The performance of stochastic gradient descent (SGD) depends critically on how learning rates are tuned and decreased over time. We propose a method to automatically adjust multiple learning rates so as to minimize the expected error at any one time. The method relies on local gradient variations across samples. In our approach, learning rates can increase as well as decrease, making it suitable for non-stationary problems. Using a number of convex and non-convex learning tasks, we show that the resulting algorithm matches the performance of SGD or other adaptive approaches with their best settings obtained through systematic search, and effectively removes the need for learning rate tuning.

Copulas allow to learn marginal distributions separately from the multivariate dependence structure (copula) that links them together into a density function. Vine factorizations ease the learning of high-dimensional copulas by constructing a hierarchy of conditional bivariate copulas. However, to simplify inference, it is common to assume that each of these conditional bivariate copulas is independent from its conditioning variables. In this paper, we relax this assumption by discovering the latent functions that specify the shape of a conditional copula given its conditioning variables. We learn these functions by following a Bayesian approach based on sparse Gaussian processes with expectation propagation for scalable, approximate inference. Experiments on real-world datasets show that, when modeling all conditional dependencies, we obtain better estimates of the underlying copula of the data.

When using deep, multi-layered architectures to build generative models of data, it is difficult to train all layers at once. We propose a layer-wise training procedure admitting a performance guarantee compared to the global optimum. It is based on an optimistic proxy of future performance, the best latent marginal. We interpret auto-encoders in this setting as generative models, by showing that they train a lower bound of this criterion. We test the new learning procedure against a state of the art method (stacked RBMs), and find it to improve performance. Both theory and experiments highlight the importance, when training deep architectures, of using an inference model (from data to hidden variables) richer than the generative model (from hidden variables to data).

Masashi Sugiyama Department of Computer Science Tokyo Institute of Technology Tokyo 152-8552, Japan sugi@cs.titech.ac.jp Abstract Hidden Markov models and their variants are the predominant sequential classification method in such domains as speech recognition, bioinformatics and natural language processing. Being generative rather than discriminative models, however, their classification performance is a drawback. In this paper we apply ideas from the field of density ratio estimation to bypass the difficult step of learning likelihood functions in HMMs. By reformulating inference and model fitting in terms of density ratios and applying a fast kernel-based estimation method, we show that it is possible to obtain a striking increase in discriminative performance while retaining the probabilistic qualities of the HMM. We demonstrate experimentally that this formulation makes more efficient use of training data than alternative approaches. 1 Introduction Inference of a sequence of estimated classes from a sequence of noisy observations is fundamental in many applications. The hidden Markov model (HMM) and its variants are the usual methods employed to do this, and have been used with conspicuous success in such domains as speech recognition, bioinformatics and natural language processing. As well as being computationally efficient, they are a popular choice due to their intuitive probabilistic interpretation.

The classical mixture of Gaussians model is related to K-means via small-variance asymptotics: as the covariances of the Gaussians tend to zero, the negative log-likelihood of the mixture of Gaussians model approaches the K-means objective, and the EM algorithm approaches the K-means algorithm. Kulis & Jordan (2012) used this observation to obtain a novel K-means-like algorithm from a Gibbs sampler for the Dirichlet process (DP) mixture. We instead consider applying small-variance asymptotics directly to the posterior in Bayesian nonparametric models. This framework is independent of any specific Bayesian inference algorithm, and it has the major advantage that it generalizes immediately to a range of models beyond the DP mixture. To illustrate, we apply our framework to the feature learning setting, where the beta process and Indian buffet process provide an appropriate Bayesian nonparametric prior. We obtain a novel objective function that goes beyond clustering to learn (and penalize new) groupings for which we relax the mutual exclusivity and exhaustivity assumptions of clustering. We demonstrate several other algorithms, all of which are scalable and simple to implement. Empirical results demonstrate the benefits of the new framework.

The application of machine learning to bioinformatics problems is well established. Less well understood is the application of bioinformatics techniques to machine learning and, in particular, the representation of non-biological data as biosequences. The aim of this paper is to explore the effects of giving amino acid representation to problematic machine learning data and to evaluate the benefits of supplementing traditional machine learning with bioinformatics tools and techniques. The signatures of 60 computer viruses and 60 computer worms were converted into amino acid representations and first multiply aligned separately to identify conserved regions across different families within each class (virus and worm). This was followed by a second alignment of all 120 aligned signatures together so that non-conserved regions were identified prior to input to a number of machine learning techniques. Differences in length between virus and worm signatures after the first alignment were resolved by the second alignment. Our first set of experiments indicates that representing computer malware signatures as amino acid sequences followed by alignment leads to greater classification and prediction accuracy. Our second set of experiments indicates that checking the results of data mining from artificial virus and worm data against known proteins can lead to generalizations being made from the domain of naturally occurring proteins to malware signatures. However, further work is needed to determine the advantages and disadvantages of different representations and sequence alignment methods for handling problematic machine learning data.

The problem of change-point estimation is considered under a general framework where the data are generated by unknown stationary ergodic process distributions. In this context, the consistent estimation of the number of change-points is provably impossible. However, it is shown that a consistent clustering method may be used to estimate the number of change points, under the additional constraint that the correct number of process distributions that generate the data is provided. This additional parameter has a natural interpretation in many real-world applications. An algorithm is proposed that estimates the number of change-points and locates the changes. The proposed algorithm is shown to be asymptotically consistent; its empirical evaluations are provided.

This paper works through the optimization of a real world planning problem, with a combination of a generative planning tool and an influence diagram solver. The problem is taken from an existing application in the domain of oil spill emergency response. The planning agent manages constraints that order sets of feasible equipment employment actions. This is mapped at an intermediate level of abstraction onto an influence diagram. In addition, the planner can apply a surveillance operator that determines observability of the state---the unknown trajectory of the oil. The uncertain world state and the objective function properties are part of the influence diagram structure, but not represented in the planning agent domain. By exploiting this structure under the constraints generated by the planning agent, the influence diagram solution complexity simplifies considerably, and an optimum solution to the employment problem based on the objective function is found. Finding this optimum is equivalent to the simultaneous evaluation of a range of plans. This result is an example of bounded optimality, within the limitations of this hybrid generative planner and influence diagram architecture.

Directed acyclic graphs have been used fruitfully to represent causal strucures (Pearl 1988). However, in the social sciences and elsewhere models are often used which correspond both causally and statistically to directed graphs with directed cycles (Spirtes 1995). Pearl (1993) discussed predicting the effects of intervention in models of this kind, so-called linear non-recursive structural equation models. This raises the question of whether it is possible to make inferences about causal structure with cycles, form sample data. In particular do there exist general, informative, feasible and reliable precedures for inferring causal structure from conditional independence relations among variables in a sample generated by an unknown causal structure? In this paper I present a discovery algorithm that is correct in the large sample limit, given commonly (but often implicitly) made plausible assumptions, and which provides information about the existence or non-existence of causal pathways from one variable to another. The algorithm is polynomial on sparse graphs.

Lower and upper probabilities, also known as Choquet capacities, are widely used as a convenient representation for sets of probability distributions. This paper presents a graphical decomposition and exact propagation algorithm for computing marginal posteriors of 2-monotone lower probabilities (equivalently, 2-alternating upper probabilities).