Michael Lewicki, Case Western Reserve University, USA: "Learning structures in natural sounds"

ABSTRACT:

Our auditory systems are wonderfully adept at processing natural sounds.  What is the information processing the auditory system performs to carry out our everyday auditory tasks?  Consider the basic coding problem of transforming the vibrations at the eardrum to the neural code at the auditory nerve.  Out of the infinite range of possible codes -- Fourier, wavelet or something more exotic -- why do biological systems use the codes they do?  Are these the result of random evolutionary adaptation or constraints imposed by biological hardware?  Or are there theories that can explain auditory coding in terms of fundamental principles?  Consider now the higher-level auditory problem of generalizing to a sound class from specific instances of audio waveforms.  For example, we easily recognize the clink of a glass, footsteps in the hall, and myriad other sounds in our natural environment, yet the individual waveforms of any given glass are all unique.  What acoustic cues tell us they are similar?  How do we learn these cues?  In this talk, I will present some solutions to these problems using statistical learning algorithms and efficient coding theory.  This optimality principle postulates that biological auditory systems are adapted to their natural acoustic environments.  When applied to natural sounds, these results provide evidence that the auditory neural code approaches an information theoretic optimum, and suggests, surprisingly, that speech itself is adapted to the coding capacity of the mammalian auditory system.  In the second part of talk, I will show that the same methods can be applied at higher levels to learn the underlying structure of natural impact sounds, so that these complex sounds can be characterized with a small number of intrinsic dimensions.  With this reduced representation, it is possible to perform accurate sound categorization from individual waveforms, and even synthesize realistic impact sounds from a small number of intrinsic values.  This is joint work with Vivienne Ming and Sofia Cavaco.

ABOUT MICHAEL LEWICKI:

Dr. Lewicki an associate professor in the Electrical Engineering and Computer Science Department at Case Western Reserve University.  For ten years he was on the faculty at Carnegie Mellon University in the Computer Science Department and the Center for the Neural Basis of Cognition.  He received his BS degree in mathematics and cognitive science from Carnegie Mellon University, his PhD degree in computation and neural systems from the California Institute of Technology, and did postdoctoral studies in the Computational Neurobiology Laboratory at the Salk Institute.  For the academic year 2008-2009, he was a fellow at the Institute for Advanced Study (Wissenschaftskolleg) in Berlin.

 

VIDEO ARCHIVE - MICHAEL LEWICKI:

 

APA video citation:

Lewicki, M. (2012, October 22). Learning structures in natural sounds -
CIRMMT Distinguished Lectures in the Science and Technology of Music. [Video file].
Retrieved from https://www.youtube.com/watch?v=UN_j04vyvS0