These schemes are often heuristic and fixed over the course of training. We empirically observe that the statistics of gradients of deep models change during the training. Motivated by this observation, we introduce two adaptive quantization schemes, ALQ and AMQ. In both schemes, processors update their compression schemes in parallel by efficiently computing sufficient statistics of a parametric distribution. We improve the validation accuracy by almost 2% on CIFAR-10 and 1% on ImageNet in challenging low-cost communication setups. Our adaptive methods are also significantly more robust to the choice of hyperparameters.

Human intelligence is characterized not only by the capacity to learn complex skills, but the ability to rapidly adapt and acquire new skills within an ever-changing environment. In this work we study how the learning of modular solutions can allow for effective generalization to both unseen and potentially differently distributed data. Our main postulate is that the combination of task segmentation, modular learning and memory-based ensembling can give rise to generalization on an exponentially growing number of unseen tasks. We provide a concrete instantiation of this idea using a combination of: (1) the Forget-Me-Not Process, for task segmentation and memory based ensembling; and (2) Gated Linear Networks, which in contrast to contemporary deep learning techniques use a modular and local learning mechanism. We demonstrate that this system exhibits a number of desirable continual learning properties: robustness to catastrophic forgetting, no negative transfer and increasing levels of positive transfer as more tasks are seen. We show competitive performance against both offline and online methods on standard continual learning benchmarks.

The most successful methods such as ReLU transfer functions, batch normalization, Xavier initialization, dropout, learning rate decay, or dynamic optimizers, have become standards in the field due, particularly, to their ability to increase the performance of Neural Networks (NNs) significantly and in almost all situations. Here we present a new method to calculate the gradients while training NNs, and show that it significantly improves final performance across architectures, data-sets, hyper-parameter values, training length, and model sizes, including when it is being combined with other common performance-improving methods (such as the ones mentioned above). Besides being effective in the wide array situations that we have tested, the increase in performance (e.g. F1) it provides is as high or higher than this one of all the other widespread performance-improving methods that we have compared against. We call our method Population Gradients (PG), and it consists on using a population of NNs to calculate a non-local estimation of the gradient, which is closer to the theoretical exact gradient (i.e. this one obtainable only with an infinitely big data-set) of the error function than the empirical gradient (i.e. this one obtained with the real finite data-set).

Most popular optimizers for deep learning can be broadly categorized as adaptive methods (e.g. Adam) and accelerated schemes (e.g. stochastic gradient descent (SGD) with momentum). For many models such as convolutional neural networks (CNNs), adaptive methods typically converge faster but generalize worse compared to SGD; for complex settings such as generative adversarial networks (GANs), adaptive methods are typically the default because of their stability.We propose AdaBelief to simultaneously achieve three goals: fast convergence as in adaptive methods, good generalization as in SGD, and training stability. The intuition for AdaBelief is to adapt the stepsize according to the "belief" in the current gradient direction. Viewing the exponential moving average (EMA) of the noisy gradient as the prediction of the gradient at the next time step, if the observed gradient greatly deviates from the prediction, we distrust the current observation and take a small step; if the observed gradient is close to the prediction, we trust it and take a large step. We validate AdaBelief in extensive experiments, showing that it outperforms other methods with fast convergence and high accuracy on image classification and language modeling. Specifically, on ImageNet, AdaBelief achieves comparable accuracy to SGD. Furthermore, in the training of a GAN on Cifar10, AdaBelief demonstrates high stability and improves the quality of generated samples compared to a well-tuned Adam optimizer. Code is available at https://github.com/juntang-zhuang/Adabelief-Optimizer

We prove a non-asymptotic distribution-independent lower bound for the expected mean squared generalization error caused by label noise in ridgeless linear regression. Our lower bound generalizes a similar known result to the overparameterized (interpolating) regime. In contrast to most previous works, our analysis applies to a broad class of input distributions with almost surely full-rank feature matrices, which allows us to cover various types of deterministic or random feature maps. Our lower bound is asymptotically sharp and implies that in the presence of label noise, ridgeless linear regression does not perform well around the interpolation threshold for any of these feature maps. We analyze the imposed assumptions in detail and provide a theory for analytic (random) feature maps. Using this theory, we can show that our assumptions are satisfied for input distributions with a (Lebesgue) density and feature maps given by random deep neural networks with analytic activation functions like sigmoid, tanh, softplus or GELU. As further examples, we show that feature maps from random Fourier features and polynomial kernels also satisfy our assumptions. We complement our theory with further experimental and analytic results.

Forecasting is challenging since uncertainty resulted from exogenous factors exists. This work investigates the rank position forecasting problem in car racing, which predicts the rank positions at the future laps for cars. Among the many factors that bring changes to the rank positions, pit stops are critical but irregular and rare. We found existing methods, including statistical models, machine learning regression models, and state-of-the-art deep forecasting model based on encoder-decoder architecture, all have limitations in the forecasting. By elaborative analysis of pit stops events, we propose a deep model, RankNet, with the cause effects decomposition that modeling the rank position sequence and pit stop events separately. It also incorporates probabilistic forecasting to model the uncertainty inside each sub-model. Through extensive experiments, RankNet demonstrates a strong performance improvement over the baselines, e.g., MAE improves more than 10% consistently, and is also more stable when adapting to unseen new data. Details of model optimization, performance profiling are presented. It is promising to provide useful forecasting tools for the car racing analysis and shine a light on solutions to similar challenging issues in general forecasting problems.

Natural Gradient Descent (NGD) helps to accelerate the convergence of gradient descent dynamics, but it requires approximations in large-scale deep neural networks because of its high computational cost. Empirical studies have confirmed that some NGD methods with approximate Fisher information converge sufficiently fast in practice. Nevertheless, it remains unclear from the theoretical perspective why and under what conditions such heuristic approximations work well. In this work, we reveal that, under specific conditions, NGD with approximate Fisher information achieves the same fast convergence to global minima as exact NGD. We consider deep neural networks in the infinite-width limit, and analyze the asymptotic training dynamics of NGD in function space via the neural tangent kernel. In the function space, the training dynamics with the approximate Fisher information are identical to those with the exact Fisher information, and they converge quickly. The fast convergence holds in layer-wise approximations; for instance, in block diagonal approximation where each block corresponds to a layer as well as in block tri-diagonal and K-FAC approximations. We also find that a unit-wise approximation achieves the same fast convergence under some assumptions. All of these different approximations have an isotropic gradient in the function space, and this plays a fundamental role in achieving the same convergence properties in training. Thus, the current study gives a novel and unified theoretical foundation with which to understand NGD methods in deep learning.

We introduce Transductive Infomation Maximization (TIM) for few-shot learning. Our method maximizes the mutual information between the query features and their label predictions for a given few-shot task, in conjunction with a supervision loss based on the support set. Furthermore, we propose a new alternating-direction solver for our mutual-information loss, which substantially speeds up transductive-inference convergence over gradient-based optimization, while yielding similar accuracy. TIM inference is modular: it can be used on top of any base-training feature extractor. Following standard transductive few-shot settings, our comprehensive experiments demonstrate that TIM outperforms state-of-the-art methods significantly across various datasets and networks, while used on top of a fixed feature extractor trained with simple cross-entropy on the base classes, without resorting to complex meta-learning schemes. It consistently brings between 2% and 5% improvement in accuracy over the best performing method, not only on all the well-established few-shot benchmarks but also on more challenging scenarios,with domain shifts and larger numbers of classes.

Recently, graph neural networks (GNNs) become a principal research direction to learn low-dimensional continuous embeddings of nodes and graphs to predict node and graph labels, respectively. However, Euclidean embeddings have high distortion when using GNNs to model complex graphs such as social networks. Furthermore, existing GNNs are not very efficient with the high number of model parameters when increasing the number of hidden layers. Therefore, we move beyond the Euclidean space to a hyper-complex vector space to improve graph representation quality and reduce the number of model parameters. To this end, we propose quaternion graph neural networks (QGNN) to generalize GCNs within the Quaternion space to learn quaternion embeddings for nodes and graphs. The Quaternion space, a hyper-complex vector space, provides highly meaningful computations through Hamilton product compared to the Euclidean and complex vector spaces. As a result, our QGNN can reduce the model size up to four times and enhance learning better graph representations. Experimental results show that the proposed QGNN produces state-of-the-art accuracies on a range of well-known benchmark datasets for three downstream tasks, including graph classification, semi-supervised node classification, and text (node) classification.

While various complexity measures for deep neural networks exist, specifying an appropriate measure capable of predicting and explaining generalization in deep networks has proven challenging. We propose Neural Complexity (NC), a meta-learning framework for predicting generalization. Our model learns a scalar complexity measure through interactions with many heterogeneous tasks in a datadriven way. The trained NC model can be added to the standard training loss to regularize any task learner in a standard supervised learning scenario. We contrast NC's approach against existing manually-designed complexity measures and other meta-learning models, and we validate NC's performance on multiple regression and classification tasks.