Soham Bakshi

Research

Modern data science and machine learning present new challenges that calls for rigorous theoretical frameworks to analyze complex, high-dimensional data and to ensure reliable inference from data driven methods. My research aims to devise innovative statistical methodologies that provide robust inferential guarantees, with applications in causal inference and contemporary data-science. My current research interests can be organized around the following themes:

  1. Explore randomization as a tool to facilitate flexible and tractable inference for a wide-range of data-driven methods

  2. Develop novel statistical frameworks to advance theoretical insight into the complexities arising in today’s AI & ML systems

Publications & Preprints

  1. Inference with Randomized Regression Trees
    Authors: Soham Bakshi, Yiling Huang, Snigdha Panigrahi, Walter Dempsey
    Year: 2024
    Link: arXiv:2412.20535

  2. Selective Inference for Time-varying Effect Moderation
    Authors: Soham Bakshi, Walter Dempsey, Snigdha Panigrahi
    Year: 2024
    Link: arXiv:2412.20535

  3. Bayes Classifier Cannot Be Learned from Noisy Responses with Unknown Noise Rates
    Authors: Soham Bakshi, Subha Maity
    ICLR 2023, Tiny Papers
    Link: OpenReview

  4. How to Approximate Irrational Numbers Nicely?
    Authors: Tirthankar Bhattacharyya, Soham Bakshi, Arka Das
    Year: 2022
    Link: arXiv:2206.12579

  5. Set of Points of Continuity and Maximally Discontinuous Extensions
    Author: Soham Bakshi
    Journal: Resonance, Volume 27, Issue 1, Pages 131–142
    Year: 2022
    Link: Springer

Selected Talks

  1. Selective Inference for Time-Varying Moderated Effects (slides)
    Statistical Analysis of Multi-Outcome Data (SAM) 2024, Salzburg, Austria

  2. Non-Identifiability of Bayes Classifier (slides)
    Joint Statistical Meet (JSM) 2023, Torronto, Canada

  3. Geomnetry in Statistics (slides)
    Math Club 2022, Indian Statistical Institute, Bangaluru Centre, India

  4. Hopf-Rinow’s theorem and Geodesic Completeness (slides)
    Visiting Students Research Program (VSRP) 2021, School of Mathematics, TIFR, India