We study an approach to offline reinforcement learning (RL) based on optimally solving finitely-represented MDPs derived from a static dataset of experience. This approach can be applied on top of any learned representation and has the potential to easily support multiple solution objectives as well as zero-shot adjustment to changing environments and goals. Our main contribution is to introduce the Deep Averagers with Costs MDP (DAC-MDP) and to investigate its solutions for offline RL. DAC-MDPs are a non-parametric model that can leverage deep representations and account for limited data by introducing costs for exploiting under-represented parts of the model. In theory, we show conditions that allow for lower-bounding the performance of DAC-MDP solutions. We also investigate the empirical behavior in a number of environments, including those with image-based observations. Overall, the experiments demonstrate that the framework can work in practice and scale to large complex offline RL problems.
There have been significant efforts to interpret the encoder of Transformer-based encoder-decoder architectures for neural machine translation (NMT); meanwhile, the decoder remains largely unexamined despite its critical role. During translation, the decoder must predict output tokens by considering both the source-language text from the encoder and the target-language prefix produced in previous steps. In this work, we study how Transformer-based decoders leverage information from the source and target languages -- developing a universal probe task to assess how information is propagated through each module of each decoder layer. We perform extensive experiments on three major translation datasets (WMT En-De, En-Fr, and En-Zh). Our analysis provides insight on when and where decoders leverage different sources. Based on these insights, we demonstrate that the residual feed-forward module in each Transformer decoder layer can be dropped with minimal loss of performance -- a significant reduction in computation and number of parameters, and consequently a significant boost to both training and inference speed.
Reward function specification can be difficult, even in simple environments. Realistic environments contain millions of states. Rewarding the agent for making a widget may be easy, but penalizing the multitude of possible negative side effects is hard. In toy environments, Attainable Utility Preservation (AUP) avoids side effects by penalizing shifts in the ability to achieve randomly generated goals. We scale this approach to large, randomly generated environments based on Conway's Game of Life. By preserving optimal value for a single randomly generated reward function, AUP incurs modest overhead, completes the specified task, and avoids side effects.
There are notable examples of online search improving over hand-coded or learned policies (e.g. AlphaZero) for sequential decision making. It is not clear, however, whether or not policy improvement is guaranteed for many of these approaches, even when given a perfect evaluation function and transition model. Indeed, simple counter examples show that seemingly reasonable online search procedures can hurt performance compared to the original policy. To address this issue, we introduce the choice function framework for analyzing online search procedures for policy improvement. A choice function specifies the actions to be considered at every node of a search tree, with all other actions being pruned. Our main contribution is to give sufficient conditions for stationary and non-stationary choice functions to guarantee that the value achieved by online search is no worse than the original policy. In addition, we describe a general parametric class of choice functions that satisfy those conditions and present an illustrative use case of the framework's empirical utility.
We present a new local entity disambiguation system. The key to our system is a novel approach for learning entity representations. In our approach we learn an entity aware extension of Embedding for Language Model (ELMo) which we call Entity-ELMo (E-ELMo). Given a paragraph containing one or more named entity mentions, each mention is first defined as a function of the entire paragraph (including other mentions), then they predict the referent entities. Utilizing E-ELMo for local entity disambiguation, we outperform all of the state-of-the-art local and global models on the popular benchmarks by improving about 0.5\% on micro average accuracy for AIDA test-b with Yago candidate set. The evaluation setup of the training data and candidate set are the same as our baselines for fair comparison.
Deep learning has emerged as a compelling solution to many NLP tasks with remarkable performances. However, due to their opacity, such models are hard to interpret and trust. Recent work on explaining deep models has introduced approaches to provide insights toward the model's behaviour and predictions, which are helpful for assessing the reliability of the model's predictions. However, such methods do not improve the model's reliability. In this paper, we aim to teach the model to make the right prediction for the right reason by providing explanation training and ensuring the alignment of the model's explanation with the ground truth explanation. Our experimental results on multiple tasks and datasets demonstrate the effectiveness of the proposed method, which produces more reliable predictions while delivering better results compared to traditionally trained models.
Reward functions are often misspecified. An agent optimizing an incorrect reward function can change its environment in large, undesirable, and potentially irreversible ways. Work on impact measurement seeks a means of identifying (and thereby avoiding) large changes to the environment. We propose a novel impact measure which induces conservative, effective behavior across a range of situations. The approach attempts to preserve the attainable utility of auxiliary objectives. We evaluate our proposal on an array of benchmark tasks and show that it matches or outperforms relative reachability, the state-of-the-art in impact measurement.
We consider the problem of explaining the decisions of deep neural networks for image recognition in terms of human-recognizable visual concepts. In particular, given a test set of images, we aim to explain each classification in terms of a small number of image regions, or activation maps, which have been associated with semantic concepts by a human annotator. This allows for generating summary views of the typical reasons for classifications, which can help build trust in a classifier and/or identify example types for which the classifier may not be trusted. For this purpose, we developed a user interface for "interactive naming," which allows a human annotator to manually cluster significant activation maps in a test set into meaningful groups called "visual concepts". The main contribution of this paper is a systematic study of the visual concepts produced by five human annotators using the interactive naming interface. In particular, we consider the adequacy of the concepts for explaining the classification of test-set images, correspondence of the concepts to activations of individual neurons, and the inter-annotator agreement of visual concepts. We find that a large fraction of the activation maps have recognizable visual concepts, and that there is significant agreement between the different annotators about their denotations. Our work is an exploratory study of the interplay between machine learning and human recognition mediated by visualizations of the results of learning.