Efficient Coding for Natural Signals

Michael Lewicki
Computer Science Department
Center For the Neural Basis of Cognition
Carnegie Mellon University

Time and Place

1305 Newell-Simon Hall
Refreshments 3:15 pm
Talk 3:30 pm

Abstract

The representation and encoding of sensor information is a necessary first step in any artificial or robotic sensory system, but the choice of which among the many strategies to use is often empirical or ad hoc. One idea for obtaining better sensory codes due to Horace Barlow is based on the idea of maximizing information transmission and eliminating statistical redundancy from the raw sensory signal. In this talk I will show how this idea can be made explicit and can be used to derive sensory codes that are optimal in the sense of Shannon statistical efficiency.

A second theme of my talk will be that this framework not only provides a means for deriving sensory codes for robotics and signal processing, but can also be used to make theoretical predictions about biological sensory systems. Much is known about how the brain encodes sensory information, but why it has evolved to use the particular coding strategies it does have been the subject of long-standing interest and debate. In this talk I will also show how efficient coding of natural images or sounds yields sensory codes similar to those observed in animals and provides a simple theoretical explanation for a large and complex set of experimental data. These results provide evidence that the neural coding of sensory signals approaches an information theoretic optimum and lends further support to the hypothesis that efficient coding is a general principle of sensory representation.

Speaker Biography

Dr. Lewicki received his B.S. degree in mathematics and cognitive science in 1989 from Carnegie Mellon University. He received his Ph.D. degree in computation and neural systems from the California Institute of Technology in 1996. From 1996 to 1998, he was a postdoctoral fellow in the Computational Neurobiology Laboratory at the Salk Institute. He is currently an assistant professor in the Computer Science Department at Carnegie Mellon University and in the CMU-University of Pittsburgh Center for the Neural Basis of Cognition. His research involves the study and development of computational approaches to the representation, processing, and learning of pattern structure in natural visual and acoustic environments.

Speaker Appointments

For appointments, please contact Yanxi Liu (yanxi@cs.cmu.edu).