Abstract: Mixture of simple machine learning models can represent complicated relations between features and labels with reasonable number of parameters in situations where single models with very large number of parameters (e.g. deep neural networks) are inadequate. The smaller number of parameters directly translate into computational efficiency: and also the simplicity of the models render them suitable for provable guarantees.
A sustainable and efficient training phase also calls for active learning frameworks where carefully designed features are queried for labels. In this talk we discuss complexities of learning mixtures of simple ML models (such as linear classification and regression) in the active learning framework. The problems have strong relations with many statistical reconstruction problems, and our solutions exploits interesting connections with properties of complex polynomials, de-mixing Gaussians and error-correcting codes.

Bio: Arya Mazumdar is an associate professor in the College of Information and Computer Sciences at the University of Massachusetts Amherst, on leave now at Amazon AI and Search Sciences. Prior to this, Arya was an assistant professor at the University of Minnesota-Twin Cities, and a postdoctoral scholar at Massachusetts Institute of Technology. Arya received his Ph.D. from the University of Maryland, College Park, where his thesis won the Distinguished Dissertation Award. Arya is a recipient of an NSF CAREER award and an IEEE ISIT Jack K. Wolf Paper Award. Arya’s research interests include machine learning, coding theory, and information theory. He is an associate editor of the IEEE Transactions on Information Theory.