We consider the problem of domain generalization, namely, how to learn representations given data from a set of domains that generalize to data from a previously unseen domain. We propose the Domain Invariant Variational Autoencoder (DIVA), a generative model that tackles this problem by learning three independent latent subspaces, one for the domain, one for the class, and one for any residual variations. We highlight that due to the generative nature of our model we can also incorporate unlabeled data from known or previously unseen domains. To the best of our knowledge this has not been done before in a domain generalization setting. This property is highly desirable in fields like medical imaging where labeled data is scarce. We experimentally evaluate our model on the rotated MNIST benchmark and a malaria cell images dataset where we show that (i) the learned subspaces are indeed complementary to each other, (ii) we improve upon recent works on this task and (iii) incorporating unlabelled data can boost the performance even further.

We consider distributed optimization under communication constraints for training deep learning models. We propose a new algorithm, whose parameter updates rely on two forces: a regular gradient step, and a corrective direction dictated by the currently best-performing worker (leader). Our method differs from the parameter-averaging scheme EASGD in a number of ways: (i) our objective formulation does not change the location of stationary points compared to the original optimization problem; (ii) we avoid convergence decelerations caused by pulling local workers descending to different local minima to each other (i.e. to the average of their parameters); (iii) our update by design breaks the curse of symmetry (the phenomenon of being trapped in poorly generalizing sub-optimal solutions in symmetric non-convex landscapes); and (iv) our approach is more communication efficient since it broadcasts only parameters of the leader rather than all workers. We provide theoretical analysis of the batch version of the proposed algorithm, which we call Leader Gradient Descent (LGD), and its stochastic variant (LSGD). Finally, we implement an asynchronous version of our algorithm and extend it to the multi-leader setting, where we form groups of workers, each represented by its own local leader (the best performer in a group), and update each worker with a corrective direction comprised of two attractive forces: one to the local, and one to the global leader (the best performer among all workers). The multi-leader setting is well-aligned with current hardware architecture, where local workers forming a group lie within a single computational node and different groups correspond to different nodes. For training convolutional neural networks, we empirically demonstrate that our approach compares favorably to state-of-the-art baselines.

Semi-supervised learning (SSL) provides a powerful framework for leveraging unlabeled data when labels are limited or expensive to obtain. SSL algorithms based on deep neural networks have recently proven successful on standard benchmark tasks. However, many of them have thus far been either inflexible, inefficient or non-scalable. This paper explores recently developed contrastive predictive coding technique to improve discriminative power of deep learning models when a large portion of labels are absent. Two models, cpc-SSL and a class conditional variant~(ccpc-SSL) are presented. They effectively exploit the unlabeled data by extracting shared information between different parts of the (high-dimensional) data. The proposed approaches are inductive, and scale well to very large datasets like ImageNet, making them good candidates in real-world large scale applications.

By seeking the narrowest prediction intervals (PIs) that satisfy the specified coverage probability requirements, the recently proposed quality-based PI learning principle can extract high-quality PIs that better summarize the predictive certainty in regression tasks, and has been widely applied to solve many practical problems. Currently, the state-of-the-art quality-based PI estimation methods are based on deep neural networks or linear models. In this paper, we propose Highest Density Interval Regression Forest (HDI-Forest), a novel quality-based PI estimation method that is instead based on Random Forest. HDI-Forest does not require additional model training, and directly reuses the trees learned in a standard Random Forest model. By utilizing the special properties of Random Forest, HDI-Forest could efficiently and more directly optimize the PI quality metrics. Extensive experiments on benchmark datasets show that HDI-Forest significantly outperforms previous approaches, reducing the average PI width by over 30\% while achieving the same or better coverage probability.

Knowledge base is one of the main forms to represent information in a structured way. A knowledge base typically consists of Resource Description Frameworks (RDF) triples which describe the entities and their relations. Generating natural language description of the knowledge base is an important task in NLP, which has been formulated as a conditional language generation task and tackled using the sequence-to-sequence framework. Current works mostly train the language models by maximum likelihood estimation, which tends to generate lousy sentences. In this paper, we argue that such a problem of maximum likelihood estimation is intrinsic, which is generally irrevocable via changing network structures. Accordingly, we propose a novel Triple-to-Text (T2T) framework, which approximately optimizes the inverse Kullback-Leibler (KL) divergence between the distributions of the real and generated sentences. Due to the nature that inverse KL imposes large penalty on fake-looking samples, the proposed method can significantly reduce the probability of generating low-quality sentences. Our experiments on three real-world datasets demonstrate that T2T can generate higher-quality sentences and outperform baseline models in several evaluation metrics.

Tensor decomposition, a collection of factorization techniques for multidimensional arrays, are among the most general and powerful tools for scientific analysis. However, because of their increasing size, today's data sets require more complex tensor decomposition involving factorization with multiple matrices and diagonal tensors such as DEDICOM or PARATUCK2. Traditional tensor resolution algorithms such as Stochastic Gradient Descent (SGD), Non-linear Conjugate Gradient descent (NCG) or Alternating Least Square (ALS), cannot be easily applied to complex tensor decomposition or often lead to poor accuracy at convergence. We propose a new resolution algorithm, called VecHGrad, for accurate and efficient stochastic resolution over all existing tensor decomposition, specifically designed for complex decomposition. VecHGrad relies on gradient, Hessian-vector product and adaptive line search to ensure the convergence during optimization. Our experiments on five real-world data sets with the state-of-the-art deep learning gradient optimization models show that VecHGrad is capable of converging considerably faster because of its superior theoretical convergence rate per step. Therefore, VecHGrad targets as well deep learning optimizer algorithms. The experiments are performed for various tensor decomposition including CP, DEDICOM and PARATUCK2. Although it involves a slightly more complex update rule, VecHGrad's runtime is similar in practice to that of gradient methods such as SGD, Adam or RMSProp.

In this manuscript we propose two objective terms for neural image compression: a compression objective and a cycle loss. These terms are applied on the encoder output of an autoencoder and are used in combination with reconstruction losses. The compression objective encourages sparsity and low entropy in the activations. The cycle loss term represents the distortion between encoder outputs computed from the original image and from the reconstructed image (code-domain distortion). We train different autoencoders by using the compression objective in combination with different losses: a) MSE, b) MSE and MSSSIM, c) MSE, MS-SSIM and cycle loss. We observe that images encoded by these differently-trained autoencoders fall into different points of the perception-distortion curve (while having similar bit-rates). In particular, MSE-only training favors low image-domain distortion, whereas cycle loss training favors high perceptual quality.

Recent methods for generating novel molecules use graph representations of molecules and employ various forms of graph convolutional neural networks for inference. However, training requires solving an expensive graph isomorphism problem, which previous approaches do not address or solve only approximately. In this work, we propose LF-MolGAN, a likelihood-free approach for de novo molecule generation that avoids explicitly computing a reconstruction loss. Our approach extends generative adversarial networks by including an adversarial cycle-consistency loss to implicitly enforce the reconstruction property. To capture properties unique to molecules, such as valence, we extend Graph Isomorphism Network to multi-graphs. To quantify the performance of models, we propose to compute the distance between distributions of physicochemical properties with the 1-Wasserstein distance. We demonstrate that LF-MolGAN more accurately learns the distribution over the space of molecules than all baselines. Moreover, it can be utilized for drug discovery by efficiently searching the space of molecules using molecules' continuous latent representation.

Attention mechanisms and non-local mean operations in general are key ingredients in many state-of-the-art deep learning techniques. In particular, the Transformer model based on multi-head self-attention has recently achieved great success in natural language processing and computer vision. However, the vanilla algorithm computing the Transformer of an image with n pixels has O(n^2) complexity, which is often painfully slow and sometimes prohibitively expensive for large-scale image data. In this paper, we propose a fast randomized algorithm --- SCRAM --- that only requires O(n log(n)) time to produce an image attention map. Such a dramatic acceleration is attributed to our insight that attention maps on real-world images usually exhibit (1) spatial coherence and (2) sparse structure. The central idea of SCRAM is to employ PatchMatch, a randomized correspondence algorithm, to quickly pinpoint the most compatible key (argmax) for each query first, and then exploit that knowledge to design a sparse approximation to non-local mean operations. Using the argmax (mode) to dynamically construct the sparse approximation distinguishes our algorithm from all of the existing sparse approximate methods and makes it very efficient. Moreover, SCRAM is a broadly applicable approximation to any non-local mean layer in contrast to some other sparse approximations that can only approximate self-attention. Our preliminary experimental results suggest that SCRAM is indeed promising for speeding up or scaling up the computation of attention maps in the Transformer.

It remains a puzzle that why deep neural networks (DNNs), with more parameters than samples, often generalize well. An attempt of understanding this puzzle is to discover implicit biases underlying the training process of DNNs, such as the Frequency Principle (F-Principle), i.e., DNNs often fit target functions from low to high frequencies. Inspired by the F-Principle, we propose an effective model of linear F-Principle (LFP) dynamics which accurately predicts the learning results of two-layer ReLU neural networks (NNs) of large widths. This LFP dynamics is rationalized by a linearized mean field residual dynamics of NNs. Importantly, the long-time limit solution of this LFP dynamics is equivalent to the solution of a constrained optimization problem explicitly minimizing an FP-norm, in which higher frequencies of feasible solutions are more heavily penalized. Using this optimization formulation, an a priori estimate of the generalization error bound is provided, revealing that a higher FP-norm of the target function increases the generalization error. Overall, by explicitizing the implicit bias of the F-Principle as an explicit penalty for two-layer NNs, our work makes a step towards a quantitative understanding of the learning and generalization of general DNNs.