50 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Flexible Neural Networks and the Frontiers of Meta-Learning
Stanford University
Recording
Abstract
Deep models can learn perception and control, but each new skill can require substantial data and human labeling. Chelsea Finn explains how training across related tasks can yield models that adapt with less experience. She surveys what meta-learning can handle and examines three practical challenges: improving adaptation speed in an online stream, constructing tasks without human supervision, and training across broad task distributions.