Aaditya Ramdas – Machine Learning for Hypothesis Testing

One of the themes of my research has been on how to use machine learning and Bayesian techniques for testing hypotheses, but with proper frequentist error guarantees. Informally, these methods usually decouple type-1 and type-2 error guarantees. The false positive rate (or equivalent) is often controlled under “minimal” assumptions, independent of which ML method or Bayesian prior/model one chose to employ. The power (or equivalent) will, of course, depend on the Bayesian model used or the ML method being employed. This idea has now appeared in several different contexts in my recent work, and I continue to explore this theme in current work.

  • On the power of conditional independence testing under model-X
    E. Katsevich, A. Ramdas       (to be submitted)   arxiv

  • Fast and powerful conditional randomization testing via distillation
    M. Liu, E. Katsevich, L. Janson, A. Ramdas       Biometrika, 2021   arxiv   code

  • Classification accuracy as a proxy for two sample testing
    I. Kim*, A. Ramdas*, A. Singh, L. Wasserman       Annals of Stat., 2021   arxiv   proc

  • Interactive martingale tests for the global null
    B. Duan, A. Ramdas, S. Balakrishnan, L. Wasserman       Electronic J of Stat., 2020   arxiv   code   proc

  • Familywise error rate control by interactive unmasking
    B. Duan, A. Ramdas, L. Wasserman       ICML, 2020   arxiv   code

  • Which Wilcoxon should we use? An interactive rank test and other alternatives
    B. Duan, A. Ramdas, L. Wasserman       arxiv

  • Simultaneous high-probability bounds on the FDP in structured, regression and online settings
    E. Katsevich, A. Ramdas       Annals of Stat., 2020   arxiv   code   proc

  • STAR: A general interactive framework for FDR control under structural constraints
    L. Lei, A. Ramdas, W. Fithian       Biometrika, 2020   arxiv   movies

  • MAB-FDR: Multi (A)rmed/(B)andit testing with online FDR control
    F. Yang, A. Ramdas, K. Jamieson, M. Wainwright       NeurIPS, 2017   arxiv   code   30-min talk   proc   (spotlight talk)