......................................................................

NJIT Mathematical Biology Seminar

Tuesday, November 22, 2011, 4:00pm
Cullimore Hall 611
New Jersey Institute of Technology

......................................................................

Identifying behavioral traits through continuous feature selection

Esteban Tabak

Courant Institute of Mathematical Sciences, New York University

We present a methodology for continuous feature selection, based on a descent algorithm that seeks the small dimensional subspace of observed features that optimally discriminates between various classes of observations. The objective function to minimize over the set of subspaces is given by the Kullback-Leibler distance between the real class assignment of the observations and their posterior probability based on density estimation in the chosen subspace. The motivation for this line of work arose from the need to identify those behavioral traits that are most affected by the inhibition or over-activation of individual classes of neurons in Drosophila larvae: one can think of behavioral traits as linear, nonlinear or evan causal combinations of the observed features. This dimensional reduction reveals inner structure of the data: not only does it separate classes whose behavior differ strongly, but it also clusters together classes that behave very similarly, which might prove useful in identifying which classes of neurons belong to the same circuits. This is joint work with Marta Zlatic, at Janelia Farm, and Rebeca Salas-Boni, at the Courant Institute.