Abstract

This talk will begin by introducing information elicitation, a research area that generalizes proper scoring rules for eliciting truthful forecasts from strategic agents. Then, we will see how results from this area impact design of loss functions for machine learning. We'll formalize what it means to have a "good" surrogate loss function, e.g. why do we use hinge or logistic loss as a surrogate for 0-1 loss? In particular, we'll focus on the dimensionality of the surrogate hypothesis space and use information elicitation techniques to give some upper and lower bounds, particularly for what we call the embedding approach.

Based on recent and ongoing work with Jessie Finocchiaro and Raf Frongillo, including: https://arxiv.org/abs/1907.07330

Biography

Bo Waggoner is an assistant professor of Computer Science in University of Colorado, Boulder researching machine learning, AI, game theory, and EconCS.

• Admission

Zoom ID: 680 3831 4216

Link: https://zoom.com.cn/j/68038314216