Speaker: Dr. Rahul Mazumder, Nanyang Technological University Associate Professor of Operations Research and Statistics, MIT Sloan School of Management

Title: Algorithmic Aspects of Statistical Learning with Combinatorial Structures

Abstract: Optimization problems involving discrete or combinatorial structure frequently arise in statistics and machine learning. Some examples include: sparse high dimensional linear models, Gaussian graphical models, decision trees and ensembles, pruning and quantization of neural networks, among others. Combinatorial structures can be useful from a statistical viewpoint (e.g, interpretation, fewer model parameters, etc) and can offer computational benefits (e.g., in downstream prediction-tasks due to fast model evaluations, improved storage, etc).  In this talk I'll discuss some of these combinatorial problems, and how tools from mathematical optimization can offer a conceptually appealing framework for modeling and computation. Large-scale computation of these problems however poses interesting challenges. In some cases, under statistical assumptions, the underlying optimization problems can be simplified. Time permitting, I'll discuss how "neural training" (e.g, end-to-end training of tree ensembles, flexible additive models) offers a flexible and practical algorithmic framework for some of these problems, though with limited theoretical guarantees.