Research Overview:

I am interested in the generalization theory of machine learning algorithms. My Ph.D. thesis provides a rigorous information-theoretic framework to characterize and improve the generalization capabilities of supervised learning algorithms. 

Through distributionally robust optimization with Kullback-Leibler divergence constraints, closed-form expressions for generalization error, model sensitivity, etc., are provided in terms of information measures. Building on a training-dependent minimax framework, an algorithm robustification method is proposed to improve the performance of a given learning algorithm w.r.t unseen data, which addresses the gap between empirical training performance and reliability under distribution shifts.

• Google Scholar

• Publications

• CV