SublinearOptML · Sublinear Optimization for Machine Learning

The explosion of data with the expansion of the World Wide Web, social networking and e-commerce is a ripe opportunity for smart data analysis and machine learning on a scale never before encountered. In many applications, making even a few passes over the data is prohibitive. Our research goal is to develop sublinear time optimization algorithms for data analysis and machine learning, i.e. algorithms that do not observe all the data even once, and yet return a provably correct solution with high probability.Previous research of the PI, conducted mainly at Princeton University and IBM Research, has produced promising preliminary results. This proposal aims to expand this line of research at the host institution.