Soumya Ghosh

Research Staff Member

I am a research scientist at IBM Research, Cambridge and the MIT-IBM Watson AI lab. I work on developing statistical models for understanding and explaining images, text, and noisy, real world healthcare data.I hold a Ph.D. in Computer Science from Brown University, where I was advised by Erik Sudderth. Before Brown, I spent a few years in beautiful Boulder getting a master’s degree from the University of Colorado. At Colorado, I was advised by Jane Mulligan. Going further back, I went to the University of Mumbai (Bombay) (KJSCE) as an undergrad. I also spent a year as a postdoctoral scientist at the now defunct Disney research, Cambridge.

Recent Highlights

  • A comprehensive overview of learning Bayesian neural networks with Horseshoe priors will appear in JMLR.
  • New code release for distance dependent Chinese restaurant processes. The code by Ishana Shastri is an efficient, python translated, version of this old MATLAB package.
  • Our Work on using Bayesian non-parametric meta models for fusing local models will appear at NeurIPS 2019.
  • Commonly used metrics such as test log likelihoods can be misleading indicators of posterior quality of BNNs. Preliminary work will appear at Uncertainty and Robustness workshop at ICML’ 19.

Top Work

SPAHM: Parameter matching for model fusion

SPAHM: Parameter matching for model fusion

Bayesian Modeling

Publications with the MIT-IBM Watson AI Lab

Model Fusion with Kullback–Leibler Divergence
Model Fusion with Kullback–Leibler Divergence
 
Approximate Cross-Validation for Structured Models
Approximate Cross-Validation for Structured Models
 
SPAHM: Parameter matching for model fusion
SPAHM: Parameter matching for model fusion
 
Bayesian Nonparametric Federated Learning of Neural Networks
Bayesian Nonparametric Federated Learning of Neural Networks
 
Unsupervised learning with contrastive latent variable models
Unsupervised learning with contrastive latent variable models