Sparse codes for natural sounds
Google Tech TalksMay, 20 2008ABSTRACTThe auditory neural code must serve a wide range of tasks thatrequire great sensitivity in time and frequency and be effective over thediverse array of sounds present in natural acoustic environments. It hasbeen suggested (Barlow, 1961; Atick, 1992; Simoncelli & Olshausen, 2001;Laughlin & Sejnowski, 2003) that sensory systems might have evolved highlyefficient coding strategies to maximize the information conveyed to thebrain while minimizing the required energy and neural resources. In thistalk, I will show that, for natural sounds, the complete acoustic waveformcan be represented efficiently with a nonlinear model based on a populationspike code. In this model, idealized spikes encode the precise temporalpositions and magnitudes of underlying acoustic features. We find that whenthe features are optimized for coding either natural sounds or speech, theyshow striking similarities to time-domain cochlear filter estimates, have afrequency-bandwidth dependence similar to that of auditory nerve fibers, andyield significantly greater coding efficiency than conventional signalrepresentations. These results indicate that the auditory code mightapproach an information theoretic optimum and that the acoustic structure ofspeech might be adapted to the coding capacity of the mammalian auditorysystem.Speaker: Vivienne MingVivienne Ming was born in 1971 in Pasadena, CA. She receivedher B.S. (2000) in Cognitive Neuroscience from UC San Diego, developing faceand expression recognition systems in the Machine Perception Lab. She earnedher M.A. (2003) and Ph.D. (2006) in Psychology from Carnegie MellonUniversity along with a doctoral training degree in computationalneuroscience from the Center for the Neural Basis of Cognition. Herdissertation, *Efficient auditory coding*, combined computational andbehavioral approaches to study the perception of natural sounds, includingspeech. Since 2006, she has worked jointly as a junior fellow andpost-doctoral researcher at the Redwood Center for Theoretical Neuroscienceat UC Berkeley and MBC/Mind, Brain & Cognition at Stanford University.