Deep neural networks have been demonstrated to be vulnerable to adversarial attacks, where small perturbations are intentionally added to the original inputs to fool the classifier. In this paper, we propose a defense method, Featurized Bidirectional Generative Adversarial Networks (FBGAN), to capture the semantic features of the input and filter the non-semantic perturbation. FBGAN is pre-trained on the clean dataset in an unsupervised manner, adversarially learning a bidirectional mapping between the high-dimensional data space and the low-dimensional semantic space, and mutual information is applied to disentangle the semantically meaningful features. After the bidirectional mapping, the adversarial data can be reconstructed to denoised data, which could be fed into the classifier for classification. We empirically show the quality of reconstruction images and the effectiveness of defense.

Learning with a {\it convex loss} function has been a dominating paradigm for many years. It remains an interesting question how non-convex loss functions help improve the generalization of learning with broad applicability. In this paper, we study a family of objective functions formed by truncating traditional loss functions, which is applicable to both shallow learning and deep learning. Truncating loss functions has potential to be less vulnerable and more robust to large noise in observations that could be adversarial. More importantly, it is a generic technique without assuming the knowledge of noise distribution. To justify non-convex learning with truncated losses, we establish excess risk bounds of empirical risk minimization based on truncated losses for heavy-tailed output, and statistical error of an approximate stationary point found by stochastic gradient descent (SGD) method. Our experiments for shallow and deep learning for regression with outliers, corrupted data and heavy-tailed noise further justify the proposed method.

In this paper we consider whether RNNs can learn functionally equivalent models of unknown computer hardware peripherals through input/output observation. Peripheral devices attach to a main computer and use both hardware within the device and driver software running on the main computer to perform a task, such as printing a page or sending a message. However, there are instances when hardware is accessible from the main system but driver software is not, rendering the peripheral unusable. This situation is prevalent in open source operating systems where driver software may not be available from the vendor. Without driver software or development documentation, it is incumbent on the system's owner to write software to make use of the peripheral. The device itself is a "black box", with no information directly available to the developer beyond a set of memory addresses to interact with the device and the observable output of the hardware itself. This leads to labor-intensive reverse engineering efforts with varying degrees of success (see e.g.

Convolution has been playing a prominent role in various applications in science and engineering for many years. It is the most important operation in convolutional neural networks. There has been a recent growth of interests of research in generalizing convolutions on curved domains such as manifolds and graphs. However, existing approaches cannot preserve all the desirable properties of Euclidean convolutions, namely compactly supported filters, directionality, transferability across different manifolds. In this paper we develop a new generalization of the convolution operation, referred to as parallel transport convolution (PTC), on Riemannian manifolds and their discrete counterparts. PTC is designed based on the parallel transportation which is able to translate information along a manifold and to intrinsically preserve directionality. PTC allows for the construction of compactly supported filters and is also robust to manifold deformations. This enables us to preform wavelet-like operations and to define deep convolutional neural networks on curved domains.

The application of deep recurrent networks to audio transcription has led to impressive gains in automatic speech recognition (ASR) systems. Many have demonstrated that small adversarial perturbations can fool deep neural networks into incorrectly predicting a specified target with high confidence. Current work on fooling ASR systems have focused on white-box attacks, in which the model architecture and parameters are known. In this paper, we adopt a black-box approach to adversarial generation, combining the approaches of both genetic algorithms and gradient estimation to solve the task. We achieve a 89.25% targeted attack similarity after 3000 generations while maintaining 94.6% audio file similarity.

Learning to control robots directly based on images is a primary challenge in robotics. However, many existing reinforcement learning approaches require iteratively obtaining millions of samples to learn a policy which can take significant time. In this paper, we focus on the problem of learning real-world robot action policies solely based on a few random off-policy samples. We learn a realistic dreaming model that can emulate samples equivalent to a sequence of images from the actual environment, and make the agent learn action policies by interacting with the dreaming model rather than the real world. We experimentally confirm that our dreaming model can learn realistic policies that transfer to the real-world.

Meta-learning approaches have been proposed to tackle the few-shot learning problem.Typically, a meta-learner is trained on a variety of tasks in the hopes of being generalizable to new tasks. However, the generalizability on new tasks of a meta-learner could be fragile when it is over-trained on existing tasks during meta-training phase. In other words, the initial model of a meta-learner could be too biased towards existing tasks to adapt to new tasks, especially when only very few examples are available to update the model. To avoid a biased meta-learner and improve its generalizability, we propose a novel paradigm of Task-Agnostic Meta-Learning (TAML) algorithms. Specifically, we present an entropy-based approach that meta-learns an unbiased initial model with the largest uncertainty over the output labels by preventing it from over-performing in classification tasks. Alternatively, a more general inequality-minimization TAML is presented for more ubiquitous scenarios by directly minimizing the inequality of initial losses beyond the classification tasks wherever a suitable loss can be defined.Experiments on benchmarked datasets demonstrate that the proposed approaches outperform compared meta-learning algorithms in both few-shot classification and reinforcement learning tasks.

The study of high-throughput genomic profiles from a pharmacogenomics viewpoint has provided unprecedented insights into the oncogenic features modulating drug response. A recent screening of ~1,000 cancer cell lines to a collection of anti-cancer drugs illuminated the link between genotypes and vulnerability. However, due to essential differences between cell lines and tumors, the translation into predicting drug response in tumors remains challenging. Here we proposed a DNN model to predict drug response based on mutation and expression profiles of a cancer cell or a tumor. The model contains a mutation and an expression encoders pre-trained using a large pan-cancer dataset to abstract core representations of high-dimension data, followed by a drug response predictor network. Given a pair of mutation and expression profiles, the model predicts IC50 values of 265 drugs. We trained and tested the model on a dataset of 622 cancer cell lines and achieved an overall prediction performance of mean squared error at 1.96 (log-scale IC50 values). The performance was superior in prediction error or stability than two classical methods and four analog DNNs of our model. We then applied the model to predict drug response of 9,059 tumors of 33 cancer types. The model predicted both known, including EGFR inhibitors in non-small cell lung cancer and tamoxifen in ER+ breast cancer, and novel drug targets. The comprehensive analysis further revealed the molecular mechanisms underlying the resistance to a chemotherapeutic drug docetaxel in a pan-cancer setting and the anti-cancer potential of a novel agent, CX-5461, in treating gliomas and hematopoietic malignancies. Overall, our model and findings improve the prediction of drug response and the identification of novel therapeutic options.

In this paper, we introduce Attentive Guidance (AG), a new mechanism to direct a sequence to sequence model equipped with attention to find more compositional solutions that generalise even in cases where the training and testing distribution strongly diverge. We test AG on two tasks, devised precisely to asses the composi- tional capabilities of neural models and show how vanilla sequence to sequence models with attention overfit the training distribution, while the guided versions come up with compositional solutions that, in some cases, fit the training and testing distributions equally well. AG is a simple and intuitive method to provide a learning bias to a sequence to sequence model without the need of including extra components, that we believe allows to inject a component in the training process which is also present in human learning: guidance.

A widespread folklore for explaining the success of convolutional neural network (CNN) is that CNN is a more compact representation than the fully connected neural network (FNN) and thus requires fewer samples for learning. We initiate the study of rigorously characterizing the sample complexity of learning convolutional neural networks. We show that for learning an $m$-dimensional convolutional filter with linear activation acting on a $d$-dimensional input, the sample complexity of achieving population prediction error of $\epsilon$ is $\widetilde{O} (m/\epsilon^2)$, whereas its FNN counterpart needs at least $\Omega(d/\epsilon^2)$ samples. Since $m \ll d$, this result demonstrates the advantage of using CNN. We further consider the sample complexity of learning a one-hidden-layer CNN with linear activation where both the $m$-dimensional convolutional filter and the $r$-dimensional output weights are unknown. For this model, we show the sample complexity is $\widetilde{O}\left((m+r)/\epsilon^2\right)$ when the ratio between the stride size and the filter size is a constant. For both models, we also present lower bounds showing our sample complexities are tight up to logarithmic factors. Our main tools for deriving these results are localized empirical process and a new lemma characterizing the convolutional structure. We believe these tools may inspire further developments in understanding CNN.