Date and Time:
A major challenge for machine learning in the next decade is the development of methods that continuously predict and learn from a stream of data. The scale of data and the non-stationary nature of the environment make the prevalent "learn from a batch of i.i.d. data" paradigm inadequate. In this talk, we will formally define the problem of sequential prediction as a (computationally difficult) optimization problem. We will present novel techniques for approximately solving it. These techniques recover many prediction algorithms known in the literature, but also yield new and surprising methods with near-optimal performance. The advances are possible due to a confluence of ideas from empirical process theory, game theory, and optimization. The field of sequential prediction is largely unexplored, and we hope to present a framework that might be of interest to the optimization community.
Alexander Rakhlin is an Assistant Professor at the Department of Statistics and the Department of Computer and Information Science (secondary) at the University of Pennsylvania. He received his bachelors degrees in Mathematics and Computer Science from Cornell University, and doctoral degree from MIT. He was a postdoc at UC Berkeley EECS before joining UPenn, where he is a co-director of the Penn Research in Machine Learning (PRiML) center. His research is in machine learning, with an emphasis on statistics and computation.