_id
stringlengths 36
36
| text
stringlengths 200
328k
| label
stringclasses 5
values |
---|---|---|
62658f52-8487-4efd-8abd-2c62b428385c | The boundary between what is considered model-free and model-based reinforcement learning is blurred when one can considers both the model network and controller network together as one giant policy that can be trained end-to-end with model-free methods. [1]} demonstrates this by training both world model and policy via evolution. [2]} explore modifying sensor information similarly to our observational dropout. Instead of performance, however, this work focus on understanding what these models learn and show there usefulness – e.g. training a policy inside the learned models.
| w |
9b51cea8-c500-473d-ae8d-e8452677e2c4 | In this work, we explore world models that emerge when training with observational dropout for several reinforcement learning tasks. In particular, we've demonstrated how effective world models can emerge from the optimization of total reward. Even on these simple environments, the emerged world models do not perfectly model the world, but they facilitate policy learning well enough to solve the studied tasks.
| d |
eb892484-f68c-4387-868e-22650a371dc8 | The deficiencies of the world models learned in this way have a consistency: the cart-pole world models learned to swing up the pole, but did not have a perfect notion of equilibrium—the grid world world models could perform reliable bit-shift maps, but only in certain directions—the car racing world model tended to ignore the forward motion of the car, unless a turn was visible to the agent (or imagined). Crucially, none of these deficiencies were catastrophic enough to cripple the agent's performance. In fact, these deficiencies were, in some cases, irrelevant to the performance of the policy. We speculate that the complexity of world models could be greatly reduced if they could fully leverage this idea: that a complete model of the world is actually unnecessary for most tasks—that by identifying the important part of the world, policies could be trained significantly more quickly, or more sample efficiently.
| d |
429fe1d7-e66d-4712-b69e-05befb649bce | In the last decades various aqm mechanisms have been proposed to minimize excessive standing queues in the Internet. One of the most influential recent efforts is CoDel [1]} whose goal is that the queuing delay at the bottleneck link is at least once under 5ms in a moving window of 100ms.
| i |
cbd7956a-aae9-4518-aa93-37118fe8990e | While it is important to keep the queuing delay constrained, it is also necessary to ensure that overly aggressive flows cannot benefit from “stealing” less aggressive flows' bandwidth. Thus researchers and engineers have developed fq mechanisms [1]}, [2]} to isolate different flows' queues so that for example a delay-sensitive live video call cannot be impaired by a concurrent bulk transfer which consumes all the available bandwidth.
| i |
c833bdaa-2d2a-4bef-92e3-9bcc0a911546 | Recent approaches have tried to combine aqm with fq. [1]} demonstrate fq_codel, a qdisc that uses fq and lets CoDel manage each queue. [2]} expand upon this and create the cake qdisc that also adds features such as not only per-flow queuing but also per-host queuing for even increased fairness. Furthermore they also include bandwidth shaping into their solution and aim to create one qdisc that is easy to configure, can be easily deployed on home routers and offers all features in one solution.
| i |
a411cfb0-98e1-482a-a791-4d46716620d1 | While we do not want to make statements about the general performance of CoDel, we show that fq_codel and cake do not optimally use available bandwidth in common network configurations for common cca. This becomes especially prevalent for links with a high bandwidth or a large rtt but is already noticeable for common scenarios, such as a link with 100Mbit/s and an rtt of 50ms. We show that the impaired performance is a result of keeping the queuing delay under 5ms, which hinders cca such as Reno or Cubic from reaching maximum throughput. Moreover, this behavior can result in unnecessary standing queues on links with very low latencies under 10ms, which occur in data centers and between users and close by servers of content delivery networks.
| i |
8717069b-804c-4fa0-b28c-7e0610acc1d4 | We show that a prototype of our solution cocoa can achieve the aforementioned objectives for the most common loss-based cca, Reno [1]} and Cubic [2]}. These cca operate by continuously increasing the number of bytes that are allowed to be in the network (congestion window) if no packet loss is experienced and by sharply decreasing this number if a packet is lost (multiplicative decrease). With Cubic being the default cc in all major OS, we especially emphasize our evaluation on improving its performance. Contrasting to the aforementioned cca, recently proposed BBR [3]} does not continuously increase and then sharply decrease when packets are lost but instead uses periodic measurements to estimate the available bandwidth as well as the minimal rtt and then tries to stay at this point of optimal bandwidth and minimal delay. BBR is thus considered to be model-based. We show that cocoa also behaves well in interaction with BBR v1.
| i |
81e2ee38-0af6-4938-82d9-a4ca5f7b5389 | We implement cocoa as an extension of the “fq” qdisc [1]}. This means that when using cocoa also all features offered by fq are available. We add three configuration parameters to the qdisc: the multiplier (default 1.25), the maximum buffer increase (default 2.0) and the maximum gi duration (default 1.0 s). We make the source code of our implementation freely available to enable reproducibility and encourage experimentation: https://github.com/CN-TU/cocoa-qdisc.
| m |
ee8de447-2420-45ac-9c6e-bce0957eb94f | We evaluate cocoa using the py-virtnet (https://github.com/CN-TU/py-virtnet) toolkit to build a virtual network using Linux's network namespaces. We initialize the buffer size for cocoa to 100 packets per flow like in fq. The testbed consists of a sender and a receiver connected via a switch. We run the qdisc being tested on the interface that connects the switch to the receiver. We introduce delay with netem and limit bandwidth with htb.
<FIGURE><FIGURE><FIGURE> | m |
5ae2cb64-c31e-4b31-9d7b-237b70f77132 | First we evaluate if cocoa is able to maintain a small buffer like it is necessary in case of small bdp. As can be seen in fig:fq the regular fq qdisc (with a standard queue size of 100 packets) maintains a standing queue and unnecessarily leads to a significant increase in rtt (100 ms minimum, when the real minimum is 10 ms). In contrast, fig:cn maintains full throughput while at the same time keeping the delay minimal and not having a standing queue. Also, when drastically changing the bandwidth by halving it (fig:cn) or doubling it (fig:cn2) cocoa rapidly adapts to the new conditions and returns to the optimum state.
<FIGURE><FIGURE><TABLE> | m |
b63a4d33-ac81-488c-94f0-6ce341b1bff3 | Next we compare cocoa against fq_codel. fig:fqCodel shows that fq_codel keeps the queuing delay under 5 ms. It also shows that keeping the queuing delay that low is detrimental for achieving full throughput when using the Cubic cca. With cocoa we achieve throughput that is more than 10% higher overall, while only slightly increasing the average delay. In addition to Cubic, we performed experiments with Reno. Here, over a 240 s flow of 100 Mbit/s with a delay of 50 ms, cocoa achieves more than 20% higher throughput than fq_codel. The average rtt for fq_codel is 51.34 ms while it is 96.78 ms for cocoa. tab:performanceresults shows that cocoa reaches link utilizations close to 100% for both Cubic and Reno in large bdp scenarios. rtt can be higher than for other qdisc if this is required to achieve optimum throughput for the cca for the given link speed and base rtt.
| m |
407979f4-c4e2-4e58-b6ec-6e6ecdc4e20e | Besides fq_codel possibly leading to lower throughput as we have shown, another problem is that for links with a small bdp fq_codel can keep a standing queue akin to fq since fq_codel only wants queuing delay to fall below 5 ms once every 100 ms. For example, on a link with a base rtt of 1 ms, fq_codel would accept a permanent standing queue of 4.9 ms, leading to overall rtt being more than doubled, which is not required for achieving optimal throughput.
| m |
9fc9a3cc-b4ce-471f-92f2-31d2d2e3ea63 | Our goal was to design a fair qdisc which achieves maximum throughput while keeping the queuing delay as small as possible. The results of the experiments, which we perform with a prototype of cocoa, show that there are certain common scenarios in which current state-of-the-art fair qdisc like fq, fq_codel and cake fall short of the optimum throughput by a significant margin while our approach succeeds in fully utilizing the bottleneck link (with a potential increase in delay) (tab:performanceresults; fig:fqCodel vs. fig:cnLong). Furthermore, the aforementioned qdisc also suffer from standing queues in scenarios with small bdp which we can also mitigate with our approach (fig:fq vs. fig:cn).
| d |
9476fde9-c1b9-4285-8ae6-7cd673c88fd5 | We see this work as an initial step towards learning and flow-adaptive fair queuing at bottlenecks and think that the most promising application domain of cocoa is at the Internet's edge. An interesting further improvement could be to not statically initialize the buffer with a constant 100 packets but instead use experience from previous flows. Moreover, we think that a promising direction for future work might be to explore the use of reinforcement learning. Such an approach would fingerprint a flow and choose the buffer size accordingly, maximizing a chosen objective such as high throughput with minimal delay.
| d |
44bffcf2-7041-402c-a435-567b28c61391 | When it comes to deploying machine learning models in real-world applications, there is often a mistaken belief that models will be used in a stationary environment. This belief assumes that the same concepts learned during the training phase will remain valid at inference time [1]}, or that training samples and production-time samples will come from the same distribution [2]}. However, in real-world scenarios this is often not true, and both situations may result in some type of drift that can end up affecting model's performance [3]}. In addition, due the high cost of collecting and labeling samples, this performance loss can often not be confirmed in many real-world problems and other methods that only rely on distribution changes must be used.
| i |
87190202-f024-4ae1-9fb4-2188b8b6b1d7 | to unify in a single library the part of concept drift detection and adaptation (traditionally researched and used for streaming/evolving data streams and incremental learning [1]}) with the research of change detection in the covariate distributions (also known as data shift, related to the field of statistical two-sample testing [2]} and methods that measure distance between distributions [3]}).
Drift detection
Traditionally there has been little consensus on the terminology and definitions of the different types of drift, as stated in [4]}. In order to adopt some clear definitions for the remainder of this paper, we apply those used in [5]} for the concept drift part, in combination with those used in [2]}'s work for detecting dataset shift using only the covariates. Therefore, we set up the following definitions assuming two different time points, \({t_{0}}\) and \({t_{1}}\) :
Concept drift.
There is a change in the joint probability \({P(X, y) = P(y|X)P(X)}\) , with or without a change in \({P(X)}\) . Thus, it can be defined as \({P_{t_{0}}(X, y) \ne P_{t_{1}}(X, y)}\) . Known as real concept drift when changes only affect \({P(y|X)}\) [5]}.
Virtual drift.
There is a change in \({P(X)}\) but does not affect the conditional probability \({P(y|X)}\) [5]}, [9]}. Thus, \({P_{t_{0}}(X) \ne P_{t_{1}}(X)}\) and \({P_{t_{0}}(y|X) = P_{t_{1}}(y|X)}\) .
Data drift.
As well as virtual drift, there is a change in \({P(X)}\) but due the fact that there is no labeled data \({y}\) available, it cannot be verified that \({P(y|X)}\) is being affected or not. Therefore, this type of drift only focuses in the distribution of the covariates \({P(X)}\) , so \({P_{t_{0}}(X) \ne P_{t_{1}}(X)}\) . Hereafter, we rename dataset shift [2]} to data drift in order to maintain consistency with the above definitions and with some of the related software mentioned in Section that also refer to it as data drift [11]}, [12]}.
Overview and design
The design and implementation of the library has been carried out with the aim of making it compatible with the use of scikit-learn, both through the use of its estimators, as well as its pipeline. We also provide transformations that can be included in a pipeline and allow to use multiple of the unsupervised methods, described in Section REF , at the same time. Additionally, we provide prequential error metrics [13]} to evaluate the performance of supervised methods described in Section REF .
Detection methods are divided in the supervised, unsupervised and semi-supervised categories depending on the type of drift they can detect, according to the definitions given in Section , and how they detect it. These categories are explained in what follows.
Supervised
These methods are aimed at detecting concept drift, so in order to update the detector they require the ground-truth labels of the predictions that have previously been made.
In terms of implementation, the detector wraps the scikit-learn estimator and receives through the update method the value that is used to perform the necessary steps (at least update detector's inner statistics and check if drift is occurring) on each iteration. In addition to receiving a scikit-learn estimator when the detector is instantiated, it receives a configuration class that contains a set of parameters that determine how it will behave.
Despite the fact that detectors can be updated by interacting with them directly (via update method), we provide the following helper classes, so-called modes, to facilitate the interaction with them.
IncrementalLearningMode works with methods that use warning and drift thresholds, and scikit-learn estimators that support partial_fit method. It acts on an instance-incremental manner by passing a tuple that contains a feature vector \(\vec{x}\) , the prediction made by the model and the ground-truth label, only when the user requires it. Additionally, a value functions that uses that tuple needs to be provided to compute a value for the detector, usually an error function. Therefore, this helper class allows to check the presence of drift and add samples to the model's decision.
NormalMode can be used by all the supervised methods without incrementally training the wrapped model. It only checks if the detector has raised the drift flag and resets its inner statistics. Like IncrementalLearningMode, a value function must be provided, but estimators that only support fit method are allowed, so it is not restricted only to those that support partial_fit.
Unsupervised
The unsupervised methods are focused on detecting data drift by considering only the covariates and regardless of the existence of labels or its lack thereof. Therefore, these algorithms try to detect changes at a feature level by comparing new data distributions against reference data distributions.
BaseEstimator and TransformerMixin classes from scikit-learn are used to implement these type of methods. The fit method stores the reference distribution and the transform method applies the corresponding algorithm to compare new samples distribution with the reference distribution. In order to check if drift has happened after calling transform method, distance or test attributes can be acceded, depending on the type of algorithm used. All the implemented methods act in batch mode, expecting to receive multiple samples each time a drift check is performed. Moreover, these detectors are implemented considering the type of data that they are expected to work with: categorical or numerical, and the number of features that can be considered: univariate or multivariate.
Semi-supervised
As with supervised methods, semi-supervised ones aim to detect concept drift by providing ground-truth labels only when they are required. Therefore, when drift is suspected (equivalent to the warning zone present in several of the supervised algorithms) it is necessary to manually provide the detector with new labeled samples coming from an external entity to verify whether drift is occurring or not. If drift is present, the model is replaced by one trained with these new samples, otherwise the algorithm returns to a control state.
As far as implementation is concerned, their way of working is similar to supervised methods with the exception of the use of the helper classes described in Section REF .
Development
With the intention of following a set of open source software development standards that allow the maintainability and extensibility of the library over time, we emphasize the following areas:
Continuous integration.
A Continuous integration workflow based on GitHub Actions ensures that new modifications easily integrate with the existing code base and that they are compatible with multiple Python versions.
Documentation.
An API documentation is provided using sphinx and hosted in Read the Docshttps://frouros.readthedocs.io website. Some basic examples on the use of these detection methods are included in the documentation.
Quality code.
In order to ensure minimum standards in terms of code quality, code coverage is set to be greater than 90% and some Python quality and style tools that comply with PEP8 standards are used, such as flake8, pylint, black and mypy.
Open source.
In addition to the source code being available on GitHubhttps://github.com/IFCA/frouros, Frouros package can be installed through the Python Package Index (PyPI)https://pypi.org/project/frouros/. In terms of licensing, it is distributed under the BSD-3-Clause license.
Comparison to related software
With regard to the concept drift detection part, MOA [14]} has most of the supervised methods that we are including, but they are implemented in Java. In Python, River [15]}, that is focused on online machine learning and streaming data, offers some supervised algorithms but only a subset of those presented here. Another Python library that contains this type of detectors is scikit-multiflow [16]}, but it has not been taken into account due to the fact that it was merged with the online machine learning library Creme [17]}, resulting in the above-mentioned River library.
For the data drift part, Alibi-detect [11]} has several algorithms related to the field of statistical two-sample hypothesis testing and some of them can act both online (single sample) and offline (batch sample). TorchDrift [19]} also implements some statistical two-sample hypothesis testing methods but in this case uses PyTorch for their implementation.
To the best of our knowledge, Menelaus [12]} is the only open source library that has both supervised and unsupervised methods, as well as a semi-supervised algorithm, although they classify them in the following types: change detection, concept drift and data drift. Supervised algorithms are implemented in such a way that the user must necessarily be in charge of controlling each iteration of the sample without offering some helper functions or classes to interact with the detector, as Frouros does and it is explained in Section REF .
Table REF provides a more detailed view of the methods implemented in each of the libraries mentioned above, as well as those included in Frouros. At the time of writing this paper, Frouros is listed as the library with the highest number of methods available with 24.
Moreover, there are several other libraries and tools that have been excluded from Table REF , due to the fact that they implement a more limited number of methods, or that are more focused on building graphical dashboard and visual representations, such as Deepchecks [21]}, Eurybia [22]}, Evidently [23]} or NannyML [24]}.
<TABLE>
Conclusion and future work
This papers presents Frouros, a Python library for drift detection in machine learning problems that can be used with scikit-learn library, both for algorithms that aim to detect concept drift and for those that try to detect data drift. Moreover, this library tries to meet some of the open source software development standards that allow to extend it, both in terms of adding new methods or modes to interact with the detectors as helper functions or classes. In view of future work, we plan to adapt supervised methods to support batch-incremental learning, as long as the nature of each algorithm allows it. We will also consider to extend the unsupervised part to include some methods that work with individual instances (online) and not only in batch mode (offline). Finally, adding new modes, as described in Section REF , to interact with the detectors would make it possible to adapt the library to handle more real-world use cases.
The authors acknowledge the funding from the Agencia Estatal de Investigación, Unidad de Excelencia María de Maeztu, ref. MDM-2017-0765.
| i |
088718a6-c0df-4432-98e5-bcfb1b2bc4c0 | This papers presents Frouros, a Python library for drift detection in machine learning problems that can be used with scikit-learn library, both for algorithms that aim to detect concept drift and for those that try to detect data drift. Moreover, this library tries to meet some of the open source software development standards that allow to extend it, both in terms of adding new methods or modes to interact with the detectors as helper functions or classes. In view of future work, we plan to adapt supervised methods to support batch-incremental learning, as long as the nature of each algorithm allows it. We will also consider to extend the unsupervised part to include some methods that work with individual instances (online) and not only in batch mode (offline). Finally, adding new modes, as described in Section REF , to interact with the detectors would make it possible to adapt the library to handle more real-world use cases.
| d |
d27fe04b-18be-40d8-ba58-d6bafb2b94a0 | With the availability of petabytes of data and the emergence of `deep' learning as an AI technique to find statistical regularities in these large quantities of data, artificial intelligence in general and machine learning in particular has arguably entered a new phase since its emergence in the 1950s. Deep learning aims to build hierarchical models representing the data, with every new layer in the hierarchy representing ever more abstract information; for example, from individual pixels to lines and curves, to geometric patterns, to features, to categories. Superficially this might be related to how the human visual cortex interprets visual stimuli and seeks to classify a picture to be that of a cat, rather than of a dog.
| i |
9d5091b5-52f8-465b-9d0a-b56a6c458dd2 | When describing in what sense a cat is different from a dog, humans may use features and categories that we agreed upon to be defining features of cats and dogs, such as whiskers, location and form of the ears, the nose, etc. The deep learning method, however, does not adhere to features we humans find to be good descriptors; it bases its decisions where and how to `carve nature's joints' solely on basis of the statistics of the data. Hence, it might very well be that the curvature of the spine (or some other apparently `random' feature) happens to be the statistically most important factor to distinguish cats from dogs. This imposes huge challenges when the machine learning algorithm is asked to justify its classification to a human user. The sub-field of explainable AI has recently emerged to study how to align statistical machine learning with informative user-based motivations or explanations.
Explainable AI, however, is not limited to deep neural network applications. Any AI application where trustworthiness is important benefits from justification and transparency of its internal process [1]}, and this includes decision support systems that are based on Bayesian networks [2]}, which is the focus of this paper. In these systems typically one is interested in the hypothesis that best explains the available evidence; for example in a medical case, the infection that is most probable given a set of health complaints and test findings [3]}.
| i |
fc9d218c-ba12-4d43-8d3d-970cced7442e | Note that `explainability' in explainable AI is in principle a triadic relationship between what needs to be explained, the explanation, and the user who seeks the explanation [1]}. An explanation will be more satisfying (`lovelier', in Peter Lipton's [2]} terms) if it allows the user to gain more understanding about the phenomenon to be explained. Lacave and Díez [3]} review explanation in the context of a Bayesian network. In their work, three major categories of explanation has been identified. For example, the focus can be on explanation of the evidence or observed variables that need to be explained, where the most probable joint value assignment (or MAP explanation) serves as this explanationIn this paper we do not touch the question whether `best' is to be identified with `most probable'. The interested reader is referred to the vast literature on inference to the best explanation such as [2]}, and more in particular to some of our earlier work [5]} that discusses the trade-off between probability and informativeness of explanations.. Alternatively, the focus can be on explanation of the model, i.e., the structure of the Bayesian network can be explained by providing conditional independence tests and justification in the form of expert assessment of probability distributions or training data sets. Finally, and for us most relevant, the focus can be on explaining the reasoning process. Here, the MAP explanation itself needs to be explained or motivated in terms of an explication of the reasoning process that needs further justification.
| i |
a45fa3c4-ac4d-42c5-9b50-8aa0d534b489 | In this paper our aim is to improve the user's understanding of a specific MAP explanation by explicating the relevant information that contributed to said MAP explanation. That is, we seek to justify the MAP explanation by zooming in on the reasoning process. In some way, in the computation of the most probable joint value assignment to the hypothesis variables given the evidence, the process of marginalizing out the intermediate variables makes the decision more opaque. Some of these variables may have a bigger impact (i.e., are more influential) on the eventual decision than others, and this information is lost in the process. For example, the absence of a specific test result (i.e., a variable we marginalize out in the MAP computation) may lead to a different explanation of the available evidence compared to when a negative (or positive) test result were present. In this situation, this variable is more influential to the eventual explanation, and hence more relevant for the justification, than if the best explanation would be the same, irrespective of whether the test result was positive, negative, or missing. Our approach in this paper is to justify an explanation by showing which of these variables were influential towards arriving at this explanation. An intermediate variable is relevant in this justification if it has potential influence on the explanation.
| i |
ec216d68-2c39-490b-acc7-897bc69b8582 | This approach complements other approaches that focus on relevance of variables for justifying decisions. For MAP (i.e., where a Bayesian network is partitioned into hypothesis, evidence, and intermediate nodes) several authors focus on minimizing the set of hypothesis variables, pruning those variables that are deemed less relevant, thereby avoiding over-specification [1]}, [2]}; see [3]} for an alternative proposal where over-specification is avoided by allowing the best explanation to be a set of joint value assignments rather than a singleton set. Others propose an assessment of the relevance of the evidence variables (rather than intermediate variables) for justifying an explanation [4]}, [5]}, [6]}. Similarly, in Bayesian classifiers (without intermediate nodes) justification is often sought in terms of the impact an observable variable may have on the classification [7]}, [8]}, [9]}. In contrast to these approaches, we assume that the partition between hypothesis variables, evidence variables, and intermediate variables is a given, and we focus on assessing the contribution of the intermediate variables towards a specific explanation.
| i |
51131dc1-e878-4194-8b51-56607646a12b | Our perspective has roots in Pearl's early work on conditional independence [1]}. Pearl suggests that human reasoning is in principle based on conditional independence: The organizational structure of human memory is such that it allows for easily retrieving context-dependent relevant information. For example (from [1]}): The color of my friend's car is normally not related to the color of my neighbour's car. However, when my friend tells me she almost mistook my neighbour's car for her own, this information suddenly becomes relevant for me to understand, and for her to explain, this mistake. That is, the color of both cars is independent but becomes conditionally dependent on the evidenceGraphically one can see this as a so-called common-effect structure, where \(C_1\) and \(C_2\) are variables that represent my car's, respectively my neighbour's car's, color; both variables have a directed edge towards the variable \(M\) that indicates whether my friend misidentified the cars or not. When \(M\) is unobserved, \(C_1\) and \(C_2\) are independent, but they become conditionally dependent on observation of \(M\) ..
| i |
1c2d62d3-bf9c-4bfe-a9ec-14f8a2f492cf | In this paper we will argue that Pearl's proposal to model context-dependent (ir)relevance as conditional (in)dependence is in fact too strict. It generally leads to too many variables that are considered to be relevant: for some it is likely the case that, while they may not be conditionally independent on the hypothesized explanation given the evidence, they do not contribute to understanding why some explanation \(h\) is better than the alternatives. That means, for justification purposes their role is limited. In the remainder of this paper we will build on Pearl's work, yet provide a stronger notion of context-dependent relevance and irrelevance of variables relative to explanations of observations. Our goal is to advance explainable AI in the context of Bayesian networks by formalizing the problem of justification of an explanation (i.e., given an AI-generated explanation, advance the user's understanding why this explanation is preferred over others) into a computational problem that captures some aspects of this justification; in particular, by opening up the `marginalization black box' and show which variables contributed to this decision. We show that this problem is intractable in the general case, but also give fixed-parameter tractability results that show what constraints are needed to render it tractable.
| i |
fbfb830d-5e51-4d9f-a451-e09ec3afca21 | To summarize, we are interested in the potential applicability of this new concept for motivation and justification of MAP explanations, with a focus on its theoretical properties. The remainder of this paper is structured as follows. In the next section we offer some preliminary background on Bayesian networks and computational complexity and share our conventions with respect to notation with the reader. In section we introduce so-called MAP-independence as an alternative to conditional independence, embed this concept in the literature, and elaborate on the potential of these computational problems for justifying explanations in Bayesian networks. In section we introduce several formal computational problems based on this notion, and give complexity proofs and fixed-parameter tractability results for these problems. We conclude in section .
| i |
212d9ba0-f5ca-44a6-aa4f-21507633cc32 | In this paper we introduced MAP-independence as a formal notion, relevant for decision support and justification of decisions. In a sense, MAP-independence is a relaxation of conditional independence, suggested by Pearl [1]} to be a scaffold for human context-dependent reasoning. We suggest that MAP-independence may be a useful notion to further explicate the variables that are relevant for the establishment of a particular MAP explanation. Establishing whether the MAP explanation is MAP-independent from a set of variables given the evidence (and so, whether these variables are relevant for justifying the MAP explanation) is a computationally intractable problem; we expect that the problem of finding the maximal set for which the MAP explanation is MAP-independent is even harder, modulo commonly assumed separation properties of the counting hierarchy. However, for a specific variable of interest \(I\) (or a small set of these variables together) the problem is tractable whenever MAP can be computed tractably; in practice, this may suffice for usability in typical decision support systems.
| d |
5d7333b7-d052-4978-99ed-31b3eaad48fd | There are many related problems of interest that one can identify, but which will be delegated to future work. For example, if the set of relevant variables is large, one might be interested in deciding whether observing one variable can bring down this set (by more than one, obviously). Initial work in this area has been proposed by [1]} who looked at several heuristic approaches (e.g., Gini-index [2]} and linear-value utility [3]}) to decide which variable may best be observed.
| d |
cc3ccdc5-69d3-4e18-b9ee-eed25ff63214 | Another related problem would be to decide upon the observations that are relevant for the MAP explanation (i.e., had we not observed \(E \in \mathbf {E}\) or had we observed a different value, would that change the MAP explanation?). This would extend previous work [1]} where the relevance of \(E\) for computing a posterior probability (conditioned on \(\mathbf {E}\) ) was established.
| d |
6d4970c0-a0d6-4f55-8bde-4ab4f06a22b8 | The notion of MAP-independence can, for practical purposes, be to strict; furthermore, when an explanation is MAP-dependent on a set of variables, the `amount of impact' can be relevant. Various authors have proposed a form of quantification of MAP-independence, for example, based on the summed probability of joint value assignments to \(\mathbf {R}\) for which the MAP explanation does not change [1]}, the proportion of such joint value assignments (ignoring their probability)
[2]}, or the average structural impact (in terms of Hamming distance) on the best explanation [3]}. A thorough overview of these and possibly other ways of quantifying MAP-independence would be welcome.
| d |
69f13adf-50d0-4a5b-a3f9-2542e884e09f | Finally, in order to test its practical usage, the formal concept introduced in this paper should be put to the empirical test in an actual decision support system to establish whether the justifications supported by the notion of MAP-independence actually help understand and appreciate the system's advise.
| d |
8d3d35ee-e880-4e73-80b9-3f1f1f26874e | This paper's algorithm, named Topic Scaling, is a two-stage learning process for time-based topics from document positions, which serve as labels for supervised topic model. Hence, a sequential scale is built with Wordfish model, and used as a label to run a supervised Latent Dirichlet Allocation that renders topic-allocation over time for historical corpora. This strategy combines two different machine learning approaches to give an augmented, automated content analysis method where documents and topics are put on a unique scale.
| m |
32431bd4-6d86-493f-a6f8-e39d58d5e2f9 |
Estimate document positions \(\hat{\psi }\) using Wordfish
Assuming: \(w_{ik} = \textrm {Poisson}(\lambda _{ik})\)
Learn via Expected Maximization (\(\alpha ,\nu ,\beta ,\psi \) ) from:
\(log(\lambda _{ik}) =\alpha _{i}+\nu _{k}+\beta _{k} \times \psi _{i}\)
Learn a Supervised LDA with an L2 regularization (shrinkage parameter \(\lambda \) )
Draw topic proportions \(\theta | \alpha \sim Dir(\alpha )\)
For each word:
Draw a topic assignment: \(z_{n}|\theta \sim Mult(\theta ) \)
Draw word \(w_{n}|z_{n},\beta _{1:K} \sim Mult(\beta _{z_{n}}) \)
Draw document scale \(\hat{\psi }|z_{1:N},\eta ,\sigma ^{2} \sim \mathcal {N}(\eta ^\intercal \bar{z},\sigma ^{2})\) with \(z \sim \mathcal {N}(0,\frac{1}{\lambda })\)
| m |
59c32023-2bba-42f9-b30d-5fc1006acccc | We present a novelty in text mining, suited to study sequential corpora and outperforming other topic models in terms of interpretation and parametrization. Topic Scaling could be seen as a dual algorithm: a supervised scaling method where topics are scaled on the same ideological dimension of documents, and a robust alternative to other sequential topic models where the estimated document scores serve as an ordered variable to retrieve topics rather than a learning process requiring a time frame. Under regularization schemes and entropy-related metrics, increasing the number of topics helps maximizing the information gain and uncovering nested structures that render information about potential embedded subtopics, thus, unveiling topics that signal important changes in the evolution of the corpus. Applied to study the party duality democratic-republican in the State Of The Union addresses, this method confirms the existence of two distinct periods correlated with the prevailing topics throughout the modern history of the Unites States, with a clear dominance of foreign affairs and business discourse in post-war addresses, while recent addresses seem to prioritize security and the economic status. Results confirm previous findings about the rhetoric as for vocabulary-rich, varying modern speeches and party-related subject preferences expressed in earlier addresses.
| d |
1e230c90-66dd-4aa2-82a6-04bd5576a8ac | In our current era of connected devices, data is often generated in a continuous manner, e.g. by mobile devices, intelligent sensors or surveillance equipment.
This creates a challenge for machine learning systems, especially when the available computational and memory resources are limited, such as when learning at the edge.
On the one hand, systems cannot store all prior data as the necessary amount of storage would grow linearly over the runtime of the system. On the other hand, the systems have to make predictions continuously, interleaved with learning, rather than following the traditional offline setting with separate training and prediction phases.
Furthermore, continuously generated data is typically non-stationary: its data distribution varies over time, e.g. due to changing environmental conditions or changing user behavior.
| i |
a5869d63-3e0d-4676-bfc6-b8b4a893ed95 | For a machine learning system, a time-varying data distribution is problematic, because it means that previously observed data might not be representative of future data.
Consequently, simply learning a single fixed model, as is usually done in the standard setting of independent and identically distributed (i.i.d.) data, leads to suboptimal results.
Alternatively, it is possible to learn multiple models, essentially starting a new one whenever the data distribution changes.
This, however, can also lead to suboptimal results when each model achieves acceptable prediction quality only after having seen enough data. Furthermore, learning multiple models leads to increased memory requirements.
| i |
e983b040-a092-439f-af6f-c286f6a09d9d | In this work, we introduce LIMES (short for Lightweight Model
Extrapolation for Streaming data)
that offers a new way for learning classifiers from
non-stationary streaming data in resource-constrained
settings.
Inspired by recent progress in meta-learning,
LIMES's core idea is to combine the resource efficiency of
having only a single set of parameters with the flexibility
of learning different models for different data distributions.
We construct a base model that can be adapted efficiently to different target distributions together with a procedure for forecasting the parameters necessary for adapting the base model to be used on future data, even before any examples of that data have been observed.
| i |
cbfe4aab-5907-4023-b0ff-804c60d8d7e4 | Specifically, we adopt the scenario of multi-class classification
under class-prior shift, in which the class frequencies
change over time, while the class-conditional data densities
remain approximately constant.
One of our contributions is to show how in this setting it is possible to learn a model
from multiple data distributions, and how make use of it for yet another data distribution, with hardly any computational or memory overhead.
Furthermore, we introduce an extrapolation step that forecasts the class distribution of future data based on the observed previous data.
| i |
7ebf4dfb-604a-40e6-bca9-baa3907b4e4d | We show experimentally on exemplary tasks using Twitter data that
LIMES achieves higher accuracy than either learning a single model or an ensemble of multiple models. The code is available at https://github.com/ptomaszewska/LIMES.
| i |
bfa3316e-e8c2-4ab4-909f-ae9d587f3d68 | Overall, one can see that the task of predicting
the country from the tweet text is more difficult
than from the user-provided meta-data. Combining
both feature types yields the highest accuracy.
In all cases, the minimal accuracy across the day
is substantially lower than the average. This
indicates that the difficulty of the classification
task varies over time, presumably due to changes
in the label distribution.
| r |
0c4e83d8-0688-4f3d-8bde-9068e72991b5 | Comparing the different methods, we observe a
ranking that is stable across all experimental
settings:
LIMES achieves the best results, followed by
incremental and random. Ensemble performs substantially worse, and restart even worse than that.
Given that incremental differs from LIMES only in
the lack of an adaptation and forecasting step,
we can conclude that the way we propose the adaptation
to the class distribution indeed has a positive effect.
For the avg-of-avg accuracy measure, the gain
due to adaptation and forecasting is the biggest
for the most difficult task: \(0.53\) –\(0.65\) % when
predicting the country based on the tweet text
itself.
It gets smaller, the easier the task becomes:
\(0.13\) –\(0.19\) % for location features, and
\(0.07\) –\(0.12\) % for the concatenation of both.
More apparent, however, is the positive effect
on the avg-of-min accuracy: for tweet
features, it is \(2.64\) –\(3.00\) %, for location
\(0.71\) –\(0.83\) % and for their concatenation
\(0.53\) –\(0.60\) %.
<FIGURE> | r |
d15a6b35-cb63-4e55-a6ab-3667e34b0b59 | Figure REF provides an illustrative
explanation of these results. Subfigure (a) shows
the four methods' accuracy curves over the first 120
hours of one exemplary learning run (tweet
features, subset early, realization 0).
First, one can clearly see the 24-hour cycle by
which the accuracies of all classifiers fluctuate.
The accuracies are highest at around 2:00 GMT. As we
know from Figure REF , this corresponds
to the time the USA and Brazil are by far the most
common countries from which tweets originate.
The accuracy is the lowest around 9:00 GMT, which is the
time when tweets come from several different countries
in comparable amounts.
| r |
699d318f-17f8-4b27-bad7-62c2c01ecf8d | Second, one can also see that ensemble, which does not
share information between different hours, starts working
only after 24 hours and afterwards increases
its prediction quality quite slowly. In contrast, the other models make reasonable predictions already after the first hour,
because they transfer information between different hours.
| r |
3c6be08f-b18e-46b6-996e-bcc0c4c2d7bc | Finally, one observes that LIMES is superior over incremental mostly in the valleys around 9:00 GMT, i.e. during the hours where the classification problem is the hardest.
Presumably, this is because incremental learns parameters that are adapted to the average data distribution, and that is closer to the simple situation with two dominant countries than to the difficult situations where many countries are similarly likely.
| r |
e9f71f02-236c-43fd-9155-02dac90f64ea | Subfigures (b) and (c) make the difference between the methods more visible by suppressing the daily fluctuations.
They show the accuracy values for all 20 days aggregated
as the per-day-average and per-day-minimum, respectively. The line plots show mean of results over all ten realizations, whereas the error bars reflect the standard deviation. As expected, standard deviation is the biggest in the case of random method as for each time step, it draws a class distribution vector from historical values. This effect of randomness is alleviated when the average per-day is computed.
| r |
a046a001-baa0-499b-a892-98bc842125bf | Incremental training is clearly effective, as can be seen
from the fact that the accuracy curves for the three
incrementally trained methods, LIMES, incremental and
ensemble, increase over time.
In contrast, restart's accuracy is approximately constant
(up to the hourly fluctuations) over the runtime of the
system.
| r |
e4e399b2-33e8-4ad5-9aea-360b731f9778 | In this work, we introduced LIMES, a lightweight method
for continuous classifier training from streaming data with class-prior shift.
Inspired by the success of recent meta-learning methods,
we introduced a system that trains multiple models, each
adapted to a specific class distribution, but requires
only a single set of parameters, as each model is
derived from a base model by analytic classifier
adaptation.
The adaptation parameters are extrapolated for the next data distribution based on the data observed so far.
Our experiments on the large-scale geo-tweets
dataset show that this process results in improved
prediction quality compared to common baselines, especially when judged by the worst-case measure of the lowest accuracy across a day.
| d |
5fd95203-ab38-487f-95ef-bc6917bb91a6 | Despite the promising results, the problem of continual
learning from data streams with time-varying distribution
is far from solved. In particular, our results suggest
that new methods are required to tackle the problem
that classifiers for certain data distributions are
harder to learn than others. This suggests that one
might have to adapt not only the model parameters,
but also the model architecture.
We plan to address this aspect in future work.
| d |
41824476-e683-40d2-83ea-d3c2e0564e1c | Extremum seeking (ES) is a popular technique for adaptive optimization of the performance of dynamic systems by tuning certain system parameters based on measurements. The main advantage of this technique is that limited or no knowledge of the plant model is required. ES is suitable for optimization of the performance of systems with complex dynamics, unavailable suitable measurements to validate the model, and time-varying disturbances that are difficult to model accurately ([1]}).
| i |
c49755b5-f320-4ebb-bf75-261ec3aa90b7 | The most common ES algorithm used in the literature is the classical band-pass filtering based one, in which the gradient of the output with respect to the input will determine the direction of adjusting the input variables. This method was successfully applied to different application areas including biochemical reactors [[1]}, [2]}], ABS control in automotive brakes ([3]}, [4]}, [5]}, [6]}, [7]}), mobile robots ([8]}, [9]}, [10]}), mobile sensor networks ([11]}, [12]}, [13]}).
| i |
9477803f-f501-40d2-b167-85ec2f44da64 | Among other types of ES algorithms, perturbation based ES relies on added perturbation signals to estimate the gradient of the output by correlating the perturbations. To overcome the implementation drawbacks of introducing perturbation signals, some methods that are free of perturbation signals have been developed by [1]}, [2]}, [3]}.
| i |
f20d7af2-37e8-438a-8b0e-61e17fe4996d | Convergence rate of conventional ES algorithms is a limiting factor in many applications. Recursive Least Squares (RLS) based estimation has significant potential in relaxing this limitation and improving robustness to measurement noises. [1]}, [2]}, [3]} used certain LS based techniques in their ES algorithms to obtain better convergence results. [1]} estimated the gradient of the
output with respect to the input using a LS based adaptive law for a class of nonlinear dynamic systems together with a sinusoidal perturbation signal. [2]} used past data of a performance map to estimate the gradient of this performance map by a first order LS fit. The proposed method used no dither signal, but utilized a time window of history data of the performance map. [3]} provided general results and a framework for the design of ES schemes applied to systems with parametric uncertainties and used LS algorithm to estimate unknown parameters of the known system.
| i |
9f37aab6-14ca-493b-88ee-15d43ba8e43e | In absence of the parameter knowledge, a series of control/optimization schemes have been proposed in the literature utilizing certain ES tools such as switching methods ([1]}), signal perturbation for persistence excitation, and band pass filtering ([1]},[3]},[4]},[5]}. [6]} and [7]} used a discrete time ES scheme to estimate the gradient as a time-varying parameter using LS like update laws. They removed the need for averaging system in order to achieve the convergence of ES. The designs are simulated for static unknown maps, systems with unknown discrete-time dynamics and sampled-data systems.
| i |
215e485d-187b-4123-940f-da9cd4694aac | In this paper, a continuous time RLS parameter estimation based ES scheme is designed and analysed for scalar parameter and vector parameter static map and dynamic systems. Asymptotic convergence to the extremum is established for each case. Numerical simulation examples are provided to validate the performance of proposed scheme comparing the results with gradient parameter estimation based one. A specific simulation example, antilock braking systems (ABS), in [1]} is studied to compare the performance of RLS estimation based ES with classical gradient based ES.
| i |
b58c0942-91eb-4271-9295-f6638acc3db3 | Contents of this paper are as follows. Section II is dedicated to the problem statement. In Section III, existing classical perturbation based ES is reviewed. Proposed RLS estimation based adaptive ES is developed for scalar parameter systems in Section IV, and for vector parameter systems in Section V. Comparative simulation examples are presented in Section VI. Finally, conclusions of the paper are given in Section VII.
| i |
2b4ba872-d74a-491d-af68-207c9a0ced10 | In this section, we present simulation results to show the validity of the proposed schemes. We will present two examples for scalar parameter and vector parameter cases with their comparison results with classical ES method in Section .
| m |
7d5869c1-ed15-4d16-af9d-8c9fafb8acac | This paper focuses on designing an RLS parameter estimation based ES scheme for scalar parameter and vector parameter static map and dynamic systems. Their stability conditions are stated for each case. The proposed ES scheme does not need perturbation signals for scalar parameter systems; however, the proposed ES scheme needs perturbation signals with different frequencies for vector parameter systems. Proposed scheme is applied to different simulation scenarios and compared to classical gradient estimation based ES under measurement noise. The results show the validity and effectiveness of RLS parameter estimation based ES scheme over gradient one.
| d |
bd70b624-2fde-4ac7-ad76-1902eebef2d2 | This work aims to study the effectiveness of a multilingual end-to-end (E2E) automatic speech recognition (ASR) system applied to three languages used in Kazakhstan: Kazakh, Russian, and English.
Kazakhstan is a multinational country where Kazakh is the official state language, whereas Russian and English are the languages of interethnic and international communication commonly used in business, science, and education. These three languages are part of a large-scale cultural project initiated by the government named “The Trinity of Languages” [1]}.
The goal of the project is the mastery of the aforementioned languages by the Kazakhstani people.
This will presumably enable citizens' successful integration into the international economic and scientific environments.
In this regard, we initiate the first study of a single joint E2E ASR model applied to simultaneously recognize the Kazakh, Russian, and English languages.
| i |
02ee2225-e8f0-401d-907f-4fb5b27c2a82 | Having a single ASR model for multiple languages considerably simplifies training, deployment, and maintenance [1]}.
In particular, this is advantageous for multilingual communities where several languages are used for communication.
A multilingual ASR system can automatically detect an input language and produce corresponding transcripts without prompting for language or requiring visual and tactile interfaces.
This becomes especially useful when ASR is employed in a pipeline of a larger system, such as message dictation, voice command recognition, virtual assistants, a transcription engine on online audio/video sharing platforms (e.g., YouTube), and so on.
| i |
ac2e1a5e-25b6-4abc-9bea-99b72fbfbfe9 | Recently presented E2E ASR architectures have been shown to be effective for the multilingual speech recognition task [1]}, [2]}, [3]}.
It has also been demonstrated that the E2E approaches achieve comparable results to the conventional deep neural network-hidden Markov model (DNN-HMM) ASR [4]}, [5]}.
Moreover, they significantly reduce the burden of developing ASR systems thanks to the encapsulation of the acoustic, pronunciation, and language models under a single network.
Importantly, the E2E ASR models obviate the need for a pronunciation model, which requires significant human effort to be constructed and often proves critical to overall performance [4]}. All these features make the E2E architectures attractive for the multilingual speech recognition task.
| i |
c7086a92-55e8-4a43-a387-858acd198cc8 | This work leverages the above-mentioned advantages and studies the multilingual E2E ASR systems applied to simultaneously recognize the Kazakh, Russian, and English languages.
Specifically, we thoroughly explore the performance of the Transformer-based E2E architecture [1]}.
To the best of our knowledge, this is the first study of multilingual E2E ASR dedicated to these languages.
We also compared the use of two different grapheme set construction methods (i.e., combined and independent).
We also analyzed the impact of language models and data augmentation techniques, such as speed perturbation [2]} and spectral augmentation [3]}.
We found that the multilingual models can achieve comparable results to strong monolingual baselines, despite having a similar number of parameters.
To enable experiment reproducibility and facilitate future research, we share our training recipes, datasets, and pre-trained models REF.
| i |
4f84bff6-3b8a-41b9-8bac-f55b7daf0fbf |
We introduce a 7-hour evaluation set of transcribed Kazakh-accented English audio recordings (i.e., native Kazakh speakers reading English sentences extracted from the SpeakingFaces dataset [1]}).
We introduce a 334-hour manually-cleaned subset of the OpenSTT dataset [2]} for the Russian language, which can also be used to train robust standalone Russian ASR systems.
| i |
e593414e-e93e-4a8f-b65b-c34c1f699069 | The rest of the paper is organized as follows:
Section briefly reviews related works on multilingual ASR.
Sections and describe the multilingual models and datasets used in our experiments, respectively.
Section presents the experimental setup and obtained results.
Section discusses the important findings and highlights potential future work.
Finally, Section concludes this paper.
| i |
4685d033-da60-4b1a-9013-51f04ec70ffc | A single model capable of recognizing multiple languages has been a long-term goal of the speech recognition community and remains an active area of research for decades [1]}, [2]}, [3]}.
The use of a single model for several languages simplifies the ASR production pipeline significantly, since maintaining one model per language becomes cumbersome as the number of languages increases.
Furthermore, multilingual ASR systems leverage cross-lingual knowledge transfer, which has been shown to improve recognition performance, especially for low-resource languages [4]}.
| w |
fcf360dd-96af-438d-9818-b02b93cd7b83 | Prior works on multilingual ASR have explored both hybrid DNN-HMM [1]} and E2E [2]} architectures.
Both small- and large-capacity multilingual models with up to ten billion parameters have been studied [3]}.
Offline and streaming speech recognition modes of multilingual ASR have also been investigated [4]}.
The authors of [5]} developed multilingual models capable of recognizing over 100 languages simultaneously. The prior works have also studied different approaches to further improve the multilingual speech recognition performance, such as multi-task [6]} and transfer [7]} learning.
In multi-task learning, a model is jointly trained with other tasks, such as language identification (LID), whereas, in transfer learning, a model pre-trained on other languages (usually high-resource ones) is fully or partially fine-tuned using the target languages.
However, to the best of our knowledge, there is no prior work dedicated to simultaneous recognition of the Kazakh, Russian, and English languages.
| w |
c33bc9ca-d458-4262-a446-c96ff551e7fa | Among the aforementioned three languages, Russian and English are considered resource-rich, i.e., a large number of annotated datasets exist [1]}, [2]}, [3]} and extensive studies have been conducted, both in monolingual and multilingual settings [4]}, [5]}, [6]}.
On the other hand, Kazakh is considered a low-resource language, where annotated datasets and speech processing research have emerged only in recent years [7]}, [8]}.
The authors of [7]} presented the first crowdsourced open-source Kazakh speech corpus and conducted initial Kazakh speech recognition experiments on both DNN-HMM and E2E architectures.
Similarly, the authors of [8]} presented the first publicly available speech synthesis dataset for Kazakh.
Previously, the Kazakh language was part of several multilingual studies under the IARPA's Babel project [11]}, [12]}, [13]}, and it was also explored in the context of Kazakh-Russian [14]}, [15]} and Kazakh-English [16]} code-switching.
| w |
bca1963a-ac28-49ee-b656-dbf8e97b832e | This work is the first to study the multilingual E2E ASR systems dedicated to Kazakh, Russian, and English, which we believe will further progress the speech processing research and advance the speech-enabled technology in Kazakhstan and its neighboring countries. These languages belong to different language families (i.e., Kazakh belongs to Turkic, Russian to Slavic, and English to Germanic), which poses an additional challenge to our work.
Therefore, we posit that our work will be of interest to the general speech research community, especially for researchers from the post-Soviet states, where Russian and English are also commonly spoken as lingua francas.
| w |
989ba089-54d2-43e1-9ced-63806296f1d2 | All E2E ASR systems were trained on the training sets, using the V100 GPUs running on an Nvidia DGX-2 server; hyper-parameters were tuned on the validation sets, and the final systems were evaluated on the test sets (see Table REF ).
For all systems, the input acoustic features were represented as 80-dimensional log Mel filter bank features with pitch computed every 10 ms over a 25 ms window, and the output units were represented using the character-level graphemes.
| m |
3ca9ef56-b699-49d2-966f-fd4e01182cc6 | To train the E2E ASR systems, we used the ESPnet toolkit [1]} and followed the Wall Street Journal (WSJ) recipe.
The E2E architecture was based on the Transformer network [2]} consisting of 12 encoder and 6 decoder blocks. It was jointly trained with the Connectionist Temporal Classification (CTC) [3]} objective function under the multi-task learning framework [4]}.
The interpolation weight for the CTC objective was set to 0.3 and 0.4 during the training and decoding stages, respectively.
For the Transformer module, we set the number of heads in the self-attention layer to 8 each with 512-dimension hidden states, and the feed-forward network dimensions to 2,048.
In addition, a VGG-like convolution module [5]} was used to pre-process the input audio features before the encoder blocks.
All models were trained for 120 epochs using the Noam optimizer [2]} with the initial learning rate of 10 and 25k warm-up steps.
We set the dropout rate and label smoothing to 0.1.
For data augmentation, we used a standard 3-way speed perturbation [7]} with factors of 0.9, 1.0, and 1.1, and the spectral augmentation [8]}.
We report results on an average model constructed using the last ten checkpoints.
| m |
dbcfb446-442c-4ca7-aa84-fee03e0a8ff3 | To evaluate the impact of language models (LM) on recognition performance, we built character-level LMs using the transcripts of the training sets.
The LMs were built as a 2-layer long short-term memory (LSTM) [1]} network with a memory cell size of 650 each.
We built both monolingual and multilingual LSTM LMs for monolingual and multilingual E2E ASRs, respectively.
The multilingual LSTM LM was trained on the combined training set.
The LSTM LMs were employed during the decoding stage using shallow fusion [2]}.
For decoding, we set the beam size to 60 and the LSTM LM interpolation weight to 0.6.
The other hyper-parameter values can be found in our GitHub repository REF.
<TABLE> | m |
c05e1d66-2aef-4641-9dcd-77527ea9b6c4 | In this paper, we explored multilingual E2E ASR applied to simultaneously recognize three languages used in Kazakhstan: Kazakh, Russian, and English.
Specifically, we developed both monolingual and multilingual E2E ASR models based on the Transformer networks and compared their performances in terms of WER.
To the best of our knowledge, this is the first multilingual E2E ASR work dedicated to these languages.
In addition, we compared the use of two different grapheme set construction methods (i.e., combined and independent).
We also evaluated the impact of language models and data augmentation techniques on the WER performances of the monolingual and multilingual models and found them extremely effective.
Additionally, we introduced two manually-transcribed datasets: OpenSTT-CS334 and test-SF.
The first one is a manually cleaned 334-hour subset extracted from the OpenSTT dataset.
The second one is a 7-hour set of Kazakh-accented English utterances designed to be used for evaluation purposes.
Given that acquiring high-quality speech data is prohibitively expensive, these datasets will be of great use for the speech community both in academia and industry.
Our experiment results show that the multilingual models achieve comparable results to the monolingual models, while having a similar number of parameters.
The best monolingual and multilingual models achieved average WERs of 20.9% and 20.5% on the test sets, respectively.
We strongly believe that the conducted experiments and reported findings will benefit researchers planning to build multilingual E2E ASR systems for similar languages, especially from the post-Soviet space.
We also hope our work will encourage future research that leverages the findings and datasets presented in this paper.
| d |
aae2013c-3d0b-4c1f-b997-48ad5233c63c | Fine-tuning pre-trained large-scale language models (LMs) is the dominant paradigm of current NLP. The LMs proved to be a versatile technology that can help to solve an array of NLP tasks, such as parsing, machine translation, text summarization, sentiment analysis, semantic similarity etc. The LMs can be used for tasks on various levels of linguistic complexity (syntactic, semantic, etc.) but also with various types of data modalities (text classification, text generation, text comparison, etc.). As such, it seems to be vital for any speech community to develop a proper model for their language and thus push the limits of what NLP can do for them.
| i |
0b9744a4-22b6-4a4e-932f-4f610ce6ee9f | In this paper, we introduce the first Slovak-only transformers-based language model trained with a non-trivial corpus called SlovakBERTAvailable at https://github.com/gerulata/slovakbert. Although several multilingual models already support Slovak, we believe that developing a Slovak-only model is still important, as it can lead to better results and more compute-wise and memory-wise efficient processing of Slovak language. SlovakBERT has RoBERTa architecture [1]} and it is trained on a Web-crawled corpus.
| i |
b4264d9d-ee97-4ca3-a341-8b6ffa517184 | We evaluate SlovakBERT on various existing datasets to study how well it handles different tasks. We also compare it to other available LMs (mainly multilingual) and other existing approaches. The tasks we study are: part-of-speech tagging, semantic textual similarity, sentiment analysis and document classification. As a by-product of our experimentation, we have developed and published the best performing models for selected tasks. These might be used by other researchers or Slovak NLP practitioners in the future as strong baseline.
| i |
eefc0315-3742-4062-9fca-9e0d98841592 |
We trained and published the first proper Slovak-only LM on a dataset of non-trivial size that we collected.
We evaluated the LM on a series of Slovak NLP tasks.
We published several fine-tuned models based on this LM, namely a part-of-speech tagger, sentiment analysis model and sentence embedding model.
We published several additional datasets for multiple tasks, namely sentiment analysis test sets and semantic similarity translated dataset.
| i |
22a7c5d1-935d-4911-8003-6f97b505f53e | The rest of this paper is structured as follows: In Section we discuss related work about language models and their language mutations. In Section we describe the corpus crawling efforts and how we train our LM with it. In Section we evaluate the model with four NLP tasks. In Section we conclude our findings.
| i |
c2df611b-b930-45f9-bd7d-a9f468a2cd1e |
Downstream performance. We fine-tune the LMs for various NLP tasks and we analyze the achieved results. We compare the results with existing solutions based on other approaches, e.g. rule-based solutions or solutions based on word embeddings.
Probing. Probing is a technique that aims to measure how useful the information contained within individual layers of the LM is. We can check how important the individual layers for various tasks are. We use simple linear probes in our work, i.e. the hidden representations from the LMs are used as features for linear classification.
| m |
2a5d82a0-cfb3-4904-8fc3-05a56cddaf35 | We conducted the evaluation on four different tasks: part-of-speech tagging, semantic textual similarity, sentiment analysis and document classification. For each task, we introduce the dataset that is used, various baselines solutions, the LM-based approach we took and the final results for the task.
| m |
11a2053d-79a9-40c2-95b7-ed22663e8bfd | We have performed a random hyperparameter search with SlovakBERT. The range of individual hyperparameters is shown in Table REF . We have found out that weight decay is a beneficial regularization technique, while label smoothing proved itself to be inappropriate for our case. Other hyperparameters showed to have very little reliable effect, apart from the learning rate, which proved to be very sensitive. We have not repeated this tuning for other LMs, instead, we only tuned the learning rate. We have found out that it is appropriate to use learning rate of \(10^{-5}\) for all the models, but XLM-R-Large. XLM-R-Large, the biggest model we tested, needs a smaller learning rate of \(10^{-6}\) .
<TABLE> | r |
7a1e7502-0b27-4fc7-974c-c58063faddf2 | The results for POS tagging are shown in Table REF . We report accuracy for both XPOS and UPOS tagsets. WikiBERT seems to be the worst-performing LM, probably because of its small training set. SlovakBERT seems to be on par with larger XLM-R-Large. Other models lag behind slightly. From existing solutions, only transformers-based Trankit seems to be able to keep up.
<TABLE> | r |
3bf58b95-c17d-4730-820c-43b8c1c05bd3 | To make sure that the results are statistically significant, we have performed McNemar test among all the models. We have used the information about whether each token was predicted correctly or incorrectly as the observed variables. The results are shown in Table REF . We can see that the differences between the top-performing models are mostly not considered statistically significant.
<TABLE> | r |
79e91aa6-d07a-4156-8200-a39968017f4e | We also analyzed the dynamics of the LM fine-tuning. We analyzed the performance for various checkpoints of our LM (checkpoints were made after 1000 training steps). We can see in Figure REF , that SlovakBERT was saturated w.r.t POS performance quite soon, after approximately 15k steps. We stopped the analysis after the first 125k steps, since the results seemed to be stable. Similar results for probing can be seen in the same figure. We show the performance for all the layers for selected checkpoints. Again, we can see a rapid saturation of the model. Note that the performance on layers peaks quite soon at layer 6 and then plateaus. The last layers even have degraded performance. This shows, that the morphosyntactic information needed for POS tagging is stored and processed mainly in the middle part of the model. This is in accord with the current knowledge about how LMs work, i.e. that they process the text in a bottom-up manner [1]}.
<FIGURE> | r |
3f5c26f6-0664-4ea0-b839-4a817fcec967 | We compare the systems using Spearman correlation between the cosine similarity of the generated sentence representations and the ground truth data. The original STS datasets are using \(\langle 0,5 \rangle \) scale. We normalize these scores to \(\langle 0,1 \rangle \) range so that they can be directly compared to the cosine similarities. We performed a hyperparameter search in this case as well. Again, we have found out that the results are quite stable across various hyperparameter values, with learning rate being the most sensitive hyperparameter. The details of the hyperparameter tuning are shown in Table REF . We show the main results in Table REF .
<TABLE> | r |
d9d55215-ba0b-414f-8b2e-4f543c30ccfc | We can see that the results are fairly similar to POS tagging w.r.t. how the LMs are relatively ordered. The existing solutions are worse, except for XLM-R trained with English data, which is actually the best performing model in our experiments. It seems that their model fine-tuned with real data without machine-translation-induced noise works better, even if it has to perform the inference cross-lingually on Slovak data.
| r |
2ca0853f-0fb5-44d7-8334-0760aa7d4c05 | We also experimented with Slovak-translated NLI data in a way where the model was first fine-tuned on NLI task and then the final STS fine-tuning was performed. However, we were not able to outperform the purely STS fine-tuning with this approach and the results remained virtually the same. This result is in contrast with the usual case for English training, where the NLI data regularly improve the results [1]}. We theorize that this effect might be caused by noisy machine translation.
| r |
9657b6c4-c1b8-4e91-a706-5585855d02db | Figure REF shows the learning dynamics of STS. On the left, we can see that the performance takes much longer to plateau than in the case of POS. This shows that the model needs longer time to learn about semantics. Still, we can see that the performance ultimately stabilizes just below \(0.8\) score. Similarly, unlike POS, we can see that the best performing layers are actually the last layers of the model. This suggests that a model with bigger capacity might have been trained successfully.
<FIGURE> | r |
a67534e9-0216-4354-ba89-eff1af53df80 | We report macro-F1 scores for all three classes as our main performance measure. The LMs were trained on the Twitter dataset. We calculate average F1 from our Reviews dataset as an additional measure.
| r |
f13d0b2c-9bb5-4737-bdfa-a18c86239817 | Again, we have performed a hyperparameter optimization of SlovakBERT. The results are similar to results from POS tagging and STS. We have found out that learning rate is the most sensitive hyperparameter and that a small amount of weight decay is a beneficial regularization. The main results are shown in Table REF . We can see that we were able to obtain better results than the results that were reported previously. However, the comparison is not perfect, as we use slightly different datasets for the aforementioned reasons.
<TABLE> | r |
e7a187fc-0ba3-4ed7-8d1b-7a7d042ff685 | The LMs are ordered in performance similarly to how they are ordered in the two previous tasks. SlovakBERT seems to be among the best performing models, along with the larger XLM-R-Large. The LMs were also able to successfully transfer their sentiment knowledge to new domains and they achieve up to 0.617 macro-F1 in the reviews as well. However, both Amazon commercial sentiment API and NLP4SK have even better scores, even though their performance on Twitter data was not very impressive. This is probably caused by the underlying training data they use in their systems, that might match our Reviews datasets more than the tweets used for our fine-tuning.
| r |
2022f946-8739-441a-8664-a9365db20fd1 | The main results from our experiment are shown in Table REF . We can see that the LMs are again the best performing approach. In this case, the results are quite similar with SlovakBERT being the best by a narrow margin. The baselines achieved significantly worse results. Note that our sentence embedding model has the worst results on this task, while it had competitive performance in sentiment classification. We theorize, that the sentence embedding model was trained on sentences and is therefore less capable of handling longer texts, typical for the dataset used here.
<TABLE> | r |
97d0eab8-8555-4beb-98ab-fd9f74271b8a | We have trained and published SlovakBERT – the first large-scale transformers-based Slovak masked language model using 19.35GB of Web-crawled Slovak text. We evaluated this model on multiple NLP tasks. We conclude, that SlovakBERT achieves state-of-the-art results on these tasks. We also release the fine-tuned models for the Slovak community. Existing multilingual models can achieve comparable results on some tasks, however they are less efficient memory-wise and/or compute-wise.
| d |
07a303db-8716-487e-8426-604a6f47f478 | Weakly supervised object localization (WSOL) aims to locate objects by using only image-level labels.
Since no expensive bounding box annotations are required, WSOL has attracted lots of attentions in various applications, lesion localization for medical image diagnosis, image-label guided retrieval, etc [1]}, [2]}, [3]}, [4]}, [5]}, [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}.
<FIGURE> | i |
9f9baa3f-8bf7-4797-a207-19587345b6fc | Existing WSOL methods are mainly based on the class activation maps (CAMs) [1]}.
However, CAM-based models are initially trained for classification, where the network objective is inconsistent with localization.
Specifically, classification prefers features with semantic meanings, usually derived from deep layers of convolutional neural networks (CNNs).
In contrast, shallow features, derived from the shallow layers of CNNs, have less semantics but are rich in details, which have clearer edges and less distortion.
Unfortunately, direct fusion of shallow and deep features is invalid for WSOL due to the following two defects:
1) The meaningful information embedded in the shallow features cannot be well captured through weak supervision, due to the large interference of background noise.
To better elaborate this statement, we illustrate the CAM-based WSOL pipeline in Fig. REF (a).
Within this pipeline, features of different layers from ResNet50 [2]} are firstly aggregated to form the fused CAMs.
As the fused CAMs obtained through addition operation shown in Fig. REF (a), the shallow features are buried and not fully utilized, leading to inferior CAMs.
To make it general, as shown in Fig. REF (b), we further test other conventional CAM fusion strategies (concatenation) for different layers, which are evaluated by GT-known Loc on CUB-200 [3]} dataset.
Regarding the quantitative and qualitative results, neither the obtained CAMs nor the overall localization accuracy has been improved when more shallow features are involved through conventional fusion strategies, addition or concatenation.
2) Another issue is that only the most discriminative regions are activated in the original CAMs [4]}, [5]}, [6]}, [7]}, [8]}, [9]}.
As shown in Fig. REF (a) and Fig. REF (b), most areas have low response except for the head region, even though the low response areas occupy most of the image and reflect the object shape.
| i |
10597cb3-d98a-46a5-970c-fe4352cf202e | To address above concerns, we propose a simple but effective Shallow feature-aware Pseudo supervised Object Localization (named SPOL) model for accurate WSOL, which makes the utmost of the low-level features embedded in shallow layers.
Our SPOL model mainly consists of two stages, CAM generation and class-agnostic segmentation.
For the CAM generation, the multiplicative feature fusion network (MFF-Net) is designed to aggregate both shallow and deep features.
Different from previous fusion methods, features in MFF-Net are treated in a synergistic way.
Namely, deep features with clear background help suppress the noise of the shallow ones while shallow features with rich local structures make object boundaries sharper, just as the multiplicative fusion CAMs shown in Fig. REF (a) and performance gains shown in Fig. REF (b).
For the class-agnostic segmentation stage, initial CAMs will be refined with the Gaussian prior pseudo label (GPPL) module, which is then regarded as the pseudo label for class-agnostic segmentation module training.
Specifically, taking full advantage of the initial entire CAM as the weighting coefficients, the mean and variance for all coordinates are calculated to obtain the object gravity.
Then, a Gaussian distribution can be generated with achieved mean and variance, called Gaussian prior pseudo label (GPPL).
As shown in Fig. REF (c), GPPL approximates the gravity center of the bird, and enhances the responses of areas inside the body.
Combining GPPL and original CAM, a better CAM can be obtained, just as Fig. REF (d) shows.
To further refine these regions, we design a class-agnostic segmentation model by using combined GPPL and CAMs as pseudo labels through another MFF-Net.
Note that during the training phase, areas with large and small CAMs responses will be binarized into foreground and background, respectively using two pre-defined thresholds, and other parts will be ignored to avoid label conflict during training.
After training, the obtained object mask will become more complete compared with initial CAMs, as shown in Fig. REF (e).
Finally, a bounding box extractor is applied to the object mask to obtain the final object localization. In summary, our contributions are three-fold:
| i |
898de4a6-9607-4687-a556-bdb0830f3da9 |
We propose a SPOL model to fully utilize the vital shallow features for WSOL, owing to the proposed multiplicative feature fusion strategy that makes the utmost of shallow features.
We further propose the Gaussian prior pseudo label (GPPL) and class-agnostic segmentation model to achieve a better object mask for WSOL.
SPOL outperforms previous methods by a large margin on both CUB-200 and ImageNet-1K benchmarks.
<FIGURE><FIGURE> | i |
b8fefeaf-66cd-4d44-ba4d-488315544055 | In this section, we first analyze the importance of shallow features and then elaborate our proposed SPOL model, which consists of two stages: CAM generation and class-agnostic segmentation, as shown in Fig. REF (a).
<FIGURE> | m |
1e079af9-38c4-46d3-8e14-dcdd3c9baa5e | Datasets. To evaluate the proposed SPOL, two datasets are adopted, including CUB-200 [1]} and ImageNet-1K [2]}. CUB-200 contains 200 categories of birds with 5,994 training images and 5,794 testing images. ImageNet-1K is a much larger dataset with 1000 classes, containing 1,281,197 training images and 50,000 validation images.
| m |
0ca80919-a7bb-4e9b-a545-83cf2f8e45d0 | Metrics. Following previous methods [1]}, [2]}, three metrics are adopted for evaluation. 1) Top-1 localization accuracy (Top-1 Loc): fraction of images with right prediction of class label and more than 50% IoU with the ground-truth box. 2) Top-5 localization accuracy (Top-5 Loc): fraction of images with class labels belonging to Top-5 predictions and more than 50% IoU with the ground-truth box. 3) GT-known localization accuracy (GT-known Loc): fraction of images for which the predicted bounding box has more than 50% IoU with the ground-truth box.
| m |
163942bc-7005-4f96-9b0d-dee6806a633b | In this paper, we propose the Shallow feature-aware Pseudo supervised Object Localization (SPOL) model for accurate WSOL.
We first analyze the importance of shallow features for object detection and then show that conventional fusion ignores the power of shallow features due to background noise interference.
Thus, a multiplicative feature fusion strategy is introduced to utilize shallow features, suppress background noise and enhance object boundaries.
Further, a class-agnostic segmentation model is trained with the pseudo labels to refine object predictions.
Extensive experiments verify the effectiveness of the proposed SPOL, which outperforms previous methods by a large margin.
| d |
11987906-f27c-46b5-8242-4ac97018f2c2 | Survivorship bias [1]} refers to systematic error about our understanding of the world, where we analyze data only based on success cases, omitting consideration (on purpose or not) of the failing cases. One of the most known cases is pictured in Fig. REF credits to By Martin Grandjean (vector), McGeddon (picture), Cameron Moll (concept) - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=102017718 when Abraham Wald attempted to reduce bomber losses due to enemy fire in World War II. During the analysis, he suggested that the bullet marks (hypothetically represented as red dots) showed only cases where the airplanes could land safely, whereas portions without red dots were due to cases when the bombers crashed. Abraham Wald recommended reinforcing portions without red dots, contrary to the original belief that the areas with red dots would need to be reinforced as they showed areas where enemy fire hit the bombers.
| i |