Modelling the behaviours of other agents (opponents) is essential for understanding how agents interact and making effective decisions. Existing methods for opponent modelling commonly assume knowledge of the local observations and chosen actions of the modelled opponents, which can significantly limit their applicability. We propose a new modelling technique based on variational autoencoders, which are trained to reconstruct the local actions and observations of the opponent based on embeddings which depend only on the local observations of the modelling agent (its observed world state, chosen actions, and received rewards). The embeddings are used to augment the modelling agent's decision policy which is trained via deep reinforcement learning; thus the policy does not require access to opponent observations. We provide a comprehensive evaluation and ablation study in diverse multi-agent tasks, showing that our method achieves comparable performance to an ideal baseline which has full access to opponent's information, and significantly higher returns than a baseline method which does not use the learned embeddings. An important aspect of autonomous decision-making agents is the ability to reason about the unknown intentions and behaviours of other agents. Much research has been devoted to this opponent modelling problem [2], with recent works focused on the use of deep learning architectures for opponent modelling and reinforcement learning (RL) [15, 27, 11, 26]. A common assumption in existing methods is that the modelling agent has access to the local trajectory of the modelled agents [2], which may include their local observations of the environment state, their past actions, and possibly their received rewards.

Pruning and quantization are widely applied techniques for compressing deep neural networks, often driven by the resource constraints of deploying models to mobile phones or embedded devices (Esteva et al., 2017; Lane & Warden, 2018). To-date, discussion around the relative merits of different compression methods has centered on the tradeoff between level of compression and top-line metrics such as top-1 and top-5 accuracy (Blalock et al., 2020). Along this dimension, compression techniques are remarkably successful. It is possible to prune the majority of weights (Gale et al., 2019; Evci et al., 2019) or heavily quantize the bit representation (Jacob et al., 2017) with negligible decreases to test-set accuracy. However, recent work by Hooker et al. (2019a) has found that the minimal changes to top-line metrics obscure critical differences in generalization between pruned and non-pruned networks. The authors establish that pruning disproportionately impacts predictive performance on a small subset of the dataset. We build upon this work and focus on the implications of these findings for a dataset with sensitive protected attributes such as gender and age. Our work addresses the question: Does compression amplify existing algorithmic bias?

We propose a method to automatically generate a domain- and task-adaptive maskings of the given text for self-supervised pre-training, such that we can effectively adapt the language model to a particular target task (e.g. question answering). Specifically, we present a novel reinforcement learning-based framework which learns the masking policy, such that using the generated masks for further pre-training of the target language model helps improve task performance on unseen texts. We use off-policy actor-critic with entropy regularization and experience replay for reinforcement learning, and propose a Transformer-based policy network that can consider the relative importance of words in a given text. We validate our Neural Mask Generator (NMG) on several question answering and text classification datasets using BERT and DistilBERT as the language models, on which it outperforms rule-based masking strategies, by automatically learning optimal adaptive maskings.

Variational autoencoders (VAEs) are one of the powerful likelihood-based generative models with applications in various domains. However, they struggle to generate high-quality images, especially when samples are obtained from the prior without any tempering. One explanation for VAEs' poor generative quality is the prior hole problem: the prior distribution fails to match the aggregate approximate posterior. Due to this mismatch, there exist areas in the latent space with high density under the prior that do not correspond to any encoded image. Samples from those areas are decoded to corrupted images. To tackle this issue, we propose an energy-based prior defined by the product of a base prior distribution and a reweighting factor, designed to bring the base closer to the aggregate posterior. We train the reweighting factor by noise contrastive estimation, and we generalize it to hierarchical VAEs with many latent variable groups. Our experiments confirm that the proposed noise contrastive priors improve the generative performance of state-of-the-art VAEs by a large margin on the MNIST, CIFAR-10, CelebA 64, and CelebA HQ 256 datasets.

Discovering concepts (or temporal abstractions) in an unsupervised manner from demonstration data in the absence of an environment is an important problem. Organizing these discovered concepts hierarchically at different levels of abstraction is useful in discovering patterns, building ontologies, and generating tutorials from demonstration data. However, recent work to discover such concepts without access to any environment does not discover relationships (or a hierarchy) between these discovered concepts. In this paper, we present a Transformer-based concept abstraction architecture UNHCLE (pronounced uncle) that extracts a hierarchy of concepts in an unsupervised way from demonstration data. We empirically demonstrate how UNHCLE discovers meaningful hierarchies using datasets from Chess and Cooking domains. Finally, we show how UNHCLE learns meaningful language labels for concepts by using demonstration data augmented with natural language for cooking and chess. All of our code is available at https://github.com/UNHCLE/UNHCLE

Microsatellite instability (MSI) is associated with several tumor types and its status has become increasingly vital in guiding patient treatment decisions. However, in clinical practice, distinguishing MSI from its counterpart is challenging since the diagnosis of MSI requires additional genetic or immunohistochemical tests. In this study, interpretable pathological image analysis strategies are established to help medical experts to automatically identify MSI. The strategies only require ubiquitous Haematoxylin and eosin-stained whole-slide images and can achieve decent performance in the three cohorts collected from The Cancer Genome Atlas. The strategies provide interpretability in two aspects. On the one hand, the image-level interpretability is achieved by generating localization heat maps of important regions based on the deep learning network; on the other hand, the feature-level interpretability is attained through feature importance and pathological feature interaction analysis. More interestingly, both from the image-level and feature-level interpretability, color features and texture characteristics are shown to contribute the most to the MSI predictions. Therefore, the classification models under the proposed strategies can not only serve as an efficient tool for predicting the MSI status of patients, but also provide more insights to pathologists with clinical understanding.

We propose OneFlow - a flow-based one-class classifier for anomaly (outliers) detection that finds a minimal volume bounding region. Contrary to density-based methods, OneFlow is constructed in such a way that its result typically does not depend on the structure of outliers. This is caused by the fact that during training the gradient of the cost function is propagated only over the points located near to the decision boundary (behavior similar to the support vectors in SVM). The combination of flow models and Bernstein quantile estimator allows OneFlow to find a parametric form of bounding region, which can be useful in various applications including describing shapes from 3D point clouds. Experiments show that the proposed model outperforms related methods on real-world anomaly detection problems.

How can multiple distributed entities collaboratively train a shared deep net on their private data while preserving privacy? This paper introduces InstaHide, a simple encryption of training images, which can be plugged into existing distributed deep learning pipelines. The encryption is efficient and applying it during training has minor effect on test accuracy. InstaHide encrypts each training image with a "one-time secret key" which consists of mixing a number of randomly chosen images and applying a random pixel-wise mask. Other contributions of this paper include: (a) Using a large public dataset (e.g. ImageNet) for mixing during its encryption, which improves security. (b) Experimental results to show effectiveness in preserving privacy against known attacks with only minor effects on accuracy. (c) Theoretical analysis showing that successfully attacking privacy requires attackers to solve a difficult computational problem. (d) Demonstrating that use of the pixel-wise mask is important for security, since Mixup alone is shown to be insecure to some some efficient attacks. (e) Release of a challenge dataset https://github.com/Hazelsuko07/InstaHide_Challenge Our code is available at https://github.com/Hazelsuko07/InstaHide

Overparametrized neural networks have infinitely many solutions that achieve zero training error, and such global minima have different generalization performance. Moreover, training a neural network is a high-dimensional nonconvex problem, which is typically intractable to solve. However, the success of deep learning indicates that first-order methods such as gradient descent or stochastic gradient descent (GD/SGD) not only (a) succeed in finding global minima, but also (b) are biased towards solutions that generalize well, which largely has remained a mystery in the literature. To explain part (a) of the phenomenon, there is a growing literature studying the convergence of GD/SGD on overparametrized neural networks (e.g., Du et al. (2018a,b); Allen-Zhu et al. (2018); Zou et al. (2018); Jacot et al. (2018); Oymak and Soltanolkotabi (2020), and many more). There are also convergence results that focus on linear networks, without nonlinear activations (Bartlett et al., 2018; Arora et al., 2019a; Wu et al., 2019; Du and Hu, 2019; Hu et al., 2020). These results typically focus on the convergence of loss, hence do not address which of the many global minima is reached.

Speaker verification (SV) has recently attracted considerable research interest due to the growing popularity of virtual assistants. At the same time, there is an increasing requirement for an SV system: it should be robust to short speech segments, especially in noisy and reverberant environments. In this paper, we consider one more important requirement for practical applications: the system should be robust to an audio stream containing long non-speech segments, where a voice activity detection (VAD) is not applied. To meet these two requirements, we introduce feature pyramid module (FPM)-based multi-scale aggregation (MSA) and self-adaptive soft VAD (SAS-VAD). We present the FPM-based MSA to deal with short speech segments in noisy and reverberant environments. Also, we use the SAS-VAD to increase the robustness to long non-speech segments. To further improve the robustness to acoustic distortions (i.e., noise and reverberation), we apply a masking-based speech enhancement (SE) method. We combine SV, VAD, and SE models in a unified deep learning framework and jointly train the entire network in an end-to-end manner. To the best of our knowledge, this is the first work combining these three models in a deep learning framework. We conduct experiments on Korean indoor (KID) and VoxCeleb datasets, which are corrupted by noise and reverberation. The results show that the proposed method is effective for SV in the challenging conditions and performs better than the baseline i-vector and deep speaker embedding systems.