Bhargob Deka's Research Page

Position

Ph.D candidate | 2018-2022

Contact

Bhargob.deka@polymtl.ca

B-431.1 Main Building

Ecole Polytechnique

Montreal, CANADA

Education

Ph.D., Civil Engineering | 2018-2022
   Polytechnique Montréal, Montréal Canada

M.Sc., Civil Engineering | 2016-2018
   McGill University, Montréal Canada
   Advisor: Luc Chouinard

B.E., Civil Engineering | 2010-2014
   Assam Engineering College, Guwahati,
   Assam, India

Research

Project

Title: Online estimation of parameters in Bayesian dynamic linear models (BDLM) and Bayesian neural networks (BNN) for time series in the context of Structural Health Monitoring

Summary: This research focuses on developing novel estimation methods for finding optimal parameters online. This will help to bypass offline optimization techniques and reduce the computational burden significantly, thereby facilitating scalable models for larger and complex datasets. Currently, we have two articles (in review): 1) Gaussian Multiplicative Approximation in state-space models (GMA): This is a novel method that provides closed-form solution for the moments of the product of two Gaussian random variables. It allows us to estimate the parameters as hidden states together with other hidden states. 2) Online Bayesian Inference of Process Noise in State-Space Models (AGVI): This method provides an analytically tractable online Bayesian inference method for the process-error covariance matrix in state-space models. Using AGVI, we can treat the error-variance-terms also as hidden states and estimate the mean as well as the variance of the error terms at any time step. 3) Heteroscedastic Aleatory Uncertainty in Regression Tasks using TAGI and AGVI (in progress): In this work, we are combining AGVI with TAGI to estimate heteroscedastic aleatory uncertainty. 

Research interests: Bayesian neural networks, Uncertainty estimation, online learning, statistical signal processing.

Publications