Stephen V. David | |
|
More ... |
I'm a postdoctoral fellow with Shihab Shamma in the Neural Systems Laboratory at the University of Maryland. In May 2004, I completed a Ph.D. in Bioengineering at the University of California, Berkeley (joint program with UCSF), studying vision and attention with Jack Gallant. Humans and other animals are exquisitely adept at recognizing and discriminating important sounds, even when masked by substantial irrelevant noise. Although this process is often effortless for animals, common sources of environmental noise severely confound automatic speech processors and distort the output of hearing aids and prosthetics. I am interested in undestanding the neurophysiological and computational processes that underlie this remarkable ability, with an aim of improving engineered systems for sensory signal processing. During normal behavior, important information can arrive from multiple sensory modalities, and the relevance of any given stimulus can change with behaviorial demands. Thus the ability to robustly identify sounds represents a combined effort of bottom-up multimodal representations that are modulated by top-down down demands for information appropriate to the task at hand. To understand these processes, I conduct experiments that manipulate auditory attention and study how the cortex functions under these different behavior conditions. Data from these studies is used to develop computational models that integrate top-down and bottom-up processing under realistic, natural conditions. I am also interested in methods for effective comparison of models for sensory processing by neurons. With the continuous increase in available computational power, we have the ability to test and compare a huge variety of models. This new potential raises new issues: What is the best way to compare functional models of neurons? How should the large and diverse neurophysiological datasets be stored so that they can be available for testing new models? The Neural Prediction Challenge, a collaboration with Jack Gallant and Frederic Theunissen at UC Berkeley, is a database of single neuron recordings from auditory and visual systems using natural stimuli. Interested researchers can download the data and compare the performance of their model against other models fit with the same data. A related project, STRFpak, is a software package providing model estimation and validation tools that can be applied to any neurophysiological data set. Recent publications (including some pdfs that can be difficult to locate):
|
