How to Learn Without Data

Classical supervised machine learning starts from a collection of input-output data pairs corresponding to a well-defined task. Predictive accuracy then scales as the number of datapoints, the size of the function class, and the number of optimization steps increase in appropriate, relative proportions. In this talk, Zaid Harchaoui will describe a new learning and prediction paradigm in which one can learn a highly predictive model without such data for the task at hand, nor even a formal definition of the actual task. This new paradigm, called zero-shot prediction, has gained a lot of momentum over the last decade. It draws its strength from a fundamental equality, which can be understood both from an information theoretic and from a operator theoretic perspective. The talk will cover the origins of this approach as well as its latest incarnations in computer vision and language modeling, and it will highlight current opportunities and challenges related to this approach.

Zaid Harchaoui is a professor in the Department of Statistics, the Allen School of Computer Science and Engineering, and the Department of Mathematics at the University of Washington in Seattle, and a senior data science fellow in the eScience Institute. He is an acting editor at the Journal of Machine Learning Research and a principal investigator and co-founder of IFML, the NSF-AI Institute on Foundations of Machine Learning. He obtained his doctoral degree from the Institut Polytechnique de Paris — Telecom Paris for research performed at CNRS — the French National Institute for Fundamental Research. He previously held appointments at the Courant Institute of Mathematical Sciences at New York University and at INRIA — the French National Institute for Research in Digital Science and Technology.

Light refreshments will be served. This seminar is sponsored by the Data Science Institute and the RISE-AI Collaboration HQ.