-9083-4dae-aa0f-a3b3ad856ae2", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "58db0b38b6326b1bf61137c8dbe62d75f8d11c21e85243ec183c9bda8fb05027", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "903896c3-e90f-4fb8-933c-3b9a419f220d", "node_type": "1", "metadata": {}, "hash": "e995007d39659f8ed1b131ab550becf6731d17bd6470483e2afe60328761ef1d", "class_name": "RelatedNodeInfo"}}, "text": "We plan to investigate this approach further in future work.\nA single convolutional layer with kernel width k < n does not connect all pairs of input and output\npositions. Doing so requires a stack of O(n/k)convolutional layers in the case of contiguous kernels,\norO(logk(n))in the case of dilated convolutions [ 18], increasing the length of the longest paths\nbetween any two positions in the network. Convolutional layers are generally more expensive than\nrecurrent layers, by a factor of k. Separable convolutions [ 6], however, decrease the complexity\nconsiderably, to O(k\u00b7n\u00b7d+n\u00b7d2). Even with k=n, however, the complexity of a separable\nconvolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer,\nthe approach we take in our model.\nAs side benefit, self-attention could yield more interpretable models. ", "start_char_idx": 520, "end_char_idx": 1371, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "903896c3-e90f-4fb8-933c-3b9a419f220d": {"__data__": {"id_": "903896c3-e90f-4fb8-933c-3b9a419f220d", "embedding": null, "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3d32a5c8f07d78a73fefaf20f8136d569775b557772bd46a2ce2d90c3365c26b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "0716be40-0d7e-4a1a-9cf0-9f14437b82f9", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f48d0ff5ec08366539996b772ac04c071acf8af22b72c11e1d8f8931409f4bcc", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "0fa8f49d-ee77-4e8b-99f9-430ef1086b18", "node_type": "1", "metadata": {}, "hash": "08b98bd856e3b568167819c21f23dfe2d27afd29dc5197004a61d3446a50b002", "class_name": "RelatedNodeInfo"}}, "text": "We inspect attention distributions\nfrom our models and present and discuss examples in the appendix. Not only do individual attention\nheads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic\nand semantic structure of the sentences.\n5 Training\nThis section describes the training regime for our models.\n5.1 Training Data and Batching\nWe trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million\nsentence pairs. Sentences were encoded using byte-pair encoding [ 3], which has a shared source-\ntarget vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT\n2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece\nvocabulary [ 38]. Sentence pairs were batched together by approximate sequence length. Each training\nbatch contained a set of sentence pairs containing approximately 25000 source tokens and 25000\ntarget tokens.\n5.2 Hardware and Schedule\nWe trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using\nthe hyperparameters described throughout the paper, each training step took about 0.4 seconds. We\ntrained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the\nbottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps\n(3.5 days).\n5.3 Optimizer\nWe used the Adam optimizer [ 20] with \u03b21= 0.9,\u03b22= 0.98and\u03f5= 10\u22129. We varied the learning\nrate over the course of training, according to the formula:\nlrate =d\u22120.5\nmodel\u00b7min(step_num\u22120.5, step _num\u00b7warmup _steps\u22121.5) (3)\nThis corresponds to increasing the learning rate linearly for the first warmup _steps training steps,\nand decreasing it thereafter proportionally to the inverse square root of the step number. We used\nwarmup _steps = 4000 .\n", "start_char_idx": 1371, "end_char_idx": 3228, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "0fa8f49d-ee77-4e8b-99f9-430ef1086b18": {"__data__": {"id_": "0fa8f49d-ee77-4e8b-99f9-430ef1086b18", "embedding": null, "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7", "node_type": "4", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3d32a5c8f07d78a73fefaf20f8136d569775b557772bd46a2ce2d90c3365c26b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "903896c3-e90f-4fb8-933c-3b9a419f220d", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "18393b3aa39c791d80bb634cbf5b3e31f34a2417b1d83fc705b84c50c0e7d67b", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "7ac597d6-9c0c-4946-b6da-2e6954dd3ff1", "node_type": "1", "metadata": {}, "hash": "2ee52b486177a249fde2286c059b570f769f98a23868517e0103c77d53fb230a", "class_name": "RelatedNodeInfo"}}, "text": "5.4 Regularization\nWe employ three types of regularization during training:\n7", "start_char_idx": 3228, "end_char_idx": 3305, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "7ac597d6-9c0c-4946-b6da-2e6954dd3ff1": {"__data__": {"id_": "7ac597d6-9c0c-4946-b6da-2e6954dd3ff1", "embedding": null, "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "0da32021-5f29-482c-ba55-b4082af76e37", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "104f868c9ec7bb6f6157b5bbd4bd5478e9edf52dbafbc104103eb9f004b9a858", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "0fa8f49d-ee77-4e8b-99f9-430ef1086b18", "node_type": "1", "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "eb179bdea2450b5906d42af007dd524a30ad1d4232843531eed033060a74a62f", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "333ede79-147a-4ddb-a536-4af80aa36813", "node_type": "1", "metadata": {}, "hash": "60789b4d682bfe6a26431e4dbe92bb206917ff4c763c2a94bd513ecdfe8967e9", "class_name": "RelatedNodeInfo"}}, "text": "Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the\nEnglish-to-German and English-to-French newstest2014 tests at a fraction of the training cost.\nModelBLEU Training Cost (FLOPs)\nEN-DE EN-FR EN-DE EN-FR\nByteNet [18] 23.75\nDeep-Att + PosUnk [39] 39.2 1.0\u00b71020\nGNMT + RL [38] 24.6 39.92 2.3\u00b710191.4\u00b71020\nConvS2S [9] 25.16 40.46 9.6\u00b710181.5\u00b71020\nMoE [32] 26.03 40.56 2.0\u00b710191.2\u00b71020\nDeep-Att + PosUnk Ensemble [39] 40.4 8.0\u00b71020\nGNMT + RL Ensemble [38] 26.30 41.16 1.8\u00b710201.1\u00b71021\nConvS2S Ensemble [9] 26.36 41.29 7.7\u00b710191.2\u00b71021\nTransformer (base model) 27.3 38.1 3.3\u00b71018\nTransformer (big) 28.4 41.8 2.3\u00b71019\nResidual Dropout We apply dropout [ 33] to the output of each sub-layer, before it is added to the\nsub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the\npositional encodings in both the encoder and decoder stacks. ", "start_char_idx": 0, "end_char_idx": 917, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "333ede79-147a-4ddb-a536-4af80aa36813": {"__data__": {"id_": "333ede79-147a-4ddb-a536-4af80aa36813", "embedding": null, "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "0da32021-5f29-482c-ba55-b4082af76e37", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "104f868c9ec7bb6f6157b5bbd4bd5478e9edf52dbafbc104103eb9f004b9a858", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "7ac597d6-9c0c-4946-b6da-2e6954dd3ff1", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "f8a87f87a229ea433cafaddb9a7b882283a1255e295a8e9f20c489580811c1b8", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "5d8d189d-424b-4a2a-b0ad-e28758dce1d4", "node_type": "1", "metadata": {}, "hash": "cc6ddb6cf1dd1fc9fd3b3083c970a010d117eee847efa7092e4894f37d0c123e", "class_name": "RelatedNodeInfo"}}, "text": "For the base model, we use a rate of\nPdrop= 0.1.\nLabel Smoothing During training, we employed label smoothing of value \u03f5ls= 0.1[36]. This\nhurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.\n6 Results\n6.1 Machine Translation\nOn the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big)\nin Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0\nBLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is\nlisted in the bottom line of Table 3. ", "start_char_idx": 917, "end_char_idx": 1515, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "5d8d189d-424b-4a2a-b0ad-e28758dce1d4": {"__data__": {"id_": "5d8d189d-424b-4a2a-b0ad-e28758dce1d4", "embedding": null, "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "0da32021-5f29-482c-ba55-b4082af76e37", "node_type": "4", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "104f868c9ec7bb6f6157b5bbd4bd5478e9edf52dbafbc104103eb9f004b9a858", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "333ede79-147a-4ddb-a536-4af80aa36813", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "847192d4d9919bf8dc8014956a35a1777b97e1b9ac70ad634785cfbb7705755a", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "caa5c999-7d0e-447b-a3ac-6e3180209f3d", "node_type": "1", "metadata": {}, "hash": "51d2c03cc287c61e329de8841e02495d7c4f75c096bd196f124b7cdf49004162", "class_name": "RelatedNodeInfo"}}, "text": "Training took 3.5days on 8P100 GPUs. Even our base model\nsurpasses all previously published models and ensembles, at a fraction of the training cost of any of\nthe competitive models.\nOn the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0,\noutperforming all of the previously published single models, at less than 1/4the training cost of the\nprevious state-of-the-art model. The Transformer (big) model trained for English-to-French used\ndropout rate Pdrop= 0.1, instead of 0.3.\nFor the base models, we used a single model obtained by averaging the last 5 checkpoints, which\nwere written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We\nused beam search with a beam size of 4and length penalty \u03b1= 0.6[38]. These hyperparameters\nwere chosen after experimentation on the development set. We set the maximum output length during\ninference to input length + 50, but terminate early when possible [38].\nTable 2 summarizes our results and compares our translation quality and training costs to other model\narchitectures from the literature. We estimate the number of floating point operations used to train a\nmodel by multiplying the training time, the number of GPUs used, and an estimate of the sustained\nsingle-precision floating-point capacity of each GPU5.\n6.2 Model Variations\nTo evaluate the importance of different components of the Transformer, we varied our base model\nin different ways, measuring the change in performance on English-to-German translation on the\n5We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.\n8", "start_char_idx": 1515, "end_char_idx": 3149, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "caa5c999-7d0e-447b-a3ac-6e3180209f3d": {"__data__": {"id_": "caa5c999-7d0e-447b-a3ac-6e3180209f3d", "embedding": null, "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "41e170cc-f118-4a60-bb0a-2ec27a0f6318", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3555b87155aec9aa8a75b706c876df3cf0558b251144393e5e38cc83b87f3d3b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "5d8d189d-424b-4a2a-b0ad-e28758dce1d4", "node_type": "1", "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "b4f25e74ebd0475b910b654d5a93c49c99526a91afd42bb12bd4b92824d630b6", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "2cab4eb7-bc6c-4c75-8dcc-763aaaceb767", "node_type": "1", "metadata": {}, "hash": "f3277f5fec6ed95385f3dbd440c02404609c1387a95b86c76a32467d8ee3a95c", "class_name": "RelatedNodeInfo"}}, "text": "Table 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base\nmodel. ", "start_char_idx": 0, "end_char_idx": 111, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "2cab4eb7-bc6c-4c75-8dcc-763aaaceb767": {"__data__": {"id_": "2cab4eb7-bc6c-4c75-8dcc-763aaaceb767", "embedding": null, "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "41e170cc-f118-4a60-bb0a-2ec27a0f6318", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3555b87155aec9aa8a75b706c876df3cf0558b251144393e5e38cc83b87f3d3b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "caa5c999-7d0e-447b-a3ac-6e3180209f3d", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a0f9febd7e187836acd4a8a2f32e8e1c8dcf61714ac8c19e23a90cc151f71418", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "7466eb16-7489-427f-9838-f8bdb2c721a7", "node_type": "1", "metadata": {}, "hash": "77c42a81387029e301b36817bcbfc5f0b5b9f6f996b7a8449cf67da0914770c2", "class_name": "RelatedNodeInfo"}}, "text": "All metrics are on the English-to-German translation development set, newstest2013. Listed\nperplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to\nper-word perplexities.\nN d model dff h d k dvPdrop \u03f5lstrain PPL BLEU params\nsteps (dev) (dev) \u00d7106\nbase 6 512 2048 8 64 64 0.1 0.1 100K 4.92 25.8 65\n(A)1 512 512 5.29 24.9\n4 128 128 5.00 25.5\n16 32 32 4.91 25.8\n32 16 16 5.01 25.4\n(B)16 5.16 25.1 58\n32 5.01 25.4 60\n(C)2 6.11 23.7 36\n4 5.19 25.3 50\n8 4.88 25.5 80\n256 32 32 5.75 24.5 28\n1024 128 128 4.66 26.0 168\n1024 5.12 25.4 53\n4096 4.75 26.2 90\n(D)0.0 5.77 24.6\n0.2 4.95 25.5\n0.0 4.67 25.3\n0.2 5.47 25.7\n(E) positional embedding instead of sinusoids 4.92 25.7\nbig 6 1024 4096 16 0.3 300K 4.33 26.4 213\ndevelopment set, newstest2013. We used beam search as described in the previous section, but no\ncheckpoint averaging. ", "start_char_idx": 111, "end_char_idx": 975, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "7466eb16-7489-427f-9838-f8bdb2c721a7": {"__data__": {"id_": "7466eb16-7489-427f-9838-f8bdb2c721a7", "embedding": null, "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "41e170cc-f118-4a60-bb0a-2ec27a0f6318", "node_type": "4", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "3555b87155aec9aa8a75b706c876df3cf0558b251144393e5e38cc83b87f3d3b", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "2cab4eb7-bc6c-4c75-8dcc-763aaaceb767", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fd1f6f3268954ed574af42f2a7adca2cb23a5b323cc99293d58a8a7f2e51f71f", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "3470fcf0-d739-471d-a9cb-46f1eaddc675", "node_type": "1", "metadata": {}, "hash": "0a4812b67002733db4519b06b1d218113d2dfef8b1bc223dd2721ebd5185cade", "class_name": "RelatedNodeInfo"}}, "text": "We present these results in Table 3.\nIn Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions,\nkeeping the amount of computation constant, as described in Section 3.2.2. While single-head\nattention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.\nIn Table 3 rows (B), we observe that reducing the attention key size dkhurts model quality. This\nsuggests that determining compatibility is not easy and that a more sophisticated compatibility\nfunction than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected,\nbigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our\nsinusoidal positional encoding with learned positional embeddings [ 9], and observe nearly identical\nresults to the base model.\n6.3 English Constituency Parsing\nTo evaluate if the Transformer can generalize to other tasks we performed experiments on English\nconstituency parsing. This task presents specific challenges: the output is subject to strong structural\nconstraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence\nmodels have not been able to attain state-of-the-art results in small-data regimes [37].\nWe trained a 4-layer transformer with dmodel = 1024 on the Wall Street Journal (WSJ) portion of the\nPenn Treebank [ 25], about 40K training sentences. We also trained it in a semi-supervised setting,\nusing the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences\n[37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens\nfor the semi-supervised setting.\nWe performed only a small number of experiments to select the dropout, both attention and residual\n(section 5.4), learning rates and beam size on the Section 22 development set, all other parameters\nremained unchanged from the English-to-German base translation model. During inference, we\n9", "start_char_idx": 975, "end_char_idx": 2969, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "3470fcf0-d739-471d-a9cb-46f1eaddc675": {"__data__": {"id_": "3470fcf0-d739-471d-a9cb-46f1eaddc675", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "7466eb16-7489-427f-9838-f8bdb2c721a7", "node_type": "1", "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ffdd464141d42dff9893063dd9743953ef246c6f3caedc88f3a461ac8d2ada65", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "9acba64d-c6f2-406b-a1c8-5da9691f5868", "node_type": "1", "metadata": {}, "hash": "2c01788355b4081bae742c8ff8e18d40d08d89aa8dcc02f0954bb8e243417195", "class_name": "RelatedNodeInfo"}}, "text": "Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23\nof WSJ)\nParser Training WSJ 23 F1\nVinyals & Kaiser el al. (2014) [37] WSJ only, discriminative 88.3\nPetrov et al. ", "start_char_idx": 0, "end_char_idx": 215, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "9acba64d-c6f2-406b-a1c8-5da9691f5868": {"__data__": {"id_": "9acba64d-c6f2-406b-a1c8-5da9691f5868", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "3470fcf0-d739-471d-a9cb-46f1eaddc675", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "2452e5b3e5ff62d94096818af5a1d3620d61489948275f630fc54745090bf3a3", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "cce9180a-7469-43e7-955c-62a11e492759", "node_type": "1", "metadata": {}, "hash": "b38cd8e544f44a26f93d4cb275a25febb200b6e4b330287b6e7435b29a0e3823", "class_name": "RelatedNodeInfo"}}, "text": "(2006) [29] WSJ only, discriminative 90.4\nZhu et al. (2013) [40] WSJ only, discriminative 90.4\nDyer et al. (2016) [8] WSJ only, discriminative 91.7\nTransformer (4 layers) WSJ only, discriminative 91.3\nZhu et al. (2013) [40] semi-supervised 91.3\nHuang & Harper (2009) [14] semi-supervised 91.3\nMcClosky et al. (2006) [26] semi-supervised 92.1\nVinyals & Kaiser el al. (2014) [37] semi-supervised 92.1\nTransformer (4 layers) semi-supervised 92.7\nLuong et al. (2015) [23] multi-task 93.0\nDyer et al. (2016) [8] generative 93.3\nincreased the maximum output length to input length + 300. We used a beam size of 21and\u03b1= 0.3\nfor both WSJ only and the semi-supervised setting.\nOur results in Table 4 show that despite the lack of task-specific tuning our model performs sur-\nprisingly well, yielding better results than all previously reported models with the exception of the\nRecurrent Neural Network Grammar [8].\nIn contrast to RNN sequence-to-sequence models [ 37], the Transformer outperforms the Berkeley-\nParser [29] even when training only on the WSJ training set of 40K sentences.\n7 Conclusion\nIn this work, we presented the Transformer, the first sequence transduction model based entirely on\nattention, replacing the recurrent layers most commonly used in encoder-decoder architectures with\nmulti-headed self-attention.\nFor translation tasks, the Transformer can be trained significantly faster than architectures based\non recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014\nEnglish-to-French translation tasks, we achieve a new state of the art. In the former task our best\nmodel outperforms even all previously reported ensembles.\nWe are excited about the future of attention-based models and plan to apply them to other tasks. We\nplan to extend the Transformer to problems involving input and output modalities other than text and\nto investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs\nsuch as images, audio and video. Making generation less sequential is another research goals of ours.\nThe code we used to train and evaluate our models is available at https://github.com/\ntensorflow/tensor2tensor .\nAcknowledgements We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful\ncomments, corrections and inspiration.\n", "start_char_idx": 215, "end_char_idx": 2526, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "cce9180a-7469-43e7-955c-62a11e492759": {"__data__": {"id_": "cce9180a-7469-43e7-955c-62a11e492759", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "9acba64d-c6f2-406b-a1c8-5da9691f5868", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ca90a964e379767202d4fd498fdb92350f9e2a87fda59e4392e11f70fd3ae227", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "50c18361-106f-4a2a-a0a5-731ef39b76f9", "node_type": "1", "metadata": {}, "hash": "40cc5a74c9b96205b43ae910c6c5ca3ea7f192127db83336ff242a9c8cfa00ad", "class_name": "RelatedNodeInfo"}}, "text": "References\n[1]Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint\narXiv:1607.06450 , 2016.\n[2]Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly\nlearning to align and translate. CoRR , abs/1409.0473, 2014.\n", "start_char_idx": 2526, "end_char_idx": 2810, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "50c18361-106f-4a2a-a0a5-731ef39b76f9": {"__data__": {"id_": "50c18361-106f-4a2a-a0a5-731ef39b76f9", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "cce9180a-7469-43e7-955c-62a11e492759", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7ebcbdca9beb0f2b2055033447e5820ee04856ee3b6f8376a7a43470e27568fc", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "322758d7-f621-4f04-abb6-3d73e08e273e", "node_type": "1", "metadata": {}, "hash": "8c3a33e9ff98dc22faaa89324e411b4fe7680f578975442fffb7fee2e3bc6f4b", "class_name": "RelatedNodeInfo"}}, "text": "[3]Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V . ", "start_char_idx": 2810, "end_char_idx": 2870, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "322758d7-f621-4f04-abb6-3d73e08e273e": {"__data__": {"id_": "322758d7-f621-4f04-abb6-3d73e08e273e", "embedding": null, "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0", "node_type": "4", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "35e605af3680c1bb6f37315710b423320292459542699c0d4753a669bdf1d252", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "50c18361-106f-4a2a-a0a5-731ef39b76f9", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "0be6230d8a0e1ec43cfcd886930a0beb9ed97ac6da6889b42bc879831d7bf862", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "087633e1-63a2-4f8e-a266-c2e92468c58a", "node_type": "1", "metadata": {}, "hash": "5b8dbb6c47ab8f1b8f392bbd43ea822198828f36d1e67f37933a787830641e9b", "class_name": "RelatedNodeInfo"}}, "text": "Le. Massive exploration of neural\nmachine translation architectures. CoRR , abs/1703.03906, 2017.\n[4]Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine\nreading. arXiv preprint arXiv:1601.06733 , 2016.\n10", "start_char_idx": 2870, "end_char_idx": 3111, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "087633e1-63a2-4f8e-a266-c2e92468c58a": {"__data__": {"id_": "087633e1-63a2-4f8e-a266-c2e92468c58a", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "322758d7-f621-4f04-abb6-3d73e08e273e", "node_type": "1", "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "5e8ee706174d8b81333912e397fce755fc14b64f18580c768109292b8da150a4", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "86eac87b-b047-43cf-8a9d-e520bd0bf31d", "node_type": "1", "metadata": {}, "hash": "e2bad031038c5330dddc478435ba8dcda240aa676cae3d3ce500b99125d0e81c", "class_name": "RelatedNodeInfo"}}, "text": "[5]Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk,\nand Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical\nmachine translation. CoRR , abs/1406.1078, 2014.\n", "start_char_idx": 0, "end_char_idx": 231, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "86eac87b-b047-43cf-8a9d-e520bd0bf31d": {"__data__": {"id_": "86eac87b-b047-43cf-8a9d-e520bd0bf31d", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "087633e1-63a2-4f8e-a266-c2e92468c58a", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "990904d016b4432da73521124cde5423e69795eb577199c4f45c7108326be663", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "490b3459-7a05-444a-b656-c6d80893a882", "node_type": "1", "metadata": {}, "hash": "75189f3288d37fc892233605f7ce71569e48d667a719fb43102c7b4bdb35f23b", "class_name": "RelatedNodeInfo"}}, "text": "[6]Francois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv\npreprint arXiv:1610.02357 , 2016.\n", "start_char_idx": 231, "end_char_idx": 355, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "490b3459-7a05-444a-b656-c6d80893a882": {"__data__": {"id_": "490b3459-7a05-444a-b656-c6d80893a882", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "86eac87b-b047-43cf-8a9d-e520bd0bf31d", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "4fced396588f621fccc8cfcd794eeb4a6de05086890061e25ca1653793c73c56", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "6960da48-4380-4f2b-a8da-62719c89a6df", "node_type": "1", "metadata": {}, "hash": "385615682866e774914e0762c469f5f7821153687d1af64ea13912819fdbfcf4", "class_name": "RelatedNodeInfo"}}, "text": "[7]Junyoung Chung, \u00c7aglar G\u00fcl\u00e7ehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation\nof gated recurrent neural networks on sequence modeling. CoRR , abs/1412.3555, 2014.\n[8]Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural\nnetwork grammars. In Proc. of NAACL , 2016.\n[9]Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolu-\ntional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2 , 2017.\n[10] Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint\narXiv:1308.0850 , 2013.\n[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im-\nage recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern\nRecognition , pages 770\u2013778, 2016.\n[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and J\u00fcrgen Schmidhuber. Gradient flow in\nrecurrent nets: the difficulty of learning long-term dependencies, 2001.\n[13] Sepp Hochreiter and J\u00fcrgen Schmidhuber. Long short-term memory. Neural computation ,\n9(8):1735\u20131780, 1997.\n[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations\nacross languages. In Proceedings of the 2009 Conference on Empirical Methods in Natural\nLanguage Processing , pages 832\u2013841. ACL, August 2009.\n[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring\nthe limits of language modeling. arXiv preprint arXiv:1602.02410 , 2016.\n", "start_char_idx": 355, "end_char_idx": 1837, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "6960da48-4380-4f2b-a8da-62719c89a6df": {"__data__": {"id_": "6960da48-4380-4f2b-a8da-62719c89a6df", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "490b3459-7a05-444a-b656-c6d80893a882", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8711b9214dabab460c0c1db37828ef546489b3e37782eb8fce220a74cd5e5f81", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "c311c6c7-0b31-4015-a322-5a6d52e7b831", "node_type": "1", "metadata": {}, "hash": "33de0027a55eeb67b971f5bb1fb80e804d230ce4ada148fcb121699ba8d27a7e", "class_name": "RelatedNodeInfo"}}, "text": "[16] \u0141ukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural\nInformation Processing Systems, (NIPS) , 2016.\n[17] \u0141ukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference\non Learning Representations (ICLR) , 2016.\n[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Ko-\nray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2 ,\n2017.\n", "start_char_idx": 1837, "end_char_idx": 2317, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "c311c6c7-0b31-4015-a322-5a6d52e7b831": {"__data__": {"id_": "c311c6c7-0b31-4015-a322-5a6d52e7b831", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "6960da48-4380-4f2b-a8da-62719c89a6df", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "a0e76f2b888e24855cd06dd7b08cfa2c499637b1331c61e5dfaad481bc9472e9", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "bad43489-2169-4a7f-86da-500c248e1461", "node_type": "1", "metadata": {}, "hash": "26d7e191978a11862ca9e0645cf86ecf622b0002daf5fbeb7d7708666b7102cd", "class_name": "RelatedNodeInfo"}}, "text": "[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks.\nInInternational Conference on Learning Representations , 2017.\n[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR , 2015.\n", "start_char_idx": 2317, "end_char_idx": 2570, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "bad43489-2169-4a7f-86da-500c248e1461": {"__data__": {"id_": "bad43489-2169-4a7f-86da-500c248e1461", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "c311c6c7-0b31-4015-a322-5a6d52e7b831", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "bc6c2e7b082588db7c8384a6a7b4eea48521d1809219b69c4422420906b566c6", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "bb5c29e7-08eb-46d6-9ca6-ce32474bc798", "node_type": "1", "metadata": {}, "hash": "56bd80b1bdf82527d61cd0d3f0bffe0c7c018356aa4616e8f0a9916690dfbab7", "class_name": "RelatedNodeInfo"}}, "text": "[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint\narXiv:1703.10722 , 2017.\n", "start_char_idx": 2570, "end_char_idx": 2692, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "bb5c29e7-08eb-46d6-9ca6-ce32474bc798": {"__data__": {"id_": "bb5c29e7-08eb-46d6-9ca6-ce32474bc798", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "bad43489-2169-4a7f-86da-500c248e1461", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "e4d727ce734f9ddfd9f0e20b4820ac060878d95d4915de9f724dbc88a3017055", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "9d77aef3-85ae-4d47-8bb1-fbdc37abacd6", "node_type": "1", "metadata": {}, "hash": "8d77459b8be2b95ef185499fabf082d83096113e57464684ceec6de056d85888", "class_name": "RelatedNodeInfo"}}, "text": "[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen\nZhou, and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint\narXiv:1703.03130 , 2017.\n[23] Minh-Thang Luong, Quoc V . Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task\nsequence to sequence learning. arXiv preprint arXiv:1511.06114 , 2015.\n[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-\nbased neural machine translation. arXiv preprint arXiv:1508.04025 , 2015.\n", "start_char_idx": 2692, "end_char_idx": 3227, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "9d77aef3-85ae-4d47-8bb1-fbdc37abacd6": {"__data__": {"id_": "9d77aef3-85ae-4d47-8bb1-fbdc37abacd6", "embedding": null, "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd", "node_type": "4", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "fdd0547ae1b1f1ae891a05da6ae2d2c306d5d85d1cef1ad885bdadf2f18f4500", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "bb5c29e7-08eb-46d6-9ca6-ce32474bc798", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "d3738f25a0cd8c4d20d3085de3e53869ef7e448baed6654c781e2885379d9496", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "614cffc6-535e-4af4-8944-86a4a3751472", "node_type": "1", "metadata": {}, "hash": "84203037f8a0c53aaeb71352e2689230828a14154526d8423435388e249f5d75", "class_name": "RelatedNodeInfo"}}, "text": "11", "start_char_idx": 947, "end_char_idx": 949, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "614cffc6-535e-4af4-8944-86a4a3751472": {"__data__": {"id_": "614cffc6-535e-4af4-8944-86a4a3751472", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "9d77aef3-85ae-4d47-8bb1-fbdc37abacd6", "node_type": "1", "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "9e6f66859ddfec72e70fb13359e1f388130a6c023a5c82cf503736651dc1f7c7", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "e29a95b6-72a8-4718-96e4-b76dc31dbb2e", "node_type": "1", "metadata": {}, "hash": "fc2346583d85860528ee82bc01f5a0c13c812330901c2f3f1b8de4b4d49fe525", "class_name": "RelatedNodeInfo"}}, "text": "[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated\ncorpus of english: The penn treebank. Computational linguistics , 19(2):313\u2013330, 1993.\n[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In\nProceedings of the Human Language Technology Conference of the NAACL, Main Conference ,\npages 152\u2013159. ACL, June 2006.\n", "start_char_idx": 0, "end_char_idx": 401, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "e29a95b6-72a8-4718-96e4-b76dc31dbb2e": {"__data__": {"id_": "e29a95b6-72a8-4718-96e4-b76dc31dbb2e", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "614cffc6-535e-4af4-8944-86a4a3751472", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "8a443340e87718cb142ec1aa7a8da55f95242662bb1cf660d0b6f3da84dd5e02", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "b3c684f3-9fb4-499b-b85d-f8c168179b98", "node_type": "1", "metadata": {}, "hash": "524e63f2027f2bdaec83b5d9375718140fb4d1b15ee36987ed4cf363a3aafbd2", "class_name": "RelatedNodeInfo"}}, "text": "[27] Ankur Parikh, Oscar T\u00e4ckstr\u00f6m, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention\nmodel. In Empirical Methods in Natural Language Processing , 2016.\n[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive\nsummarization. arXiv preprint arXiv:1705.04304 , 2017.\n[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact,\nand interpretable tree annotation. In Proceedings of the 21st International Conference on\nComputational Linguistics and 44th Annual Meeting of the ACL , pages 433\u2013440. ACL, July\n2006.\n[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv\npreprint arXiv:1608.05859 , 2016.\n[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words\nwith subword units. arXiv preprint arXiv:1508.07909 , 2015.\n[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton,\nand Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts\nlayer. arXiv preprint arXiv:1701.06538 , 2017.\n", "start_char_idx": 401, "end_char_idx": 1502, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "b3c684f3-9fb4-499b-b85d-f8c168179b98": {"__data__": {"id_": "b3c684f3-9fb4-499b-b85d-f8c168179b98", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "e29a95b6-72a8-4718-96e4-b76dc31dbb2e", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "cb2a3128ee8cd9093dd904f00de96c9f1420b1eef8ae224a6d32107045e1dfa6", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "7bec4fde-0642-4600-ac12-4829fe5048ae", "node_type": "1", "metadata": {}, "hash": "3bccf818cb61eb7052c2a89b9fae81422dd21a412f5fff0a0c9d69eea5650e49", "class_name": "RelatedNodeInfo"}}, "text": "[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi-\nnov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine\nLearning Research , 15(1):1929\u20131958, 2014.\n", "start_char_idx": 1502, "end_char_idx": 1735, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "7bec4fde-0642-4600-ac12-4829fe5048ae": {"__data__": {"id_": "7bec4fde-0642-4600-ac12-4829fe5048ae", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "b3c684f3-9fb4-499b-b85d-f8c168179b98", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7b4cd38ff6e38b2dabbd1edd476a34157af122470c247af2dd9d4affce2bef75", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "afce34e9-ad90-40d0-b91f-22b98b3e629e", "node_type": "1", "metadata": {}, "hash": "42f93162cc74963f7dcc86c2334f13671856c452eebd670250bb1e7afd830a07", "class_name": "RelatedNodeInfo"}}, "text": "[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory\nnetworks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors,\nAdvances in Neural Information Processing Systems 28 , pages 2440\u20132448. Curran Associates,\nInc., 2015.\n[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural\nnetworks. In Advances in Neural Information Processing Systems , pages 3104\u20133112, 2014.\n[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna.\nRethinking the inception architecture for computer vision. CoRR , abs/1512.00567, 2015.\n", "start_char_idx": 1735, "end_char_idx": 2380, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "afce34e9-ad90-40d0-b91f-22b98b3e629e": {"__data__": {"id_": "afce34e9-ad90-40d0-b91f-22b98b3e629e", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "7bec4fde-0642-4600-ac12-4829fe5048ae", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c044cdc18704ffff9d448325add32008d4dcd520fa42070a35cdd83e64df20d5", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "93a25878-f8c9-4040-a40c-c4adb172d3cc", "node_type": "1", "metadata": {}, "hash": "8d68d3f6744e1b97e31e9af6a78f4a7184143d70a4d8d4a132a2cf3671f401bb", "class_name": "RelatedNodeInfo"}}, "text": "[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In\nAdvances in Neural Information Processing Systems , 2015.\n[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang\nMacherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google\u2019s neural machine\ntranslation system: Bridging the gap between human and machine translation. arXiv preprint\narXiv:1609.08144 , 2016.\n[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with\nfast-forward connections for neural machine translation. CoRR , abs/1606.04199, 2016.\n", "start_char_idx": 2380, "end_char_idx": 2994, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "93a25878-f8c9-4040-a40c-c4adb172d3cc": {"__data__": {"id_": "93a25878-f8c9-4040-a40c-c4adb172d3cc", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "afce34e9-ad90-40d0-b91f-22b98b3e629e", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "73b3e9c645e02dc8dfc1bc6e3be0cdfceaadbc390d6a4535f9a1a66dab3f750c", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "c7a5188e-cdff-430d-aa72-592fba5a4eb2", "node_type": "1", "metadata": {}, "hash": "a5db1cde4cdbcdaa1ca213a650e7168ef85409b04f996221bf72423f6d52b8f4", "class_name": "RelatedNodeInfo"}}, "text": "[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate\nshift-reduce constituent parsing. In Proceedings of the 51st Annual Meeting of the ACL (Volume\n1: Long Papers) , pages 434\u2013443. ", "start_char_idx": 2994, "end_char_idx": 3209, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "c7a5188e-cdff-430d-aa72-592fba5a4eb2": {"__data__": {"id_": "c7a5188e-cdff-430d-aa72-592fba5a4eb2", "embedding": null, "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e", "node_type": "4", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "affe6912ae90c80377419f523ef69a96efc3f9928b0a3f97059ab1bcd45ca42d", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "93a25878-f8c9-4040-a40c-c4adb172d3cc", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "26f318ce06dcf322c1b7c7ba4e474ada94075c50979ee14b6336fb46c88464d1", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "9f775881-1987-4c1e-8a89-b2d53bf06cb3", "node_type": "1", "metadata": {}, "hash": "1c18b295c041c2822394737be0afb1c51f9258ddc315596ebdddc4bc593a1d8c", "class_name": "RelatedNodeInfo"}}, "text": "ACL, August 2013.\n12", "start_char_idx": 3209, "end_char_idx": 3229, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "9f775881-1987-4c1e-8a89-b2d53bf06cb3": {"__data__": {"id_": "9f775881-1987-4c1e-8a89-b2d53bf06cb3", "embedding": null, "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7cfca8ec-9d55-4349-9796-102e53179a86", "node_type": "4", "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "83f916af064b4eef81dc72612f14a9e99b432e342bad25bb3ee9f350fa0caae4", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "c7a5188e-cdff-430d-aa72-592fba5a4eb2", "node_type": "1", "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "39b99350971078f7fc3d0418b95292c0a299be0c6cc4df484ac63d18ad9c48db", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "fa150e9a-d081-4a51-97d0-be00e5f33d26", "node_type": "1", "metadata": {}, "hash": "e90d26ea6082900350e55977cebfd27272137fcae3ff1643187482884be70731", "class_name": "RelatedNodeInfo"}}, "text": "Attention Visualizations\nInput-Input Layer5\nIt\nis\nin\nthis\nspirit\nthat\na\nmajority\nof\nAmerican\ngovernments\nhave\npassed\nnew\nlaws\nsince\n2009\nmaking\nthe\nregistration\nor\nvoting\nprocess\nmore\ndifficult\n.\n<EOS>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\nIt\nis\nin\nthis\nspirit\nthat\na\nmajority\nof\nAmerican\ngovernments\nhave\npassed\nnew\nlaws\nsince\n2009\nmaking\nthe\nregistration\nor\nvoting\nprocess\nmore\ndifficult\n.\n<EOS>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\n<pad>\nFigure 3: An example of the attention mechanism following long-distance dependencies in the\nencoder self-attention in layer 5 of 6. Many of the attention heads attend to a distant dependency of\nthe verb \u2018making\u2019, completing the phrase \u2018making...more difficult\u2019. ", "start_char_idx": 0, "end_char_idx": 694, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "fa150e9a-d081-4a51-97d0-be00e5f33d26": {"__data__": {"id_": "fa150e9a-d081-4a51-97d0-be00e5f33d26", "embedding": null, "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "7cfca8ec-9d55-4349-9796-102e53179a86", "node_type": "4", "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "83f916af064b4eef81dc72612f14a9e99b432e342bad25bb3ee9f350fa0caae4", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "9f775881-1987-4c1e-8a89-b2d53bf06cb3", "node_type": "1", "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "0f66dd1fa94e42224747a720646f13aa6361d47db9f333ac9611fef2be76bbc1", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "8ce05fff-a173-46d6-955c-0349cc70ba53", "node_type": "1", "metadata": {}, "hash": "82178916cb1f647546c46f05856544259e8a5253367f062c9d726d3f2bd111ee", "class_name": "RelatedNodeInfo"}}, "text": "Attentions here shown only for\nthe word \u2018making\u2019. Different colors represent different heads. Best viewed in color.\n13", "start_char_idx": 694, "end_char_idx": 812, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "8ce05fff-a173-46d6-955c-0349cc70ba53": {"__data__": {"id_": "8ce05fff-a173-46d6-955c-0349cc70ba53", "embedding": null, "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6d6d77db-69ad-441f-8a04-34ee3cdef806", "node_type": "4", "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1aee818b1ec4a25980c87581f500a4cd556a90dc8e0a52ad4f6aae08ff93e6ef", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "fa150e9a-d081-4a51-97d0-be00e5f33d26", "node_type": "1", "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "7da9e349a032cba09bcb86f8bf8a29ffa22831d0745372e1fb2ffa7716163ea3", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "fc9c3269-3a82-478c-bdbd-a0f5999ccce6", "node_type": "1", "metadata": {}, "hash": "5258a15c954ee82b06d1aae6640117e9336cbedce629d4663d56af2389879f66", "class_name": "RelatedNodeInfo"}}, "text": "Input-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>Figure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top:\nFull attentions for head 5. ", "start_char_idx": 0, "end_char_idx": 675, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "fc9c3269-3a82-478c-bdbd-a0f5999ccce6": {"__data__": {"id_": "fc9c3269-3a82-478c-bdbd-a0f5999ccce6", "embedding": null, "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "6d6d77db-69ad-441f-8a04-34ee3cdef806", "node_type": "4", "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "1aee818b1ec4a25980c87581f500a4cd556a90dc8e0a52ad4f6aae08ff93e6ef", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "8ce05fff-a173-46d6-955c-0349cc70ba53", "node_type": "1", "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "abdba424cd7be782708c822c2e7f151604871a6898738214cb86b325d9430ff3", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "81109a7f-a6b1-4ef5-bec5-cf4dfa1e26f7", "node_type": "1", "metadata": {}, "hash": "74e3304cb376cacaafb53934946d5a9be2ae44e4f093003f9bc27f8cfcdca178", "class_name": "RelatedNodeInfo"}}, "text": "Bottom: Isolated attentions from just the word \u2018its\u2019 for attention heads 5\nand 6. Note that the attentions are very sharp for this word.\n14", "start_char_idx": 675, "end_char_idx": 814, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "81109a7f-a6b1-4ef5-bec5-cf4dfa1e26f7": {"__data__": {"id_": "81109a7f-a6b1-4ef5-bec5-cf4dfa1e26f7", "embedding": null, "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "421d32ad-3674-426c-88dd-7d44873ee0b7", "node_type": "4", "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ced7a073f0f0f4013feee9dcf052485cd29c25a4c52712cfe6f08c0d17107802", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "fc9c3269-3a82-478c-bdbd-a0f5999ccce6", "node_type": "1", "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "c60f4996f8dcb81dd4154fa77aa567c50325ae00304d6361c33aee7aef1b34a6", "class_name": "RelatedNodeInfo"}, "3": {"node_id": "16a597b5-9745-49c7-925a-c37a02d1848f", "node_type": "1", "metadata": {}, "hash": "23c3dcefc794b12005578dcccfd379b5caefc4110796fdf1bf866a478d202c07", "class_name": "RelatedNodeInfo"}}, "text": "Input-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nInput-Input Layer5\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>\nThe\nLaw\nwill\nnever\nbe\nperfect\n,\nbut\nits\napplication\nshould\nbe\njust\n-\nthis\nis\nwhat\nwe\nare\nmissing\n,\nin\nmy\nopinion\n.\n<EOS>\n<pad>Figure 5: Many of the attention heads exhibit behaviour that seems related to the structure of the\nsentence. We give two such examples above, from two different heads from the encoder self-attention\nat layer 5 of 6. The heads clearly learned to perform different tasks.\n", "start_char_idx": 0, "end_char_idx": 815, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}, "16a597b5-9745-49c7-925a-c37a02d1848f": {"__data__": {"id_": "16a597b5-9745-49c7-925a-c37a02d1848f", "embedding": null, "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "excluded_embed_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["file_name", "file_type", "file_size", "creation_date", "last_modified_date", "last_accessed_date"], "relationships": {"1": {"node_id": "421d32ad-3674-426c-88dd-7d44873ee0b7", "node_type": "4", "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "ced7a073f0f0f4013feee9dcf052485cd29c25a4c52712cfe6f08c0d17107802", "class_name": "RelatedNodeInfo"}, "2": {"node_id": "81109a7f-a6b1-4ef5-bec5-cf4dfa1e26f7", "node_type": "1", "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}, "hash": "638e638c4f36246e4e4551570c51e672d237533837ff9d3c7328588ef0741f8d", "class_name": "RelatedNodeInfo"}}, "text": "15", "start_char_idx": 815, "end_char_idx": 817, "text_template": "{metadata_str}\n\n{content}", "metadata_template": "{key}: {value}", "metadata_seperator": "\n", "class_name": "TextNode"}, "__type__": "1"}}, "docstore/metadata": {"3431a00b-b2c3-4aa3-9597-d5c79a81f9c5": {"doc_hash": "e5fad06d02d35abea195f627bac19ad00bca2e1036c8f8810f3b4ee531ca14c7", "ref_doc_id": "6d6fd0f4-5a6d-47a0-b304-59417f3cadb6"}, "135d7103-a9ad-43cd-b4e3-619fe02e901f": {"doc_hash": "ff03134021863208563480fbef84ac81faeb0087f4be4f89a65e000553e9e08e", "ref_doc_id": "6d6fd0f4-5a6d-47a0-b304-59417f3cadb6"}, "7fa60ea2-8f3d-403a-91ab-ca2409c0a30f": {"doc_hash": "c622a92b6bc9e23792425b98d86dd29232b783cd3e5c099cd76330a5790730db", "ref_doc_id": "6d6fd0f4-5a6d-47a0-b304-59417f3cadb6"}, "3ea43674-021d-4a53-bf50-8b7ba146a699": {"doc_hash": "2461968fbfdb377e396e4fcd7628cdff5e9e23bfaf2994f850fef85b09157907", "ref_doc_id": "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5"}, "40ceaf90-165a-4ba9-91cb-e441a1b06a33": {"doc_hash": "7de9d0ce7f1f67f6746d3dd0073dcfce83e23a2d541737cfb5a5e72c7056933e", "ref_doc_id": "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5"}, "4e00cc5e-34a8-4d28-96cd-521a8613082e": {"doc_hash": "ee9d9ba0f9c4dee86ba934c329cbd5c2561d61f5da47753803c99c6e53e586b4", "ref_doc_id": "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5"}, "72dfb8e6-9f71-4211-a72d-5ec77bd4b2fe": {"doc_hash": "8f3f71fcecd60992e39150337420806a2a1342fa9d895e95c3c1d6319d861979", "ref_doc_id": "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5"}, "bb6aa732-aa8c-440f-86bb-de07bd602ea1": {"doc_hash": "ec844fda5ac16742fcb4be7508ed70c02440962916828d5582d53939c797fd65", "ref_doc_id": "10071188-42e8-4929-98dd-642ee4131169"}, "536fb5fb-63e5-4170-80b6-739234a9149f": {"doc_hash": "2e731d358bfcfda1c02958e2f95c5290561829d2a32e82ac769b3dda94ffaf86", "ref_doc_id": "10071188-42e8-4929-98dd-642ee4131169"}, "fce102f2-322f-4e80-8130-96ff394dd00d": {"doc_hash": "1ae967901d22e2c27573427e0a60288b2a0d6e75e5a20f39487c84d8f6faa5db", "ref_doc_id": "10071188-42e8-4929-98dd-642ee4131169"}, "637186e0-6b70-47af-9395-a5f23c9922c7": {"doc_hash": "25d2f70ef16cdd181bcf39f32b2fa03c410a8ea0d81c8a4e7350866c7d4e3f0a", "ref_doc_id": "de79a486-751a-414d-b21e-078ab065ed82"}, "422f9866-8f04-463f-8d50-eab9a4886128": {"doc_hash": "6d8d07e874fb92b09011ea27bcd9ef831e9cabce9634f81d8862de406b9ab9ed", "ref_doc_id": "de79a486-751a-414d-b21e-078ab065ed82"}, "8687e9ed-235b-4278-a496-45395cddaf37": {"doc_hash": "26988312de96dce95a0094d5e86a130f91e5f69e4d6a720c5a581491f23ed818", "ref_doc_id": "de79a486-751a-414d-b21e-078ab065ed82"}, "6524bd60-c48e-4539-b3e1-78075b31bac0": {"doc_hash": "23f9f422947d199bca0707b35985de63fb61e9b64a4622365a90a1cb8d4f5cbe", "ref_doc_id": "1006cf3f-32fc-4c2e-b31a-f519be43fd1b"}, "412d7574-4fb3-4985-8745-7494188dd8e6": {"doc_hash": "6f7ef2d54429da3b4bfe849e20e66927288c8f200a426a593a4ce40fced9bb5e", "ref_doc_id": "1006cf3f-32fc-4c2e-b31a-f519be43fd1b"}, "ff341210-7431-4b1b-a4d2-0b144b35d985": {"doc_hash": "9df99719393b4ec72ccf855808242c10151dcb245937fcda88f7e8456ec7e8fb", "ref_doc_id": "1006cf3f-32fc-4c2e-b31a-f519be43fd1b"}, "abdc2077-fd4a-4184-bd93-07b2088061de": {"doc_hash": "ed5894d3e4cc2756a80d2170b70d9eb799ba0dd9bd09c0986a226bfedba5444a", "ref_doc_id": "1006cf3f-32fc-4c2e-b31a-f519be43fd1b"}, "67d23967-8054-4c86-9a3f-213604c7f8d6": {"doc_hash": "4149d0733a2077929fc0fc7832725ed2c7850610d31f1fa2e6770bea04f69a4d", "ref_doc_id": "e95894c0-aec1-4614-b98f-d141b63efd5c"}, "67607d1c-0394-46f4-8a3b-6bb6cff1bfe1": {"doc_hash": "087f13d1ed39ef41492b3d272f5c1b21b4b783865a4247e17bc158debda912f5", "ref_doc_id": "e95894c0-aec1-4614-b98f-d141b63efd5c"}, "84010686-fc60-4ee0-b307-f08a0e19cfe6": {"doc_hash": "ef3f6df3b02956ee87ad5c023687353efda09b51ec3d0682d81ad9e3c0af0e23", "ref_doc_id": "e95894c0-aec1-4614-b98f-d141b63efd5c"}, "bbfd17cd-9afc-4fae-9279-52965ee289d1": {"doc_hash": "930cc5f06b3219abf8c66d02b61b945a23f8268250497902ad05ed77a71d848f", "ref_doc_id": "e95894c0-aec1-4614-b98f-d141b63efd5c"}, "0c7f8f31-d6ab-4f30-803d-7a0056b15963": {"doc_hash": "58db0b38b6326b1bf61137c8dbe62d75f8d11c21e85243ec183c9bda8fb05027", "ref_doc_id": "ab05337f-6fcc-408d-8685-7654ec1f476d"}, "07227a15-44fc-4950-9f7d-7c1461364365": {"doc_hash": "f48d0ff5ec08366539996b772ac04c071acf8af22b72c11e1d8f8931409f4bcc", "ref_doc_id": "ab05337f-6fcc-408d-8685-7654ec1f476d"}, "6c2e74fb-f80a-4074-97a1-2f78727f1fc7": {"doc_hash": "18393b3aa39c791d80bb634cbf5b3e31f34a2417b1d83fc705b84c50c0e7d67b", "ref_doc_id": "ab05337f-6fcc-408d-8685-7654ec1f476d"}, "612fbc2e-e4ab-4d6e-b3d6-af6dc250aaa8": {"doc_hash": "eb179bdea2450b5906d42af007dd524a30ad1d4232843531eed033060a74a62f", "ref_doc_id": "ab05337f-6fcc-408d-8685-7654ec1f476d"}, "826d10fd-380f-47e8-b970-51cd5c3e9270": {"doc_hash": "f8a87f87a229ea433cafaddb9a7b882283a1255e295a8e9f20c489580811c1b8", "ref_doc_id": "8df67ec6-1e0f-4188-9687-0323556948cb"}, "b3bfb5b0-4f15-45b3-970b-ea71328a74cf": {"doc_hash": "847192d4d9919bf8dc8014956a35a1777b97e1b9ac70ad634785cfbb7705755a", "ref_doc_id": "8df67ec6-1e0f-4188-9687-0323556948cb"}, "42d035cb-f399-442d-b7da-260217dd8a0c": {"doc_hash": "b4f25e74ebd0475b910b654d5a93c49c99526a91afd42bb12bd4b92824d630b6", "ref_doc_id": "8df67ec6-1e0f-4188-9687-0323556948cb"}, "aef4eca4-aecd-4f06-99ee-3c35d2427597": {"doc_hash": "a0f9febd7e187836acd4a8a2f32e8e1c8dcf61714ac8c19e23a90cc151f71418", "ref_doc_id": "8e0c2ca4-0215-412a-ba2e-f675e6a861af"}, "9c91e9f4-91a7-403b-9a4e-6f517810bcad": {"doc_hash": "fd1f6f3268954ed574af42f2a7adca2cb23a5b323cc99293d58a8a7f2e51f71f", "ref_doc_id": "8e0c2ca4-0215-412a-ba2e-f675e6a861af"}, "739d85c4-e25b-49ae-a049-96c25851a609": {"doc_hash": "ffdd464141d42dff9893063dd9743953ef246c6f3caedc88f3a461ac8d2ada65", "ref_doc_id": "8e0c2ca4-0215-412a-ba2e-f675e6a861af"}, "d463eec0-8dbb-4558-b9e9-5b3ff0ed9137": {"doc_hash": "2452e5b3e5ff62d94096818af5a1d3620d61489948275f630fc54745090bf3a3", "ref_doc_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce"}, "1f60dd32-a23b-4678-880c-cee36938e787": {"doc_hash": "ca90a964e379767202d4fd498fdb92350f9e2a87fda59e4392e11f70fd3ae227", "ref_doc_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce"}, "680e94b6-62fb-4536-a7ff-32dc41f99c89": {"doc_hash": "7ebcbdca9beb0f2b2055033447e5820ee04856ee3b6f8376a7a43470e27568fc", "ref_doc_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce"}, "3a6477bc-b068-4c75-90c1-c406affa9259": {"doc_hash": "0be6230d8a0e1ec43cfcd886930a0beb9ed97ac6da6889b42bc879831d7bf862", "ref_doc_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce"}, "51d4b860-054a-4489-b47a-281b1b4f11ad": {"doc_hash": "5e8ee706174d8b81333912e397fce755fc14b64f18580c768109292b8da150a4", "ref_doc_id": "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce"}, "27a20239-3af4-4aac-841c-c53dbde13a14": {"doc_hash": "990904d016b4432da73521124cde5423e69795eb577199c4f45c7108326be663", "ref_doc_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3"}, "d5d45e29-a0d7-4bb4-9cc7-c8d088f4c963": {"doc_hash": "4fced396588f621fccc8cfcd794eeb4a6de05086890061e25ca1653793c73c56", "ref_doc_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3"}, "ad73e473-afa4-41aa-a795-c7a9b1082fa9": {"doc_hash": "8711b9214dabab460c0c1db37828ef546489b3e37782eb8fce220a74cd5e5f81", "ref_doc_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3"}, "de4d13e0-dfe7-4887-a0f6-e087cbcda550": {"doc_hash": "a0e76f2b888e24855cd06dd7b08cfa2c499637b1331c61e5dfaad481bc9472e9", "ref_doc_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3"}, "6e5ad201-a5fe-4090-bfbb-17ef7fcb9bb8": {"doc_hash": "bc6c2e7b082588db7c8384a6a7b4eea48521d1809219b69c4422420906b566c6", "ref_doc_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3"}, "b4bdc4ea-ca2a-496d-b377-238c8447c25b": {"doc_hash": "e4d727ce734f9ddfd9f0e20b4820ac060878d95d4915de9f724dbc88a3017055", "ref_doc_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3"}, "6533451f-b299-4d94-ab0e-f50c4bb006cf": {"doc_hash": "d3738f25a0cd8c4d20d3085de3e53869ef7e448baed6654c781e2885379d9496", "ref_doc_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3"}, "cdf35e7c-fe91-4b76-83b7-a59e88b47277": {"doc_hash": "9e6f66859ddfec72e70fb13359e1f388130a6c023a5c82cf503736651dc1f7c7", "ref_doc_id": "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3"}, "91571f0f-6bcf-4172-870b-8a9ff8c2f42e": {"doc_hash": "8a443340e87718cb142ec1aa7a8da55f95242662bb1cf660d0b6f3da84dd5e02", "ref_doc_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac"}, "fdb3ba2e-2595-4e34-b4ff-55ddb6157c27": {"doc_hash": "cb2a3128ee8cd9093dd904f00de96c9f1420b1eef8ae224a6d32107045e1dfa6", "ref_doc_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac"}, "d36501b8-45cd-4893-8869-e0deadd4d8f7": {"doc_hash": "7b4cd38ff6e38b2dabbd1edd476a34157af122470c247af2dd9d4affce2bef75", "ref_doc_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac"}, "df80c204-cf8e-47e7-a08e-ef0f828244ac": {"doc_hash": "c044cdc18704ffff9d448325add32008d4dcd520fa42070a35cdd83e64df20d5", "ref_doc_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac"}, "0e66ad56-e2bc-41d3-9693-771e1655bd78": {"doc_hash": "73b3e9c645e02dc8dfc1bc6e3be0cdfceaadbc390d6a4535f9a1a66dab3f750c", "ref_doc_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac"}, "fe56f7b0-b0aa-4bab-b28c-c8eb52d2dab1": {"doc_hash": "26f318ce06dcf322c1b7c7ba4e474ada94075c50979ee14b6336fb46c88464d1", "ref_doc_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac"}, "357c6c9e-81fe-40af-ac73-8af5a4a0797c": {"doc_hash": "39b99350971078f7fc3d0418b95292c0a299be0c6cc4df484ac63d18ad9c48db", "ref_doc_id": "edd3667f-1e40-42fa-ad91-ef19863b7aac"}, "f6bd0eec-8952-4d79-b677-d7589a0c9025": {"doc_hash": "0f66dd1fa94e42224747a720646f13aa6361d47db9f333ac9611fef2be76bbc1", "ref_doc_id": "21c4166d-6661-46b2-955e-0e7a112b0a80"}, "79a1ef13-5aca-4612-8f62-bfcf74a5f248": {"doc_hash": "7da9e349a032cba09bcb86f8bf8a29ffa22831d0745372e1fb2ffa7716163ea3", "ref_doc_id": "21c4166d-6661-46b2-955e-0e7a112b0a80"}, "84327c1d-50ac-4c2c-bca4-773e888100f1": {"doc_hash": "abdba424cd7be782708c822c2e7f151604871a6898738214cb86b325d9430ff3", "ref_doc_id": "d167b925-01c2-44fa-aad1-97716bc024bc"}, "2f750689-4461-4756-8fde-01726308eaf3": {"doc_hash": "c60f4996f8dcb81dd4154fa77aa567c50325ae00304d6361c33aee7aef1b34a6", "ref_doc_id": "d167b925-01c2-44fa-aad1-97716bc024bc"}, "16dc7d44-1a9e-4ed7-8959-6798f8529cac": {"doc_hash": "638e638c4f36246e4e4551570c51e672d237533837ff9d3c7328588ef0741f8d", "ref_doc_id": "4bc641df-b009-48a6-8ba9-9d001e5879b7"}, "f8bb10ed-3dd9-4d4d-9c27-d66565142fbd": {"doc_hash": "ffe3659172531da0d7313fc185d6509a6a6b811247ac5d61833440c86e1d0b99", "ref_doc_id": "4bc641df-b009-48a6-8ba9-9d001e5879b7"}, "4ae62500-f4b7-4576-bfd1-dd7a206dd0b7": {"doc_hash": "854e1173728be619baa11d2a14d9a0da66822261ba6501caedeb38a116ae1359", "ref_doc_id": "8c7447f6-6a91-435c-80ce-786595a1941b"}, "03db1e24-2be5-417d-aca0-fd45ec65c9df": {"doc_hash": "f112420a13a842d1bffbdde4d53bca9b21072e7c99ec41fb4b01dfac76d7cf16", "ref_doc_id": "8c7447f6-6a91-435c-80ce-786595a1941b"}, "ce5cdd06-2fdf-47b3-a43a-47e144522450": {"doc_hash": "fb990e5b8378b925ee40845ae6e5c5674c2ebf9234e1b737f8cb3965a8d28924", "ref_doc_id": "8c7447f6-6a91-435c-80ce-786595a1941b"}, "5337277a-d5ed-4ff1-b72e-4e76d6d7b65e": {"doc_hash": "3178c3ccad23d13d83701bd4fd1a779c64ad3ee09c190028be13ad4f6026b3a8", "ref_doc_id": "8c7447f6-6a91-435c-80ce-786595a1941b"}, "1dae63a4-7407-4fec-911b-2befb2937f12": {"doc_hash": "0f85c02e4238f7eee72e00b3b4d3915fbcac0812d45b87fe943f6f4faa40fc72", "ref_doc_id": "7c1efa7b-b862-45f1-8f56-75407df01893"}, "3dd4b47e-81ae-41a1-9efb-277f7ee3c064": {"doc_hash": "7b856afe11149f92c9cdeb662455d2e1537d3a7b0b851df90019d56ec9783d35", "ref_doc_id": "7c1efa7b-b862-45f1-8f56-75407df01893"}, "7181f2d0-c9a5-45a7-86cb-3ee42f1833c0": {"doc_hash": "050d25b181f9347622d6be2e67d5acc2f749ece9eb7e540841ac987c17e2c91a", "ref_doc_id": "7c1efa7b-b862-45f1-8f56-75407df01893"}, "0f5f3713-6d91-42ee-a1d0-6f6d439ea009": {"doc_hash": "2c03e6fdfaa7612e38e9c3530588951a3928a7b29384e995cfbc73e040516232", "ref_doc_id": "7c1efa7b-b862-45f1-8f56-75407df01893"}, "e339daf1-51f9-4679-a79b-40564c5ccf0b": {"doc_hash": "c0dd363079f7880ac69a36b1905de189031268f45d010ff3872722418c10a53e", "ref_doc_id": "7c1efa7b-b862-45f1-8f56-75407df01893"}, "e7c5b44d-5ae0-4cc6-9d00-dee9cb2a8814": {"doc_hash": "e635cc1759213b8b444e6fd44991b0c21a24965f0ef7546c05f02f3131d9ee70", "ref_doc_id": "495b8b80-518b-4889-9658-6901e0229bc0"}, "285134f3-afad-4ca8-a9a3-d4d13b2438c1": {"doc_hash": "4f472d203cf887d9754200c958ebdff874b6839cfb4e60123d4c1a5617b6cd23", "ref_doc_id": "495b8b80-518b-4889-9658-6901e0229bc0"}, "f4519fee-459f-4522-b07d-82747620a572": {"doc_hash": "efb449c09a7384b0bc47ed7d0f48350c3fadcc92d4d4f85565f6f62347dfe23e", "ref_doc_id": "495b8b80-518b-4889-9658-6901e0229bc0"}, "ccc697df-34bf-4228-a9c7-5834f08cd529": {"doc_hash": "80de42688b76a29cabb2b2b7194c9405c5e85541b3938c88f8fdc08d20b19bef", "ref_doc_id": "495b8b80-518b-4889-9658-6901e0229bc0"}, "2eadfedb-30e8-41e5-8a28-756700b37955": {"doc_hash": "692bca53190ce8ef17128416e23cb12445935daa8c9eabae4be5b8832fe1b1c0", "ref_doc_id": "6ad9b48a-2a55-40b1-9808-07cd8e439c05"}, "6bc1dfa7-bb04-4d5f-ae1e-90f9723a74a7": {"doc_hash": "e60ee4ab698375af92d3103104908c062b4d9c1219e48a8b14e0342feac2dc3e", "ref_doc_id": "6ad9b48a-2a55-40b1-9808-07cd8e439c05"}, "1c0272bc-94c8-41e3-8fe7-a93957197edd": {"doc_hash": "7f62a1e26d1b342046d0f61f84de2ccddb69c779de57200bdf914b84b062fd6c", "ref_doc_id": "6ad9b48a-2a55-40b1-9808-07cd8e439c05"}, "23cc386c-71cf-4b1a-bdcd-d2fd8be427cc": {"doc_hash": "63beb233a676f65a336090ad756d1aadd9007a32cc59922d110161a79dae3d09", "ref_doc_id": "6ad9b48a-2a55-40b1-9808-07cd8e439c05"}, "1223e2d0-f3f1-45f1-9962-24e1d2a3139c": {"doc_hash": "f3388435b0e94eae6a26de8ab56b586ab5cf73364bc23877542c4435763ff135", "ref_doc_id": "501f9969-eca6-4e92-9080-b03e896be613"}, "92b507ee-ce43-4dd8-b017-7dcbdf070435": {"doc_hash": "53f59ef1f5e610b5e4f0b15dba8440677d52d2b910c550957109b3b818164709", "ref_doc_id": "501f9969-eca6-4e92-9080-b03e896be613"}, "02afc34b-32f4-4c24-b536-349dd39356e3": {"doc_hash": "4625d48a162c194f959462b3cdd39131c9ec62082453cbd70eb6d44456f37301", "ref_doc_id": "501f9969-eca6-4e92-9080-b03e896be613"}, "65332716-e35c-496a-b0bc-e6a4ed0cc983": {"doc_hash": "517bf1cba2d2b86e1d2be7ec59c2f325c0bd949a649cdc5e7ec09bf76fe75057", "ref_doc_id": "501f9969-eca6-4e92-9080-b03e896be613"}, "8c2c7fe2-0b52-4c53-81be-65d29a7cb679": {"doc_hash": "b1afd56513692b317ef9d86b9ca4217f0a6a5fb1e7b233805ed2b61f00aecf83", "ref_doc_id": "defb123a-181a-439f-9b2a-c8baddeb016e"}, "15d23f05-58ab-4ae9-92c7-9ea57abb2ae3": {"doc_hash": "fe85b2dec32804668809c04e83c6101883c7e74867fb62ca340cc951bd96d445", "ref_doc_id": "defb123a-181a-439f-9b2a-c8baddeb016e"}, "e01a3c34-8381-4a3d-aad8-4d3de10bd55f": {"doc_hash": "05aaf00d208a1e422e5acdfd67c245231de26c7b5319e94c561f252054141c72", "ref_doc_id": "defb123a-181a-439f-9b2a-c8baddeb016e"}, "8040ef1b-6d72-49bd-a743-013c777a8bc0": {"doc_hash": "4fbc6ad5867b4893855fbcca008106baa59d145d72a240cb8329be7aaa6e0afc", "ref_doc_id": "defb123a-181a-439f-9b2a-c8baddeb016e"}, "d0f23a06-1048-4e37-9e50-ca0035756ec6": {"doc_hash": "337a4eb6f4d8a41034a11178e27d97f95fb25d446ec9bdfed3c48cbe28449728", "ref_doc_id": "defb123a-181a-439f-9b2a-c8baddeb016e"}, "282e536f-fe88-4f7a-a179-754220179b24": {"doc_hash": "68cad5e0fa0b9cb32544333f236c6e2c818681cecc1e163dfa3b271d43b3259e", "ref_doc_id": "defb123a-181a-439f-9b2a-c8baddeb016e"}, "b5e4fffe-eede-4a33-ade6-23fcded34c28": {"doc_hash": "ce4e3dcb8c15078726ada683b9ea412d68c56d23e0dde61984a972a2aad5950a", "ref_doc_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633"}, "e2ca30f9-a9f6-409f-8018-03870a276b86": {"doc_hash": "5b5b2b6282b576441f9e3a864ae84fe24d90d51014a640d99901bd6d1862208f", "ref_doc_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633"}, "8a29b80e-3241-4249-8c79-3058f85811ed": {"doc_hash": "2a3332e07c8b4384a5e09cd9baf05824754cd6f744efc084d523e8c344218d00", "ref_doc_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633"}, "2069f9a4-c773-4744-a410-67dfc05b1dc5": {"doc_hash": "448e25e6e5abb25955acdd85ff93893027a1d5d37ebf100a352d14e512b395d9", "ref_doc_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633"}, "612ad73d-a6a1-434e-a9c1-cb0f768657ed": {"doc_hash": "4e66e924330425b7075d15666249089f298e839f4728f88c8f1d692a92255ab8", "ref_doc_id": "3d538daf-6e8d-4715-a7c7-c3256db4d633"}, "1bced689-bd10-4d07-8faf-c77ead5b12ca": {"doc_hash": "6e9d6dc7a3209b90923ecae233a451fdbe1a6e75b97027ba106c1e45befd0459", "ref_doc_id": "55f2872f-47ad-45b9-92b9-21548eaf276e"}, "afe7667f-e82d-42fb-8fde-0cb1954cf5d1": {"doc_hash": "a8f4dda62982dac0fd4f0e9647019b3e17cdc6a880e77ac15e1ea873efafea64", "ref_doc_id": "55f2872f-47ad-45b9-92b9-21548eaf276e"}, "41e711b7-ffd9-4a80-9d2f-3b24107c2763": {"doc_hash": "3a90b5c06bcb8aa84e7a8e1fe848e63012a968fe47a11b799f462e3df1e56eb7", "ref_doc_id": "55f2872f-47ad-45b9-92b9-21548eaf276e"}, "9afdc6f5-b962-4b05-9964-28346b31319d": {"doc_hash": "29e1bb31e456e1fd8fa3330301c5ba66e57efa8e218c991ff57daabc4d8edb64", "ref_doc_id": "55f2872f-47ad-45b9-92b9-21548eaf276e"}, "356b6eab-f375-4784-a9a9-b346e1d248f9": {"doc_hash": "5a21f50f8a04ce6f841393d90b70bbe9b4b5fd3ee85c723b0b6959a547e17eaf", "ref_doc_id": "55f2872f-47ad-45b9-92b9-21548eaf276e"}, "5fffd5af-e267-451e-bdef-ce3e373ee153": {"doc_hash": "0125b97faa3a21500503bcd9c08f5156102f9490b91608d41ad74c4ecfa6297a", "ref_doc_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db"}, "6ef9eca0-2c8a-4e3e-85d4-d052726b1d8f": {"doc_hash": "779d1528dcec97358a3fa5f23046f233b6c261856ef6d1fcd1a4e6e839476553", "ref_doc_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db"}, "124d2351-91ef-48a0-b5b0-eb5a7e3b0a78": {"doc_hash": "476c9198f0b338b489bbf71a984c347f5981ca9fa093c928b3ed320d89df1211", "ref_doc_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db"}, "1b37c06a-7824-45fd-8f61-d101b5a9df19": {"doc_hash": "fb215b41abeba0959faad13b87990e002b2b0ff18c6a11f58746800e049f8e61", "ref_doc_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db"}, "8b1fa614-496b-43af-993f-498d707efb50": {"doc_hash": "16ad24743741abee31f0dc2cc3941345ed47e0809edc75e056b78cd3cc5c6274", "ref_doc_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db"}, "7124a746-c0f8-4291-88f5-da1376271eb5": {"doc_hash": "baa95f8666ed8ec5fdc1d41cc7ecb8fff97570b650fc04273f0e88af51e425ce", "ref_doc_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db"}, "5979ee85-867e-4749-b395-408a98e32df4": {"doc_hash": "b1cc529152403413241e213ac736e951831bbdc314ee74ef7a806438f8950711", "ref_doc_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db"}, "fcc49761-9190-43c7-b273-101c413fb492": {"doc_hash": "c70af929ec42a31913065dfe5a801e860551412546424915ac47c385af840428", "ref_doc_id": "50e3ac80-5506-4247-9e25-6a2f3c5f33db"}, "97c2d55b-945c-4c01-9dc0-86bd520257d0": {"doc_hash": "d3eb79886d6a4e5e44a9ea6f8a6e1e2834fb824b5af37a63ee29f7d6f7ce1401", "ref_doc_id": "4a57759d-5ee3-4c11-8861-9bcad409f797"}, "6a2b5ec3-b6df-4951-9797-fac4f6881808": {"doc_hash": "3ac213c6f19275b96c8066f3305c620b9e7136bf65feacd99ffa4784b9fe1d49", "ref_doc_id": "4a57759d-5ee3-4c11-8861-9bcad409f797"}, "38ac2daf-0a08-4e26-b28c-cc5f9a69f395": {"doc_hash": "70d4592261cdda5ef9025cb2d6f6e59d3b5b84174cde2260476d5360b6c97f35", "ref_doc_id": "597b8534-da28-40e6-af8b-49dc5f661cad"}, "b7c64a3c-4735-4672-815f-7cf7cd07c193": {"doc_hash": "e5fad06d02d35abea195f627bac19ad00bca2e1036c8f8810f3b4ee531ca14c7", "ref_doc_id": "7bcf88e4-6827-4be9-988b-55bb1b49d948"}, "d5f956f2-7566-4ce2-b90c-9b1972c8a919": {"doc_hash": "ff03134021863208563480fbef84ac81faeb0087f4be4f89a65e000553e9e08e", "ref_doc_id": "7bcf88e4-6827-4be9-988b-55bb1b49d948"}, "2741d814-901a-40c1-b959-552157129a9d": {"doc_hash": "c622a92b6bc9e23792425b98d86dd29232b783cd3e5c099cd76330a5790730db", "ref_doc_id": "7bcf88e4-6827-4be9-988b-55bb1b49d948"}, "2923d6a1-d583-496a-81ed-0ab01dd2aaaf": {"doc_hash": "2461968fbfdb377e396e4fcd7628cdff5e9e23bfaf2994f850fef85b09157907", "ref_doc_id": "0ec692ee-efde-4036-bc5b-51ba9d6eca6c"}, "bfa3ce6d-c588-4f78-90f0-f0245cc7e6ba": {"doc_hash": "7de9d0ce7f1f67f6746d3dd0073dcfce83e23a2d541737cfb5a5e72c7056933e", "ref_doc_id": "0ec692ee-efde-4036-bc5b-51ba9d6eca6c"}, "300e8947-8fa3-417d-9a79-9ae7c242f4ef": {"doc_hash": "ee9d9ba0f9c4dee86ba934c329cbd5c2561d61f5da47753803c99c6e53e586b4", "ref_doc_id": "0ec692ee-efde-4036-bc5b-51ba9d6eca6c"}, "93964b05-4478-455c-87d2-d75c334e4c09": {"doc_hash": "8f3f71fcecd60992e39150337420806a2a1342fa9d895e95c3c1d6319d861979", "ref_doc_id": "0ec692ee-efde-4036-bc5b-51ba9d6eca6c"}, "d5bcec10-5d92-4d3b-93e1-07345863d62e": {"doc_hash": "ec844fda5ac16742fcb4be7508ed70c02440962916828d5582d53939c797fd65", "ref_doc_id": "c8a07910-0192-481a-bcf1-1a58d71a8f38"}, "dc04e657-c9d4-4ae4-9608-d506dcdba62d": {"doc_hash": "2e731d358bfcfda1c02958e2f95c5290561829d2a32e82ac769b3dda94ffaf86", "ref_doc_id": "c8a07910-0192-481a-bcf1-1a58d71a8f38"}, "019fd1ce-c068-408b-8d2d-58d3a5d48b96": {"doc_hash": "1ae967901d22e2c27573427e0a60288b2a0d6e75e5a20f39487c84d8f6faa5db", "ref_doc_id": "c8a07910-0192-481a-bcf1-1a58d71a8f38"}, "c304a4ca-c2b7-4211-9bc8-2a5108b28e07": {"doc_hash": "25d2f70ef16cdd181bcf39f32b2fa03c410a8ea0d81c8a4e7350866c7d4e3f0a", "ref_doc_id": "cd926aa7-9c47-4517-abf3-097ddfe907f2"}, "bdf8130c-d544-420b-9005-33872a3f74c1": {"doc_hash": "6d8d07e874fb92b09011ea27bcd9ef831e9cabce9634f81d8862de406b9ab9ed", "ref_doc_id": "cd926aa7-9c47-4517-abf3-097ddfe907f2"}, "ab54a858-bc18-4ee5-b02f-1f549c149532": {"doc_hash": "26988312de96dce95a0094d5e86a130f91e5f69e4d6a720c5a581491f23ed818", "ref_doc_id": "cd926aa7-9c47-4517-abf3-097ddfe907f2"}, "b334377e-b057-43e0-b76f-17f19781fc1d": {"doc_hash": "23f9f422947d199bca0707b35985de63fb61e9b64a4622365a90a1cb8d4f5cbe", "ref_doc_id": "e1e2ea31-60e3-4801-805b-c4df3bc294dd"}, "b3784fbe-100b-49fc-b7f7-1d16b9296a1e": {"doc_hash": "6f7ef2d54429da3b4bfe849e20e66927288c8f200a426a593a4ce40fced9bb5e", "ref_doc_id": "e1e2ea31-60e3-4801-805b-c4df3bc294dd"}, "897256a3-8c19-4b9b-8f51-204cff114902": {"doc_hash": "9df99719393b4ec72ccf855808242c10151dcb245937fcda88f7e8456ec7e8fb", "ref_doc_id": "e1e2ea31-60e3-4801-805b-c4df3bc294dd"}, "989eb0dd-6927-47e5-91e3-d6f522a4bd2a": {"doc_hash": "ed5894d3e4cc2756a80d2170b70d9eb799ba0dd9bd09c0986a226bfedba5444a", "ref_doc_id": "e1e2ea31-60e3-4801-805b-c4df3bc294dd"}, "38465ff0-f5e8-43e3-a581-ff76dc2cea52": {"doc_hash": "4149d0733a2077929fc0fc7832725ed2c7850610d31f1fa2e6770bea04f69a4d", "ref_doc_id": "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5"}, "c1fac594-7f52-4e5f-abbd-a8b9a2385ea1": {"doc_hash": "087f13d1ed39ef41492b3d272f5c1b21b4b783865a4247e17bc158debda912f5", "ref_doc_id": "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5"}, "abe19b59-2b09-4ccc-ad63-74dc10dd0989": {"doc_hash": "ef3f6df3b02956ee87ad5c023687353efda09b51ec3d0682d81ad9e3c0af0e23", "ref_doc_id": "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5"}, "700226de-87f3-4ff6-8023-2e460ab3599d": {"doc_hash": "930cc5f06b3219abf8c66d02b61b945a23f8268250497902ad05ed77a71d848f", "ref_doc_id": "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5"}, "5d14e9e0-9083-4dae-aa0f-a3b3ad856ae2": {"doc_hash": "58db0b38b6326b1bf61137c8dbe62d75f8d11c21e85243ec183c9bda8fb05027", "ref_doc_id": "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7"}, "0716be40-0d7e-4a1a-9cf0-9f14437b82f9": {"doc_hash": "f48d0ff5ec08366539996b772ac04c071acf8af22b72c11e1d8f8931409f4bcc", "ref_doc_id": "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7"}, "903896c3-e90f-4fb8-933c-3b9a419f220d": {"doc_hash": "18393b3aa39c791d80bb634cbf5b3e31f34a2417b1d83fc705b84c50c0e7d67b", "ref_doc_id": "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7"}, "0fa8f49d-ee77-4e8b-99f9-430ef1086b18": {"doc_hash": "eb179bdea2450b5906d42af007dd524a30ad1d4232843531eed033060a74a62f", "ref_doc_id": "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7"}, "7ac597d6-9c0c-4946-b6da-2e6954dd3ff1": {"doc_hash": "f8a87f87a229ea433cafaddb9a7b882283a1255e295a8e9f20c489580811c1b8", "ref_doc_id": "0da32021-5f29-482c-ba55-b4082af76e37"}, "333ede79-147a-4ddb-a536-4af80aa36813": {"doc_hash": "847192d4d9919bf8dc8014956a35a1777b97e1b9ac70ad634785cfbb7705755a", "ref_doc_id": "0da32021-5f29-482c-ba55-b4082af76e37"}, "5d8d189d-424b-4a2a-b0ad-e28758dce1d4": {"doc_hash": "b4f25e74ebd0475b910b654d5a93c49c99526a91afd42bb12bd4b92824d630b6", "ref_doc_id": "0da32021-5f29-482c-ba55-b4082af76e37"}, "caa5c999-7d0e-447b-a3ac-6e3180209f3d": {"doc_hash": "a0f9febd7e187836acd4a8a2f32e8e1c8dcf61714ac8c19e23a90cc151f71418", "ref_doc_id": "41e170cc-f118-4a60-bb0a-2ec27a0f6318"}, "2cab4eb7-bc6c-4c75-8dcc-763aaaceb767": {"doc_hash": "fd1f6f3268954ed574af42f2a7adca2cb23a5b323cc99293d58a8a7f2e51f71f", "ref_doc_id": "41e170cc-f118-4a60-bb0a-2ec27a0f6318"}, "7466eb16-7489-427f-9838-f8bdb2c721a7": {"doc_hash": "ffdd464141d42dff9893063dd9743953ef246c6f3caedc88f3a461ac8d2ada65", "ref_doc_id": "41e170cc-f118-4a60-bb0a-2ec27a0f6318"}, "3470fcf0-d739-471d-a9cb-46f1eaddc675": {"doc_hash": "2452e5b3e5ff62d94096818af5a1d3620d61489948275f630fc54745090bf3a3", "ref_doc_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0"}, "9acba64d-c6f2-406b-a1c8-5da9691f5868": {"doc_hash": "ca90a964e379767202d4fd498fdb92350f9e2a87fda59e4392e11f70fd3ae227", "ref_doc_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0"}, "cce9180a-7469-43e7-955c-62a11e492759": {"doc_hash": "7ebcbdca9beb0f2b2055033447e5820ee04856ee3b6f8376a7a43470e27568fc", "ref_doc_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0"}, "50c18361-106f-4a2a-a0a5-731ef39b76f9": {"doc_hash": "0be6230d8a0e1ec43cfcd886930a0beb9ed97ac6da6889b42bc879831d7bf862", "ref_doc_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0"}, "322758d7-f621-4f04-abb6-3d73e08e273e": {"doc_hash": "5e8ee706174d8b81333912e397fce755fc14b64f18580c768109292b8da150a4", "ref_doc_id": "10a87b3a-d733-48e8-ae5c-72eacb885ea0"}, "087633e1-63a2-4f8e-a266-c2e92468c58a": {"doc_hash": "990904d016b4432da73521124cde5423e69795eb577199c4f45c7108326be663", "ref_doc_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd"}, "86eac87b-b047-43cf-8a9d-e520bd0bf31d": {"doc_hash": "4fced396588f621fccc8cfcd794eeb4a6de05086890061e25ca1653793c73c56", "ref_doc_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd"}, "490b3459-7a05-444a-b656-c6d80893a882": {"doc_hash": "8711b9214dabab460c0c1db37828ef546489b3e37782eb8fce220a74cd5e5f81", "ref_doc_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd"}, "6960da48-4380-4f2b-a8da-62719c89a6df": {"doc_hash": "a0e76f2b888e24855cd06dd7b08cfa2c499637b1331c61e5dfaad481bc9472e9", "ref_doc_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd"}, "c311c6c7-0b31-4015-a322-5a6d52e7b831": {"doc_hash": "bc6c2e7b082588db7c8384a6a7b4eea48521d1809219b69c4422420906b566c6", "ref_doc_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd"}, "bad43489-2169-4a7f-86da-500c248e1461": {"doc_hash": "e4d727ce734f9ddfd9f0e20b4820ac060878d95d4915de9f724dbc88a3017055", "ref_doc_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd"}, "bb5c29e7-08eb-46d6-9ca6-ce32474bc798": {"doc_hash": "d3738f25a0cd8c4d20d3085de3e53869ef7e448baed6654c781e2885379d9496", "ref_doc_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd"}, "9d77aef3-85ae-4d47-8bb1-fbdc37abacd6": {"doc_hash": "9e6f66859ddfec72e70fb13359e1f388130a6c023a5c82cf503736651dc1f7c7", "ref_doc_id": "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd"}, "614cffc6-535e-4af4-8944-86a4a3751472": {"doc_hash": "8a443340e87718cb142ec1aa7a8da55f95242662bb1cf660d0b6f3da84dd5e02", "ref_doc_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e"}, "e29a95b6-72a8-4718-96e4-b76dc31dbb2e": {"doc_hash": "cb2a3128ee8cd9093dd904f00de96c9f1420b1eef8ae224a6d32107045e1dfa6", "ref_doc_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e"}, "b3c684f3-9fb4-499b-b85d-f8c168179b98": {"doc_hash": "7b4cd38ff6e38b2dabbd1edd476a34157af122470c247af2dd9d4affce2bef75", "ref_doc_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e"}, "7bec4fde-0642-4600-ac12-4829fe5048ae": {"doc_hash": "c044cdc18704ffff9d448325add32008d4dcd520fa42070a35cdd83e64df20d5", "ref_doc_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e"}, "afce34e9-ad90-40d0-b91f-22b98b3e629e": {"doc_hash": "73b3e9c645e02dc8dfc1bc6e3be0cdfceaadbc390d6a4535f9a1a66dab3f750c", "ref_doc_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e"}, "93a25878-f8c9-4040-a40c-c4adb172d3cc": {"doc_hash": "26f318ce06dcf322c1b7c7ba4e474ada94075c50979ee14b6336fb46c88464d1", "ref_doc_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e"}, "c7a5188e-cdff-430d-aa72-592fba5a4eb2": {"doc_hash": "39b99350971078f7fc3d0418b95292c0a299be0c6cc4df484ac63d18ad9c48db", "ref_doc_id": "f069cd89-6d10-41d9-bee8-9513d9bf6d4e"}, "9f775881-1987-4c1e-8a89-b2d53bf06cb3": {"doc_hash": "0f66dd1fa94e42224747a720646f13aa6361d47db9f333ac9611fef2be76bbc1", "ref_doc_id": "7cfca8ec-9d55-4349-9796-102e53179a86"}, "fa150e9a-d081-4a51-97d0-be00e5f33d26": {"doc_hash": "7da9e349a032cba09bcb86f8bf8a29ffa22831d0745372e1fb2ffa7716163ea3", "ref_doc_id": "7cfca8ec-9d55-4349-9796-102e53179a86"}, "8ce05fff-a173-46d6-955c-0349cc70ba53": {"doc_hash": "abdba424cd7be782708c822c2e7f151604871a6898738214cb86b325d9430ff3", "ref_doc_id": "6d6d77db-69ad-441f-8a04-34ee3cdef806"}, "fc9c3269-3a82-478c-bdbd-a0f5999ccce6": {"doc_hash": "c60f4996f8dcb81dd4154fa77aa567c50325ae00304d6361c33aee7aef1b34a6", "ref_doc_id": "6d6d77db-69ad-441f-8a04-34ee3cdef806"}, "81109a7f-a6b1-4ef5-bec5-cf4dfa1e26f7": {"doc_hash": "638e638c4f36246e4e4551570c51e672d237533837ff9d3c7328588ef0741f8d", "ref_doc_id": "421d32ad-3674-426c-88dd-7d44873ee0b7"}, "16a597b5-9745-49c7-925a-c37a02d1848f": {"doc_hash": "ffe3659172531da0d7313fc185d6509a6a6b811247ac5d61833440c86e1d0b99", "ref_doc_id": "421d32ad-3674-426c-88dd-7d44873ee0b7"}}, "docstore/ref_doc_info": {"6d6fd0f4-5a6d-47a0-b304-59417f3cadb6": {"node_ids": ["3431a00b-b2c3-4aa3-9597-d5c79a81f9c5", "135d7103-a9ad-43cd-b4e3-619fe02e901f", "7fa60ea2-8f3d-403a-91ab-ca2409c0a30f"], "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "21e5a458-23ba-4bdb-a3b2-3f352fd07ae5": {"node_ids": ["3ea43674-021d-4a53-bf50-8b7ba146a699", "40ceaf90-165a-4ba9-91cb-e441a1b06a33", "4e00cc5e-34a8-4d28-96cd-521a8613082e", "72dfb8e6-9f71-4211-a72d-5ec77bd4b2fe"], "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "10071188-42e8-4929-98dd-642ee4131169": {"node_ids": ["bb6aa732-aa8c-440f-86bb-de07bd602ea1", "536fb5fb-63e5-4170-80b6-739234a9149f", "fce102f2-322f-4e80-8130-96ff394dd00d"], "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "de79a486-751a-414d-b21e-078ab065ed82": {"node_ids": ["637186e0-6b70-47af-9395-a5f23c9922c7", "422f9866-8f04-463f-8d50-eab9a4886128", "8687e9ed-235b-4278-a496-45395cddaf37"], "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "1006cf3f-32fc-4c2e-b31a-f519be43fd1b": {"node_ids": ["6524bd60-c48e-4539-b3e1-78075b31bac0", "412d7574-4fb3-4985-8745-7494188dd8e6", "ff341210-7431-4b1b-a4d2-0b144b35d985", "abdc2077-fd4a-4184-bd93-07b2088061de"], "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "e95894c0-aec1-4614-b98f-d141b63efd5c": {"node_ids": ["67d23967-8054-4c86-9a3f-213604c7f8d6", "67607d1c-0394-46f4-8a3b-6bb6cff1bfe1", "84010686-fc60-4ee0-b307-f08a0e19cfe6", "bbfd17cd-9afc-4fae-9279-52965ee289d1"], "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "ab05337f-6fcc-408d-8685-7654ec1f476d": {"node_ids": ["0c7f8f31-d6ab-4f30-803d-7a0056b15963", "07227a15-44fc-4950-9f7d-7c1461364365", "6c2e74fb-f80a-4074-97a1-2f78727f1fc7", "612fbc2e-e4ab-4d6e-b3d6-af6dc250aaa8"], "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "8df67ec6-1e0f-4188-9687-0323556948cb": {"node_ids": ["826d10fd-380f-47e8-b970-51cd5c3e9270", "b3bfb5b0-4f15-45b3-970b-ea71328a74cf", "42d035cb-f399-442d-b7da-260217dd8a0c"], "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "8e0c2ca4-0215-412a-ba2e-f675e6a861af": {"node_ids": ["aef4eca4-aecd-4f06-99ee-3c35d2427597", "9c91e9f4-91a7-403b-9a4e-6f517810bcad", "739d85c4-e25b-49ae-a049-96c25851a609"], "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "ffbe95b2-20f2-4ab0-ac28-8c3458d946ce": {"node_ids": ["d463eec0-8dbb-4558-b9e9-5b3ff0ed9137", "1f60dd32-a23b-4678-880c-cee36938e787", "680e94b6-62fb-4536-a7ff-32dc41f99c89", "3a6477bc-b068-4c75-90c1-c406affa9259", "51d4b860-054a-4489-b47a-281b1b4f11ad"], "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "09c38cd0-0cea-4d3c-8eb7-23fd063e80a3": {"node_ids": ["27a20239-3af4-4aac-841c-c53dbde13a14", "d5d45e29-a0d7-4bb4-9cc7-c8d088f4c963", "ad73e473-afa4-41aa-a795-c7a9b1082fa9", "de4d13e0-dfe7-4887-a0f6-e087cbcda550", "6e5ad201-a5fe-4090-bfbb-17ef7fcb9bb8", "b4bdc4ea-ca2a-496d-b377-238c8447c25b", "6533451f-b299-4d94-ab0e-f50c4bb006cf", "cdf35e7c-fe91-4b76-83b7-a59e88b47277"], "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "edd3667f-1e40-42fa-ad91-ef19863b7aac": {"node_ids": ["91571f0f-6bcf-4172-870b-8a9ff8c2f42e", "fdb3ba2e-2595-4e34-b4ff-55ddb6157c27", "d36501b8-45cd-4893-8869-e0deadd4d8f7", "df80c204-cf8e-47e7-a08e-ef0f828244ac", "0e66ad56-e2bc-41d3-9693-771e1655bd78", "fe56f7b0-b0aa-4bab-b28c-c8eb52d2dab1", "357c6c9e-81fe-40af-ac73-8af5a4a0797c"], "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "21c4166d-6661-46b2-955e-0e7a112b0a80": {"node_ids": ["f6bd0eec-8952-4d79-b677-d7589a0c9025", "79a1ef13-5aca-4612-8f62-bfcf74a5f248"], "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "d167b925-01c2-44fa-aad1-97716bc024bc": {"node_ids": ["84327c1d-50ac-4c2c-bca4-773e888100f1", "2f750689-4461-4756-8fde-01726308eaf3"], "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "4bc641df-b009-48a6-8ba9-9d001e5879b7": {"node_ids": ["16dc7d44-1a9e-4ed7-8959-6798f8529cac", "f8bb10ed-3dd9-4d4d-9c27-d66565142fbd"], "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "8c7447f6-6a91-435c-80ce-786595a1941b": {"node_ids": ["4ae62500-f4b7-4576-bfd1-dd7a206dd0b7", "03db1e24-2be5-417d-aca0-fd45ec65c9df", "ce5cdd06-2fdf-47b3-a43a-47e144522450", "5337277a-d5ed-4ff1-b72e-4e76d6d7b65e"], "metadata": {"page_label": "1", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "7c1efa7b-b862-45f1-8f56-75407df01893": {"node_ids": ["1dae63a4-7407-4fec-911b-2befb2937f12", "3dd4b47e-81ae-41a1-9efb-277f7ee3c064", "7181f2d0-c9a5-45a7-86cb-3ee42f1833c0", "0f5f3713-6d91-42ee-a1d0-6f6d439ea009", "e339daf1-51f9-4679-a79b-40564c5ccf0b"], "metadata": {"page_label": "2", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "495b8b80-518b-4889-9658-6901e0229bc0": {"node_ids": ["e7c5b44d-5ae0-4cc6-9d00-dee9cb2a8814", "285134f3-afad-4ca8-a9a3-d4d13b2438c1", "f4519fee-459f-4522-b07d-82747620a572", "ccc697df-34bf-4228-a9c7-5834f08cd529"], "metadata": {"page_label": "3", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "6ad9b48a-2a55-40b1-9808-07cd8e439c05": {"node_ids": ["2eadfedb-30e8-41e5-8a28-756700b37955", "6bc1dfa7-bb04-4d5f-ae1e-90f9723a74a7", "1c0272bc-94c8-41e3-8fe7-a93957197edd", "23cc386c-71cf-4b1a-bdcd-d2fd8be427cc"], "metadata": {"page_label": "4", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "501f9969-eca6-4e92-9080-b03e896be613": {"node_ids": ["1223e2d0-f3f1-45f1-9962-24e1d2a3139c", "92b507ee-ce43-4dd8-b017-7dcbdf070435", "02afc34b-32f4-4c24-b536-349dd39356e3", "65332716-e35c-496a-b0bc-e6a4ed0cc983"], "metadata": {"page_label": "5", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "defb123a-181a-439f-9b2a-c8baddeb016e": {"node_ids": ["8c2c7fe2-0b52-4c53-81be-65d29a7cb679", "15d23f05-58ab-4ae9-92c7-9ea57abb2ae3", "e01a3c34-8381-4a3d-aad8-4d3de10bd55f", "8040ef1b-6d72-49bd-a743-013c777a8bc0", "d0f23a06-1048-4e37-9e50-ca0035756ec6", "282e536f-fe88-4f7a-a179-754220179b24"], "metadata": {"page_label": "6", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "3d538daf-6e8d-4715-a7c7-c3256db4d633": {"node_ids": ["b5e4fffe-eede-4a33-ade6-23fcded34c28", "e2ca30f9-a9f6-409f-8018-03870a276b86", "8a29b80e-3241-4249-8c79-3058f85811ed", "2069f9a4-c773-4744-a410-67dfc05b1dc5", "612ad73d-a6a1-434e-a9c1-cb0f768657ed"], "metadata": {"page_label": "7", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "55f2872f-47ad-45b9-92b9-21548eaf276e": {"node_ids": ["1bced689-bd10-4d07-8faf-c77ead5b12ca", "afe7667f-e82d-42fb-8fde-0cb1954cf5d1", "41e711b7-ffd9-4a80-9d2f-3b24107c2763", "9afdc6f5-b962-4b05-9964-28346b31319d", "356b6eab-f375-4784-a9a9-b346e1d248f9"], "metadata": {"page_label": "8", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "50e3ac80-5506-4247-9e25-6a2f3c5f33db": {"node_ids": ["5fffd5af-e267-451e-bdef-ce3e373ee153", "6ef9eca0-2c8a-4e3e-85d4-d052726b1d8f", "124d2351-91ef-48a0-b5b0-eb5a7e3b0a78", "1b37c06a-7824-45fd-8f61-d101b5a9df19", "8b1fa614-496b-43af-993f-498d707efb50", "7124a746-c0f8-4291-88f5-da1376271eb5", "5979ee85-867e-4749-b395-408a98e32df4", "fcc49761-9190-43c7-b273-101c413fb492"], "metadata": {"page_label": "9", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "4a57759d-5ee3-4c11-8861-9bcad409f797": {"node_ids": ["97c2d55b-945c-4c01-9dc0-86bd520257d0", "6a2b5ec3-b6df-4951-9797-fac4f6881808"], "metadata": {"page_label": "10", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "597b8534-da28-40e6-af8b-49dc5f661cad": {"node_ids": ["38ac2daf-0a08-4e26-b28c-cc5f9a69f395"], "metadata": {"page_label": "11", "file_name": "1502.03167v3.pdf", "file_path": "/kaggle/input/sample-pepers/1502.03167v3.pdf", "file_type": "application/pdf", "file_size": 173548, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "7bcf88e4-6827-4be9-988b-55bb1b49d948": {"node_ids": ["b7c64a3c-4735-4672-815f-7cf7cd07c193", "d5f956f2-7566-4ce2-b90c-9b1972c8a919", "2741d814-901a-40c1-b959-552157129a9d"], "metadata": {"page_label": "1", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "0ec692ee-efde-4036-bc5b-51ba9d6eca6c": {"node_ids": ["2923d6a1-d583-496a-81ed-0ab01dd2aaaf", "bfa3ce6d-c588-4f78-90f0-f0245cc7e6ba", "300e8947-8fa3-417d-9a79-9ae7c242f4ef", "93964b05-4478-455c-87d2-d75c334e4c09"], "metadata": {"page_label": "2", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "c8a07910-0192-481a-bcf1-1a58d71a8f38": {"node_ids": ["d5bcec10-5d92-4d3b-93e1-07345863d62e", "dc04e657-c9d4-4ae4-9608-d506dcdba62d", "019fd1ce-c068-408b-8d2d-58d3a5d48b96"], "metadata": {"page_label": "3", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "cd926aa7-9c47-4517-abf3-097ddfe907f2": {"node_ids": ["c304a4ca-c2b7-4211-9bc8-2a5108b28e07", "bdf8130c-d544-420b-9005-33872a3f74c1", "ab54a858-bc18-4ee5-b02f-1f549c149532"], "metadata": {"page_label": "4", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "e1e2ea31-60e3-4801-805b-c4df3bc294dd": {"node_ids": ["b334377e-b057-43e0-b76f-17f19781fc1d", "b3784fbe-100b-49fc-b7f7-1d16b9296a1e", "897256a3-8c19-4b9b-8f51-204cff114902", "989eb0dd-6927-47e5-91e3-d6f522a4bd2a"], "metadata": {"page_label": "5", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "3d21e68d-f59b-4dff-9e5f-8a1cd924edb5": {"node_ids": ["38465ff0-f5e8-43e3-a581-ff76dc2cea52", "c1fac594-7f52-4e5f-abbd-a8b9a2385ea1", "abe19b59-2b09-4ccc-ad63-74dc10dd0989", "700226de-87f3-4ff6-8023-2e460ab3599d"], "metadata": {"page_label": "6", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "ef3c2142-fd1c-4ca6-bd8d-cb41f2c41bb7": {"node_ids": ["5d14e9e0-9083-4dae-aa0f-a3b3ad856ae2", "0716be40-0d7e-4a1a-9cf0-9f14437b82f9", "903896c3-e90f-4fb8-933c-3b9a419f220d", "0fa8f49d-ee77-4e8b-99f9-430ef1086b18"], "metadata": {"page_label": "7", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "0da32021-5f29-482c-ba55-b4082af76e37": {"node_ids": ["7ac597d6-9c0c-4946-b6da-2e6954dd3ff1", "333ede79-147a-4ddb-a536-4af80aa36813", "5d8d189d-424b-4a2a-b0ad-e28758dce1d4"], "metadata": {"page_label": "8", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "41e170cc-f118-4a60-bb0a-2ec27a0f6318": {"node_ids": ["caa5c999-7d0e-447b-a3ac-6e3180209f3d", "2cab4eb7-bc6c-4c75-8dcc-763aaaceb767", "7466eb16-7489-427f-9838-f8bdb2c721a7"], "metadata": {"page_label": "9", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "10a87b3a-d733-48e8-ae5c-72eacb885ea0": {"node_ids": ["3470fcf0-d739-471d-a9cb-46f1eaddc675", "9acba64d-c6f2-406b-a1c8-5da9691f5868", "cce9180a-7469-43e7-955c-62a11e492759", "50c18361-106f-4a2a-a0a5-731ef39b76f9", "322758d7-f621-4f04-abb6-3d73e08e273e"], "metadata": {"page_label": "10", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "ef983cb8-fbf1-4ed3-a5f8-fb2dc78246cd": {"node_ids": ["087633e1-63a2-4f8e-a266-c2e92468c58a", "86eac87b-b047-43cf-8a9d-e520bd0bf31d", "490b3459-7a05-444a-b656-c6d80893a882", "6960da48-4380-4f2b-a8da-62719c89a6df", "c311c6c7-0b31-4015-a322-5a6d52e7b831", "bad43489-2169-4a7f-86da-500c248e1461", "bb5c29e7-08eb-46d6-9ca6-ce32474bc798", "9d77aef3-85ae-4d47-8bb1-fbdc37abacd6"], "metadata": {"page_label": "11", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "f069cd89-6d10-41d9-bee8-9513d9bf6d4e": {"node_ids": ["614cffc6-535e-4af4-8944-86a4a3751472", "e29a95b6-72a8-4718-96e4-b76dc31dbb2e", "b3c684f3-9fb4-499b-b85d-f8c168179b98", "7bec4fde-0642-4600-ac12-4829fe5048ae", "afce34e9-ad90-40d0-b91f-22b98b3e629e", "93a25878-f8c9-4040-a40c-c4adb172d3cc", "c7a5188e-cdff-430d-aa72-592fba5a4eb2"], "metadata": {"page_label": "12", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "7cfca8ec-9d55-4349-9796-102e53179a86": {"node_ids": ["9f775881-1987-4c1e-8a89-b2d53bf06cb3", "fa150e9a-d081-4a51-97d0-be00e5f33d26"], "metadata": {"page_label": "13", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "6d6d77db-69ad-441f-8a04-34ee3cdef806": {"node_ids": ["8ce05fff-a173-46d6-955c-0349cc70ba53", "fc9c3269-3a82-478c-bdbd-a0f5999ccce6"], "metadata": {"page_label": "14", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}, "421d32ad-3674-426c-88dd-7d44873ee0b7": {"node_ids": ["81109a7f-a6b1-4ef5-bec5-cf4dfa1e26f7", "16a597b5-9745-49c7-925a-c37a02d1848f"], "metadata": {"page_label": "15", "file_name": "1706.03762.pdf", "file_path": "/kaggle/input/sample-pepers/1706.03762.pdf", "file_type": "application/pdf", "file_size": 2215244, "creation_date": "2024-03-15", "last_modified_date": "2024-03-15"}}}}
\ No newline at end of file