PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Nonparametric independent process analysis
Zoltan Szabo and Barnabas Poczos
In: 19th European Signal Processing Conference (EUSIPCO), 29 Aug - 02 Sep 2011, Barcelona, Spain.

This is the latest version of this eprint.

Abstract

Linear dynamical systems are widely used tools to model stochastic time processes, but they have severe limitations; they assume linear dynamics with Gaussian driving noise. Independent component analysis (ICA) aims to weaken these limitations by allowing independent, non-Gaussian sources in the model. Independent subspace analysis (ISA), an important generalization of ICA, has proven to be successful in many source separation applications. Still, the general ISA problem of separating sources with nonparametric dynamics has been hardly touched in the literature yet. The goal of this paper is to extend ISA to the case of (i) nonparametric, asymptotically stationary source dynamics and (ii) unknown source component dimensions. We make use of functional autoregressive (fAR) processes to model the temporal evolution of the hidden sources. An extension of the well-known ISA separation principle is derived for the solution of the introduced fAR independent process analysis (fAR-IPA) task. By applying fAR identification we reduce the problem to ISA. The Nadaraya-Watson kernel regression technique is adapted to obtain strongly consistent fAR estimation. We illustrate the efficiency of the fAR-IPA approach by numerical examples and demonstrate that in this framework our method is superior to standard linear dynamical system based estimators.

PDF - PASCAL Members only - Requires Adobe Acrobat Reader or other PDF viewer.
PDF - Requires Adobe Acrobat Reader or other PDF viewer.
EPrint Type:Conference or Workshop Item (Paper)
Additional Information:http://www.eurasip.org/Proceedings/Eusipco/Eusipco2011/papers/1569422585.pdf
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Computational, Information-Theoretic Learning with Statistics
Learning/Statistics & Optimisation
Theory & Algorithms
ID Code:9537
Deposited By:Zoltan Szabo
Deposited On:30 May 2012

Available Versions of this Item