In many models, variances are assumed to be constant although this assumption is often unrealistic in practice. Joint modelling of means and variances is di#cult in many learning approaches, because it can lead into infinite probability densities. We show that a Bayesian variational technique which is sensitive to probability mass instead of density is able to jointly model both variances and means. We consider a model structure where a Gaussian variable, called variance node, controls the variance of another Gaussian variable. Variance nodes make it possible to build hierarchical models for both variances and means. We report experiments with artificial data which demonstrate the ability of the learning algorithm to find variance sources explaining and characterizing well the variances in the multidimensional data. Experiments with biomedical MEG data show that variance sources are present in real-world signals.

A new algorithmic framework called denoising source separation (DSS) is introduced. The main benefit of this framework is that it allows for easy development of new source separation algorithms which are optimised for specific problems. In this framework, source separation algorithms are constucted around denoising procedures. The resulting algorithms can range from almost blind to highly specialised source separation algorithms. Both simple linear and more complex nonlinear or adaptive denoising schemes are considered. Some existing independent component analysis algorithms are reinterpreted within DSS framework and new, robust blind source separation algorithms are suggested. Although DSS algorithms need not be explicitly based on objective functions, there is often an implicit objective function that is optimised. The exact relation between the denoising procedure and the objective function is derived and a useful approximation of the objective function is presented. In the experimental section, various DSS schemes are applied extensively to artificial data, to real magnetoencephalograms and to simulated CDMA mobile network signals. Finally, various extensions to the proposed DSS algorithms are considered. These include nonlinear observation mappings, hierarchical models and overcomplete, nonorthogonal feature spaces. With these extensions, DSS appears to have relevance to many existing models of neural information processing.

The building blocks introduced earlier by us in [1] are used for constructing a hierarchical nonlinear model for nonlinear factor analysis. We call the resulting method hierarchical nonlinear factor analysis (HNFA). The variational Bayesian learning algorithm used in this method has a linear computational complexity, and it is able to infer the structure of the model in addition to estimating the unknown parameters. We show how nonlinear mixtures can be separated by first estimating a nonlinear subspace using HNFA and then rotating the subspace using linear independent component analysis. Experimental results show that the cost function minimised during learning predicts well the quality of the estimated subspace.

The appearance of dynamic scenes is often largely governed by a latent low-dimensional dynamic process. We show how to learn a mapping from video frames to this lowdimensional representation by exploiting the temporal coherence between frames and supervision from a user. This function maps the frames of the video to a low-dimensional sequence that evolves according to Markovian dynamics. This ensures that the recovered low-dimensional sequence represents a physically meaningful process. We relate our algorithm to manifold learning, semi-supervised learning, and system identification, and demonstrate it on the tasks of tracking 3D rigid objects, deformable bodies, and articulated bodies. We also show how to use the inverse of this mapping to manipulate video.

Abstract — In this paper, we briefly review recent advances in blind source separation (BSS) for nonlinear mixing models. After a general introduction to the nonlinear BSS and ICA (independent Component Analysis) problems, we discuss in more detail uniqueness issues, presenting some new results. A fundamental difficulty in the nonlinear BSS problem and even more so in the nonlinear ICA problem is that they are nonunique without extra constraints, which are often implemented by using a suitable regularization. Post-nonlinear mixtures are an important special case, where a nonlinearity is applied to linear mixtures. For such mixtures, the ambiguities are essentially the same as for the linear ICA or BSS problems. In the later part of this paper, various separation techniques proposed for post-nonlinear mixtures and general nonlinear mixtures are reviewed. I. THE NONLINEAR ICA AND BSS PROBLEMS Consider Æ samples of the observed data vector Ü, modeled by

In this paper we present a framework for using multi-layer perceptron (MLP) networks in nonlinear generative models trained by variational Bayesian learning. The nonlinearity is handled by linearizing it using a Gauss–Hermite quadrature at the hidden neurons. This yields an accurate approximation for cases of large posterior variance. The method can be used to derive nonlinear counterparts for linear algorithms such as factor analysis, independent component/factor analysis and state-space models. This is demonstrated with a nonlinear factor analysis experiment in which even 20 sources can be estimated from a real world speech data set. 1

Abstract. We show that the choice of posterior approximation affects the solution found in Bayesian variational learning of linear independent component analysis models. Assuming the sources to be independent a posteriori favours a solution which has orthogonal mixing vectors. Linear mixing models with either temporally correlated sources or non-Gaussian source models are considered but the analysis extends to nonlinear mixtures as well.

Abstract—The bits-back coding first introduced by Wallace in 1990 and later by Hinton and van Camp in 1993 provides an interesting link between Bayesian learning and information-theoretic minimum-description-length (MDL) learning approaches. The bits-back coding allows interpreting the cost function used in the variational Bayesian method called ensemble learning as a code length in addition to the Bayesian view of misfit of the posterior approximation and a lower bound of model evidence. Combining these two viewpoints provides interesting insights to the learning process and the functions of different parts of the model. In this paper, the problem of variational Bayesian learning of hierarchical latent variable models is used to demonstrate the benefits of the two views. The code-length interpretation provides new views to many parts of the problem such as model comparison and pruning and helps explain many phenomena occurring in learning. Index Terms—Bits-back coding, ensemble learning, hierarchical latent variable models, minimum description length, variational Bayesian learning. I.

Abstract. A popular strategy for dealing with large parameter estimation problems is to split the problem into manageable subproblems and solve them cyclically one by one until convergence. A well-known drawback of this strategy is slow convergence in low noise conditions. We propose using so-called pattern searches which consist of an exploratory phase followed by a line search. During the exploratory phase, a search direction is determined by combining the individual updates of all subproblems. The approach can be used to speed up several well-known learning methods such as variational Bayesian learning (ensemble learning) and expectation-maximization algorithm with modest algorithmic modifications. Experimental results show that the proposed method is able to reduce the required convergence time by 60–85 % in realistic variational Bayesian learning problems.

this paper, we first consider a static nonlinear mixing model, with a successful application to real-world speech data compression. Then we discuss extraction of sources from nonlinear dynamic processes, and detection of abrupt changes in the process dynamics. In a difficult test problem with chaotic data, our approach clearly outperforms currently available nonlinear prediction and change detection techniques. The proposed methods are computationally demanding, but they can be applied to blind nonlinear problems of higher dimensions than other existing approaches