In this paper we investigate the application of natural gradient descent to Bellman error based reinforcement learning algorithms. This combination is interesting because natural gradient descent is invariant to the parameterization of the value function. This invariance property means that natural gradient descent adapts its update directions to correct for poorly conditioned representations. We present and analyze quadratic and linear time natural temporal difference learning algorithms, and prove that they are covariant. We conclude with experiments which suggest that the natural algorithms can match or outperform their non-natural counterparts using linear function approximation, and drastically improve upon their non-natural counterparts when using non-linear function approximation.

In many real-world applications of learning, the environment is open and changes gradually, which requires the learning system to have the ability of detecting and adapting to the changes. Class-incremental learning (C-IL) is an important and practical problem where data from unseen augmented classes are fed, but has not been studied well in the past. In C-IL, the system should beware of predicting instances from augmented classes as a seen class, and thus faces the challenge that no such instances were observed during training stage. In this paper, we tackle the challenge by using unlabeled data, which can be cheaply collected in many real-world applications. We propose the LACU framework as well as the LACU-SVM approach to learn the concept of seen classes while incorporating the structure presented in the unlabeled data, so that the misclassification risks among the seen classes as well as between the augmented and the seen classes are minimized simultaneously. Experiments on diverse datasets show the effectiveness of the proposed approach.

Dropout and other feature noising schemes have shown promising results in controlling over-fitting by artificially corrupting the training data. Though extensive theoretical and empirical studies have been performed for generalized linear models, little work has been done for support vector machines (SVMs), one of the most successful approaches for supervised learning. This paper presents dropout training for linear SVMs. To deal with the intractable expectation of the non-smooth hinge loss under corrupting distributions, we develop an iteratively re-weighted least square (IRLS) algorithm by exploring data augmentation techniques. Our algorithm iteratively minimizes the expectation of a re-weighted least square problem, where the re-weights have closed-form solutions. The similar ideas are applied to develop a new IRLS algorithm for the expected logistic loss under corrupting distributions. Our algorithms offer insights on the connection and difference between the hinge loss and logistic loss in dropout training. Empirical results on several real datasets demonstrate the effectiveness of dropout training on significantly boosting the classification accuracy of linear SVMs.

The local neighborhood selection plays a crucial role for most representation based manifold learning algorithms. This paper reveals that an improper selection of neighborhood for learning representation will introduce negative components in the learnt representations. Importantly, the representations with negative components will affect the intrinsic manifold structure preservation. In this paper, a local non-negative pursuit (LNP) method is proposed for neighborhood selection and non-negative representations are learnt. Moreover, it is proved that the learnt representations are sparse and convex. Theoretical analysis and experimental results show that the proposed method achieves or outperforms the state-of-the-art results on various manifold learning problems.

Restricted Boltzmann machines (RBMs) are a powerful generative modeling technique, based on a complex graphical model of hidden (latent) variables. Conditional RBMs (CRBMs) are an extension of RBMs tailored to modeling temporal data. A drawback of CRBMs is their consideration of linear temporal dependencies, which limits their capability to capture complex temporal structure. They also require many variables to model long temporal dependencies, a fact that might provoke overfitting proneness. To resolve these issues, in this paper we propose the echo-state CRBM (ES-CRBM): our model uses an echo-state network reservoir in the context of CRBMs to efficiently capture long and complex temporal dynamics, with much fewer trainable parameters compared to conventional CRBMs. In addition, we introduce an (implicit) mixture of ES-CRBM experts (im-ES-CRBM) to enhance even further the capabilities of our ES-CRBM model. The introduced im-ES-CRBM allows for better modeling temporal observations which might comprise a number of latent or observable subpatterns that alternate in a dynamic fashion. It also allows for performing sequence segmentation using our framework. We apply our methods to sequential data modeling and classification experiments using public datasets. As we show, our approach outperforms both existing RBM-based approaches as well as related state-of-the-art methods, such as conditional random fields.

Collaborative filtering algorithms generally rely on the assumption that user preference patterns remain stationary. However, real-world relational data are seldom stationary. User preference patterns may change over time, giving rise to the requirement of designing collaborative filtering systems capable of detecting and adapting to preference pattern shifts. Motivated by this observation, in this paper we propose a dynamic Bayesian probabilistic matrix factorization model, designed for modeling time-varying distributions. Formulation of our model is based on imposition of a dynamic hierarchical Dirichlet process (dHDP) prior over the space of probabilistic matrix factorization models to capture the time-evolving statistical properties of modeled sequential relational datasets. We develop a simple Markov Chain Monte Carlo sampler to perform inference. We present experimental results to demonstrate the superiority of our temporal model.

Given a CNF formula and a weight for each assignment of values tovariables, two natural problems are weighted model counting anddistribution-aware sampling of satisfying assignments. Both problems have a wide variety of important applications. Due to the inherentcomplexity of the exact versions of the problems, interest has focusedon solving them approximately. Prior work in this area scaled only tosmall problems in practice, or failed to provide strong theoreticalguarantees, or employed a computationally-expensive most-probable-explanation ({\MPE}) queries that assumes prior knowledge of afactored representation of the weight distribution. We identify a novel parameter,\emph{tilt}, which is the ratio of the maximum weight of satisfying assignment to minimum weightof satisfying assignment and present anovel approach that works with a black-box oracle for weights ofassignments and requires only an {\NP}-oracle (in practice, a {\SAT}-solver) to solve both thecounting and sampling problems when the tilt is small. Our approach provides strong theoretical guarantees, and scales toproblems involving several thousand variables. We also show that theassumption of small tilt can be significantly relaxed while improving computational efficiency if a factored representation of the weights is known.

We consider the problem of learning soft assignments of N items to K categories given two sources of information: an item-category similarity matrix, which encourages items to be assigned to categories they are similar to (and to not be assigned to categories they are dissimilar to), and an item-item similarity matrix, which encourages similar items to have similar assignments. We propose a simple quadratic programming model that captures this intuition. We give necessary conditions for its solution to be unique, define an out-of-sample mapping, and derive a simple, effective training algorithm based on the alternating direction method of multipliers. The model predicts reasonable assignments from even a few similarity values, and can be seen as a generalization of semisupervised learning. It is particularly useful when items naturally belong to multiple categories, as for example when annotating documents with keywords or pictures with tags, with partially tagged items, or when the categories have complex interrelations (e.g. hierarchical) that are unknown.

We introduce the problem of PAC rank elicitation, which consists of sorting a given set of options based on adaptive sampling of stochastic pairwise preferences. More specifically, we assume the existence of a ranking procedure, such as Copeland's method, that determines an underlying target order of the options. The goal is to predict a ranking that is sufficiently close to this target order with high probability, where closeness is measured in terms of a suitable distance measure. We instantiate this setting with combinations of two different distance measures and ranking procedures. For these instantiations, we devise efficient strategies for sampling pairwise preferences and analyze the corresponding sample complexity. We also present first experiments to illustrate the practical performance of our methods.

We present a novel approach to low-dimensional neighbor embedding for visualization, based on formulating an information retrieval based neighborhood preservation cost function as Maximum satisfiability on a discretized output display. The method has a rigorous interpretation as optimal visualization based on the cost function. Unlike previous low-dimensional neighbor embedding methods, our formulation is guaranteed to yield globally optimal visualizations, and does so reasonably fast. Unlike previous manifold learning methods yielding global optima of their cost functions, our cost function and method are designed for low-dimensional visualization where evaluation and minimization of visualization errors are crucial. Our method performs well in experiments, yielding clean embeddings of datasets where a state-of-the-art comparison method yields poor arrangements. In a real-world case study for semi-supervised WLAN signal mapping in buildings we outperform state-of-the-art methods.