Jump to content,
COLLEGE OF ENGINEERING
COLLEGE OF COMPUTING
Search GT | Contact Us | BuzzPort
GT Home > Bioengineering Home > News and Events
Bioengineering Header Georgia Tech Bioengineering Home

Upcoming/Recent Graduate student presentations, proposals and defenses

Jeremi Lewi - Ph.D. Proposal Presentation, 2:00 PM, Thurdsay May 22, 2008, Room, UAW 2110
Co-Advisor: Robert Butera, (Georgia Institute of Technology)
Co-Advisor: Liam Paninski, (Columbia University)
Committee:  Brani Vidakovic, (Georgia Institute of Technology), Garrett Stanley, (Georgia Institute of Technology), Charles Isbell, (Georgia Institute of Technology), Chris Rozell, (Georgia Institute of Technology)

TITLE:  "Sequential Optimal Design of Neurophysiology Experiments"

In most neurophysiology experiments, data is collected according to a design that is finalized before the experiment begins. During the experiment, the data already collected is rarely analyzed to evaluate the quality of the design. The data already gathered, however, often contains information which could be used to redesign our experiments to better test our hypotheses. Adaptive experimental designs are particularly valuable in domains where data is expensive and/or limited. In neuroscience, experiments often require training and caring for animals which can be time-consuming and costly. As a result, neuroscientists are often unable to collect enough data to investigate high-dimensional, complex neural systems. By using adaptive experimental designs, neuroscientists could potentially collect data more efficiently. I will outline our preliminary efforts to develop an efficient algorithm for real-time, sequential, optimization, of neurophysiology experiments. I consider experiments where the goal is to identify the conditional response function of a neuron by recording the neuron’s response to various stimuli. Stimuli are selected by optimizing an objective function which quantifies the expected reduction in uncertainty about the unknown response function. Our objective function is based on mutual information and leads to an optimality criterion known as D-optimality. Our implementation overcomes the computational hurdles of sequential optimal experimental design in this setting.

Last revised on May 13th, 2008