A grand challenge in representation learning is to learn the different explanatory factors of variation behind the high dimen- sional data. Encoder models are often determined to optimize performance on training data when the real objective is to generalize well to unseen data. Although there is enough numerical evidence suggesting that noise injection (during training) at the representation level might improve the generalization ability of encoders, an information-theoretic understanding of this principle remains elusive. This paper presents a sample-dependent bound on the generalization gap of the cross-entropy loss that scales with the information complexity (IC) of the representations, meaning the mutual information between inputs and their representations. The IC is empirically investigated for standard multi-layer neural networks with SGD on MNIST and CIFAR-10 datasets; the behaviour of the gap and the IC appear to be in direct correlation, suggesting that SGD selects encoders to implicitly minimize the IC. We specialize the IC to study the role of Dropout on the generalization capacity of deep encoders which is shown to be directly related to the encoder capacity, being a measure of the distinguishability among samples from their representations. Our results support some recent regularization methods.

Multiple modalities often co-occur when describing natural phenomena. Learning a joint representation of these modalities should yield deeper and more useful representations. Previous work have proposed generative models to handle multi-modal input. However, these models either do not learn a joint distribution or require complex additional computations to handle missing data. Here, we introduce a multimodal variational autoencoder that uses a product-of-experts inference network and a sub-sampled training paradigm to solve the multi-modal inference problem. Notably, our model shares parameters to efficiently learn under any combination of missing modalities, thereby enabling weakly-supervised learning. We apply our method on four datasets and show that we match state-of-the-art performance using many fewer parameters. In each case our approach yields strong weakly-supervised results. We then consider a case study of learning image transformations---edge detection, colorization, facial landmark segmentation, etc.---as a set of modalities. We find appealing results across this range of tasks.

From left to right, given a graph G with a set of node features F and edge weights Y, the encoder aggregates information from a different number of hops j K away for each nodev G into an embedding vectorc v(j). To do so, it uses a feedforward network to propagate information between different search depths, which is parametrized by a set of weight matrices W j . This embedding vectors are then fed into a differentiable functionφ enc, which sets the parameters,µ k andσ k, of several multidimensional Gaussian distributionsq φ, from where the latent representation of each node in the input graph are sampled from. Variational autoencoders are characterized by a probabilistic generative modelp θ(x z) of the observed variablesx R N given the latent variablesz R M, a prior distribution over the latent variablesp(z) and an approximate probabilistic inference modelq φ (z x). In this characterization,p θ and q φ are arbitrary distributions parametrized by two (deep) neural networksθ and φ and one can think of the generative model as a probabilistic decoder, which decodes latent variables into observed variables, and the inference model as a probabilistic encoder, which encodes observed variables into latent variables. Ideally, if we use the maximum likelihood principle to train a variational autoencoder, we should optimize the marginal log-likelihood of the observed data, i.e., E D [log p θ(x)], wherep D is the data distribution. Unfortunately, computing logp θ(x) requires marginalization with respect to the the latent variablez, which is typically intractable. Therefore, one resorts to maximizing a variational lower bound or evidence lower bound (ELBO) of the log-likelihood the observed data, i.e., max θ max φ E D [ KL(q φ (z x) p(z)) E q φ (z x)log p θ(x z)] . Finally, note that the quality of this variational lower bound (and thus the quality of the resulting V AE) depends on the expressive ability of the approximate inference modelq φ (z x), which is typically assumed to be a normal distribution whose mean and variance are parametrized by a (deep) neural networkφ with the observed datax as an input.

Historically, machine learning in computer security has prioritized defense: think intrusion detection systems, malware classification, and botnet traffic identification. Offense can benefit from data just as well. Social networks, with their access to extensive personal data, bot-friendly APIs, colloquial syntax, and prevalence of shortened links, are the perfect venues for spreading machine-generated malicious content. We aim to discover what capabilities an adversary might utilize in such a domain. We present a long short-term memory (LSTM) neural network that learns to socially engineer specific users into clicking on deceptive URLs. The model is trained with word vector representations of social media posts, and in order to make a click-through more likely, it is dynamically seeded with topics extracted from the target's timeline. We augment the model with clustering to triage high value targets based on their level of social engagement, and measure success of the LSTM's phishing expedition using click-rates of IP-tracked links. We achieve state of the art success rates, tripling those of historic email attack campaigns, and outperform humans manually performing the same task.

Deep convolutional network has been the state-of-the-art approach for a wide variety of tasks over the last few years. Its successes have, in many cases, turned it into the default model in quite a few domains. In this work we will demonstrate that convolutional networks have limitations that may, in some cases, hinder it from learning properties of the data, which are easily recognizable by traditional, less demanding, models. To this end, we present a series of competitive analysis studies on image recognition and text analysis tasks, for which convolutional networks are known to provide state-of-the-art results. In our studies, we inject a truth-reveling signal, indiscernible for the network, thus hitting time and again the network's blind spots. The signal does not impair the network's existing performances, but it does provide an opportunity for a significant performance boost by models that can capture it. The various forms of the carefully designed signals shed a light on the strengths and weaknesses of convolutional network, which may provide insights for both theoreticians that study the power of deep architectures, and for practitioners that consider to apply convolutional networks to the task at hand.

We propose a new integrated method of exploiting model, batch and domain parallelism for the training of deep neural networks (DNNs) on large distributed-memory computers using minibatch stochastic gradient descent (SGD). Our goal is to find an efficient parallelization strategy for a fixed batch size using $P$ processes. Our method is inspired by the communication-avoiding algorithms in numerical linear algebra. We see $P$ processes as logically divided into a $P_r \times P_c$ grid where the $P_r$ dimension is implicitly responsible for model/domain parallelism and the $P_c$ dimension is implicitly responsible for batch parallelism. In practice, the integrated matrix-based parallel algorithm encapsulates these types of parallelism automatically. We analyze the communication complexity and analytically demonstrate that the lowest communication costs are often achieved neither with pure model nor with pure data parallelism. We also show how the domain parallel approach can help in extending the the theoretical scaling limit of the typical batch parallel method.

Many deployed learned models are black boxes: given input, returns output. Internal information about the model, such as the architecture, optimisation procedure, or training data, is not disclosed explicitly as it might contain proprietary information or make the system more vulnerable. This work shows that such attributes of neural networks can be exposed from a sequence of queries. This has multiple implications. On the one hand, our work exposes the vulnerability of black-box neural networks to different types of attacks -- we show that the revealed internal information helps generate more effective adversarial examples against the black box model. On the other hand, this technique can be used for better protection of private content from automatic recognition models using adversarial examples. Our paper suggests that it is actually hard to draw a line between white box and black box models.

Thousands of security vulnerabilities are discovered in production software each year, either reported publicly to the Common Vulnerabilities and Exposures database or discovered internally in proprietary code. Vulnerabilities often manifest themselves in subtle ways that are not obvious to code reviewers or the developers themselves. With the wealth of open source code available for analysis, there is an opportunity to learn the patterns of bugs that can lead to security vulnerabilities directly from data. In this paper, we present a data-driven approach to vulnerability detection using machine learning, specifically applied to C and C++ programs. We first compile a large dataset of hundreds of thousands of open-source functions labeled with the outputs of a static analyzer. We then compare methods applied directly to source code with methods applied to artifacts extracted from the build process, finding that source-based models perform better. We also compare the application of deep neural network models with more traditional models such as random forests and find the best performance comes from combining features learned by deep models with tree-based models. Ultimately, our highest performing model achieves an area under the precision-recall curve of 0.49 and an area under the ROC curve of 0.87.

Social media has grown to be a crucial information source for pharmacovigilance studies where an increasing number of people post adverse reactions to medical drugs that are previously unreported. Aiming to effectively monitor various aspects of Adverse Drug Reactions (ADRs) from diversely expressed social medical posts, we propose a multi-task neural network framework that learns several tasks associated with ADR monitoring with different levels of supervisions collectively. Besides being able to correctly classify ADR posts and accurately extract ADR mentions from online posts, the proposed framework is also able to further understand reasons for which the drug is being taken, known as'indications', from the given social media post. A coverage-based attention mechanism is adopted in our framework to help the model properly identify'phrasal' ADRs and Indications that are attentive to multiple words in a post. Our framework is applicable in situations where limited parallel data for different pharmacovigilance tasks are available. We evaluate the proposed framework on real-world Twitter datasets, where the proposed model outperforms the state-of-the-art alternatives of each individual task consistently.

We demonstrate that state-of-the-art optical character recognition (OCR) based on deep learning is vulnerable to adversarial images. Minor modifications to images of printed text, which do not change the meaning of the text to a human reader, cause the OCR system to "recognize" a different text where certain words chosen by the adversary are replaced by their semantic opposites. This completely changes the meaning of the output produced by the OCR system and by the NLP applications that use OCR for preprocessing their inputs.