October 24, 2019: Seminar by Prof. Kaushik Majumdar, SSIU, ISI, Bangalore

A novel estimation of mutual information
by
Prof. Kaushik Majumdar
Systems Science and Informatics Unit
Indian Statistical Unit, Bangalore

Date: Thursday, October 24, 2019
Time: 15:00 HRS
Venue: ECSU Seminar Room, 9th Floor, S N Bose Bhavan  (Library Building)

Abstract

Mutual information is perhaps the most important interdependence measure among random variables and one of the two most widely used information theoretic measures outside the information theory (the other is entropy, which is closely related to mutual information). However, computing mutual information requires the knowledge of joint and marginal probability distributions. Here, we will take a nonparametric approach to estimate mutual information among discrete random variables. First, we will introduce a model based approach to information encoding by deformation of manifold. In the current work it is one dimensional, but multidimensional extension is straight forward. Invoking a discrete version of the scheme we will prove that there are precisely 13 number of 3-point motifs in which any discrete random variable can be decomposed. We have computed mutual information across random variables essentially by 1 nearest neighbor method in terms of the 3-point motifs, which of course can be extended to k (k > 1) nearest neighbor method, but only with marginal improvements in the outcome. The 3-point motifs are the smallest to capture nonlinearity.

 

 

All are cordially invited.

 

Dipti Prasad Mukherjee
Head,
Electronics and Communication Sciences Unit