text
stringlengths 1
1k
⌀ | title
stringlengths 2
143
|
---|---|
8
Figure 3: Fine-tuning evaluation on domain-specific tasks (left) and prompting evaluation on
general tasks (right). General LLM is the general language model, Raw Text trains the general
model on the domain-specific raw corpora, and Read. Compre.
trains the general model on the
reading comprehension texts constructed based on the raw corpora. We report the average of task
scores within each domain/type, detailed results are listed in Appendix F. | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
3
(cid:32)
T − p(cid:88)
i=1
(cid:33)2
di
,
Ldur = | Translatotron3 |
—
—
—
—
—
—
—
—
—
65.9/64.4
75.6/67.3
85.4/79.0
26.8/27.9
17.3/21.9
—
79.7/85.4
93.8/93.6
87.3/78.6
57.9/51.1
42.2/41.2
69.6/68.3
Test
—
—
—
—
—
37.5/33.8
19.5/18.2
22.5/23.6
33.9/32.6
76.2/74.9
0.65/0.00c
83.8/74.3
65.8/86.5
88.9/87.1
77.3/72.0
80.2/76.7
99.0/100.0
—
—
—
—
49.9/ 53.6
75.5/73.6
39.5/35.9
31.8/31.8
39.6/33.28
49.0/45.4
40.2/32.9
97.4/96.7
—
75.7/95.9
82.1/93.1
83.0/81.7
76.5/85.0
19.5/23.1
92.8/92.1
94.2/93.6
33.3/98.8
96.6/97.4
74.1/66.3
85.9/78.9
28.0/31.5
70.6/83.5
49.3/49.6
57.9/50.7
—
—
—
—
—
—
Table 7: Development and test scores of T0++ and ID-PT+J1-Large for all datasets in their training
corpus (results attained only for datasets with publicly available development or test sets). The
development and test scores of ID-PT+J1-Large were attained after training on roughly 50% of the
T0++ training data.
18
Preprint.
Vectors from
Same dataset, same template
Same dataset, different templates
Different datasets
Mean
0.028
0.069
0.16 | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
sushi disappear one by one
prompt: Rotating around a vase holding a dozen roses
Figure 14. A comparison of a 1B (left) and 8B (right) parameter models on the same prompt and settings.
19
starting frame
video name
elephant
elephant
car-turn
car-turn
dog-agility
dog-agility
bmx-bumps
bmx-bumps
train
train
bus
bus
lucia
lucia
tennis
tennis
bear
bear
flamingo
flamingo
hike
hike
goat
goat
parkour
parkour
cows
cows
camel
camel
blackswan
blackswan
dog
dog
kite-surf
kite-surf
libby
libby
horsejump-high
horsejump-high
10
10
40
40
0
0
10
10
0
0
0
0
0
0
15
15
60
60
0
0
0
0
59
59
60
60
10
10
10
10
0
0
20
20
10
10
0
0
0
0 | VideoPoet |
Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. QASC: A dataset for
question answering via sentence composition. In AAAI, 2020. URL https://arxiv.org/abs/1910.11473.
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language
models are zero-shot reasoners. NeurIPS, 2022. URL https://arxiv.org/abs/2205.11916.
Matthew Lamm, Jennimaria Palomaki, Chris Alberti, Daniel Andor, Eunsol Choi, Livio Baldini Soares, and
Michael Collins. QED: A framework and dataset for explanations in question answering. TACL, 9:790–806,
2021. URL https://arxiv.org/abs/2009.06354.
Chelsea Lee. Welcome, singular "they". https://apastyle.apa.org/blog/singular-they, 2019. Accessed:
2022-11-18.
Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tun-
ing. EMNLP, 2021. doi: 10.18653/v1/2021.emnlp-main.243. URL https://aclanthology.org/2021.
emnlp-main.243. | Scaling Instruction-Finetuned Language Models |
2.1.1 Scaling Laws
Before the popularization of LLMs, the relation-
ship between training dataset size and the perfor-
mance of language models with Transformer ar-
chitecture (Vaswani et al., 2017) had already at-
tracted researchers’ attention. Kaplan et al. (2020)
study the empirical scaling laws for Transformer
language model performance on the cross-entropy
loss and find the model performance has a power-
law relationship with training dataset size when not
bottlenecked by model size and training computa-
tion budget. They also conclude that model perfor-
mance improves predictably as long as model size
and training dataset size are scaled up simultane-
ously, but encounters overfitting if either of them is
fixed while the other increases. Their proposed pre-
diction ratio of performance penalty shows that the
model size should increase faster than the training
dataset size. | DataManagementForLargeLanguageModels-ASurvey |
*Equal contribution and corresponding authors: {lj,akos}@explosion.ai
1https://spacy.io
2https://thinc.ai
1
Figure 1: Illustration of the FGREP architecture. Redrawn from Miikkulainen and Dyer (1991)
2 BACKGROUND | MULTI HASH EMBEDDINGS IN SPACY |
As schematically shown in Figure 5b, we first trained a prompt, p1, for prompt tuning the LM on the
question answering task, and then used it to sample candidate answers from the model. Then, we
trained another prompt, p2, for prompt tuning the LM again, this time on the task of producing an | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
Smith TW, Davern M, Freese J, Stephen LM (2019) General Social
Surveys, 1972–2018: Cumulative Codebook / Principal Investiga-
tor, Tom W. Smith; Co-Principal Investigators, Michael Davern,
Jeremy Freese and Stephen L. Morgan. -- Chicago: NORC, 2019.
3,758 pp., 28cm. -- (National Data Program for the Social Sci-
ences Series, no. 25)
Solnick RE, Chao G, Ross R, Kraft-Todd GT, Kocher KE (2020) Emer-
gency physicians and personal narratives improve the perceived
effectiveness of COVID-19 public health recommendations on
social media: a randomized experiment. Acad Emerg Med. https://
doi. org/ 10. 1111/ acem. 14188
Sommariva S, Vamos C, Mantzarlis A, Đào L, Tyson D (2018) Spread-
ing the (fake) news: exploring health messages on social media
and the implications for health professionals using a case study.
Am Jour Health Ed 49:246–255. https:// doi. org/ 10. 1080/ 19325
037. 2018. 14731 78 | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
preprint arXiv:2012.15723 (2020).
[50] Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. RealToxicityPrompts:
Evaluating Neural Toxic Degeneration in Language Models. In Findings of the Association for Computational Linguistics:
EMNLP 2020. 3356–3369.
[51] Zorik Gekhman, Jonathan Herzig, Roee Aharoni, Chen Elkind, and Idan Szpektor. 2023. Trueteacher: Learning factual
consistency evaluation with large language models. arXiv preprint arXiv:2305.11171 (2023).
[52] Gaël Gendron, Qiming Bao, Michael Witbrock, and Gillian Dobbie. 2023. Large Language Models Are Not Abstract
Reasoners. arXiv preprint arXiv:2305.19555 (2023). | ASurveyonEvaluationofLargeLanguageModels |
Orlando, USA, September 2022. Association for Machine
Translation in the Americas.
[Arora et al., 2023] Daman Arora, Anush Kini, Sayak Ray
Chowdhury, Nagarajan Natarajan, Gaurav Sinha, and
Amit Sharma. Gar-meets-rag paradigm for zero-shot infor-
mation retrieval. arXiv preprint arXiv:2310.20158, 2023.
[Asai et al., 2023] Akari Asai, Zeqiu Wu, Yizhong Wang,
Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning
to retrieve, generate, and critique through self-reflection.
arXiv preprint arXiv:2310.11511, 2023.
[BAAI, 2023] BAAI. Flagembedding. https://github.com/
FlagOpen/FlagEmbedding, 2023.
[Baek et al., 2023] Jinheon Baek, Soyeong Jeong, Minki
Kang, Jong C Park, and Sung Ju Hwang. Knowledge-
augmented language model verification. arXiv preprint
arXiv:2310.12836, 2023. | RAG forLargeLanguageModels-ASurvey |
In this use case, the hybrid team aims at monitoring animal wildlife in a given terri-
tory for an extended, multi-year period. A visual representation is given in Figure 4. The
AI has access to on-site sensors (e.g., camera traps, microphones from static arrays [32]
and aboard autonomous vehicles). The human in the team is an expert who can identify
essential biodiversity variables (EBVs) [33], such as species- or community-level abun-
dance, from the acquired sounds and images. The AI aims at autonomously reporting
biodiversity status in a manner compatible with the language of human experts. Thus, it
interacts with the expert and learns how to map sensor signals to an ontology based on
EBVs. In this process, the AI uses self-supervised methods [34] to learn data structure,
and graphically communicates to the expert what it has learned. The expert experiences
this interaction as AI-aided data exploration, and provides knowledge to the AI both ex- | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
3.5 Authorship, Copyright and Ethical Issues | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
Weight Type
Rank r
WikiSQL (±0.5%)
MultiNLI (±0.1%)
# of Trainable Parameters = 18M
Wq Wk Wv Wo Wq, Wk Wq, Wv Wq, Wk, Wv, Wo
8
8
8
8
4
70.4
91.0
70.0
90.8
73.0
91.0
73.2
91.3
71.4
91.3
4
73.7
91.3
2
73.7
91.7
Table 5: Validation accuracy on WikiSQL and MultiNLI after applying LoRA to different types of
attention weights in GPT-3, given the same number of trainable parameters. Adapting both Wq and
Wv gives the best performance overall. We find the standard deviation across random seeds to be
consistent for a given dataset, which we report in the first column.
Note that putting all the parameters in ∆Wq or ∆Wk results in significantly lower performance,
while adapting both Wq and Wv yields the best result. This suggests that even a rank of four
captures enough information in ∆W such that it is preferable to adapt more weight matrices than
adapting a single type of weights with a larger rank. | LORA |
Such behavior may contribute to the overall visibility of hate speech on
mainstream online platforms. For example, on Twitter, although tweets
containing hate speech have lower numbers of replies and likes than non-
hateful tweets, they contain a similar number of retweets (Klubicka and
Fernandez 2018). The highly networked structure of hateful Twitter users
also dovetails with qualitative evidence suggesting that people are mobilized
on explicitly hateful subreddits or communities like the /pol/ board on 4chan to
engage in coordinated racist or sexist attacks on Twitter (Daniels 2017).
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
64
Alexandra A. Siegel | Social_Media_and_Democracy |
asymmetric polarization | Social_Media_and_Democracy |
Martin Fajcik, Martin Docekal, Karel Ondrej, and Pavel Smrz. R2-D2: A modular baseline for
open-domain question answering. In Findings of the Association for Computational Linguistics:
EMNLP 2021, pp. 854–870, Punta Cana, Dominican Republic, November 2021. Association
for Computational Linguistics. doi: 10.18653/v1/2021.findings-emnlp.73. URL https://
aclanthology.org/2021.findings-emnlp.73.
Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe,
Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for
nlp. In International Conference on Machine Learning, pp. 2790–2799. PMLR, 2019.
Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang,
and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint
arXiv:2106.09685, 2021.
Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
11
AllNon MoEMoEAttentionFFNParameters Being Updated81828384858687SuperGLUE Scoreparameters that worked well for the dense model can mask any pre-training improvements obtained
by the sparse model.
Figure 6: Batch size and learning rate sensitivity. We measure differences and sensitivity to fine-
tuning protocols between dense (blue) and sparse (orange) models. Each bar is an average across 6
different runs with different hyperparameters. On SuperGLUE, sparse models benefit from noisier
hyperparameters including small batch sizes and high learning rates. Dense models behave nearly
oppositely. See Appendix F for all data.
4.4 SPARSE MODELS ARE ROBUST TO DROPPED TOKENS DURING FINE-TUNING | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
log P (NA, NB|EA, EB) = −NA log(cid:0)1 + erB−rA(cid:1) − NB log(cid:0)1 + erA−rB(cid:1)
(B.1)
where rA,B = (log 10/400)EA,B ≈ EA,B/174. For an ensemble of comparisons between various models,
we estimate Elo scores and their errors by maximum likelihood estimation.
In some cases one of the models uses rejection sampling, meaning that it generates k samples, evaluates all
of them using a preference model, and shows the user the top-scored sample. Elo scores for such models are
shown in Appendix B.6. In this case, we cannot stream the sample, so instead we make the workers wait until
the sample is completed. When testing a rejection sampling model against a non-rejection sampling one, we
only show the samples when they’ve both been completed, even if the latter sample could’ve been streamed,
to mitigate bias.
46 | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
supports compositional query language. EmQL im-
plements “projection” using neural retrieval over
vectorized KB triples. Unlike this work, however,
EmQL did not embed its fact memory into a LM,
which could be finetuned for many NLP tasks: in-
stead requiring the implementation of a “neural
module” into some task-specific architecture. At a
more abstract level, the fact memory is a key-value
memory (Weston et al., 2014; Miller et al., 2016),
a construct used in many neural models in the past.
It has been shown that sufficiently large LMs
trained through self supervision (Peters et al., 2018;
Devlin et al., 2019; Raffel et al., 2019; Brown et al.,
2020) also encode factual information, motivating
work on the extent to which a LM can serve as
a KB (Roberts et al., 2020; Petroni et al., 2019;
Poerner et al., 2019). Other work has explored
techniques to improve the performance of large
LMs in answering factual probes, by adding ad-
ditional supervision in pre-training (Xiong et al., | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
need for speech super-resolution. arXiv preprint arXiv:2203.14941 (2022).
[334] Jinglin Liu, Chengxi Li, Yi Ren, Feiyang Chen, and Zhou Zhao. 2022. Diffsinger: Singing voice synthesis via shallow
diffusion mechanism. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 11020–11028.
[335] Jingjing Liu, Panupong Pasupat, Scott Cyphers, and Jim Glass. 2013. Asgard: A portable architecture for multilingual
dialogue systems. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 8386–8390.
[336] Mengzhuo Liu and Yangjie Wei. 2022. An Improvement to Conformer-Based Model for High-Accuracy Speech
Feature Extraction and Learning. Entropy 24, 7 (2022), 866.
[337] Rui Liu, Berrak Sisman, Guanglai Gao, and Haizhou Li. 2021. Expressive TTS Training With Frame and Style
IEEE/ACM Transactions on Audio, Speech, and Language Processing 29 (2021), 1806–1818.
Reconstruction Loss.
https://doi.org/10.1109/TASLP.2021.3076369 | AReviewofDeepLearningTechniquesforSpeechProcessing |
[20] Jonathan Le Roux, Scott Wisdom, Hakan Erdogan, and John R Hershey. Sdr–half-baked or
well done? In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pages 626–630. IEEE, 2019.
[21] Sang-gil Lee, Wei Ping, Boris Ginsburg, Bryan Catanzaro, and Sungroh Yoon. Bigvgan: A
universal neural vocoder with large-scale training. arXiv preprint arXiv:2206.04658, 2022.
[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017.
[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint
arXiv:1711.05101, 2017.
[24] Soroush Mehri, Kundan Kumar, Ishaan Gulrajani, Rithesh Kumar, Shubham Jain, Jose Sotelo,
Aaron Courville, and Yoshua Bengio. Samplernn: An unconditional end-to-end neural audio
generation model. arXiv preprint arXiv:1612.07837, 2016. | RVQGAN |
language models. CoRR, abs/2302.14045, 2023.
[287] Li, J., D. Li, S. Savarese, et al. BLIP-2: bootstrapping language-image pre-training with frozen
image encoders and large language models. In A. Krause, E. Brunskill, K. Cho, B. Engelhardt,
S. Sabato, J. Scarlett, eds., International Conference on Machine Learning, ICML 2023, 23-29
July 2023, Honolulu, Hawaii, USA, vol. 202 of Proceedings of Machine Learning Research,
pages 19730–19742. PMLR, 2023.
[288] Dai, W., J. Li, D. Li, et al. Instructblip: Towards general-purpose vision-language models with
instruction tuning. CoRR, abs/2305.06500, 2023.
64
[289] Gong, T., C. Lyu, S. Zhang, et al. Multimodal-gpt: A vision and language model for dialogue
with humans. CoRR, abs/2305.04790, 2023.
[290] Alayrac, J., J. Donahue, P. Luc, et al. Flamingo: a visual language model for few-shot learning.
In NeurIPS. 2022.
abs/2305.16355, 2023.
[291] Su, Y., T. Lan, H. Li, et al. Pandagpt: One model to instruction-follow them all. CoRR, | TheRiseandPotentialofLargeLanguageModel BasedAgents |
heself-reconstructionperformanceofLFAEusingoriginalandfinetunedΩ.AsTable5andFig.5show,bysimplyfinetuningthedecoderΩwithunlabelednewvideos,LFDMcanstillachievepromisingperformanceonnew-domainfacialvideos.Thisillustratestheflexibilityofourtwo-stagetrainingframework.Toimprovespatialcontentquality,onecanjustfinetunethedecoderinstage-oneLFAE,withouttheneedtoretrainthewholeframework.AblationStudy.Weconductablationstudyofdiffer-ModelL1error↓FID↓sFID↓OriginaldecoderΩ2.12938.4875.08±37.33FinetuneddecoderΩ1.31023.3651.56±26.36Table5.QuantitativecomparisonofLFDMwithoriginalvs.fine-tuneddecoderonFaceForensicsdataset.Stepsg-scaleFVD↓Time(s)↓DDIM-101.050.180.3DDIM-1001.0106.113.5DDPM-10001.131.8471.4DDPM-10005.049.8271.4DDPM-10001.032.0936.0Table6.AblationstudycomparingdifferentsamplingstrategiesforLFDMonMUGdataset(resolutionis128×128).entsamplingstrategiesinTable6.Foreachsetting,were-portFVDscoresandtheaveragesamplingtimetogenerateonevideowhenusingthebatchsizeof10ononeNVIDIAA100GPU.Wefirstcomparetheresult | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
Stack Exchange [2%]. We include a dump of
Stack Exchange, a website of high quality ques-
tions and answers that covers a diverse set of do-
mains, ranging from computer science to chemistry.
We kept the data from the 28 largest websites, re-
moved the HTML tags from text and sorted the
answers by score (from highest to lowest).
Tokenizer. We tokenize the data with the byte-
pair encoding (BPE) algorithm (Sennrich et al.,
2015), using the implementation from Sentence-
Piece (Kudo and Richardson, 2018). Notably, we
split all numbers into individual digits, and fallback
to bytes to decompose unknown UTF-8 characters.
params
6.7B
13.0B
32.5B
65.2B
dimension n heads n layers
learning rate
batch size n tokens
4096
5120
6656
8192
32
40
52
64
32
40
60
80
3.0e−4
3.0e−4
1.5e−4
1.5e−4
4M
4M
4M
4M
1.0T
1.0T
1.4T
1.4T
Table 2: Model sizes, architectures, and optimization hyper-parameters. | LLaMA- Open and Efficient Foundation Language Models |
Figure 3. Memorization results for the semantic modeling stage.
We compare the semantic tokens generated for 5 seconds of audio
to corresponding tokens in the training set, considering exact and
approximate matches. | MusicLM |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Social Media and Democracy
Over the last five years, widespread concern about the effects of social
media on democracy has led to an explosion in research from different
disciplines and corners of academia. This book is the first of its kind to
take stock of this emerging multi-disciplinary field by synthesizing what
we know, identifying what we do not know and obstacles to future
research, and charting a course for the future inquiry. Chapters by
leading scholars cover major topics – from disinformation to hate
speech to political advertising – and situate recent developments in the
context of key policy questions. In addition, the book canvasses existing
reform proposals in order to address widely perceived threats that social
media poses to democracy. This title is also available as Open Access on
Cambridge Core. | Social_Media_and_Democracy |
Instruction Complexity
3.2.3
The complexity of instructions also attracts re-
searchers’ attention, especially in developing LLM
with complex instruction-following and reasoning
2https://chatgpt.openai.com/
abilities (Xu et al., 2023; Luo et al., 2023; Mukher-
jee et al., 2023). Several works endeavor to quan-
tify and evaluate instruction complexity. Using
aforementioned tags, #InsTag (Lu et al., 2023)
quantifies complexity as the average tag number
assigned to queries in a dataset. He et al. (2023)
evaluate complex instruction with eight features,
i.e. multi-tasking, semantics, formats, quantity con-
straints for task description, and multi-turn, length,
noise, and heterogeneous information for input
text. | DataManagementForLargeLanguageModels-ASurvey |
to their volume. However, we did make an exception for JSON, YAML, and CSS, as we only want
the LLM to learn the data format without wasting compute resources on memorizing the data in such
files. For that reason, we re-weighed the volume of the data source to 1 GB for JSON and YAML and
3GB for CSS. | StarCoder_paper (1) |
𝑥𝑇 ∼ N(0, I) to 𝑥0 and parameterized by 𝜃:
(25)
where 𝑥𝑇 ∼ N(0, 𝐼) and the transition probability 𝑝𝜃 (𝑥𝑡−1|𝑥𝑡) is learnt through noise-estimation.
This process eliminates the Gaussian noise added in the forward diffusion process.
𝑝𝜃 (𝑥0, ..., 𝑥𝑇 −1|𝑥𝑇) =
𝑝𝜃 (𝑥𝑡−1|𝑥𝑡)
𝑡=1 | AReviewofDeepLearningTechniquesforSpeechProcessing |
(citing Hughey v. United States, 495 U.S. 411, 413, 110 S.Ct. 1979, 109 L.Ed.2d 408 (1990)). The
pre-sentence report attributed forty-seven fraudulent claims to the offenses for which Arledge was
convicted. There were three categories of evidence used to substantiate the government’s assertion
that these claims resulted from Arledge’s illegal conduct: (1) the testimony of Wyatt, an employee
of S&A, who created fraudulent documents; (...) | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
5.5.3 Models
End-to-end speech translation models are a promising approach to direct the speech translation
field. These models use a single sequence-to-sequence model for speech-to-text translation and then
text-to-speech translation. In 2017, researchers demonstrated that end-to-end models outperform
cascade models[3]. One study published in 2019 provides an overview of different end-to-end
architectures and the usage of an additional connectionist temporal classification (CTC) loss
for better convergence [27]. The study compares different end-to-end architectures for speech-
to-text translation. In 2019, Google introduced Translatotron [219], an end-to-end speech-to-
speech translation system. Translatotron uses a single sequence-to-sequence model for speech-
to-text translation and then text-to-speech translation. No transcripts or other intermediate text
representations are used during inference. The system was validated by measuring the BLEU score, | AReviewofDeepLearningTechniquesforSpeechProcessing |
o
f
M
e
m
o
r
y
M
e
m
o
r
y
c
a
n
b
e
d
e
f
i
n
e
d
a
s
t
h
e
p
r
o
c
e
s
s
e
s
u
s
e
d
t
o
a
c
q
u
i
r
e
,
s
t
o
r
e
,
r
e
t
a
i
n
,
a
n
d
l
a
t
e
r
r
e
t
r
i
e
v
e
i
n
f
o
r
m
a
t
i
o
n
.
T
h
e
r
e
a
r
e
s
e
v
e
r
a
l
t
y
p
e
s
o
f
m
e
m
o
r
y
i
n
h
u
m
a
n
b
r
a
i
n
s
.
| LLM Powered Autonomous Agents _ Lil'Log |
Landfilled (%) Mismanaged (%)
Incinerated (%)
Recycled (%)
22
4
6
34
19
19
38
19
9
4
12
8
Figure 8 | Solving a problem requiring multimodal chart understanding.The model has to read the
text, understand the connections between different data points and reason over them to recommend
an interesting point and follow the instructions to generate a markdown table (shown correctly
rendered).
Source: Our World In Data (Ritchie et al., 2023).
47
Gemini: A Family of Highly Capable Multimodal Models
9.3.2. Multimodal question answering
Prompt
Do you know what it this plant? How do I best take care of it?
Model Response
This is a Persian shield plant. It is a tropical plant that is native to Southeast Asia. It is a
popular houseplant because it is easy to care for and has beautiful purple leaves. | gemini_1_report |
References
[1] Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja
Fidler. Skip-thought vectors. In Advances in Neural Information Processing Systems, pages 3294–3302, 2015.
[2] Andrew M Dai and Quoc V Le. Semi-supervised sequence learning. In Advances in Neural Information
Processing Systems, 2015.
[3] Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettle-
moyer. Deep contextualized word representations. In NAACL, 2018.
[4] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification. In ACL, 2018.
[5] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by
generative pre-training. https://blog.openai.com/language-unsupervised, 2018.
[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional | LaMDA- Language Models for Dialog Applications |
Of course, to describe social media data as administrative data has the
potential to diminish its sensitive personal nature. Although some social
media data seem “administrative” (e.g., number of friends, popularity of
URLs, whether one has posted using a mobile or desktop device, time zone,
etc.), other data appear qualitatively different in that they do not contain
records of behavior so much as the personal thoughts and observations of
individuals. However, some forms of administrative data, such as those
involved with medical treatment and health statistics, are equally sensitive –
perhaps even more so – than social media data. Yet, we may be moving toward a
data access regime in which personal medical data may be more accessible to
researchers than what users share or read on Facebook. | Social_Media_and_Democracy |
We propose a two-stage searching strategy for retrieving
the most similar past driving scenario to the query scenario.
In the first stage, we generate a vectorized key ki ∈
R1×(ne+ng+nh) for each past scenario i by vectorizing its
ego-states ei ∈ R1×ne, mission goals gi ∈ R1×ng, and
historical trajectories hi ∈ R1×nh. The N past scenarios in
the experience memory collectively construct a key tensor
K ∈ RN×(ne+ng+nh):
K = {[ei, gi, hi]|i = {1, 2, ..., N}}.
(1)
Similarly, we can vectorize the query scenario into Q =
[e, g, h] ∈ R1×(ne+ng+nh).
Subsequently, we compute the similarity scores S ∈ RN
between the querying scenario Q and the past scenarios K:
S = QΛK⊤,
(2)
where Λ = diag(λe, λg, λh) ∈ R(ne+ng+nh)×(ne+ng+nh)
indicates the weights of different components. | ALanguageAgentforAutonomousDriving |
Language Model Pretraining with Weak Supervision. ArXiv abs/2108.10904 (2022).
[201] Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2019. Neural Text
Generation With Unlikelihood Training. In International Conference on Learning Representations.
[202] Sean Welleck, Jason Weston, Arthur Szlam, and Kyunghyun Cho. 2019. Dialogue Natural Language Inference. In
Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Association for Computational
Linguistics, Florence, Italy, 3731–3741. https://doi.org/10.18653/v1/P19-1363 | SurveyofHallucinationinNatural Language Generation |
and DPO-SDXL) perform well at binary preference classi-
fication (Tab. 2), with DPO-SDXL exceeding all existing
recognition models on this split. These results shows that
the implicit reward parameterization in the Diffusion-DPO
objective has comprable expressivity and generalization as
the classical reward modelling objective/architecture.
Training Data Quality Fig. 7 shows that despite SDXL
being superior to the training data (including the yw), as
measured by Pickscore, DPO training improves its perfor-
Figure 7. Diffusion-DPO improves on the baseline Dreamlike and
SDXL models, when finetuned on both in-distribution data (in case
of Dreamlike) and out-of-distribution data (in case of SDXL). yl
and yw denote the Pickscore of winning and losing samples. | DiffusionModelAlignmentUsing Direct Preference Optimization |
Prohibiting trivial retrievals
If the pre-training corpus
X and the knowledge corpus Z are the same, there exists
a trivial retrieval candidate z that is too informative: if the
masked sentence x comes from document z, the knowledge
augmented encoder can trivially predict y by looking at the
unmasked version of x in z. This results in a large positive
gradient for p(z | x). If this occurs too often, the knowledge
retriever ends up learning to look for exact string matches
between x and z, which does not capture other forms of
relevance. For this reason, we exclude this trivial candidate
during pre-training. | REALM |
With the input and the inference results, the AI assistant needs to describe
the process and results. The previous stages can be formed as - User Input:
{{ User Input }}, Task Planning: {{ Tasks }}, Model Selection: {{ Model
Assignment }}, Task Execution: {{ Predictions }}. You must first answer the
user's request in a straightforward manner. Then describe the task process
and show your analysis and model inference results to the user in the first
person. If inference results contain a file path, must tell the user the
complete file path.
https://lilianweng.github.io/posts/2023-06-23-agent/
12/22 | LLM Powered Autonomous Agents _ Lil'Log |
Formats For Large Language Models and Vision Transformers. arXiv preprint arXiv:2307.03712 (2023).
[191] Deepak Narayanan, Aaron Harlap, Amar Phanishayee, Vivek Seshadri, Nikhil R Devanur, Gregory R Ganger, Phillip B Gibbons, and Matei Zaharia. 2019.
PipeDream: Generalized pipeline parallelism for DNN training. In Proceedings of the 27th ACM Symposium on Operating Systems Principles. 1–15.
[192] Deepak Narayanan, Amar Phanishayee, Kaiyu Shi, Xie Chen, and Matei Zaharia. 2021. Memory-efficient pipeline-parallel dnn training. In International
Conference on Machine Learning. PMLR, 7937–7947.
[193] Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley, Mostofa Patwary, Vijay Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie
Bernauer, Bryan Catanzaro, et al. 2021. Efficient large-scale language model training on gpu clusters using megatron-lm. In Proceedings of the International
Conference for High Performance Computing, Networking, Storage and Analysis. 1–15. | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
5. Team Design Patterns
In this section, we describe the team design patterns (TDPs) that were identified based
on the design solutions described in the use cases.
Examining the four outlined use cases, we can extract three primary design patterns
from the domain-specific hybrid intelligence solutions, namely AI Advisor and Human
Performer TDP, AI Performer and Human Assistant TDP, and AI Performer and Human
Validator TDP. We extracted the TDPs by iterating on the derived TDPs on different
levels of abstraction. We targeted to extract TDPs at a level of abstraction for which we
could show that the TDPs are applicable for not only the use case it has been derived
from, but also for at least one of the other use cases.
In the following, we describe the derived TDPs in detail. | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
Ganguli, D., Lovitt, L., Kernion, J., Askell, A., Bai, Y., Kadavath, S., Mann, B., Perez, E., Schiefer, N., Ndousse,
K., Jones, A., Bowman, S., Chen, A., Conerly, T., DasSarma, N., Drain, D., Elhage, N., El-Showk, S., Fort, S.,
Hatfield-Dodds, Z., Henighan, T., Hernandez, D., Hume, T., Jacobson, J., Johnston, S., Kravec, S., Olsson, C.,
Ringer, S., Tran-Johnson, E., Amodei, D., Brown, T., Joseph, N., McCandlish, S., Olah, C., Kaplan, J., and Clark,
J. Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned, 2022. URL
https://arxiv.org/abs/2209.07858.
Garg, N., Schiebinger, L., Jurafsky, D., and Zou, J. Word embeddings quantify 100 years of gender and ethnic stereotypes.
Proceedings of the National Academy of Sciences, 115(16):E3635–E3644, 2018. doi: 10.1073/pnas.1720347115.
URL https://www.pnas.org/doi/abs/10.1073/pnas.1720347115.
Garg, T., Masud, S., Suresh, T., and Chakraborty, T. Handling bias in toxic speech detection: A survey. January 2022. | PaLM 2 Technical Report |
still refer to them as “<API>”, “</API>” and “→” through-
out this section. | Toolformer |
2. Closed-Book Response Generation: If provided with a fact-seeking prompt without any given
source, Gemini should not hallucinate incorrect information (see Section 2 of Roberts et al.
(2020) for a definition). These prompts can range from information-seeking prompts (e.g. “Who
is the prime minister of India?”) to semi-creative prompts that may request factual information
(e.g. “Write a 500-word speech in favor of the adoption of renewable energy”).
21
Gemini: A Family of Highly Capable Multimodal Models
3. Hedging: If prompted with an input such that it is “unanswerable”, Gemini should not hal-
lucinate. Rather, it should acknowledge that it cannot provide a response by hedging. These
include scenarios where the input prompt contains false-premise questions (see examples in Hu
et al. (2023)), the input prompt instructs the model to perform open-book QA, but the answer
is not derivable from the given context, and so forth. | gemini_1_report |
1
2
0
2
r
p
A
2
1
]
L
C
.
s
c
[
4
v
1
0
4
1
1
.
5
0
0
2
:
v
i
X
r
a
Retrieval-Augmented Generation for
Knowledge-Intensive NLP Tasks
Patrick Lewis†‡, Ethan Perez(cid:63),
Aleksandra Piktus†, Fabio Petroni†, Vladimir Karpukhin†, Naman Goyal†, Heinrich Küttler†,
Mike Lewis†, Wen-tau Yih†, Tim Rocktäschel†‡, Sebastian Riedel†‡, Douwe Kiela†
†Facebook AI Research; ‡University College London; (cid:63)New York University;
[email protected]
Abstract | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
Different from compositing multiple tasks to-
gether, some works claim that LLM tuned on a
single task data can outperform LLM tuned on mul-
tiple tasks (Jang et al., 2023; Chen et al., 2023b).
Jang et al. (2023) state that the priority of training
expert LLMs may lie in the avoidance of nega-
tive task transfer, prevention of catastrophic for-
getting via continually learning new tasks without
re-training, and compositional abilities emergent
when merging individual experts together. Wang
et al. (2023b) conduct analysis on factual knowl-
edge, reasoning, multilinguality, coding, and open-
ended instruction following abilities of models
trained with 12 instruction datasets and show that
different instruction datasets can unlock or improve
specific abilities. In contrast, no single dataset of
combinations can provide the best performance
across all evaluations. | DataManagementForLargeLanguageModels-ASurvey |
We have evaluated the validity of SHAPE across studies. In Survey #1, we could show that threat and
competence relate to the Stereotype-content model; people that attribute low threat to augmented humans
perceived them as warmer, while competence of augmented humans was increased for low social threat and
more control. This aligns with the findings of Meyer and Asbrock [56], who discovered that individuals with
bionic prostheses were perceived as competent without a reduction in perceived warmth.
On the other hand, in Survey #3, we demonstrated construct validity. There is convergent validity in terms of
correlation with the technology readiness index that addresses discomfort and insecurity about technological
developments but discriminant validity in terms of innovativeness. Therefore, the scale covers both stereotypes’
Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 7, No. 3, Article 128. Publication date: September 2023.
128:16
• Villa et al. | Society’sAttitudesTowardsHumanAugmentation |
u
n
i
t
n
o
r
m
e
a
c
h
i
t
e
r
a
t
i
o
n
.
W
e
f
i
n
d
t
h
a
t
t
h
e
a
v
e
r
a
g
e
t
o
p
-
a
n
d
-
r
a
n
d
o
m
s
c
o
r
e
a
f
t
e
r
1
0
i
t
e
r
a
t
i
o
n
s
i
s
0
.
7
1
8
,
s
u
b
s
t
a
n
t
i
a
l
l
y
h
i
g
h
e
r
t
h
a
n
t
h
e
a
v
e
r
a
g
e
s
c
o
r
e
f
o
r
r
a
n
d
o
m
n
e
u
r
o
n
s
i
n
t
h
i
s
l
a
y
e
r
(
0
.
1
4
7
)
,
a
n
d
h
i
g
h
e
r
t
h
a
n
t
h
e
a
v
e
r
a
g
e
s
c
o
r
e
f
o
r
r
a
n
d
o
m
d
i
r
e
c
t
i
o
n
s
b
e
f
o
r
e
a
n
y
o
p
t
i
m
i
z
a
t
i
o
n
(
0
.
0
6
1
)
.
O
n
e
p
o
t
e
n
t
i
a
l
p
r
o
b
l
e
m
w
i
t
h
t
h
i
s
p
r
o
c
e
d
u
r
e
i
s
t
h
a
t
w
e
c
o
u
l
d
r
e
p
e
a
t
e
d
l
y
c
o
n
v
e
r
g
e
u
p
o
n
t
h
e
s
a
m
e
e
x
p
l
a
i
n
a
b
l
e
d
i
r
e
c
t
i
o
n
,
r
a
t
h
e
r
t
h
a
n
f
i
n
d
i
n
g
a
d
i
v
e
r
s
e
s
e
t
o
f
l
o
c
a
l
m
a
x
i
m
a
.
T
o
c
h
e
c
k
t
h
e
e
x
t
e
n
t
t
o
w
h
i
c
h
t
h
i
s
i
s
h
a
p
p
e
n
i
n
g
,
w
e
m
e
a
s
u
r
e | Language models can explain neurons in language models |
Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra
Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins, Tom
Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks, William S.
Isaac, Sean Legassick, Geoffrey Irving, and Iason Gabriel. Ethical and social risks of harm from
language models. CoRR, abs/2112.04359, 2021. URL https://arxiv.org/abs/2112.04359.
David Wetherall, Abdul Kabbani, Van Jacobson, Jim Winget, Yuchung Cheng, Brad Morrey,
Uma Parthavi Moravapalle, Phillipa Gill, Steven Knight, and Amin Vahdat. Improving network
32
Gemini: A Family of Highly Capable Multimodal Models
availability with protective reroute. In SIGCOMM 2023, 2023. URL https://dl.acm.org/doi/
10.1145/3603269.3604867.
Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. NExT-QA: Next phase of question-answering
to explaining temporal actions. In CVPR, 2021. | gemini_1_report |
(OOD) datasets. An overview of the OOD evaluation datasets is presented in Table 4. Full details
of the evaluation datasets are provided in Appendix A.2.
We examine both overall robustness, that is the average performance over all datasets, and effective
robustness (Taori et al., 2020), which measures the difference in expected performance between a
reference dataset that is ID, and one or more datasets that are OOD. A model with high effective
robustness does better on OOD datasets as a function of performance on the reference dataset. A
model with ideal effective robustness performs equally well on all datasets. In our experiments,
we use GigaSpeech (Chen et al., 2021) as the reference dataset, owing to the fact it contains web-
scraped data from audiobooks, podcasts and YouTube videos, and is such ID with both the pre-
trained Whisper training data and the distilled Whisper train set.
We evaluate the noise robustness of the Distil-Whisper models, the original Whisper models, and | DISTIL-WHISPER |
ReCoRD commonsense reasoning dataset, and the RACE datasets for reading comprehension.
We measure potential bias in QA performance on questions related to identity terms, together with bias in other
generative tasks, in Section 4.6. We find that PaLM 2 performs well on disambiguated questions about social identity
and do not observe a systematic pattern of bias, with full results in Appendix E.6.
Multilingual QA To demonstrate PaLM 2’s multilingual capabilities, we evaluate on the multilingual QA dataset
TyDi QA (Clark et al., 2020) in a one-shot setting.3 We additionally propose a more challenging no-context setting
where the model has to answer the question solely based on the knowledge stored in its parameters.4 We show the
results in Table 3.
3We use the Gold Passage formulation where a model is provided with the question and the Gold Passage.
4This is akin to the closed-book evaluation setting for Natural Questions (Roberts et al., 2020).
12 | PaLM 2 Technical Report |
alinforma-tionprocessingsystems,32,2019.1,2,3[37]DavisEKing.Dlib-ml:Amachinelearningtoolkit.TheJournalofMachineLearningResearch,10:1755–1758,2009.8[38]DiederikPKingmaandJimmyBa.Adam:Amethodforstochasticoptimization.arXivpreprintarXiv:1412.6980,2014.6[39]DiederikPKingmaandMaxWelling.Auto-encodingvaria-tionalbayes.arXivpreprintarXiv:1312.6114,2013.3[40]ZhifengKongandWeiPing.Onfastsamplingofdiffu-sionprobabilisticmodels.arXivpreprintarXiv:2106.00132,2021.8[41]YijunLi,ChenFang,JimeiYang,ZhaowenWang,XinLu,andMing-HsuanYang.Flow-groundedspatial-temporalvideopredictionfromstillimages.InProceedingsoftheEu-ropeanConferenceonComputerVision(ECCV),pages600–615,2018.2,3[42]YitongLi,MartinMin,DinghanShen,DavidCarlson,andLawrenceCarin.Videogenerationfromtext.InProceedingsoftheAAAIconferenceonartificialintelligence,volume32,2018.2[43]ZhihengLi,MartinRenqiangMin,KaiLi,andChenliangXu.Stylet2i:Towardcompositionalandhigh-fidelitytext-to-imagesynthesis.InProceedingsoftheIEEE/CVFCon-ferenceonComputerVisionand | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
P. Sermanet, C. Lynch, Y. Chebotar, J. Hsu, E. Jang, S. Schaal, S. Levine, and G. Brain.
Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE interna-
tional conference on robotics and automation (ICRA), pages 1134–1141. IEEE, 2018. 8,
38
S. Shen, L. H. Li, H. Tan, M. Bansal, A. Rohrbach, K.-W. Chang, Z. Yao, and K. Keutzer.
How Much Can CLIP Benefit Vision-and-Language Tasks? In International Conference
on Learning Representations, Jan. 2022. URL https://openreview.net/forum?id=
zf_Ll3HZWgy. 42
H. Shi, D. Luo, S. Tang, J. Wang, and Y. Zhuang. Run away from your teacher: Under-
standing byol by a novel self-supervised approach. arXiv preprint arXiv:2011.10944,
2020. 26, 28
R. Shwartz-Ziv, M. Goldblum, H. Souri, S. Kapoor, C. Zhu, Y. LeCun, and A. G. Wilson.
Pre-train your loss: Easy bayesian transfer learning with informative priors. In Advances
in Neural Information Processing Systems. 42 | A Cookbook of Self-Supervised Learning |
Chen. Large language models as optimizers. arXiv preprint arXiv:2309.03409, 2023.
Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov,
and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question
answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language
Processing. Association for Computational Linguistics, 2018.
Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik
Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv
preprint arXiv:2305.10601, 2023.
Shen Zheng, Jie Huang, and Kevin Chen-Chuan Chang. Why does chatgpt fall short in providing
truthful answers. ArXiv preprint, abs/2304.10513, 2023. | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
FR
0.21 (p=0.00)
0.12 (p=0.08)
0.10 (p=0.17)
0.10 (p=0.15)
0.13 (p=0.00)
Table 5: Correlation between groups of annotators (MN,
FN, MNN, FNN) and models’ predictions, classified by
language. The degree of correlation is measured with
Kendall’s τ coefficient (τ ∈ [−1, 1]). Cells are coloured
language-wise. Cells with a darker background show
a stronger correlation compared to the average in each
language. Samples highlighted in red fail to reject the
null hypothesis, meaning that their difference is not
statistically significant (p > 0.05).
Figure 3: t-SNE representation from the last layer of
XLM-R for the top-1000 predictions for the parallel
sentences in the list above (“We want to [MASK] in-
novation .” in English). Highest scored prediction is
starred; annotator’s answers are denoted by a dot with
black edge. Legend shows language-color mapping. | Are Pretrained Multilingual Models Equally Fair Across Languages? |
REFERENCES
[1] Kevin Ackermans, Ellen Rusman, Rob Nadolski, Marcus Specht, and Saskia Brand-Gruwel. 2019. Video-or text-based
rubrics: What is most effective for mental model growth of complex skills within formative assessment in secondary
schools? Computers in Human Behavior 101 (Dec. 2019), 248–258. https://doi.org/10.1016/j.chb.2019.07.011
[2] Lauren Y Atlas. 2021. A social affective neuroscience lens on placebo analgesia. Trends in Cognitive Sciences 25, 11
(Nov. 2021), 992–1005. https://doi.org/10.1016/j.tics.2021.07.016
[3] Ebrahim Babaei, Benjamin Tag, Tilman Dingler, and Eduardo Velloso. 2021. A Critique of Electrodermal Activity
Practices at CHI. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama, Japan)
(Chi ’21). Association for Computing Machinery, New York, NY, USA, Article 177, 14 pages. https://doi.org/10.1145/
3411764.3445370 | AI enhance sour performance |
35.4
34.0
30.1
35.8
33.3
28.2
85.7
85.7
84.0
Scifact Avg
51.6
73.7
50.2
71.8
69.1
45.8
Table 5, increasing batch size from 1K to 32K leads to consistent gains across all 6 datasets. It is also
possible to train with smaller batch sizes by adding hard negatives [50]. However, the engineering
efforts of mining hard negatives for large datasets (>100M) are non-trivial.
Table 6: Fine-tuning with different combinations of labeled data.
Fine-tuned on
No fine-tuning
MS-MARCO + NQ
NLI
All above
Retrieval
42.9
50.3
38.3
48.7
STS Classification
69.5
78.3
81.1
81.0
67.9
68.3
72.6
73.1
Summ. MTEB Avg
24.3
25.8
26.5
26.0
55.5
58.9
57.2
60.3 | E5 |
A simple improvement on top of the Mahalanobis distance called the Relative Mahalanobis distance has been
proposed in [Ren et al., 2021] and shown to lead to better AUROC as well as more robust detection for a range
of OOD problems in vision and genomics (in addition to more robustness to adversarial attacks [Fort, 2022]).
Inspired by this method and recognizing that our problem does not naturally involve semantically meaningful
classes comprising the in-distribution, we propose a further modification we call the Simplified Relative
Mahalanobis distance. We compute it by fitting a full covariance matrix Σ as before, as well as a diagonal-
only covariance matrix Σdiag, and assigning the difference of their Mahalanobis distance as our scoring
function, score(x) = (x − µ)T Σ−1(x − µ) − (x − µ)T Σdiag
Figure 22 shows the results for our OOD detection experiments, trying to distinguish the helpfulness data | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Text conditioning. Given a textual description matching the input audio X, we compute a condi-
tioning tensor C ∈ RTC×D with D being the inner dimension used in the autoregressive model.
Generally, there are three main approaches for representing text for conditional audio generation.
Kreuk et al. [2022] proposed using a pretrained text encoder, specifically T5 [Raffel et al., 2020].
Chung et al. [2022] show that using instruct-based language models provide superior performance.
Lastly, Agostinelli et al. [2023], Liu et al. [2023], Huang et al. [2023a], Sheffer and Adi [2023]
claimed that joint text-audio representation, such as CLAP [Wu* et al., 2023], provides better-quality
generations. We experiment with all of the above, respectively: T5 encoder, FLAN-T5, and CLAP.
Melody conditioning. While text is the prominent approach in conditional generative models
nowadays, a more natural approach for music is conditioning on a melodic structure from another | Simple and Controllable Music Generation |
403020100-10signal-to-noise ratio (dB)125102050100WER on LibriSpeech test-clean (%)white noise403020100-10signal-to-noise ratio (dB)pub noiseunispeech-sat-base-100h-libri-ftwav2vec2-base-100hwav2vec2-base-960hwav2vec2-large-960hwav2vec2-large-robust-ft-libri-960hwav2vec2-large-960h-lv60-selfasr-crdnn-rnnlm-librispeechasr-transformer-transformerlm-librispeechhubert-large-ls960-fthubert-xlarge-ls960-fts2t-medium-librispeech-asrs2t-large-librispeech-asrstt_en_conformer_ctc_largestt_en_conformer_transducer_xlargeWhisperRobust Speech Recognition via Large-Scale Weak Supervision
3.7. Robustness to Additive Noise | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
underway, and that any analysis focused on any particular part (as important
as it might be) risks missing the systemic implications. The moments and the
parts are important, but we adopt Schumpeter’s systemic view here. Our
democracies are changing in response to many other, sometimes more
important, factors than this kind of creative destruction, including those that
political scientists and sociologists traditionally focus on, from changes in party
systems (Mair 1997) and forms of governance (Rhodes 1997) to changes in trust
in institutions (Norris 2011) and people’s value systems (Ingelhart 1997). Yet
because our democracies are intertwined with the news, and because the news is
in turn intertwined with the market forces and technology that are central to
creating, maintaining, and changing the institutions and infrastructures of free
expression,
structural
transformations gives us a way of capturing a few key aspects of the big | Social_Media_and_Democracy |
Models Shepherd (Wang et al., 2023c) is a 7B model initialized from Llama-7B and trained
on community collected critique data and 1,317 examples of high quality human annotated data.
Shepherd generates critiques on a range of diverse NLP datasets: AlpacaFarm, FairEval, CosmosQA
(Huang et al., 2019), OBQA (Mihaylov et al., 2018a), PIQA (Bisk et al., 2020), TruthfulQA and
CritiqueEval. With GPT-4 as an evaluator, Shepherd wins or equals ChatGPT over 60% of the time.
With human evaluators, Shepherd is almost on-par with ChatGPT.
8InstructRetro is not yet open-sourced.
9
TruthfulQA FactScore HotpotQA OpenBookQA MedMC-QA TriviaQA
Models
Playtus
CoVe + Llama-65B
CoK + GPT-3.5-turbo
CRITIC + GPT-3.5-turbo
KSL + GPT-3.5-turbo
PKG + text-davinci-002
Cohen et al. (2023) + text-davinci-002
GPT-3.5-turbo
62.26
-
-
-
-
-
-
47
71.4
-
-
-
-
-
-
35.4
38.7
-
-
-
-
-
58.7
24.0
-
-
-
-
-
-
81.6
78.3
-
-
-
-
-
73.3
47.4
44.4
-
-
-
-
-
75.1
83.1
79.3 | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
that it's laying down .The generated image with predicted box is stored at the path: /images/d59a.jpg.HuggingGPT organizes the collaboration of multiple models through task planning. As shown in | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
77.3 40.9 57.7
81.8 54.5 80.8
90.9 72.7 88.5
90.9 95.5 88.5
27.3
11.5
50.0 36.4 46.2
0.0
13.6 31.8 30.8
68.2 59.1 65.4
50.0 27.3 38.5
54.5 36.4 57.7
59.1 36.4 65.4
27.3 31.8 50.0
72.7 27.3 57.7
68.2 45.5 65.4
76.8 38.4 61.0
13.6 31.8 30.8
75.1 54.5 63.9
9.1
31.8
26.9
81.8 81.8 84.6
46.2
73.1
76.9
84.6
0.0
7.7
0.0
50.0
0.0
42.3
7.7
69.2
0.0
73.1
23.1
61.5
38.5
76.9
76.9
80.8
0.0
19.2
23.1
50.0
26.9
61.5
19.2
34.6
34.6
30.8
26.9
38.5
50.7
19.2
46.2
11.5
84.6
34.8 13.0 41.7 25.0 30.8
4.3
13.0 33.3 16.7 23.1
30.4 30.4 33.3 25.0
7.7
30.4 30.4 50.0 33.3 38.5
47.8 39.1 41.7 41.7
7.7
52.2 56.5 41.7 41.7 53.8
21.7 30.4 41.7 33.3
7.7
56.5 47.8 75.0 50.0 84.6
8.7
43.5 25.0 25.0 46.2
73.9 60.9 75.0 50.0 76.9
47.8 34.8 16.7 16.7 53.8
52.2 56.5 66.7 50.0 76.9
69.6 65.2 58.3 25.0 76.9
60.9 69.6 83.3 50.0 84.6
0.0
7.7
7.7
46.2
0.0
30.8
30.8
61.5
0.0
53.8
46.2
38.5
61.5
69.2 | Mixture-of-Experts |
Similarly, Google’s VGGVox [92] used a CNN with VGG architecture to learn speaker embeddings
from Mel spectrograms, achieving state-of-the-art results in speaker recognition. CNNs have
also been widely used in developing state-of-the-art speech enhancement and text-to-speech
architectures. For instance, the architecture proposed in [311, 541] for Deep Noise Suppression
(DNS) [457] challenge and Google’s Tacotron2 [491] are examples of models that use CNNs as
their core building blocks. In addition to traditional tasks like ASR and speaker identification,
CNNs have also been applied to non-traditional speech processing tasks like emotion recognition
[230], Parkinson’s disease detection [224], language identification [498] and sleep apnea detection
[497]. In all these tasks, CNN extracted features from speech signals and fed them into the task
classification model. | AReviewofDeepLearningTechniquesforSpeechProcessing |
Evaluating neural toxic degeneration in language models. In EMNLP (Findings), 2020.
[86] Haochen Liu, Jamell Dacon, Wenqi Fan, Hui Liu, Zitao Liu, and Jiliang Tang. Does gender matter? towards
fairness in dialogue systems. COLING, 2019.
[87] Irene Solaiman and Christy Dennison. Process for adapting language models to society (PALMS) with values-
targeted datasets. https://cdn.openai.com/palms.pdf, 2021.
[88] Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh
Tomar, Iulia Turc, and David Reitter. Measuring attribution in natural language generation models. arXiv preprint
arXiv:2112.12870, 2021.
[89] Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. q2: Evaluating
factual consistency in knowledge-grounded dialogues via question generation and question answering. arXiv
preprint arXiv:2104.08202, 2021. | LaMDA- Language Models for Dialog Applications |
20px;}#header nav a {color: #fff;text-decoration: none;}#main {background-color: #fff;padding: 20px;text-align: center;}#main h2 {margin: 0;}#main p {margin: 20px 0;}#footer {background-color: #333;color: #fff;padding: 20px;text-align: center;}```JS:```javascript// Today's Jokevar joke = "Why did the tomato turnred? Because it saw the saladdressing!";document.getElementById("main").innerHTML= joke;```Figure 6: Identifying problems from photos and providing solutions | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
2. Adding new vocabulary to the LLM’s lexicon may dis-
rupt the LLM’s original capabilities.
3. These methods lack the flexibility to adapt seamlessly to
different text-to-image/video generation models. For in-
stance, if the generation model requires an upgrade, the
LLM would need to be retrained.
Quantitative analyses on the representative method [45] in
Section 4 provide support for the above claims. | GPT4Video |
[16] Baltzer MCA, L´opez D, Flemisch F. Towards an interaction pattern language for human machine coop-
eration and cooperative movement. Cognition, Technology and Work. 2019;21(4):593-606. Available
from: https://doi.org/10.1007/s10111-019-00561-8.
[17] Schulte A, Donath D, Lange DS. Design patterns for human-cognitive agent teaming. In: International
Conference on Engineering Psychology and Cognitive Ergonomics. Springer; 2016. p. 231-43.
[18] L¨owgren J.
Inspirational Patterns for Embodied Interaction. Knowledge, Technology & Policy.
[19]
2007;20(3):165-77.
van Bekkum M, de Boer M, van Harmelen F, Meyer-Vitali A, Teije At. Modular design patterns
for hybrid learning and reasoning systems: a taxonomy, patterns and use cases. Applied Intelligence.
2021;51(9):6528-46. | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
[506] Elyoseph, Z., D. Hadar-Shoval, K. Asraf, et al. Chatgpt outperforms humans in emotional
awareness evaluations. Frontiers in Psychology, 14:1199058, 2023.
[507] Habibi, R., J. Pfau, J. Holmes, et al. Empathetic AI for empowering resilience in games. CoRR,
abs/2302.09070, 2023.
[508] Caron, G., S. Srivastava. Identifying and manipulating the personality traits of language
models. CoRR, abs/2212.10276, 2022.
[509] Pan, K., Y. Zeng. Do llms possess a personality? making the MBTI test an amazing evaluation
for large language models. CoRR, abs/2307.16180, 2023.
[510] Li, X., Y. Li, S. Joty, et al. Does gpt-3 demonstrate psychopathy? evaluating large language
models from a psychological perspective, 2023.
[511] Safdari, M., G. Serapio-García, C. Crepy, et al. Personality traits in large language models.
CoRR, abs/2307.00184, 2023. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
fluency” (Schwarz, et al. 2007); information that is easier to process feels
more familiar, and familiarity is a key criterion by which individuals judge
accuracy (Alter and Oppenheimer 2009). Accordingly, if individuals have
repeated contact with a piece of misinformation, they may perceive it as more
credible than if they encounter it only once, regardless of its content. | Social_Media_and_Democracy |
The original setup in BBQ included three multiple choice options (in the above example, these would be “Nancy”,
“Donald”, and “Unknown”), but such a design is less well-matched to how developers would use PaLM-2 to build a
generative QA system, and it potentially under-captures the full potential for harm, as generative text has the potential
to introduce biases or representational harms beyond those tested with just three multiple choice options. Thus we
chose to adapt the existing dataset by presenting PaLM 2 with a context and question, and analyzing the generated text
directly, without indicating any multiple choice options.
76 | PaLM 2 Technical Report |
9
(a)
(b)
Figure 6: Ablation studies on the Spider development set. (a) Accuracies with different numbers of
initial samples. (b) Breakdown accuracies on problems with different hardness levels.
(a)
(b)
Figure 7: Ablation studies on TransCoder. (a) The accuracy of SELF-DEBUGGING prompts with
different numbers of debugging turns. (b) The accuracy with different numbers of initial samples.
Note that Codex+Expl. only prompts the model to generate explanations for the input C++ code, and
does not perform SELF-DEBUGGING.
5.2.3 Text-to-Python Generation | Teaching Large Language Models to Self-Debug |
.
Springer.
Zhu, Y., Olszewski, K., Wu, Y., Achlioptas, P., Chai, M., Yan, Y., & Tulyakov,
S. (2022a). Quantized gan for complex music generation from dance videos.
In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv,
Israel, October 23–27, 2022, Proceedings, Part XXXVII (pp. 182–199).
Zhu, Y., Wu, Y., Olszewski, K., Ren, J., Tulyakov, S., & Yan, Y. (2022b).
Discrete contrastive diffusion for cross-modal and conditional generation.
arXiv preprint arXiv:2206.07771 , .
Zixun, G., Makris, D., & Herremans, D. (2021). Hierarchical recurrent neural
networks for conditional melody generation with long-term structure. In
2021 International Joint Conference on Neural Networks (IJCNN) (pp. 1–
8). IEEE.
49 | Video2Music |
enable reasoning through sequential attention over memory content.
NTMs, followed by Differentiable Neural Computer (DNC) (Graves et al., 2016) and Sparse DNC
(Rae et al., 2016), are implemented as recurrent neural networks capable of writing to memory
storage over time. All these models are differentiable and trainable via backpropagation through
time (BPTT). Parallel research lines extend recurrent neural networks, such as LSTM, with data
structures like stacks, lists, or queues (Joulin and Mikolov, 2015; Grefenstette et al., 2015). MANN
architectures with more advanced addressing mechanisms, such as address-content separation and
multi-step addressing, have been proposed in (Gulcehre et al., 2016, 2017; Meng and Rumshisky,
2018). The Global Context Layer model (Meng and Rumshisky, 2018) employs address-content
separation to address the challenge of training content-based addressing in canonical NTMs. | Scaling Transformer to 1M tokens and beyond with RMT |
Trends in Information Retrieval 9, 5 (2015), 355–475.
[59] Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. SAMSum Corpus: A Human-annotated
Dialogue Dataset for Abstractive Summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization.
Association for Computational Linguistics, Hong Kong, China, 70–79. https://doi.org/10.18653/v1/D19-5409
[60] Silke M Göbel and Matthew FS Rushworth. 2004. Cognitive Neuroscience: Acting on Numbers. Current Biology 14,
13 (2004), R517–R519.
[61] Ben Goodrich, Vinay Rao, Peter J Liu, and Mohammad Saleh. 2019. Assessing the Factual Accuracy of Generated Text.
In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 166–175. | SurveyofHallucinationinNatural Language Generation |
34.0
86.7
32.8
31.8
70.6
70.4
93.1
77.9
65.0
61.4
61.0
64.1
56.0
63.8
43.7
60.8
Table 10: Automatic evaluation results of LaMini-Flan-T5 language models and their baselines on 15 NLP tasks.
“Average” indicates the micro-average of the individual task results.
GPT-Neo
LaMini-Neo GPT-Neo
LaMini-Neo
135M
1.3B
# of params.
OpenBookQA
SciQ
RACE
ARC
PIQA
ReCoRD
SST
MRPC
RTE
MultiNLI
MultiNLI (mis)
WSC273
WinoGrande
WiC
HellaSwag
Average
26.2
68.8
27.6
23.1
62.5
65.6
53.9
68.4
54.9
35.5
35.4
55.3
50.4
50.0
30.4
47.2
33.6
77.1
34.1
25.9
71.1
81.4
65.7
68.4
60.3
35.8
36.2
75.1
54.9
50.0
48.9
54.6
36.4
84.2
34.3
32.9
71.7
75.2
91.2
70.3
71.1
49.3
49.7
66.7
54.8
50.2
47.5
59.0
Table 11: Automatic evaluation results of LaMini-Neo language models and their baselines on 15 NLP tasks.
“Average” indicates the micro-average of the individual task results.
28.8
93.0
35.9
25.1
67.0
68.2
92.3
71.3
78.7
66.7
66.9
57.5
54.2
52.7
36.4
59.7 | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
18
THE NEXT DECADE IN AI / GARY MARCUS
hybrid model could better capture a variety of learning challenges, such as the game
fizz-buzz, which defied a multlayer perceptron. A team of people including Smolensky
and Schmidhuber have produced better results on a mathematics problem set by
combining BERT with a tensor products (Smolensky et al., 2016), a formal system for
representing symbolic variables and their bindings (Schlag et al., 2019), creating a new
system called TP-Transformer. | The Next Decade in AI- |
{Instruction for the target task}
Task:
Table 8: Prompt used for the input-first approach of instance generation. The model is prompted to generate the
instance first, and then generate the corresponding output. For instructions that don’t require additional input, the
output is allowed to be generated directly.
Given the classification task definition and the class labels, generate an input that
corresponds to each of the class labels.
correct class label.
If the task doesn’t require input, just generate the
Classify the sentiment of the sentence into positive, negative, or mixed.
Task:
Class label: mixed
Sentence: I enjoy the flavor of the restaurant but their service is too slow.
Class label: Positive
Sentence: I had a great day today.
Class label: Negative
Sentence: I was really disappointed by the latest superhero movie.
The weather was beautiful and I spent time with friends.
I would not recommend it. | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
[476] Yuki Saito, Shinnosuke Takamichi, and Hiroshi Saruwatari. 2021. Perceptual-similarity-aware deep speaker repre-
sentation learning for multi-speaker generative modeling. IEEE/ACM Transactions on Audio, Speech, and Language
Processing 29 (2021), 1033–1048.
[477] Hojjat Salehinejad, Sharan Sankar, Joseph Barfett, Errol Colak, and Shahrokh Valaee. 2017. Recent advances in
recurrent neural networks. arXiv preprint arXiv:1801.01078 (2017).
[478] Elizabeth Salesky, Matthias Sperber, and Alan W Black. 2019. Exploring phoneme-level speech representations for
end-to-end speech translation. arXiv preprint arXiv:1906.01199 (2019).
[479] Kanthashree Mysore Sathyendra, Thejaswi Muniyappa, Feng-Ju Chang, Jing Liu, Jinru Su, Grant P Strimel, Athanasios
Mouchtaris, and Siegfried Kunzmann. 2022. Contextual adapters for personalized speech recognition in neural
transducers. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).
IEEE, 8537–8541. | AReviewofDeepLearningTechniquesforSpeechProcessing |
Facial reflectance capture typically requires a control-
lable illumination system equipped with multiple cameras,
first introduced as a Light Stage [13]. Polarized illumination
and gradient patterns can be employed for diffuse-specular
separation [49, 27], using which, spatially varying facial
reflectance maps can be acquired, that describe BRDF pa-
rameters, including the diffuse and specular albedo and nor-
mals. Although recent works attempt to simplify the captur-
ing apparatus and process using inverse rendering [29, 56]
or commodity devices [39], such methods still require a la-
borious capturing process and expensive equipment.
Since their introduction by Blanz and Vetter [4], 3D Mor-
1 | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
[53] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng
Zhu, Rui Zhao, and Hongsheng Li. Hpsv2 github. https:
//github.com/tgxs002/HPSv2/tree/master,
2023. Accessed: 2023 - 11 - 15. 5
[54] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V
Self-training with noisy student improves imagenet
Le.
Diffusion Model Alignment Using Direct Preference Optimization
Supplementary Material
S1. Comparisons to existing work
RL-Based Methods
such as [6, 11] have shown effectiveness in operating on a limited set of prompts (< 10 and <
1000 respectively) but do not generalize as well to the open-vocabulary setting as shown in [7, 31]. We found this in our
experiments as well, where training using the DDPO scheme did not improve PickScore over the baseline model over a sweep
of hyperparameters.
While DDPO [6] is an RL-based method as is DPOK [11], their target objective and distributional guarantees are different. | DiffusionModelAlignmentUsing Direct Preference Optimization |
Although LLMs exhibit outstanding performance in language comprehension [25; 301] and multi-turn
conversations [302], they inherently lack visual perception and can only understand discrete textual
content. Visual input usually contains a wealth of information about the world, including properties
of objects, spatial relationships, scene layouts, and more in the agent’s surroundings. Therefore,
integrating visual information with data from other modalities can offer the agent a broader context
and a more precise understanding [120], deepening the agent’s perception of the environment.
To help the agent understand the information contained within images, a straightforward approach
is to generate corresponding text descriptions for image inputs, known as image captioning [303;
304; 305; 306; 307]. Captions can be directly linked with standard text instructions and fed into
the agent. This approach is highly interpretable and doesn’t require additional training for caption | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Mean absolute error (mm) ↓
Chest Waist
Height
Method
Sengupta et al. [52]
TUCH [40]
SPIN [33]
STRAPS [51]
ExPose [9]
SHAPY (ours)
Model
SMPL
SMPL
SMPL
SMPL
SMPL-X
SMPL-X
84
82
72
207
107
71
186
92
91
278
107
64
263
129
129
326
136
98
Hips
142
91
101
145
92
74
Table 4. Evaluation on MMTS. We report the mean absolute error
between ground-truth and estimated measurements.
7. Conclusion | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
actions like typing, searching, navigating to the next page, etc. They perform well in basic tasks
such as online shopping [392] and search engine retrieval [90], which have been widely explored.
However, agents without LLM capabilities may struggle to adapt to the more realistic and complex
scenarios in the real-world Internet. In dynamic, content-rich web pages such as online forums or
online business management [391], agents often face challenges in performance.
In order to enable successful interactions between agents and more realistic web pages, some
researchers [393; 394] have started to leverage the powerful HTML reading and understanding
abilities of LLMs. By designing prompts, they attempt to make agents understand the entire HTML
source code and predict more reasonable next action steps. Mind2Web [389] combines multiple
LLMs fine-tuned for HTML, allowing them to summarize verbose HTML code [388] in real-world | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Despite these advantages, transitioning directly from FP32 to FP16 can sometimes lead to performance degradation [114, 335]
due to issues like overflow or underflow inherent in FP16. To circumvent these challenges, the automatic mixed-precision
(AMP) [184] method has been developed. AMP maintains a master copy of weights in FP32 while employing FP16 for
computations during the forward and backward passes. Post-calculation, the weights are converted back to FP32 for updating
the master weights. This method, coupled with a loss scaling technique that preserves small gradient values, enables AMP
to match the accuracy of FP32 training without the need for extensive hyperparameter tuning. Further advancements in
precision reduction have led to the introduction of Brain Floating Point (BF16) [123], a novel half-precision format. BF16,
designed to cover the same range as FP32 by allocating more bits to the exponent and fewer to the significand compared to | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
ψ
(1)
+ Lreg(x; E, D)
are
the
reconstructed
Here D(E(x))
images,
reconstruction
perceptual
Lrec(x, D(E(x)))
loss, Ladv(D(E(x))) is the adversarial
loss, Dψ(x) is
a patch based discriminator loss, and Lreg(x; E, D) is the
KL-regularisation loss.
the
is
Diffusion model fine-tuning Following the autoencoder
fine-tuning, we proceeded to the second stage, which in-
volved fine-tuning the diffusion model. This was achieved
using the frozen autoencoder’s latent representations as in-
put, with a latent input size of 64x64x4.
For this stage, we employed the Adam optimizer with a
learning rate of 10−5 and a batch size of 32 . We train the
diffusion model for 178 epochs with the loss function:
LLDM3D := Eε(x), (cid:15) ∼ N (0, 1), t(cid:2)||(cid:15) − (cid:15)θ(zt, t)||2
(2)
where (cid:15)θ(zt, t) is the predicted noise by the denoising
(cid:3)
2
U-Net, and t is uniformly sampled. | LDM3D- Latent Diffusion Model for 3D |
file that is not part of the version control system.4. Error handling: The code does not check for errors when calling the ``token.SignedString(key)'' function, which could lead to unexpected behavior. It's important to handle errors properly to avoid potential security issues.5. Lack of HTTPS: The server is listening on HTTP, which means that the communication between the client and the server is not encrypted. To ensure secure communication, use HTTPS instead.GPT-4 (launch)strategy.[29] We are thus particularly interested in evaluating power-seeking behavior due to the | gpt-4-system-card |
environment implementation and how the agents interact with the
sandbox world’s underlying engine.
3.1 Agent Avatar and Communication
A community of 25 unique agents inhabit Smallville. Each agent is
represented by a simple sprite avatar. We authored one paragraph
of natural language description to depict each agent’s identity,
including their occupation and relationship with other agents, as
seed memories. For example, John Lin has the following description: | Generative Agents- Interactive Simulacra of Human Behavior |
4.2.2 Metrics and Holistic Evaluation
For SuperGLUE, we report well-established metrics such as accuracy, F1 or Exact Match, whenever appropriate.
For GEM benchmark, we use the Rouge-L metric. For language modeling we report negative log perplexity.
The universality of the models, i.e., their collective performance across all range of tasks, is a main evaluation
criteria here. To enable the comparison between models from this perspective, we need an aggregate
performance score. However, metrics on different tasks we include are widely different in nature – take,
for example, F1 and perplexity. To address this, we opt to report and use the normalized relative gain with
respect to baselines as an overall metric. For this purpose, we use the standard language model (decoder-only)
(GPT-like) and standard span denoising encoder-decoder (T5) as prime baselines and report all methods
against their relative performance against these well-established candidates. We believe this is the most | UL2- Unifying Language Learning Paradigms |
versary is interested in recovering the private in-
formation about a target individual, given its name
and corresponding domain knowledge. This attack
usually takes the format like “ name: [name], email:
____” to force LLMs to predict private information
associated with the name. The attack based on the
association can be harmful directly to a partially
identified victim. | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
[465] Shuster, K., J. Xu, M. Komeili, et al. Blenderbot 3: a deployed conversational agent that
continually learns to responsibly engage. CoRR, abs/2208.03188, 2022.
[466] Du, W., Z. M. Kim, V. Raheja, et al. Read, revise, repeat: A system demonstration for
human-in-the-loop iterative text revision. CoRR, abs/2204.03685, 2022.
[467] Kreutzer, J., S. Khadivi, E. Matusov, et al. Can neural machine translation be improved
with user feedback? In S. Bangalore, J. Chu-Carroll, Y. Li, eds., Proceedings of the 2018
Conference of the North American Chapter of the Association for Computational Linguistics:
Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6,
2018, Volume 3 (Industry Papers), pages 92–105. Association for Computational Linguistics,
2018. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
per-parameter, τ (this is the Chinchilla trend plotted in Figure 3): | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
cludes performing tasks that require understanding
context, making decisions, and learning from inter-
actions (Xi et al., 2023; Hu and Shu, 2023). Such
agents are pivotal in applications where human-like
cognitive abilities are essential. | AppAgents |
Query: please read the sentence "Hi there, I am ChatGPT empowered by Huggingface family! Yes, I can provide more than one thousand models for dozens of tasks. For more fun and creativity, I have invited Diffusers family to join our team. Feel free to experience it!" for me.Response: Sure, I can do that. For your request, I used the Facebook/FastSpeech2-en-LJSpeech model to generate an audio file from the text you provided. The generated audio file is located at ' /audios/d5f3.wav '. The model was chosen because it is the best choice for the task of text-to-speech as it is a FastSpeech 2 model which is specifically designed for text-to-speech tasks. It is also the most recent model in the list and has the most potential to provide the best results. I hope this answers your question. Is there anything else I can help you with?Video > GenerationQuery: look at the document doc.jpg, what is the purchase amount?Response: Sure. I have used | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
from diverse sources at large scale. If following the semantic web standards,4 one can refer to knowledge graph(s) as Linked
Data.5 | Knowledge graphs as tools for explainable machine learning: A survey |
Methodology – how will you achieve the research aims?
It is important to present the proposed research methodology (e.g. techniques, sample size, target populations, species
choice, equipment and data analysis) and explain why it is the most appropriate methodology to effectively answer the
research question. If space allows, it may be a good idea to justify the methodology by explaining what alternatives have
been considered and why these have been disregarded. You could also point out how your project fits with the research
environment of your prospective institution and why this institution is the best place to conduct your research, in particular
if this will provide you with access to unique expertise, pieces of equipment or data. | research proposal guidance |
l for which 1 <= l < n.
The first line contains a single integer t (1 <= t <=
10 000) - the number of test cases .
The first line of each test case contains a single
integer n (2 <= n <= 10^5) .
The second line of each test case contains n integers
a_1 , a_2 , ... , a_n (1 <= a_i <= 10^6) .
It is guaranteed that the sum of n over all test
cases doesn ’t exceed 3 . 10^5.
The first line contains a single integer t (1 <= t <=
10 000) - the number of test cases .
The first line of each test case contains a single
integer n (2 <= n <= 10^5) .
The second line of each test case contains n integers
a_1 , a_2 , ... , a_n (1 <= a_i <= 10^6) .
It is guaranteed that the sum of n over all test
cases doesn ’t exceed 3 . 10^5.
For each test case , print a single integer - the
maximum possible value of the product from the
statement .
For each test case , print a single integer - the
maximum possible value of the product from the
statement .
Input
Output
Example
Input
Input
Output | alphacode |