John Duchi, University of California, Berkeley
Wednesday, February 19, 4:30pm
Computer Science 105
How can we maximally leverage available resources--such as computation, communication, multi-processors, or even privacy--when performing machine learning? In this talk, I will suggest statistical risk (a rigorous notion of the accuracy of learning procedures) as a way to incorporate such criteria in a framework for development of algorithms. In particular, we follow a two-pronged approach, where we (1) study the fundamental difficulties of problems, bringing in tools from optimization, information theory, and statistical minimax theory, and (2) develop algorithms that optimally trade among multiple criteria for improved performance. The resulting algorithms are widely applied in industrial and academic settings, giving up to order of magnitude improvements in speed and accuracy for several problems. To illustrate the practical benefits that a focus on the tradeoffs of statistical learning procedures brings, we explore examples from computer vision, speech recognition, document classification, and web search.