The Bayesian paradigm provides a natural and effective means of exploiting priorknowledge concerning the time-frequency structure of sound signals such as speech and music--something which has often been overlooked intraditional audio signal processing approaches. Here, after constructing aBayesian model and prior distributions capable of taking into account the time-frequency characteristics of typical audio waveforms, we apply Markov chain Monte Carlo methods in order to sample from the resultant posterior distribution of interest. We present speech enhancement resultswhich compare favourably in objective terms with standard time-varying filtering techniques (and in several cases yield superior performance, bothobjectively and subjectively); moreover, in contrast to such methods, our results are obtained without an assumption of prior knowledge of the noise power.

Standard value function approaches to finding policies for Partially Observable Markov Decision Processes (POMDPs) are intractable for large models. The intractability ofthese algorithms is due to a great extent to their generating an optimal policy over the entire belief space. However, in real POMDP problems most belief states are unlikely, and there is a structured, low-dimensional manifold of plausible beliefs embedded in the high-dimensional belief space. We introduce a new method for solving large-scale POMDPs by taking advantage of belief space sparsity. We reduce the dimensionality of the belief space by exponential family Principal Components Analysis [1], which allows us to turn the sparse, highdimensional beliefspace into a compact, low-dimensional representation in terms of learned features of the belief state. We then plan directly on the low-dimensional belief features. By planning in a low-dimensional space, we can find policies for POMDPs that are orders of magnitude larger than can be handled by conventional techniques. We demonstrate the use of this algorithm on a synthetic problem and also on a mobile robot navigation task.

Particle filters estimate the state of dynamical systems from sensor information. Inmany real time applications of particle filters, however, sensor information arrives at a significantly higher rate than the update rate of the filter. The prevalent approach to dealing with such situations is to update the particle filter as often as possible and to discard sensor information that cannot be processed in time. In this paper we present real-time particle filters, whichmake use of all sensor information even when the filter update rate is below the update rate of the sensors. This is achieved by representing posteriorsas mixtures of sample sets, where each mixture component integrates one observation arriving during a filter update. The weights of the mixture components are set so as to minimize the approximation error introduced by the mixture representation. Thereby, our approach focuses computational resources (samples) on valuable sensor information. Experiments usingdata collected with a mobile robot show that our approach yields strong improvements over other approaches.

In the Bayesian approach, regularizationis achieved by specifying a prior distribution over the parameters and subsequently averaging over the posterior distribution. This regularization provides not only smoother estimates of the parameters compared to maximum likelihood but also guides the selection of model structures. It was pointed out in [6] that a very large scale parameter of the Dirichlet prior can degrade predictive accuracy due to severe regularization of the parameter estimates. We complement this discussion here and show that a very small scale parameter can lead to poor over-regularized structures when a product of (conjugate) Dirichlet priors is used over multinomial conditional distributions (Section 3). Section 4 demonstrates the effect of the scale parameter and how it can be calibrated. We focus on the class of Bayesian network models throughout this paper.

A standard view of memory consolidation is that episodes are stored temporarily inthe hippocampus, and are transferred to the neocortex through replay. Various recent experimental challenges to the idea of transfer, particularly for human memory, are forcing its reevaluation. However, although there is independent neurophysiological evidence for replay, short of transfer, there are few theoretical ideas for what it might be doing. We suggest and demonstrate two important computational roles associated with neocortical indices.

We report a system that classifies and can learn to classify patterns of visual motion online. The complete system is described by the dynamics ofits physical network architectures. The combination of the following propertiesmakes the system novel: Firstly, the front-end of the system consists of an aVLSI optical flow chip that collectively computes 2-D global visual motion in real-time [1]. Secondly, the complexity of the classification task is significantly reduced by mapping the continuous motiontrajectories to sequences of'motion events'. And thirdly, all the network structures are simple and with the exception of the optical flow chip based on a Winner-Take-All (WTA) architecture. We demonstrate theapplication of the proposed generic system for a contactless man-machine interface that allows to write letters by visual motion. Regarding thelow complexity of the system, its robustness and the already existing front-end, a complete aVLSI system-on-chip implementation is realistic, allowing various applications in mobile electronic devices.

Population-based Incremental Learning is shown require very sensitive scalingof its learning rate. The learning rate must scale with the system size in a problem-dependent way. This is shown in two problems: the needle-in-a haystack, in which the learning rate must vanish exponentially in the system size, and in a smooth function in which the learning rate must vanish like the square root of the system size. Two methods are proposed for removing this sensitivity. Alearning dynamics which obeys detailed balance is shown to give consistent performance over the entire range of learning rates. An analog of mutation is shown to require a learning rate which scales as the inverse system size, but is problem independent.

Recent algorithms for sparse coding and independent component analysis (ICA)have demonstrated how localized features can be learned from natural images. However, these approaches do not take image transformations intoaccount. As a result, they produce image codes that are redundant because the same feature is learned at multiple locations. We describe an algorithm for sparse coding based on a bilinear generative model of images. By explicitly modeling the interaction between image featuresand their transformations, the bilinear approach helps reduce redundancy in the image code and provides a basis for transformationinvariant vision.We present results demonstrating bilinear sparse coding of natural images. We also explore an extension of the model that can capture spatial relationships between the independent features of an object, therebyproviding a new framework for parts-based object recognition.

We present the bump mixture model, a statistical model for analog data where the probabilistic semantics, inference, and learning rules derive from low-level transistor behavior. The bump mixture model relies on translinear circuits to perform probabilistic inference, andfloating-gate devices to perform adaptation. This system is low power, asynchronous, and fully parallel, and supports various on-chiplearning algorithms. In addition, the mixture model can perform several tasks such as probability estimation, vector quantization, classification,and clustering. We tested a fabricated system on clustering, quantization, and classification of handwritten digits and show performance comparable to the EM algorithm on mixtures ofGaussians.

Boosting algorithms and successful applications thereof abound for classification andregression learning problems, but not for unsupervised learning. We propose a sequential approach to adding features to a random fieldmodel by training them to improve classification performance between the data and an equal-sized sample of "negative examples" generated fromthe model's current estimate of the data density.