Learning theory for exponential families

Provable, sample-efficient recovery of high-dimensional distributions.

Exponential-family distributions are the beating heart of many modern AI/ML systems, including that of the poor AI agent that coded this website. My research here deals with fundamental problems in learning these distributions in parametric settings. When possible, we try to prove non-asymptotic sample complexity results and try to keep our computational complexities polynomial. From a physics perspective, all of these works are very closely related to inverse problems in statistical physics. Below you will find results about learning Ising models, Potts models, Phi-4 theories, and even systems with Grassmann variables.

Papers

  1. Learning of discrete graphical models with neural networks
    Abhijith Jayakumar, Andrey Lokhov, Sidhant Misra , and 1 more author
    Advances in Neural Information Processing Systems, 2020
  2. Discrete distributions are learnable from metastable samples
    Abhijith Jayakumar, Andrey Y Lokhov, Sidhant Misra , and 1 more author
    Nature Communications, 2026
  3. Finite Sample Bounds for Learning with Score Matching
    Devin Smedira, Abhijith Jayakumar, Sidhant Misra , and 2 more authors
    In The Thirty Ninth Annual Conference on Learning Theory , 2026
  4. Efficient learning of lattice gauge theories with fermions
    Shreya Shukla, Yukari Yamauchi, Andrey Y. Lokhov , and 2 more authors
    Phys. Rev. D, Jun 2026