We are seeking Ph.D. students in Machine Learning (particularly analyses and practical techniques for deep learning, generative models, and LLMs), Signal Processing, and Quantum Information and Computing. Welcome candidates from any background, e.g., EECS, math, physics.
Call for papers: Conference on the Mathematical Theory of Deep Neural Networks, manuscript due: November 14-15, 2024, Philadelphia
Call for papers: Conference on Parsimony and Learning (CPAL), March 2025, Stanford
News:
[Sep 2020] Our paper has been accepted at NeurIPS as spotlight (top 4%), which characterizes implicit bias with discrepant learning rates and builds connections between over-parameterization, RPCA, and deep neural networks.
[Jun 2020] Two papers about over-parameterization are on arXiv: one studies the benefit of over-realized model in dictionary learning, another one characterizes implicit bias with discrepant learning rates and builds connection between over-parameterization, RPCA, and deep neural networks.
[Jan 2020] Co-organized with Qing and Shuyang, our two-session mini-symposium ‘‘Recent Advances in Optimization Methods for Signal Processing and Machine Learning’’ has been accepted by the inaugural SIAM Conference on Mathematics of Data Science. See you at Cincinnati, Ohio in May!
[Nov 2019] Our paper (with Xiao, Shixiang, Zengde, Qing, and Anthony) ‘‘Nonsmooth Optimization over Stiefel Manifold: Riemannian Subgradient Methods’’ is on arxiv. This work provides (first) explicit convergence rate guarantees for a family of Riemannian subgradient methods when used to optimize nonsmooth functions (that are weakly convex in the Euclidean space) over then Stiefel manifold.
[Oct 2019] Attended the Computational Imaging workshop at IMA, University of Minnesota, and presented our work on ‘‘A Linearly Convergent Method for Non-smooth Non-convex Optimization on Grassmannian with Applications to Robust Subspace and Dictionary Learning’’.
[Aug 2019] Our paper (with Xiao, Anthony, Jason) ‘‘Incremental Methods for Weakly Convex Optimization’’ is on arxiv. This work provides (first) convergence guarantee for incrememtal algorithms and their random shuffling version (including the incremental subgradient method which is the work-horse of deep learning) in solving weakly convex optimization problems which could be nonconvex and nonsmooth.
|