Key features of mental illnesses are reflected in speech. Our research focuses on designing a multimodal deep learning structure that automatically extracts salient features from recorded speech samples for predicting various mental disorders including depression, bipolar, and schizophrenia. We adopt a variety of pre-trained models to extract embeddings from both audio and text segments. We use several state-of-the-art embedding techniques including BERT, FastText, and Doc2VecC for the text representation learning and WaveNet and VGG-ish models for audio encoding. We also leverage huge auxiliary emotion-labeled text and audio corpora to train emotion-specific embeddings and use transfer learning in order to address the problem of insufficient annotated multimodal data available. All these embeddings are then combined into a joint representation in a multimodal fusion layer and finally a recurrent neural network is used to predict the mental disorder. Our results show that mental disorders can be predicted with acceptable accuracy through multimodal analysis of clinical interviews.

We prove the precise scaling, at finite depth and width, for the mean and variance of the neural tangent kernel (NTK) in a randomly initialized ReLU network. The standard deviation is exponential in the ratio of network depth to width. Thus, even in the limit of infinite overparameterization, the NTK is not deterministic if depth and width simultaneously tend to infinity. Moreover, we prove that for such deep and wide networks, the NTK has a non-trivial evolution during training by showing that the mean of its first SGD update is also exponential in the ratio of network depth to width. This is sharp contrast to the regime where depth is fixed and network width is very large. Our results suggest that, unlike relatively shallow and wide networks, deep and wide ReLU networks are capable of learning data-dependent features even in the so-called lazy training regime.

Deep neural networks are often called black-boxes due to their difficult-to-interpret decisions. This is characteristic of a deeper trend in machine learning, where predictive performance typically comes at the cost of interpretability. In some domains, such as image-based diagnostic tasks, understanding the reasons behind machine generated predictions is vital in assessing trust. In this study, we introduce novel designs of capsule networks to provide explainable diagnoses. Our proposed deep explainable capsule architecture, called DX-Caps, can encode high-level visual attributes within the vectors of capsules in order to simultaneously produce malignancy predictions for lung cancer as well as approximations of six visually-interpretable attributes, used by radiologists to explain their predictions. To reduce parameter and memory burden of this deeper network, we introduce a new capsule-average pooling function. With this simple, but fundamental addition, capsule networks can be designed in a deeper fashion than was possible before. Our overall approach can be characterized as multi-task learning; we learn to approximate the six high-level visual attributes of a lung nodule within the vectors of our uniquely constructed deep capsule network, while simultaneously segmenting the nodule and predicting its malignancy potential (diagnosis). Tested on over 1000 CT scans, our experimental results show that our proposed algorithm can approximate the visual attributes of lung nodules far better than a deep multi-path dense 3D CNN. The proposed network also achieves higher diagnostic accuracy than a baseline explainable capsule network X-Caps and CapsNet when applied to this task for the first time as well. To the best of our knowledge, this is the first study to investigate capsule networks for visual attribute prediction in general, and explainable medical image diagnosis in particular.

One of the central themes in the classification task is the estimation of class posterior probability at a new point $\bf{x}$. The vast majority of classifiers output a score for $\bf{x}$, which is monotonically related to the posterior probability via an unknown relationship. There are many attempts in the literature to estimate this latter relationship. Here, we provide a way to estimate the posterior probability without resorting to using classification scores. Instead, we vary the prior probabilities of classes in order to derive the ratio of pdf's at point $\bf{x}$, which is directly used to determine class posterior probabilities. We consider here the binary classification problem.

As modern deep networks become more complex, and get closer to human-like capabilities in certain domains, the question arises of how the representations and decision rules they learn compare to the ones in humans. In this work, we study representations of sentences in one such artificial system for natural language processing. We first present a diagnostic test dataset to examine the degree of abstract composable structure represented. Analyzing performance on these diagnostic tests indicates a lack of systematicity in the representations and decision rules, and reveals a set of heuristic strategies. We then investigate the effect of the training distribution on learning these heuristic strategies, and study changes in these representations with various augmentations to the training set. Our results reveal parallels to the analogous representations in people. We find that these systems can learn abstract rules and generalize them to new contexts under certain circumstances -- similar to human zero-shot reasoning. However, we also note some shortcomings in this generalization behavior -- similar to human judgment errors like belief bias. Studying these parallels suggests new ways to understand psychological phenomena in humans as well as informs best strategies for building artificial intelligence with human-like language understanding.

Off-policy evaluation (OPE) in reinforcement learning is notoriously difficult in long- and infinite-horizon settings due to diminishing overlap between behavior and target policies. In this paper, we study the role of Markovian, time-invariant, and ergodic structure in efficient OPE. We first derive the efficiency limits for OPE when one assumes each of these structures. This precisely characterizes the curse of horizon: in time-variant processes, OPE is only feasible in the near-on-policy setting, where behavior and target policies are sufficiently similar. But, in ergodic time-invariant Markov decision processes, our bounds show that truly-off-policy evaluation is feasible, even with only just one dependent trajectory, and provide the limits of how well we could hope to do. We develop a new estimator based on Double Reinforcement Learning (DRL) that leverages this structure for OPE. Our DRL estimator simultaneously uses estimated stationary density ratios and $q$-functions and remains efficient when both are estimated at slow, nonparametric rates and remains consistent when either is estimated consistently. We investigate these properties and the performance benefits of leveraging the problem structure for more efficient OPE.

There is a growing interest in large-scale machine learning and optimization over decentralized networks, e.g. in the context of multi-agent learning and federated learning. Due to the imminent need to alleviate the communication burden, the investigation of communication-efficient distributed optimization algorithms --- particularly for empirical risk minimization --- has flourished in recent years. A large faction of these algorithms have been developed for the master/slave setting, relying on the presence of a central parameter server that can communicate with all agents. This paper focuses on distributed optimization over the network-distributed or the decentralized setting, where each agent is only allowed to aggregate information from its neighbors over a network (namely, no centralized coordination is present). By properly adjusting the global gradient estimate via a tracking term, we develop a communication-efficient approximate Newton-type method, called Network-DANE, which generalizes DANE [Shamir et al., 2014] for decentralized networks. We establish linear convergence of Network-DANE for quadratic losses, which shed light on the impact of data homogeneity and network connectivity upon the rate of convergence. Our key algorithmic ideas can be applied, in a systematic manner, to obtain decentralized versions of other master/slave distributed algorithms. A notable example is our development of Network-SVRG, which employs stochastic variance reduction [Johnson and Zhang, 2013] at each agent to accelerate local computation. The proposed algorithms are built upon the primal formulation without resorting to the dual. Numerical evidence is provided to demonstrate the appealing performance of our algorithms over competitive baselines, in terms of both communication and computation efficiency.

Markov switching models (MSMs) are probabilistic models that employ multiple sets of parameters to describe different dynamic regimes that a time series may exhibit at different periods of time. The switching mechanism between regimes is controlled by unobserved random variables that form a first-order Markov chain. Explicit-duration MSMs contain additional variables that explicitly model the distribution of time spent in each regime. This allows to define duration distributions of any form, but also to impose complex dependence between the observations and to reset the dynamics to initial conditions. Models that focus on the first two properties are most commonly known as hidden semi-Markov models or segment models, whilst models that focus on the third property are most commonly known as changepoint models or reset models. In this monograph, we provide a description of explicit-duration modelling by categorizing the different approaches into three groups, which differ in encoding in the explicit-duration variables different information about regime change/reset boundaries. The approaches are described using the formalism of graphical models, which allows to graphically represent and assess statistical dependence and therefore to easily describe the structure of complex models and derive inference routines. The presentation is intended to be pedagogical, focusing on providing a characterization of the three groups in terms of model structure constraints and inference properties. The monograph is supplemented with a software package that contains most of the models and examples described. The material presented should be useful to both researchers wishing to learn about these models and researchers wishing to develop them further.

We study channel number reduction in combination with weight binarization (1-bit weight precision) to trim a convolutional neural network for a keyword spotting (classification) task. We adopt a group-wise splitting method based on the group Lasso penalty to achieve over 50 % channel sparsity while maintaining the network performance within 0.25 % accuracy loss. We show an effective three-stage procedure to balance accuracy and sparsity in network training. Keywords: Convolutional Neural Network · Channel Pruning · Weight Binarization · Classification. 1 Introduction Reducing complexity of neural networks while maintaining their performance is both fundamental and practical for resource limited platforms such as mobile phones. In this paper, we integrate two methods, namely channel pruning and weight quantization, to trim down the number of parameters for a keyword spotting convolutional neural network (CNN, [4]).

In recent years, Deep Neural Networks (DNNs) have had a drama tic impact on a variety of problems that were long considered very difficult, e. g., image classification and automatic language translation to name just a few. T he accuracy of modern DNNs in classification tasks is remarkable indeed. At the same time, attackers have devised powerful methods to construct specially-craf ted malicious inputs (often referred to as adversarial examples) that can trick DNNs into mis-classifying them. What is worse is that despite the many defense mechanis ms proposed to protect DNNs against adversarial attacks, attackers are of ten able to circumvent these defenses, rendering them useless. This state of affai rs is extremely worrying, especially since machine learning systems get adopted at sc ale. In this paper, we propose a scientific evaluation methodolog y aimed at assessing the quality, efficacy, robustness and efficiency of randomiz ed defenses to protect DNNs against adversarial examples. Using this methodology, we evaluate a variety of defense mechanisms. In addition, we also propose a d efense mechanism we call Randomly Perturbed Ensemble Neural Networks (RPENN s). W e provide a thorough and comprehensive evaluation of the considered d efense mechanisms against a white-box attacker model, six different adversar ial attack methods and using the ILSVRC2012 validation data set.