2015.04.08_byounghoon kim

CNIR Neuro@noon
2015. 4. 8, 12:00 PM
Room 85777, Sungkyunkwan University, Suwon
Cracking Neuronal Coding: Probabilistic approach
Byounghoon Kim/ Ph.D.
Baylor College of Medicine
Abstract
The brain is the ultimate center of information processing which helps animals survive in harsh
environments and make us thrive on the earth. It makes a choice among multiple options, control
movements, memorize and recall facts when in need, and navigate through the rough outside world.
For neuroscientists, it has been a fundamental quest to understand how the brain works, and studying
the brain could open up doors to a better understanding of who we are: humanity.
Conventional electrophysiological studies mainly report the neuronal activities on the single neuron
level as evidence of the localized function of specific brain areas, such as retinotopically mapped
saccade related burst cells in superior colliculus, place cells and head direction cells in the limbic
system and so on. The neuronal activities are analyzed based on the averaged firing rate. Most studies
focus on correlations between the change of averaged neuronal firing rate and observed behavior such
as the neuronal tuning function versus psychometric function. Generally, variance of neuronal firing
rate in trials has been considered as a nuisance variable that needs to be minimized to increase the
significance of statistical tests. However recent studies with probabilistic approaches shed a new light
on the variance along with the averaged firing rate. In this presentation, two studies will be presented to
show how the probabilistic approaches are implemented to help us understand the brain based on the
neuronal responses and determine how to decode the neuronal responses on the population level and
how much information is carried in the neuronal responses. First, we will review a study which reveals
that the Bayesian inference (maximum a posteriori estimate: MAP) predicted a selecting action better
than other linear models in the superior colliculus. In this study, MAP was calculated from posterior
probabilities which were derived empirically from the neuronal responses. The second study shows that,
even in tightly coupled behaviors, the Fisher information analysis could be implemented to determine
which behavior is more efficiently encoded in the neuronal responses. As Fisher information is
determined by the discriminability (slope of tuning function) and the consistency of neuronal responses
(variance across trials), it provides testable measures of encoding accuracy in neuronal responses.