The Rockefeller University
Calendar of Events
Download iCalendar File Download iCalendar File Print this page
Center for Studies in Physics and Biology Seminars
Type Center for Studies in Physics and Biology Seminars
Title Optimal Coding Strategies in the Peripheral Olfactory Systems: Compressed Sensing for an Array of Nonlinear Olfactory Receptor Neurons with and without Spontaneous Activity
Date Tuesday, October 1, 2019
Time 4:00 p.m. - 5:00 p.m.
Location A Level Physics Seminar Room, Room A30, Smith Hall Annex (CRC)
Speaker(s)

Yuhai Tu, Ph.D., permanent research staff member, manager, Thomas J. Watson Research Center, IBM

Speaker Bio The peripheral olfactory systems are capable of coding sparse mixtures of a few odorants in a high dimensional space (the number of possible odorant molecules is huge) by using a relatively small number of olfactory receptor neurons. Data compression is also an important problem in computer science, where powerful algorithms have been developed for compressing sparse high-dimensional data. In particular, the compressed sensing (CS) theory has been successfully used to compress high-dimensional information efficiently by exploiting the sparsity of the signal. However, the much-celebrated CS theory/algorithm requires the sensors to be linear. For neural sensory systems such as the olfactory system, the receptor neurons (sensors) respond nonlinearly to odorant concentration and have a finite response range. Therefore, the CS algorithm does not apply to sensory systems directly, and the question on how olfactory systems compress information remains open. In this talk, Tu will present some recent results on how a relatively small number of nonlinear sensors each with a limited response range can optimize transmission of high dimensional sparse odor mixture information. For neurons without spontaneous activity, Tu found that the optimal coding matrix is sparseā —only a subset of neurons respond to a given odorant with their sensitivities following a broad (such as log-normal) distribution matching the odor mixture statistics. Tu showed that this maximum entropy code enhances the performances of the downstream reconstruction and classification tasks. For neurons with a finite spontaneous (basal) activity, Tu's study showed that introducing odor-evoked inhibition further enhances coding capacity and the fraction of inhibitory interactions for each neuron increases with its basal activity. Comparisons with available experiments in olfactory systems are consistent with his team's theory.
Open To Public
Contact Melanie Lee
Phone 212-327-8636
CLOSE