The use of synthetic data generated by Generative Adversarial Networks (GANs) has become quite a popular method to do data augmentation for many applications. While practitioners celebrate this as an economical way to get more synthetic data that can be used to train downstream classifiers, it is not clear that they recognize the inherent pitfalls of this technique. In this paper, we aim to exhort practitioners against deriving any false sense of security against data biases based on data augmentation. To drive this point home, we show that starting with a dataset consisting of head-shots of engineering researchers, GAN-based augmentation "imagines" synthetic engineers, most of whom have masculine features and white skin color (inferred from a human subject study conducted on Amazon Mechanical Turk). This demonstrates how biases inherent in the training data are reinforced, and sometimes even amplified, by GAN-based data augmentation; it should serve as a cautionary tale for the lay practitioners.

Data parallelism can boost the training speed of convolutional neural networks (CNN), but could suffer from significant communication costs caused by gradient aggregation. To alleviate this problem, several scalar quantization techniques have been developed to compress the gradients. But these techniques could perform poorly when used together with decentralized aggregation protocols like ring all-reduce (RAR), mainly due to their inability to directly aggregate compressed gradients. In this paper, we empirically demonstrate the strong linear correlations between CNN gradients, and propose a gradient vector quantization technique, named GradiVeQ, to exploit these correlations through principal component analysis (PCA) for substantial gradient dimension reduction. GradiVeQ enables direct aggregation of compressed gradients, hence allows us to build a distributed learning system that parallelizes GradiVeQ gradient compression and RAR communications. Extensive experiments on popular CNNs demonstrate that applying GradiVeQ slashes the wall-clock gradient aggregation time of the original RAR by more than 5X without noticeable accuracy loss, and reduces the end-to-end training time by almost 50%. The results also show that GradiVeQ is compatible with scalar quantization techniques such as QSGD (Quantized SGD), and achieves a much higher speed-up gain under the same compression ratio.

Because combat environments change over time and technology upgrades are widespread for ground vehicles, a large number of vehicles and equipment become quickly obsolete. A possible solution for the U.S. Army is to develop fleets of modular military vehicles, which are built by interchangeable substantial components also known as modules. One of the typical characteristics of module is their ease of assembly and disassembly through simple means such as plug-in/pull-out actions, which allows for real-time fleet reconfiguration to meet dynamic demands. Moreover, military demands are time-varying and highly stochastic because commanders keep reacting to enemy's actions. To capture these characteristics, we formulated an intelligent agent-based model to imitate decision making process during fleet operation, which combines real-time optimization with artificial intelligence. The agents are capable of inferring enemy's future move based on historical data and optimize dispatch/operation decisions accordingly. We implement our model to simulate an attacker-defender game between two adversarial and intelligent players, representing the commanders from modularized fleet and conventional fleet respectively. Given the same level of combat resources and intelligence, we highlight the tactical advantages of fleet modularity in terms of win rate, unpredictability and suffered damage.

While machine learning (ML) models are being increasingly trusted to make decisions in different and varying areas, the safety of systems using such models has become an increasing concern. In particular, ML models are often trained on data from potentially untrustworthy sources, providing adversaries with the opportunity to manipulate them by inserting carefully crafted samples into the training set. Recent work has shown that this type of attack, called a poisoning attack, allows adversaries to insert backdoors or trojans into the model, enabling malicious behavior with simple external backdoor triggers at inference time and only a blackbox perspective of the model itself. Detecting this type of attack is challenging because the unexpected behavior occurs only when a backdoor trigger, which is known only to the adversary, is present. Model users, either direct users of training data or users of pre-trained model from a catalog, may not guarantee the safe operation of their ML-based system. In this paper, we propose a novel approach to backdoor detection and removal for neural networks. Through extensive experimental results, we demonstrate its effectiveness for neural networks classifying text and images. To the best of our knowledge, this is the first methodology capable of detecting poisonous data crafted to insert backdoors and repairing the model that does not require a verified and trusted dataset.

Volatility is a quantity of measurement for the price movements of stocks or options which indicates the uncertainty within financial markets. As an indicator of the level of risk or the degree of variation, volatility is important to analyse the financial market, and it is taken into consideration in various decision-making processes in financial activities. On the other hand, recent advancement in deep learning techniques has shown strong capabilities in modelling sequential data, such as speech and natural language. In this paper, we empirically study the applicability of the latest deep structures with respect to the volatility modelling problem, through which we aim to provide an empirical guidance for the theoretical analysis of the marriage between deep learning techniques and financial applications in the future. We examine both the traditional approaches and the deep sequential models on the task of volatility prediction, including the most recent variants of convolutional and recurrent networks, such as the dilated architecture. Accordingly, experiments with real-world stock price datasets are performed on a set of 1314 daily stock series for 2018 days of transaction. The evaluation and comparison are based on the negative log likelihood (NLL) of real-world stock price time series. The result shows that the dilated neural models, including dilated CNN and Dilated RNN, produce most accurate estimation and prediction, outperforming various widely-used deterministic models in the GARCH family and several recently proposed stochastic models. In addition, the high flexibility and rich expressive power are validated in this study.

State-of-the-art machine learning models frequently misclassify inputs that have been perturbed in an adversarial manner. Adversarial perturbations generated for a given input and a specific classifier often seem to be effective on other inputs and even different classifiers. In other words, adversarial perturbations seem to transfer between different inputs, models, and even different neural network architectures. In this work, we show that in the context of linear classifiers and two-layer ReLU networks, there provably exist directions that give rise to adversarial perturbations for many classifiers and data points simultaneously. We show that these "transferable adversarial directions" are guaranteed to exist for linear separators of a given set, and will exist with high probability for linear classifiers trained on independent sets drawn from the same distribution. We extend our results to large classes of two-layer ReLU networks. We further show that adversarial directions for ReLU networks transfer to linear classifiers while the reverse need not hold, suggesting that adversarial perturbations for more complex models are more likely to transfer to other classifiers. We validate our findings empirically, even for deeper ReLU networks.

Annotating the right data for training deep neural networks is an important challenge. Active learning using uncertainty estimates from Bayesian Neural Networks (BNNs) could provide an effective solution to this. Despite being theoretically principled, BNNs require approximations to be applied to large-scale problems, and have not been used widely by practitioners. In this paper, we introduce Deep Probabilistic Ensembles (DPEs), a scalable technique that uses a regularized ensemble to approximate a deep BNN. We conduct a series of active learning experiments to evaluate DPEs on classification with the CIFAR-10, CIFAR-100 and ImageNet datasets, and semantic segmentation with the BDD100k dataset. Our models consistently outperform baselines and previously published methods, requiring significantly less training data to achieve competitive performances.

Layer-wise Relevance Propagation (LRP) and saliency maps have been recently used to explain the predictions of Deep Learning models, specifically in the domain of text classification. Given different attribution-based explanations to highlight relevant words for a predicted class label, experiments based on word deleting perturbation is a common evaluation method. This word removal approach, however, disregards any linguistic dependencies that may exist between words or phrases in a sentence, which could semantically guide a classifier to a particular prediction. In this paper, we present a feature-based evaluation framework for comparing the two attribution methods on customer reviews (public data sets) and Customer Due Diligence (CDD) extracted reports (corporate data set). Instead of removing words based on the relevance score, we investigate perturbations based on embedded features removal from intermediate layers of Convolutional Neural Networks. Our experimental study is carried out on embedded-word, embedded-document, and embedded-ngrams explanations. Using the proposed framework, we provide a visualization tool to assist analysts in reasoning toward the model's final prediction.

Time series prediction with deep learning methods, especially long short-term memory neural networks (LSTMs), have scored significant achievements in recent years. Despite the fact that the LSTMs can help to capture long-term dependencies, its ability to pay different degree of attention on sub-window feature within multiple time-steps is insufficient. To address this issue, an evolutionary attention-based LSTM training with competitive random search is proposed for multivariate time series prediction. By transferring shared parameters, an evolutionary attention learning approach is introduced to the LSTMs model. Thus, like that for biological evolution, the pattern for importance-based attention sampling can be confirmed during temporal relationship mining. To refrain from being trapped into partial optimization like traditional gradient-based methods, an evolutionary computation inspired competitive random search method is proposed, which can well configure the parameters in the attention layer. Experimental results have illustrated that the proposed model can achieve competetive prediction performance compared with other baseline methods.

It has been common to argue or imply that a regularizer can be used to alter a statistical property of a hidden layer's representation and thus improve generalization or performance of deep networks. For instance, dropout has been known to improve performance by reducing co-adaptation, and representational sparsity has been argued as a good characteristic because many data-generation processes have a small number of factors that are independent. In this work, we analytically and empirically investigate the popular characteristics of learned representations, including correlation, sparsity, dead unit, rank, and mutual information, and disprove many of the \textit{conventional wisdom}. We first show that infinitely many Identical Output Networks (IONs) can be constructed for any deep network with a linear layer, where any invertible affine transformation can be applied to alter the layer's representation characteristics. The existence of ION proves that the correlation characteristics of representation is irrelevant to the performance. Extensions to ReLU layers are provided, too. Then, we consider sparsity, dead unit, and rank to show that only loose relationships exist among the three characteristics. It is shown that a higher sparsity or additional dead units do not imply a better or worse performance when the rank of representation is fixed. We also develop a rank regularizer and show that neither representation sparsity nor lower rank is helpful for improving performance even when the data-generation process has a small number of independent factors. Mutual information $I(\mathbf{z}_l;\mathbf{x})$ and $I(\mathbf{z}_l;\mathbf{y})$ are investigated, and we show that regularizers can affect $I(\mathbf{z}_l;\mathbf{x})$ and thus indirectly influence the performance. Finally, we explain how a rich set of regularizers can be used as a powerful tool for performance tuning.