Topic: “Mixture Approaches and Steady-State Analysis for Universal Source Separation”
Presenter: Mr. Keith D. Gilbert, ECE Ph.D. Student, University of Massachusetts Dartmouth
Advisor: Dr. Karen L. Payton
Blind source separation (BSS) is the task of disentangling a set of source signals that have been combined into a set of noisy observations when nothing is known about the combination process or even the sources themselves. BSS is a mature field of research at this point, and each fundamental approach taken to solve this problem makes critical assumptions about the sources and/or mixing conditions. For any class of BSS algorithm, there are many parameters that need to be chosen a priori, and the blindness of a particular method decreases with the decrease in number of free parameters. The approach taken here is to make as few assumptions beforehand by running fixed source separation methods in parallel and adaptively combing the outputs. This mixture approach to source separation (MASS) follows from universal mixture approaches employed within the context of source coding, prediction, and more recently, adaptive filtering, and the ultimate goal of this research is to realize universal source separation. Current work focuses on using an adaptive filtering energy conservation relation to analyze the steady-state performance in supervised separation of critically-determined, instantaneous mixtures of stationary, white sources. When the source separation problem is solved via a network of adaptive filters, this steady-state analysis provides a necessary tool to derive a mixture method for universal demixing system identification, while also pointing to energy-constrained BSS as an avenue for studying transient and steady-state behavior in more complicated scenarios, e.g. convolutive mixtures, non-stationary and/or non-white sources, etc. The presentation will give a brief overview of the fundamental research topics, i.e. source separation, mixture adaptive filtering, and universality, and will then focus on the current work and its implications for future research of MASS.
Keith D. Gilbert received the B.S. degree in Physics from the University of Virginia in 1998, the M.S. degree in Electrical Engineering from the University of Massachusetts Dartmouth in 2010, and is currently pursuing a Ph.D. in Electrical Engineering at the University of Massachusetts Dartmouth. He has been a graduate student member of the Institute for Electronic and Electrical Engineers (IEEE) and the Acoustical Society of America (ASA) since 2007, and a member of Eta Kappa Nu (EKN) since 2009. His interests include statistical signal processing, acoustics, information theory, and adaptive filtering with special attention given to blind methods, universal algorithms, and the source deconvolution and separation problems.
The seminar is open to the public free of charge.