Title: Towards Provable and Scalable Machine Learning
PhD Candidate: Jin Lu
Major Advisor: Jinbo Bi
Associate Advisors: Sanguthevar Rajasekaran and Alexander Russell
Date/Time: Monday, March 25, 2019 1:00PM
Location: HBL 1947 Conference Room
Abstract:
In the recent decades, machine learning has been substantially developed and they have demonstrated great success in various domains such as web search, computer vision, natural language processing, and more. Despite of its practical success, many of these applications involve solving NP-hard problems in average cases based on heuristics. Theoretically analyzing the heuristics is very challenging. In this proposal, we will show that it is possible to solve the problem efficiently and effectively in a different approach: assuming the data has these properties, we identify natural structures of the data, then design new models and algorithms that provably works. The first part of the proposal, we propose a new method for matrix completion with side information. We are going to examine that under a reasonable assumption, it can be achievable to provably recover the missing values within a partially observed matrix, with a much lower sampling rate comparing with the classic matrix completion methods. The second part of the proposal provides ideas for provably learning generative adversarial networks with capacity control. Assuming the input data has clustering structures, we propose a novel Kolmogorov-width Coupled Nets using Surrogate Kolmogorov-width, and we are going to examine the provable guarantees for recovering each cluster distribution and the global distribution.