Technical Program

Paper Detail

Paper:	MLSP-L4.2
Session:	Blind Source Separation I
Time:	Thursday, May 18, 16:50 - 17:10
Presentation:	Lecture
Topic:	Machine Learning for Signal Processing: Blind Signal Separation and Independent Component Analysis
Title:	New Algorithms for Non-Negative Matrix Factorization in Applications to Blind Source Separation
Authors:	Andrzej Cichocki, Rafal Zdunek, Shun-Ichi Amari, RIKEN Brain Science Institute, Japan
Abstract:	In this paper we develop several novel algorithms for non-negative matrix factorization (NMF) in applications to blind (or semi blind) source separation (BSS) when sources are generally statistically dependent under conditions that additional constraints are imposed such nonnegativity, sparsity, smoothness, lower complexity or better predictability. We express the non-negativity constraints using a wide class of loss (cost) functions and we reach an extended class of multiplicative algorithms with regularization. We review several approaches which allows us to obtain generalized forms of classical NMF multiplicative algorithms and unify methods for extend these algorithms. We give the relaxed form of the NMF algorithms to increase convergence speed and impose some desired constraints. Moreover, the effects of various regularization and constraints are clearly shown. The scope results is vast since discussed loss functions include quite large number of useful cost functions such as squared Euclidean distance, relative entropy, Kullback Leibler divergence, Mahalanobis distance and generalized Hellinger/Pearson distance, etc.