Hi, my name is
Soham Bonnerjee.
I think about
Email: sohamb@wharton.upenn.edu
Office: Academic Research Building, 265 South 37th Street, Philadelphia, PA 19104
I am a postdoctoral researcher at the Department of Statistics & Data Science at the The Wharton School, University of Pennsylvania. I completed my PhD in the Department of Statistics at the University of Chicago, where I was fortunate to be advised by Prof. Wei Biao Wu. Prior to my doctoral studies, I obtained my Bachelor and Masters in Statistics degree from the Indian Statistical Institute, Kolkata.
News
- (July, 2026) Our paper on fast segmentation on spatial anomalous patches received a revision request from Biometrika.
- (June, 2026) Our paper on differentially private in-context learning has been accepted at TMLR.
- (April, 2026) Our paper on Algorithmic stability analysis under unbounded loss has been accepted at ICML 2026.
- (April, 2026) I feel grateful to receive the IMS New Researcher Travel Award for our work on Watermark segmentation in Large Language Model. Cheers to all other recipients!
- (March, 2026) We introduce SPLADE: a fast, provably consistent algorithm to localize rectangular anomalous patches in presence of spatial dependency. Check the paper and the code implementation out.
- (Feb., 2026) Our paper on Watermark segmentation in Large Language Model received a Major Revision request from Journal of American Statistical Association (JASA).
- (Jan., 2026) Check out our new prints on Edge-of-Stability analysis of SGD, which establishes first statistically valid estimator for the largest step-size you can choose without exploding the chain, and Algorithmic stability analysis under unbounded loss for ERM, transductive regression and meta-learning based on weak Lp-Lipschitz stability assumptions.
- (Jan., 2026) Our paper on Statistical inference with LD2Z learning rate for Q-learning has been accepted at ICLR 2026.
- (Sept., 2025) Check out our new preprint on Watermark segmentation in Large Language Models. Our novel, epidemic change-point perspective obtains the first ever theoretically valid and computationally efficient algorithm to solve watermark segmentation.
- (Sept., 2025) Our papers on Gaussian approximations of Decentralized Federated Learning, and Asymptotic theory of SGD with general learning rate, have been accepted at NeurIPS 2025 as Spotlight (Top 3%) and Poster respectively. This was my first submission and acceptance at NeurIPS.
- (Sept., 2025) Our paper on Synchronization of change-points received a Major Revision request from Biometrika.