System identification of complex and nonlinear systems is a central problem for model predictive control and model-based reinforcement learning. Despite their complexity, such systems can often be approximated well by a set of linear dynamical systems if broken into appropriate subsequences. This mechanism not only helps us find good approximations of dynamics, but also gives us deeper insight into the underlying system. Leveraging Bayesian inference, Variational Autoencoders and Concrete relaxations, we show how to learn a richer and more meaningful state space, e.g. encoding joint constraints and collisions with walls in a maze, from partial and high-dimensional observations. This representation translates into a gain of accuracy of learned dynamics showcased on various simulated tasks.

Recurrent neural networks (RNNs) are omnipresent in sequence modeling tasks. Practical models usually consist of several layers of hundreds or thousands of neurons which are fully connected. This places a heavy computational and memory burden on hardware, restricting adoption in practical low-cost and low-power devices. Compared to fully convolutional models, the costly sequential operation of RNNs severely hinders performance on parallel hardware. This paper challenges the convention of full connectivity in RNNs. We study structurally sparse RNNs, showing that they are well suited for acceleration on parallel hardware, with a greatly reduced cost of the recurrent operations as well as orders of magnitude less recurrent weights. Extensive experiments on challenging tasks ranging from language modeling and speech recognition to video action recognition reveal that structurally sparse RNNs achieve competitive performance as compared to fully-connected networks. This allows for using large sparse RNNs for a wide range of real-world tasks that previously were too costly with fully connected networks.

Reduced precision computation for deep neural networks is one of the key areas addressing the widening'compute gap' driven by an exponential growth in model size. In recent years, deep learning training has largely migrated to 16-bit precision, with significant gains in performance and energy efficiency. However, attempts to train DNNs at 8-bit precision have met with significant challenges because of the higher precision and dynamic range requirements of back-propagation. In this paper, we propose a method to train deep neural networks using 8-bit floating point representation for weights, activations, errors, and gradients. In addition to reducing compute precision, we also reduced the precision requirements for the master copy of weights from 32-bit to 16-bit. We demonstrate state-of-the-art accuracy across multiple data sets (imagenet-1K, WMT16) and a broader set of workloads (Resnet-18/34/50, GNMT, Transformer) than previously reported. We propose an enhanced loss scaling method to augment the reduced subnormal range of 8-bit floating point for improved error propagation. We also examine the impact of quantization noise on generalization and propose a stochastic rounding technique to address gradient noise. As a result of applying all these techniques, we report slightly higher validation accuracy compared to full precision baseline.

We devise a learning algorithm for possibly nonsmooth deep neural networks featuring inertia and Newtonian directional intelligence only by means of a back-propagation oracle. Our algorithm, called INDIAN, has an appealing mechanical interpretation, making the role of its two hyperparameters transparent. An elementary phase space lifting allows both for its implementation and its theoretical study under very general assumptions. We handle in particular a stochastic version of our method (which encompasses usual mini-batch approaches) for nonsmooth activation functions (such as ReLU). Our algorithm shows high efficiency and reaches state of the art on image classification problems.

Many applications of machine learning in science and medicine, including molecular property and protein function prediction, can be cast as problems of predicting some properties of graphs, where having good graph representations is critical. However, two key challenges in these domains are (1) extreme scarcity of labeled data due to expensive lab experiments, and (2) needing to extrapolate to test graphs that are structurally different from those seen during training. In this paper, we explore pre-training to address both of these challenges. In particular, working with Graph Neural Networks (GNNs) for representation learning of graphs, we wish to obtain node representations that (1) capture similarity of nodes' network neighborhood structure, (2) can be composed to give accurate graph-level representations, and (3) capture domain-knowledge. To achieve these goals, we propose a series of methods to pre-train GNNs at both the node-level and the graph-level, using both unlabeled data and labeled data from related auxiliary supervised tasks. We perform extensive evaluation on two applications, molecular property and protein function prediction. We observe that performing only graph-level supervised pre-training often leads to marginal performance gain or even can worsen the performance compared to non-pre-trained models. On the other hand, effectively combining both node- and graph-level pre-training techniques significantly improves generalization to out-of-distribution graphs, consistently outperforming non-pre-trained GNNs across 8 datasets in molecular property prediction (resp. 40 tasks in protein function prediction), with the average ROC-AUC improvement of 7.2% (resp. 11.7%).

Low bit-width integer weights and activations are very important for efficient inference, especially with respect to lower power consumption. We propose Monte Carlo methods to quantize the weights and activations of pre-trained neural networks without any re-training. By performing importance sampling we obtain quantized low bit-width integer values from full-precision weights and activations. The precision, sparsity, and complexity are easily configurable by the amount of sampling performed. Our approach, called Monte Carlo Quantization (MCQ), is linear in both time and space, with the resulting quantized, sparse networks showing minimal accuracy loss when compared to the original full-precision networks. Our method either outperforms or achieves competitive results on multiple benchmarks compared to previous quantization methods that do require additional training.

In this work, we propose a Factorized Disentangler-Entangler Network (FDEN) that learns to decompose a latent representation into two mutually independent factors, namely, identity and style. Given a latent representation, the proposed framework draws a set of interpretable factors aligned to identity of an observed data and learns to maximize the independency between these factors. Our work introduces an idea for a plug-in method to disentangle latent representations of already learned deep models with no affect to the model. In doing so, it brings the possibilities of extending state-of-the-art models to solve different tasks and also maintain the performance of its original task. Thus, FDEN is naturally applicable to jointly perform multiple tasks such as few-shot learning and image-to-image translation in a single framework. We show the effectiveness of our work in disentangling a latent representation in two parts. First, to evaluate the alignment of factor to an identity, we perform few-shot learning using only the aligned factor. Then, to evaluate the effectiveness of decomposition of latent representation and to show that plugin method does not affect the deep model in its performance, we perform image-to-image style transfer by mixing factors of different images. These evaluations show, qualitatively and quantitatively, that our proposed framework can indeed disentangle a latent representation.

Batch norm is a standard component of modern deep neural networks, and tends to make the training process less sensitive to the choice of hyperparameters in many cases [13]. While ease of training is desirable for model developers, an important concern among stakeholders is that of model robustness to plausible, previously unseen inputs during deployment. The adversarial examples phenomenon has exposed unstable predictions across state-of-the-art models [27]. This has led to a variety of methods that aim to improve robustness, but doing so effectively remains a challenge [1, 20, 11, 14]. We believe that a prerequisite to developing methods that increase robustness is an understanding of factors that reduce it. Approaches for improving robustness often begin with existing neural network architectures--that use batch norm--and patching them against specific attacks, e.g.,

Overparameterized models that interpolate training data often display surprisingly good generalization properties. Specifically, minimum norm solutions have been shown to generalize well in the overparameterized, interpolating regime. This paper introduces a new framework for active learning based on the notion of minimum norm interpolators. We analytically study its properties and behavior in the kernel-based setting, and present experimental studies with kernel methods and neural networks. In general, active learning algorithms adaptively select examples for labeling that (1) rule-out as many (incompatible) classifiers as possible at each step and/or (2) discover cluster structure in unlabeled data and label representative examples from each cluster. We show that our new active learning approach based on a minimum norm heuristic automatically exploits both these strategies. The success of deep learning systems has sparked interest in understanding how and why overparameterized models that interpolate the training data often display surprisingly good generalization properties [27, 11, 38, 12, 8, 1, 10, 23]. Notably, it is now understood that mininum norm solutions have the potential to generalize well in the overparameterized, interpolating regime [11, 9, 23, 25].

We focus on the important problem of emergency evacuation, which clearly could benefit from reinforcement learning that has been largely unaddressed. Emergency evacuation is a complex task which is difficult to solve with reinforcement learning, since an emergency situation is highly dynamic, with a lot of changing variables and complex constraints that makes it difficult to train on. In this paper, we propose the first fire evacuation environment to train reinforcement learning agents for evacuation planning. The environment is modelled as a graph capturing the building structure. It consists of realistic features like fire spread, uncertainty and bottlenecks. We have implemented the environment in the OpenAI gym format, to facilitate future research. We also propose a new reinforcement learning approach that entails pretraining the network weights of a DQN based agents to incorporate information on the shortest path to the exit. We achieved this by using tabular Q-learning to learn the shortest path on the building model's graph. This information is transferred to the network by deliberately overfitting it on the Q-matrix. Then, the pretrained DQN model is trained on the fire evacuation environment to generate the optimal evacuation path under time varying conditions. We perform comparisons of the proposed approach with state-of-the-art reinforcement learning algorithms like PPO, VPG, SARSA, A2C and ACKTR. The results show that our method is able to outperform state-of-the-art models by a huge margin including the original DQN based models. Finally, we test our model on a large and complex real building consisting of 91 rooms, with the possibility to move to any other room, hence giving 8281 actions. We use an attention based mechanism to deal with large action spaces. Our model achieves near optimal performance on the real world emergency environment.