Exploiting redundancy to construct listening systems
Paris Smaragdis
(Mitsubishi Electric Research Labs)

In this talk I'll present a series of projects that showcase the design of systems exhibiting listening-like behavior. These systems include a cochlea-like preprocessor, a gestalt rule grouper and an auditory scene analysis system. However, more interesting than the projects themselves is the fact that although these systems perform seemingly very different functions they are all based on the same conceptual (and computational) core. By noting how redundancy reduction can define these perceptual operations and by performing it through the use of ICA, I'll show how we can start explaining a broad section of listening using very compact descriptions. To further drive the point I will present how the same rules can apply equally successfully to other perceptual domains, as well as in multimodal perception.

Relevant material:

Barlow, H. (2001). The exploitation of regularities in the environment by the brain. Behav Brain Sci, 24(4), 602-607; discussion 652-671. http://www.physiol.cam.ac.uk/staff/barlow/Hbb_sgl.html
Bell A.J. and Sejnowski T.J. 1995. "An information maximisation approach to blind separation and blind deconvolution", Neural Computation, 7, 6, 1129-1159. ftp://ftp.cnl.salk.edu/pub/tony/bell.blind.ps
Smaragdis, P. 2001, "Redundancy Reduction for Computational Audition, a Unifying Approach.", Ph.D. dissertation, Massachusetts Institute of Technology, Media Laboratory. http://web.media.mit.edu/~paris/phd/paris-phd.pdf