We study and quantify the generalization patterns of multitask learning (MTL) models for sequence labeling tasks. MTL models are trained to optimize a set of related tasks jointly. Although multitask learning has achieved improved performance in some problems, there are also tasks that lose performance when trained together. These mixed results motivate us to study the factors that impact the performance of MTL models. We note that theoretical bounds and convergence rates for MTL models exist, but they rely on strong assumptions such as task relatedness and the use of balanced datasets. To remedy these limitations, we propose the creation of a task simulator and the use of Symbolic Regression to learn expressions relating model performance to possible factors of influence. For MTL, we study the model performance against the number of tasks (T), the number of samples per task (n) and the task relatedness measured by the adjusted mutual information (AMI). In our experiments, we could empirically find formulas relating model performance with factors of sqrt(n), sqrt(T), which are equivalent to sound mathematical proofs in Maurer[2016], and we went beyond by discovering that performance relates to a factor of sqrt(AMI).

Undertaking causal inference with observational data is extremely useful across a wide range of domains including the development of medical treatments, advertisements and marketing, and policy making. There are two main challenges associated with undertaking causal inference using observational data: treatment assignment heterogeneity (i.e., differences between the treated and untreated groups), and an absence of counterfactual data (i.e. not knowing what would have happened if an individual who did get treatment, were instead to have not been treated). We address these two challenges by combining structured inference and targeted learning. To our knowledge, Targeted Variational AutoEncoder (TVAE) is the first method to incorporate targeted learning into deep latent variable models. Results demonstrate competitive and state of the art performance.

In this paper, we propose Minimalist Transfer Learning (MinTL) to simplify the system design process of task-oriented dialogue systems and alleviate the over-dependency on annotated data. MinTL is a simple yet effective transfer learning framework, which allows us to plug-and-play pre-trained seq2seq models, and jointly learn dialogue state tracking and dialogue response generation. Unlike previous approaches, which use a copy mechanism to "carryover" the old dialogue states to the new one, we introduce Levenshtein belief spans (Lev), that allows efficient dialogue state tracking with a minimal generation length. We instantiate our learning framework with two pre-trained backbones: T5 and BART, and evaluate them on MultiWOZ. Extensive experiments demonstrate that: 1) our systems establish new state-of-the-art results on end-to-end response generation, 2) MinTL-based systems are more robust than baseline methods in the low resource setting, and they achieve competitive results with only 20\% training data, and 3) Lev greatly improves the inference efficiency.

We develop a recurrent gamma belief network (rGBN) for radar automatic target recognition (RATR) based on high-resolution range profile (HRRP), which characterizes the temporal dependence across the range cells of HRRP. The proposed rGBN adopts a hierarchy of gamma distributions to build its temporal deep generative model. For scalable training and fast out-of-sample prediction, we propose the hybrid of a stochastic-gradient Markov chain Monte Carlo (MCMC) and a recurrent variational inference model to perform posterior inference. To utilize the label information to extract more discriminative latent representations, we further propose supervised rGBN to jointly model the HRRP samples and their corresponding labels. Experimental results on synthetic and measured HRRP data show that the proposed models are efficient in computation, have good classification accuracy and generalization ability, and provide highly interpretable multi-stochastic-layer latent structure.

Performing efficient inference on high dimensional discrete Bayesian Networks (BNs) is challenging. When using exact inference methods the space complexity can grow exponentially with the tree-width, thus making computation intractable. This paper presents a general purpose approximate inference algorithm, based on a new region belief approximation method, called Triplet Region Construction (TRC). TRC reduces the cluster space complexity for factorized models from worst-case exponential to polynomial by performing graph factorization and producing clusters of limited size. Unlike previous generations of region-based algorithms, TRC is guaranteed to converge and effectively addresses the region choice problem that bedevils other region-based algorithms used for BN inference. Our experiments demonstrate that it also achieves significantly more accurate results than competing algorithms.

Mixture models, as exemplified by the Gaussian mixture model (GMM), are widely used for approximating complex multi-modal distributions. They can also be viewed as a form of latent variable models that provide a flexible approach for statistical inference with heterogeneous data. To estimate the parameters of GMM, a standard approach is via the maximum likelihood principle. When the global optimum of the likelihood function can be computed, the statistical properties of the maximum likelihood estimator is relatively well studied, including its asymptotic consistency [25] and finite-sample error rates [5, 21, 14]. Much less understood are the computational challenges associated with estimating GMMs. The negative log-likelihood function of GMM is nonconvex and in general has multiple local minima. Standard iterative algorithms, such as Expectation-Maximization (EM) [9], are only guaranteed to converge to a local minimum [27, 16].

We address the question of characterizing and finding optimal representations for supervised learning. Traditionally, this question has been tackled using the Information Bottleneck, which compresses the inputs while retaining information about the targets, in a decoder-agnostic fashion. In machine learning, however, our goal is not compression but rather generalization, which is intimately linked to the predictive family or decoder of interest (e.g. linear classifier). We propose the Decodable Information Bottleneck (DIB) that considers information retention and compression from the perspective of the desired predictive family. As a result, DIB gives rise to representations that are optimal in terms of expected test performance and can be estimated with guarantees. Empirically, we show that the framework can be used to enforce a small generalization gap on downstream classifiers and to predict the generalization ability of neural networks.

We propose an Anderson Acceleration (AA) scheme for the adaptive Expectation-Maximization (EM) algorithm for unsupervised learning a finite mixture model from multivariate data (Figueiredo and Jain 2002). The proposed algorithm is able to determine the optimal number of mixture components autonomously, and converges to the optimal solution much faster than its non-accelerated version. The success of the AA-based algorithm stems from several developments rather than a single breakthrough (and without these, our tests demonstrate that AA fails catastrophically). To begin, we ensure the monotonicity of the likelihood function (a the key feature of the standard EM algorithm) with a recently proposed monotonicity-control algorithm (Henderson and Varahdan 2019), enhanced by a novel monotonicity test with little overhead. We propose nimble strategies for AA to preserve the positive definiteness of the Gaussian weights and covariance matrices strictly, and to conserve up to the second moments of the observed data set exactly. Finally, we employ a K-means clustering algorithm using the gap statistic to avoid excessively overestimating the initial number of components, thereby maximizing performance. We demonstrate the accuracy and efficiency of the algorithm with several synthetic data sets that are mixtures of Gaussians distributions of known number of components, as well as data sets generated from particle-in-cell simulations. Our numerical results demonstrate speed-ups with respect to non-accelerated EM of up to 60X when the exact number of mixture components is known, and between a few and more than an order of magnitude with component adaptivity.

It is well known that adding any skew symmetric matrix to the gradient of Langevin dynamics algorithm results in a non-reversible diffusion with improved convergence rate. This paper presents a gradient algorithm to adaptively optimize the choice of the skew symmetric matrix. The resulting algorithm involves a non-reversible diffusion algorithm cross coupled with a stochastic gradient algorithm that adapts the skew symmetric matrix. The algorithm uses the same data as the classical Langevin algorithm. A weak convergence proof is given for the optimality of the choice of the skew symmetric matrix. The improved convergence rate of the algorithm is illustrated numerically in Bayesian learning and tracking examples.

A common defect found when reproducing old vinyl and gramophone recordings with mechanical devices are the long pulses with significant low-frequency content caused by the interaction of the arm-needle system with deep scratches or even breakages on the media surface. Previous approaches to their suppression on digital counterparts of the recordings depend on a prior estimation of the pulse location, usually performed via heuristic methods. This paper proposes a novel Bayesian approach capable of jointly estimating the pulse location; interpolating the almost annihilated signal underlying the strong discontinuity that initiates the pulse; and also estimating the long pulse tail by a simple Gaussian Process, allowing its suppression from the corrupted signal. The posterior distribution for the model parameters as well for the pulse is explored via Markov-Chain Monte Carlo (MCMC) algorithms. Controlled experiments indicate that the proposed method, while requiring significantly less user intervention, achieves perceptual results similar to those of previous approaches and performs well when dealing with naturally degraded signals.