Data collection for natural language (NL) understanding tasks has increasingly included human explanations alongside data points, allowing past works to introduce models that both perform a task and generate NL explanations for their outputs. Yet to date, model-generated explanations have been evaluated on the basis of surface-level similarities to human explanations, both through automatic metrics like BLEU and human evaluations. We argue that these evaluations are insufficient, since they fail to indicate whether explanations support actual model behavior (faithfulness), rather than simply match what a human would say (plausibility). In this work, we address the problem of evaluating explanations from the model simulatability perspective. Our contributions are as follows: (1) We introduce a leakage-adjusted simulatability (LAS) metric for evaluating NL explanations, which measures how well explanations help an observer predict a model's output, while controlling for how explanations can directly leak the output. We use a model as a proxy for a human observer, and validate this choice with two human subject experiments. (2) Using the CoS-E and e-SNLI datasets, we evaluate two existing generative graphical models and two new approaches; one rationalizing method we introduce achieves roughly human-level LAS scores. (3) Lastly, we frame explanation generation as a multi-agent game and optimize explanations for simulatability while penalizing label leakage, which can improve LAS scores. We provide code for the experiments in this paper at https://github.com/peterbhase/LAS-NL-Explanations

The number of parameters in state of the art neural networks has drastically increased in recent years. This surge of interest in large scale neural networks has motivated the development of new distributed training strategies enabling such models. One such strategy is model-parallel distributed training. Unfortunately, model-parallelism suffers from poor resource utilisation, which leads to wasted resources. In this work, we improve upon recent developments in an idealised model-parallel optimisation setting: local learning. Motivated by poor resource utilisation, we introduce a class of intermediary strategies between local and global learning referred to as interlocking backpropagation. These strategies preserve many of the compute-efficiency advantages of local optimisation, while recovering much of the task performance achieved by global optimisation. We assess our strategies on both image classification ResNets and Transformer language models, finding that our strategy consistently out-performs local learning in terms of task performance, and out-performs global learning in training efficiency.

A common approach for compressing NLP networks is to encode the embedding layer as a matrix $A\in\mathbb{R}^{n\times d}$, compute its rank-$j$ approximation $A_j$ via SVD, and then factor $A_j$ into a pair of matrices that correspond to smaller fully-connected layers to replace the original embedding layer. Geometrically, the rows of $A$ represent points in $\mathbb{R}^d$, and the rows of $A_j$ represent their projections onto the $j$-dimensional subspace that minimizes the sum of squared distances ("errors") to the points. In practice, these rows of $A$ may be spread around $k>1$ subspaces, so factoring $A$ based on a single subspace may lead to large errors that turn into large drops in accuracy. Inspired by \emph{projective clustering} from computational geometry, we suggest replacing this subspace by a set of $k$ subspaces, each of dimension $j$, that minimizes the sum of squared distances over every point (row in $A$) to its \emph{closest} subspace. Based on this approach, we provide a novel architecture that replaces the original embedding layer by a set of $k$ small layers that operate in parallel and are then recombined with a single fully-connected layer. Extensive experimental results on the GLUE benchmark yield networks that are both more accurate and smaller compared to the standard matrix factorization (SVD). For example, we further compress DistilBERT by reducing the size of the embedding layer by $40\%$ while incurring only a $0.5\%$ average drop in accuracy over all nine GLUE tasks, compared to a $2.8\%$ drop using the existing SVD approach. On RoBERTa we achieve $43\%$ compression of the embedding layer with less than a $0.8\%$ average drop in accuracy as compared to a $3\%$ drop previously. Open code for reproducing and extending our results is provided.

The field of adversarial robustness has attracted significant attention in machine learning. Contrary to the common approach of training models that are accurate in average case, it aims at training models that are accurate for worst case inputs, hence it yields more robust and reliable models. Put differently, it tries to prevent an adversary from fooling a model. The study of adversarial robustness is largely focused on $\ell_p-$bounded adversarial perturbations, i.e. modifications of the inputs, bounded in some $\ell_p$ norm. Nevertheless, it has been shown that state-of-the-art models are also vulnerable to other more natural perturbations such as affine transformations, which were already considered in machine learning within data augmentation. This project reviews previous work in spatial robustness methods and proposes evolution strategies as zeroth order optimization algorithms to find the worst affine transforms for each input. The proposed method effectively yields robust models and allows introducing non-parametric adversarial perturbations.

End-to-end training with back propagation is the standard method for training deep neural networks. However, as networks become deeper and bigger, endto-end training becomes more challenging: highly non-convex models gets stuck easily in local optima, gradients signals are prone to vanish or explode during back-propagation, training requires computational resources and time. In this work, we propose to break away from the end-to-end paradigm in the context of Knowledge Distillation. Instead of distilling a model end-to-end, we propose to split it into smaller sub-networks - also called neighbourhoods - that are then trained independently. We empirically show that distilling networks in a non endto-end fashion can be beneficial in a diverse range of use cases. First, we show that it speeds up Knowledge Distillation by exploiting parallelism and training on smaller networks. Second, we show that independently distilled neighbourhoods may be efficiently reused for Neural Architecture Search. Finally, because smaller networks model simpler functions, we show that they are easier to train with synthetic data than their deeper counterparts. As Deep Neural Networks improve on challenging tasks, they also become deeper and bigger. Convolutional neural networks for Image Classification grew from 5 layers in LeNet (LeCun et al., 1998) to more than a 100 in the latest ResNet models (He et al., 2016). However, as models grow in size, training by back propagating gradients through the entire network becomes more challenging and computationally expensive.

Time series anomaly detection has been a perennially important topic in data science, with papers dating back to the 1950s. However, in recent years there has been an explosion of interest in this topic, much of it driven by the success of deep learning in other domains and for other time series tasks. Most of these papers test on one or more of a handful of popular benchmark datasets, created by Yahoo, Numenta, NASA, etc. In this work we make a surprising claim. The majority of the individual exemplars in these datasets suffer from one or more of four flaws. Because of these four flaws, we believe that many published comparisons of anomaly detection algorithms may be unreliable, and more importantly, much of the apparent progress in recent years may be illusionary. In addition to demonstrating these claims, with this paper we introduce the UCR Time Series Anomaly Datasets. We believe that this resource will perform a similar role as the UCR Time Series Classification Archive, by providing the community with a benchmark that allows meaningful comparisons between approaches and a meaningful gauge of overall progress.

Inferring interactions from multi-agent trajectories has broad applications in physics, vision and robotics. Neural relational inference (NRI) is a deep generative model that can reason about relations in complex dynamics without supervision. In this paper, we take a careful look at this approach for relational inference in multi-agent trajectories. First, we discover that NRI can be fundamentally limited without sufficient long-term observations. Its ability to accurately infer interactions degrades drastically for short output sequences. Next, we consider a more general setting of relational inference when interactions are changing overtime. We propose an extension ofNRI, which we call the DYnamic multi-AgentRelational Inference (DYARI) model that can reason about dynamic relations. We conduct exhaustive experiments to study the effect of model architecture, under-lying dynamics and training scheme on the performance of dynamic relational inference using a simulated physics system. We also showcase the usage of our model on real-world multi-agent basketball trajectories.

Code similarity systems are integral to a range of applications from code recommendation to automated software defect correction. We argue that code similarity is now a first-order problem that must be solved. To begin to address this, we present machine Inferred Code Similarity (MISIM), a novel end-to-end code similarity system that consists of two core components. First, MISIM uses a novel context-aware semantic structure, which is designed to aid in lifting semantic meaning from code syntax. Second, MISIM provides a neural-based code similarity scoring algorithm, which can be implemented with various neural network architectures with learned parameters. We compare MISIM to three state-of-the-art code similarity systems: (i)code2vec, (ii)Neural Code Comprehension, and (iii)Aroma. In our experimental evaluation across 328,155 programs (over 18 million lines of code), MISIM has 1.5x to 43.4x better accuracy than all three systems.

In this paper, we study distributed algorithms for large-scale AUC maximization with a deep neural network as a predictive model. Although distributed learning techniques have been investigated extensively in deep learning, they are not directly applicable to stochastic AUC maximization with deep neural networks due to its striking differences from standard loss minimization problems (e.g., cross-entropy). Towards addressing this challenge, we propose and analyze a communication-efficient distributed optimization algorithm based on a {\it non-convex concave} reformulation of the AUC maximization, in which the communication of both the primal variable and the dual variable between each worker and the parameter server only occurs after multiple steps of gradient-based updates in each worker. Compared with the naive parallel version of an existing algorithm that computes stochastic gradients at individual machines and averages them for updating the model parameters, our algorithm requires a much less number of communication rounds and still achieves a linear speedup in theory. To the best of our knowledge, this is the \textbf{first} work that solves the {\it non-convex concave min-max} problem for AUC maximization with deep neural networks in a communication-efficient distributed manner while still maintaining the linear speedup property in theory. Our experiments on several benchmark datasets show the effectiveness of our algorithm and also confirm our theory.

Research challenges encountered across science, engineering, and economics can frequently be formulated as optimization tasks. In chemistry and materials science, recent growth in laboratory digitization and automation has sparked interest in optimization-guided autonomous discovery and closed-loop experimentation. Experiment planning strategies based on off-the-shelf optimization algorithms can be employed in fully autonomous research platforms to achieve desired experimentation goals with the minimum number of trials. However, the experiment planning strategy that is most suitable to a scientific discovery task is a priori unknown while rigorous comparisons of different strategies are highly time and resource demanding. As optimization algorithms are typically benchmarked on low-dimensional synthetic functions, it is unclear how their performance would translate to noisy, higher-dimensional experimental tasks encountered in chemistry and materials science. We introduce Olympus, a software package that provides a consistent and easy-to-use framework for benchmarking optimization algorithms against realistic experiments emulated via probabilistic deep-learning models. Olympus includes a collection of experimentally derived benchmark sets from chemistry and materials science and a suite of experiment planning strategies that can be easily accessed via a user-friendly python interface. Furthermore, Olympus facilitates the integration, testing, and sharing of custom algorithms and user-defined datasets. In brief, Olympus mitigates the barriers associated with benchmarking optimization algorithms on realistic experimental scenarios, promoting data sharing and the creation of a standard framework for evaluating the performance of experiment planning strategies