Soumya Ghosh
VP, Foundation AI, WHOOP
I am a VP of Foundation AI at WHOOP. Before this I was a Director of machine learning at Merck Research Labs and before that I was a researcher at IBM Research, Cambridge and the MIT-IBM Watson AI lab. I work on building both foundation models and statistical machine learning models to understand and explain images, text, and real-world healthcare data. My work also explores the robustness of these models to modeling and data perturbations.
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.
- all data
- exact refit
- jackknife estimate
Drop a training point and the fit moves. The infinitesimal jackknife approximates the new fit without actually refitting. From Approximate Cross-Validation for Structured Models, NeurIPS 2020.
Recent highlights
- ICML 2024 paper on calibrating large language models. It requires only a single forward pass through the LLM, and can learn to calibrate without labeled data. MIT News wrote a high-level gist of the work.
- NeurIPS 2022 paper on improving the fairness of pre-trained classifiers by detecting and dropping training instances that contribute to unfairness. There is a short video and a blog post describing it.
- Are Gaussian process predictions sensitive to the choice of kernel? Sometimes. We show how to check a GP-based analysis for that sensitivity in this AISTATS paper.
- A comprehensive toolbox for uncertainty quantification, described here.
- MLHC 2020 work on statistical models of Parkinson’s disease progression, covered by DigitalTrends, VentureBeat and TechRepublic.
