IBN: 9634314 PI: John Clark The research will seek a deeper understanding of how neural systems represent and process information about the world, by applying established principles of information theory and statistical inference in the context of a large body of neurobiological observations. The basic hypothesis to be explored is that probabilistic information about meaningful variables (such as measures of visual depth, optical flow, limb orientation, etc.) is encoded in the firing rates of an ensemble of neurons. Bayesian inference then provides an optimal framework for inference in the presence of uncertainty, given certain prior knowledge that may be determined genetically or through learning processes. This approach leads in a natural way to neuronal circuits that are computationally more general than the standard input-output maps of neural-network theory. For example, an explanation of experiments demonstrating contextual or attentional influences on neuronal responses may be formulated in terms of a network involving multiplicative synaptic interactions. Specific thrusts of the research will include (i) the study of small-scale neuronal circuits that illustrate the computational power of the model and furnish prototypes for large-scale simulation of important brain subsystems and (ii) the incorporation of learning rules that permit naive circuits to modify their structure and optimize task performance. Cross-disciplinary in nature, the project will join the expertise of a computational neuroscientist and a theoretical physicist in work to be performed primarily in the laboratory of a prominent neurophysiologist. Central to the project is the involvement of physics graduate students in a new interdisciplinary training program in computational neuroscience.
