Density estimation, compression and data generation are crucial tasks in artificial intelligence. Variational Auto-Encoders (VAEs) constitute a single framework to achieve these goals. Here, we present a novel class of generative models, called self-supervised Variational Auto-Encoder (selfVAE), that utilizes deterministic and discrete variational posteriors. This class of models allows to perform both conditional and unconditional sampling, while simplifying the objective function. First, we use a single self-supervised transformation as a latent variable, where a transformation is either downscaling or edge detection. Next, we consider a hierarchical architecture, i.e., multiple transformations, and we show its benefits compared to the VAE. The flexibility of selfVAE in data reconstruction finds a particularly interesting use case in data compression tasks, where we can trade-off memory for better data quality, and vice-versa. We present performance of our approach on three benchmark image data (Cifar10, Imagenette64, and CelebA).

A formal description of the compositionality of neural networks is associated directly with the formal grammar-structure of the objects it seeks to represent. This formal grammar-structure specifies the kind of components that make up an object, and also the configurations they are allowed to be in. In other words, objects can be described as a parse-tree of its components -- a structure that can be seen as a candidate for building connection-patterns among neurons in neural networks. We present a formal grammar description of convolutional neural networks and capsule networks that shows how capsule networks can enforce such parse-tree structures, while CNNs do not. Specifically, we show that the entropy of routing coefficients in the dynamic routing algorithm controls this ability. Thus, we introduce the entropy of routing weights as a loss function for better compositionality among capsules. We show by experiments, on data with a compositional structure, that the use of this loss enables capsule networks to better detect changes in compositionality. Our experiments show that as the entropy of the routing weights increases, the ability to detect changes in compositionality reduces. We see that, without routing, capsule networks perform similar to convolutional neural networks in that both these models perform badly at detecting changes in compositionality. Our results indicate that routing is an important part of capsule networks -- effectively answering recent work that has questioned its necessity. We also, by experiments on SmallNORB, CIFAR-10, and FashionMNIST, show that this loss keeps the accuracy of capsule network models comparable to models that do not use it .

Sequential sensor data is generated in a wide variety of practical applications. A fundamental challenge involves learning effective classifiers for such sequential data. While deep learning has led to impressive performance gains in recent years in domains such as speech, this has relied on the availability of large datasets of sequences with high-quality labels. In many applications, however, the associated class labels are often extremely limited, with precise labelling/segmentation being too expensive to perform at a high volume. However, large amounts of unlabeled data may still be available. In this paper we propose a novel framework for semi-supervised learning in such contexts. In an unsupervised manner, change point detection methods can be used to identify points within a sequence corresponding to likely class changes. We show that change points provide examples of similar/dissimilar pairs of sequences which, when coupled with labeled, can be used in a semi-supervised classification setting. Leveraging the change points and labeled data, we form examples of similar/dissimilar sequences to train a neural network to learn improved representations for classification. We provide extensive synthetic simulations and show that the learned representations are superior to those learned through an autoencoder and obtain improved results on both simulated and real-world human activity recognition datasets.

This paper studies the density priors for independent vector analysis (IVA) with convolutive speech mixture separation as the exemplary application. Most existing source priors for IVA are too simplified to capture the fine structures of speeches. Here, we first time show that it is possible to efficiently estimate the derivative of speech density with universal approximators like deep neural networks (DNN) by optimizing certain proxy separation related performance indices. Experimental results suggest that the resultant neural network density priors consistently outperform previous ones in convergence speed for online implementation and signal-to-interference ratio (SIR) for batch implementation.

Discriminative deep neural networks (DNNs) do well at classifying input associated with the classes they have been trained on. However, out-of-distribution (OOD) input poses a great challenge to such models and consequently represents a major risk when these models are used in safety-critical systems. In the last two years, extensive research has been performed in the domain of OOD detection. This research has relied mainly on training the model with OOD data or requiring additional computation for OOD detection. Such methods may not be applicable in many real world use cases. In this paper, we propose GLOD -- Gaussian likelihood out of distribution detector -- an extended DNN classifier capable of efficiently detecting OOD samples with no additional runtime overhead and without auxiliary training data. GLOD uses a layer that models the Gaussian density function of the trained classes. The layer outputs are used to estimate a Log-Likelihood Ratio which is employed to detect OOD samples. We evaluate GLOD's detection performance on SVHN, CIFAR-10 and CIFAR-100.

Recurrent neural networks (RNNs) due to their ability to process sequences of data have found applications in many fields of science, engineering and humanities, including speech, handwriting and human action recognition, automatic translation, robot control etc. One of the RNN application is time series prediction used in analysis of business and financial data, anomaly detection, weather forecast. A large number of different architectures were discussed recently and the flow of new modifications of standard RNN continues to increase and all these architectures share some common features inherited from the basic systems. Trajectory prediction based on incomplete or noisy data is one of the most amazing features of organism brains that allows living creatures to survive in complex and mostly unfriendly environment. A large number of mathematical algorithms developed for this purpose have many applications in multiple engineering field, e.g., development of guidance systems, self-driving vehicles, motor control etc. [1]. It was shown that when the input signal represents a chaotic dynamics (in discrete or discretized continuous setting) RNNs indeed predict chaotic attractor for some number of steps and then the predicted trajectories diverge from the actual ones [2-4]. This result seems natural as it reflects an important property of chaotic dynamics - extremely high sensitivity of chaotic systems to small perturbations in initial conditions. What does happen when a trajectory is perturbed by external noise of specific statistics, e.g., white noise?

The last decade has witnessed a boom of deep learning research and applications achieving state-of-the-art results in various domains. However, most advances have been established empirically, and their theoretical analysis remains lacking. One major issue is that our current interpretation of neural networks (NNs) as function approximators is too generic to support in-depth analysis. In this paper, we remedy this by proposing an alternative interpretation that identifies NNs as chain graphs (CGs) and feed-forward as an approximate inference procedure. The CG interpretation specifies the nature of each NN component within the rich theoretical framework of probabilistic graphical models, while at the same time remains general enough to cover real-world NNs with arbitrary depth, multi-branching and varied activations, as well as common structures including convolution / recurrent layers, residual block and dropout. We demonstrate with concrete examples that the CG interpretation can provide novel theoretical support and insights for various NN techniques, as well as derive new deep learning approaches such as the concept of partially collapsed feed-forward inference. It is thus a promising framework that deepens our understanding of neural networks and provides a coherent theoretical formulation for future deep learning research.

We address the problem of learning reusable state representations from streaming high-dimensional observations. This is important for areas like Reinforcement Learning (RL), which yields non-stationary data distributions during training. We make two key contributions. First, we propose an evaluation suite that measures alignment between latent and true low-dimensional states. We benchmark several widely used unsupervised learning approaches. This uncovers the strengths and limitations of existing approaches that impose additional constraints/objectives on the latent space. Our second contribution is a unifying mathematical formulation for learning latent relations. We learn analytic relations on source domains, then use these relations to help structure the latent space when learning on target domains. This formulation enables a more general, flexible and principled way of shaping the latent space. It formalizes the notion of learning independent relations, without imposing restrictive simplifying assumptions or requiring domain-specific information. We present mathematical properties, concrete algorithms for implementation and experimental validation of successful learning and transfer of latent relations.

Fine-tuning pre-trained transformer-based language models such as BERT has become a common practice dominating leaderboards across various NLP benchmarks. Despite the strong empirical performance of fine-tuned models, fine-tuning is an unstable process: training the same model with multiple random seeds can result in a large variance of the task performance. Previous literature (Devlin et al., 2019; Lee et al., 2020; Dodge et al., 2020) identified two potential reasons for the observed instability: catastrophic forgetting and small size of the fine-tuning datasets. In this paper, we show that both hypotheses fail to explain the fine-tuning instability. We analyze BERT, RoBERTa, and ALBERT, finetuned on three commonly used datasets from the GLUE benchmark, and show that the observed instability is caused by optimization difficulties that lead to vanishing gradients. Additionally, we show that the remaining variance of the downstream task performance can be attributed to differences in generalization where fine-tuned models with the same training loss exhibit noticeably different test performance. Based on our analysis, we present a simple but strong baseline that makes fine-tuning BERTbased models significantly more stable than the previously proposed approaches. Pre-trained transformer-based masked language models such as BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and ALBERT (Lan et al., 2020) have had a dramatic impact on the NLP landscape in the recent year. The standard recipe for using such models typically involves training a pretrained model for a few epochs on a supervised downstream dataset, which is known as fine-tuning.

We consider the problem of lossy image compression with deep latent variable models. State-of-the-art methods build on hierarchical variational autoencoders (VAEs) and learn inference networks to predict a compressible latent representation of each data point. Drawing on the variational inference perspective on compression, we identify three approximation gaps which limit performance in the conventional approach: (i) an amortization gap, (ii) a discretization gap, and (iii) a marginalization gap. We propose improvements to each of these three shortcomings based on ideas related to iterative inference, stochastic annealing for discrete optimization, and bits-back coding, resulting in the first application of bits-back coding to lossy compression. In our experiments, which include extensive baseline comparisons and ablation studies, we achieve new state-of-the-art performance on lossy image compression using an established VAE architecture, by changing only the inference method.