Code search and comprehension have become more difficult in recent years due to the rapid expansion of available source code. Current tools lack a way to label arbitrary code at scale while maintaining up-to-date representations of new programming languages, libraries, and functionalities. Comprehensive labeling of source code enables users to search for documents of interest and obtain a high-level understanding of their contents. We use Stack Overflow code snippets and their tags to train a language-agnostic, deep convolutional neural network to automatically predict semantic labels for source code documents. On Stack Overflow code snippets, we demonstrate a mean area under ROC of 0.957 over a long-tailed list of 4,508 tags. We also manually validate the model outputs on a diverse set of unlabeled source code documents retrieved from Github, and we obtain a top-1 accuracy of 86.6%. This strongly indicates that the model successfully transfers its knowledge from Stack Overflow snippets to arbitrary source code documents.

Self-attention networks (SAN) have attracted a lot of interests due to their high parallelization and strong performance on a variety of NLP tasks, e.g. machine translation. Due to the lack of recurrence structure such as recurrent neural networks (RNN), SAN is ascribed to be weak at learning positional information of words for sequence modeling. However, neither this speculation has been empirically confirmed, nor explanations for their strong performances on machine translation tasks when "lacking positional information" have been explored. To this end, we propose a novel word reordering detection task to quantify how well the word order information learned by SAN and RNN. Specifically, we randomly move one word to another position, and examine whether a trained model can detect both the original and inserted positions. Experimental results reveal that: 1) SAN trained on word reordering detection indeed has difficulty learning the positional information even with the position embedding; and 2) SAN trained on machine translation learns better positional information than its RNN counterpart, in which position embedding plays a critical role. Although recurrence structure make the model more universally-effective on learning word order, learning objectives matter more in the downstream tasks such as machine translation.

Frequency estimation is a fundamental problem in signal processing, with applications in radar imaging, underwater acoustics, seismic imaging, and spectroscopy. The goal is to estimate the frequency of each component in a multisinusoidal signal from a finite number of noisy samples. A recent machine-learning approach uses a neural network to output a learned representation with local maxima at the position of the frequency estimates. In this work, we propose a novel neural-network architecture that produces a significantly more accurate representation, and combine it with an additional neural-network module trained to detect the number of frequencies. This yields a fast, fully-automatic method for frequency estimation that achieves state-of-the-art results. In particular, it outperforms existing techniques by a substantial margin at medium-to-high noise levels.

We introduce a lifelong language learning setup where a model needs to learn from a stream of text examples without any dataset identifier. We propose an episodic memory model that performs sparse experience replay and local adaptation to mitigate catastrophic forgetting in this setup. Experiments on text classification and question answering demonstrate the complementary benefits of sparse experience replay and local adaptation to allow the model to continuously learn from new datasets. We also show that the space complexity of the episodic memory module can be reduced significantly ( 50-90%) by randomly choosing which examples to store in memory with a minimal decrease in performance. We consider an episodic memory component as a crucial building block of general linguistic intelligence and see our model as a first step in that direction.

We propose to train Bayesian Neural Networks (BNNs) by empirical Bayes as an alternative to posterior weight inference. By approximately marginalizing out an i.i.d.\ realization of a finite number of sibling weights per data-point using the Central Limit Theorem (CLT), we attain a scalable and effective Bayesian deep predictor. This approach directly models the posterior predictive distribution, by-passing the intractable posterior weight inference step. However, it introduces a prohibitively large number of hyperparameters for stable training. As the prior weights are marginalized and hyperparameters are optimized, the model also no longer provides a means to incorporate prior knowledge. We overcome both of these drawbacks by deriving a trivial PAC bound that comprises the marginal likelihood of the predictor and a complexity penalty. The outcome integrates organically into the prior networks framework, bringing about an effective and holistic Bayesian treatment of prediction uncertainty. We observe on various regression, classification, and out-of-domain detection benchmarks that our scalable method provides an improved model fit accompanied with significantly better uncertainty estimates than the state-of-the-art.

The neural machine translation model has suffered from the lack of large-scale parallel corpora. In contrast, we humans can learn multi-lingual translations even without parallel texts by referring our languages to the external world. To mimic such human learning behavior, we employ images as pivots to enable zero-resource translation learning. However, a picture tells a thousand words, which makes multi-lingual sentences pivoted by the same image noisy as mutual translations and thus hinders the translation model learning. In this work, we propose a progressive learning approach for image-pivoted zero-resource machine translation. Since words are less diverse when grounded in the image, we first learn word-level translation with image pivots, and then progress to learn the sentence-level translation by utilizing the learned word translation to suppress noises in image-pivoted multi-lingual sentences. Experimental results on two widely used image-pivot translation datasets, IAPR-TC12 and Multi30k, show that the proposed approach significantly outperforms other state-of-the-art methods.

This work investigates how a naive agent can acquire its own body image in a self-supervised way, based on the predictability of its sensorimotor experience. Our working hypothesis is that, due to its temporal stability, an agent's body produces more consistent sensory experiences than the environment, which exhibits a greater variability. Given its motor experience, an agent can thus reliably predict what appearance its body should have. This intrinsic predictability can be used to automatically isolate the body image from the rest of the environment. We propose a two-branches deconvolutional neural network to predict the visual sensory state associated with an input motor state, as well as the prediction error associated with this input. We train the network on a dataset of first-person images collected with a simulated Pepper robot, and show how the network outputs can be used to automatically isolate its visible arm from the rest of the environment. Finally, the quality of the body image produced by the network is evaluated.

Whatever information a Deep Neural Network has gleaned from past data is encoded in its weights. How this information affects the response of the network to future data is largely an open question. In fact, even how to define and measure information in a network is still not settled. We introduce the notion of Information in the Weights as the optimal trade-off between accuracy of the network and complexity of the weights, relative to a prior. Depending on the prior, the definition reduces to known information measures such as Shannon Mutual Information and Fisher Information, but affords added flexibility that enables us to relate it to generalization, via the PAC-Bayes bound, and to invariance. This relation hinges not only on the architecture of the model, but surprisingly on how it is trained. We then introduce a notion of effective information in the activations, which are deterministic functions of future inputs, resolving inconsistencies in prior work. We relate this to the Information in the Weights, and use this result to show that models of low complexity not only generalize better, but are bound to learn invariant representations of future inputs.

Gated recurrent units (GRUs) are specialized memory elements for building recurrent neural networks. Despite their incredible success in natural language, speech, and video processing, little is understood about the specific dynamics representable in a GRU network, along with the constraints these dynamics impose when generalizing a specific task. As a result, it is difficult to know a priori how successful a GRU network will perform on a given task. Using a continuous time analysis, we gain intuition on the inner workings of GRU networks. We restrict our presentation to low dimensions to allow for a comprehensive visualization. We found a surprisingly rich repertoire of dynamical features that includes stable limit cycles (nonlinear oscillations), multi-stable dynamics with various topologies, and homoclinic orbits. We contextualize the usefulness of the different kinds of dynamics and experimentally test their existence.

We propose an approach to self-supervised representation learning based on maximizing mutual information between features extracted from multiple views of a shared context. For example, a context could be an image from ImageNet, and multiple views of the context could be generated by repeatedly applying data augmentation to the image. Following this approach, we develop a new model which maximizes mutual information between features extracted at multiple scales from independently-augmented copies of each input. Our model significantly outperforms prior work on the tasks we consider. Most notably, it achieves over 60% accuracy on ImageNet using the standard linear evaluation protocol. This improves on prior results by over 4% (absolute). On Places205, using the representations learned on ImageNet, our model achieves 50% accuracy. This improves on prior results by 2% (absolute). When we extend our model to use mixture-based representations, segmentation behaviour emerges as a natural side-effect.