We analyze state-of-the-art deep learning models for three tasks: question answering on (1) images, (2) tables, and (3) passages of text. Using the notion of \emph{attribution} (word importance), we find that these deep networks often ignore important question terms. Leveraging such behavior, we perturb questions to craft a variety of adversarial examples. Our strongest attacks drop the accuracy of a visual question answering model from $61.1\%$ to $19\%$, and that of a tabular question answering model from $33.5\%$ to $3.3\%$. Additionally, we show how attributions can strengthen attacks proposed by Jia and Liang (2017) on paragraph comprehension models. Our results demonstrate that attributions can augment standard measures of accuracy and empower investigation of model performance. When a model is accurate but for the wrong reasons, attributions can surface erroneous logic in the model that indicates inadequacies in the test data.

Motivations like domain adaptation, transfer learning, and feature learning have fueled interest in inducing embeddings for rare or unseen words, n-grams, synsets, and other textual features. This paper introduces à la carte embedding, a simple and general alternative to the usual word2vec-based approaches for building such representations that is based upon recent theoretical results for GloVe-like embeddings. Our method relies mainly on a linear transformation that is efficiently learnable using pretrained word vectors and linear regression. This transform is applicable "on the fly" in the future when a new text feature or rare word is encountered, even if only a single usage example is available. We introduce a new dataset showing how the à la carte method requires fewer examples of words in context to learn high-quality embeddings and we obtain state-of-the-art results on a nonce task and some unsupervised document classification tasks.

Over the past years, distributed representations have proven effective and flexible keepers of prior knowledge to be integrated into downstream applications. This survey is focused on semantic representation of meaning. We start from the theoretical background behind word vector space models and highlight one of their main limitations: the meaning conflation deficiency, which arises from representing a word with all its possible meanings as a single vector. Then, we explain how this deficiency can be addressed through a transition from word level to the more fine-grained level of word senses (in its broader acceptation) as a method for modelling unambiguous lexical meaning. We present a comprehensive overview of the wide range of techniques in the two main branches of sense representation, i.e., unsupervised and knowledge-based. Finally, this survey covers the main evaluation procedures and provides an analysis of five important aspects: interpretability, sense granularity, adaptability to different domains, compositionality and integration into downstream applications.

ABSTRACT Regularization is a big issue for training deep neural networks. In this paper, we propose a new information-theory-based regularization scheme named SHADE for SHAnnon DEcay. The originality of the approach is to define a prior based on conditional entropy, which explicitly decouples the learning of invariant representations in the regularizer and the learning of correlations between inputs and labels in the data fitting term. Our second contribution is to derive a stochastic version of the regularizer compatible with deep learning, resulting in a tractable training scheme. We empirically validate the efficiency of our approach to improve classification performances compared to standard regularization schemes on several standard architectures.

This project combines recent advances in experience replay techniques, namely, Combined Experience Replay (CER), Prioritized Experience Replay (PER), and Hindsight Experience Replay (HER). We show the results of combinations of these techniques with DDPG and DQN methods. CER always adds the most recent experience to the batch. PER chooses which experiences should be replayed based on how beneficial they will be towards learning. HER learns from failure by substituting the desired goal with the achieved goal and recomputing the reward function. The effectiveness of combinations of these experience replay techniques is tested in a variety of OpenAI gym environments.

We propose a confidence scoring mechanism for multi-layer neural networks based on a paradigm of a base model and a meta-model. The confidence score is learned by the meta-model using features derived from the base model -- a deep multi-layer neural network -- considered a whitebox. As features, we investigate linear classifier probes inserted between the various layers of the base model and trained using each layer's intermediate activations. Experiments show that this approach outperforms various baselines in a filtering task, i.e., task of rejecting samples with low confidence. Experimental results are presented using CIFAR-10 and CIFAR-100 dataset with and without added noise exploring various aspects of the method.

In recent times, many of the breakthroughs in various vision-related tasks have revolved around improving learning of deep models; these methods have ranged from network architectural improvements such as Residual Networks, to various forms of regularisation such as Batch Normalisation. In essence, many of these techniques revolve around better conditioning, allowing for deeper and deeper models to be successfully learned. In this paper, we look towards better conditioning Generative Adversarial Networks (GANs) in an unsupervised learning setting. Our method embeds the powerful discriminating capabilities of a decision forest into the discriminator of a GAN. This results in a better conditioned model which learns in an extremely stable way. We demonstrate empirical results which show both clear qualitative and quantitative evidence of the effectiveness of our approach, gaining significant performance improvements over several popular GAN-based approaches on the Oxford Flowers and Aligned Celebrity Faces datasets.

The success of deep neural networks (DNNs) is heavily dependent on the availability of labeled data. However, obtaining labeled data is a big challenge in many real-world problems. In such scenarios, a DNN model can leverage labeled and unlabeled data from a related domain, but it has to deal with the shift in data distributions between the source and the target domains. In this paper, we study the problem of classifying social media posts during a crisis event (e.g., Earthquake). For that, we use labeled and unlabeled data from past similar events (e.g., Flood) and unlabeled data for the current event. We propose a novel model that performs adversarial learning based domain adaptation to deal with distribution drifts and graph based semi-supervised learning to leverage unlabeled data within a single unified deep learning framework. Our experiments with two real-world crisis datasets collected from Twitter demonstrate significant improvements over several baselines.

Automatic sleep staging is a challenging problem and state-of-the-art algorithms have not yet reached satisfactory performance to be used instead of manual scoring by a sleep technician. Much research has been done to find good feature representations that extract the useful information to correctly classify each epoch into the correct sleep stage. While many useful features have been discovered, the amount of features have grown to an extent that a feature reduction step is necessary in order to avoid the curse of dimensionality. One reason for the need of such a large feature set is that many features are good for discriminating only one of the sleep stages and are less informative during other stages. This paper explores how a second feature representation over a large set of pre-defined features can be learned using an auto-encoder with a selective attention for the current sleep stage in the training batch. This selective attention allows the model to learn feature representations that focuses on the more relevant inputs without having to perform any dimensionality reduction of the input data. The performance of the proposed algorithm is evaluated on a large data set of polysomnography (PSG) night recordings of patients with sleep-disordered breathing. The performance of the auto-encoder with selective attention is compared with a regular auto-encoder and previous works using a deep belief network (DBN).

We propose the method to sanitize the privacy of the IFM(Input Feature Map)s that are fed into the layers of CNN(Convolutional Neural Network)s. The method introduces the degree of the sanitization that makes the application using a CNN be able to control the privacy loss represented as the ratio of the probabilistic accuracies for original IFM and sanitized IFM. For the sanitization of an IFM, the sample-and-hold based approximation scheme is devised to satisfy an application-specific degree of the sanitization. The scheme approximates an IFM by replacing all the samples in a window with the non-zero sample closest to the mean of the sampling window. It also removes the dependency on CNN configuration by unfolding multi-dimensional IFM tensors into one-dimensional streams to be approximated.