We present an approach to domain adaptation, addressing the case where data from the source domain is abundant, labelled data from the target domain is limited or non-existent, and a small amount of paired source-target data is available. The method is designed for developing deep learning models that detect the presence of medical conditions based on data from consumer-grade portable biosensors. It addresses some of the key problems in this area, namely, the difficulty of acquiring large quantities of clinically labelled data from the biosensor, and the noise and ambiguity that can affect the clinical labels. The idea is to pre-train an expressive model on a large dataset of labelled recordings from a sensor modality for which data is abundant, and then to adapt the model's lower layers so that its predictions on the target modality are similar to the original model's on paired examples from the source modality. We show that the pre-trained model's predictions provide a substantially better learning signal than the clinician-provided labels, and that this teacher-student technique significantly outperforms both a naive application of supervised deep learning and a label-supervised version of domain adaptation on a synthetic dataset and in a real-world case study on sleep apnea. By reducing the volume of data required and obviating the need for labels, our approach should reduce the cost associated with developing high-performance deep learning models for biosensors.

Learning to control robots without requiring models has been a long-term goal, promising diverse and novel applications. Yet, reinforcement learning has only achieved limited impact on real-time robot control due to its high demand of real-world interactions. In this work, by leveraging a learnt probabilistic model of drone dynamics, we achieve human-like quadrotor control through model-based reinforcement learning. No prior knowledge of the flight dynamics is assumed; instead, a sequential latent variable model, used generatively and as an online filter, is learnt from raw sensory input. The controller and value function are optimised entirely by propagating stochastic analytic gradients through generated latent trajectories. We show that "learning to fly" can be achieved with less than 30 minutes of experience with a single drone, and can be deployed solely using onboard computational resources and sensors, on a self-built drone.

A key difficulty of cooperative decentralized planning lies in making accurate predictions about the decisions of other agents. In this paper we present a policy improvement operator for learning to plan in iterated cooperative multi-agent scenarios. At each application of our method, a selected agent learns an approximation of policies of its teammates from data from past simulations. Under the assumption of ideal function approximation, successive iterations of our algorithm are guaranteed to improve the policies, and eventually lead to convergence to a Nash equilibrium in a coordinate ascent manner. We combine the policy improvement operator with the decentralized Monte Carlo Tree Search planning method and demonstrate the application of the algorithm on several scenarios in the spatial task allocation problem introduced in (Claes et al., 2015). We show that deep learning and convolutional neural networks can be efficiently employed to produce policy approximators which exploit the spatial features of the problem, and that the proposed algorithm improves over the baseline planning performance for particularly challenging domain configurations.

Event extraction (EE) is one of the core information extraction tasks, whose purpose is to automatically identify and extract information about incidents and their actors from texts. This may be beneficial to several domains such as knowledge bases, question answering, information retrieval and summarization tasks, to name a few. The problem of extracting event information from texts is longstanding and usually relies on elaborately designed lexical and syntactic features, which, however, take a large amount of human effort and lack generalization. More recently, deep neural network approaches have been adopted as a means to learn underlying features automatically. However, existing networks do not make full use of syntactic features, which play a fundamental role in capturing very long-range dependencies. Also, most approaches extract each argument of an event separately without considering associations between arguments which ultimately leads to low efficiency, especially in sentences with multiple events. To address the two above-referred problems, we propose a novel joint event extraction framework that aims to extract multiple event triggers and arguments simultaneously by introducing shortest dependency path (SDP) in the dependency graph. We do this by eliminating irrelevant words in the sentence, thus capturing long-range dependencies. Also, an attention-based graph convolutional network is proposed, to carry syntactically related information along the shortest paths between argument candidates that captures and aggregates the latent associations between arguments; a problem that has been overlooked by most of the literature. Our results show a substantial improvement over state-of-the-art methods.

How well does a classic deep net architecture like AlexNet or VGG19 classify on a standard dataset such as CIFAR-10 when its "width"-- namely, number of channels in convolutional layers, and number of nodes in fully-connected internal layers -- is allowed to increase to infinity? Such questions have come to the forefront in the quest to theoretically understand deep learning and its mysteries about optimization and generalization. They also connect deep learning to notions such as Gaussian processes and kernels. A recent paper [Jacot et al., 2018] introduced the Neural Tangent Kernel (NTK) which captures the behavior of fully-connected deep nets in the infinite width limit trained by gradient descent; this object was implicit in some other recent papers. An attraction of such ideas is that a pure kernel-based method is used to capture the power of a fully-trained deep net of infinite width.

Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it was designed from first principles to support an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance.

As deep learning applications continue to become more diverse, an interesting question arises: Can general problem solving arise from jointly learning several such diverse tasks? To approach this question, deep multi-task learning is extended in this paper to the setting where there is no obvious overlap between task architectures. The idea is that any set of (architecture,task) pairs can be decomposed into a set of potentially related subproblems, whose sharing is optimized by an efficient stochastic algorithm. The approach is first validated in a classic synthetic multi-task learning benchmark, and then applied to sharing across disparate architectures for vision, NLP, and genomics tasks. It discovers regularities across these domains, encodes them into sharable modules, and combines these modules systematically to improve performance in the individual tasks.

Residual Network (ResNet) is undoubtedly a milestone in deep learning. ResNet is equipped with shortcut connections between layers, and exhibits efficient training using simple first order algorithms. Despite of the great empirical success, the reason behind is far from being well understood. In this paper, we study a two-layer non-overlapping convolutional ResNet. Training such a network requires solving a non-convex optimization problem with a spurious local optimum.

We propose to represent shapes as the deformation and combination of learnt elementary 3D structures. We demonstrate this decomposition in learnt elementary 3D structures is highly interpretable and leads to clear improvements in 3D shape generation and matching. More precisely, we present two complementary approaches to learn elementary structures in a deep learning framework: (i) continuous surface deformation learning and (ii) 3D structure points learning. Both approaches can be extended to abstract structures of higher dimensions for improved results. We evaluate our method on two very different tasks: ShapeNet objects reconstruction and dense correspondences estimation between human scans.

Efforts to understand the generalization mystery in deep learning have led to the belief that gradient-based optimization induces a form of implicit regularization, a bias towards models of low "complexity." We study the implicit regularization of gradient descent over deep linear neural networks for matrix completion and sensing, a model referred to as deep matrix factorization. Our first finding, supported by theory and experiments, is that adding depth to a matrix factorization enhances an implicit tendency towards low-rank solutions, oftentimes leading to more accurate recovery. Secondly, we present theoretical and empirical arguments questioning a nascent view by which implicit regularization in matrix factorization can be captured using simple mathematical norms. Our results point to the possibility that the language of standard regularizers may not be rich enough to fully encompass the implicit regularization brought forth by gradient-based optimization.