Machine Learning Lunch Meeting

General: MLLM is a cross-discipline weekly seminar open to all where UW-Madison professors present their research in machine learning, both theory and applications.  The goal is to promote the great work at UW-Madison and stimulate collaborations. Please see our website for more information.

Speaker: Fred Sala (CS)

Abstract: The weak-to-strong generalization phenomenon is the driver for important machine learning applications including highly data-efficient learning and, most recently, performing superalignment. While decades of research have resulted in numerous algorithms that produce strong empirical performance, understanding what aspects of data enable weak-to-strong generalization has been understudied. In this talk, I will describe a simple data-centric mechanism that characterizes weak-to-strong generalization, which we term the overlap density. Intuitively, generalization tracks the number of points that contain overlaps, i.e., both easy patterns (learnable by a weak model) and challenging patterns (only learnable by a stronger model), as with such points, weak predictions can be used to learn challenging patterns by stronger models. I will present a practical overlap detection algorithm to find overlap density from data and an approach to learn, among multiple sources of data, which to query when seeking to maximize overlap density and so enhance weak-to-strong generalization. I will show empirical and theoretical validations in a diverse array of settings. Finally, I will describe some of the implications of these ideas to future aspects of model supervision, including superalignment.