GRASP Seminar: Guy Hacohen, Hebrew University, “Learning how neural networks learn: Which examples are important to neural networks, and when?”

*This was a HYBRID Event with in-person attendance in Levine 512 and virtual attendance…

ABSTRACT

In this talk, I’ll present my Ph.D. research at Daphna Weinshall’s lab. Our research focus is on the learning dynamics of neural networks. In this talk, we’ll draw inspiration from human learning, discussing how to divide the learning process of neural networks into several qualitatively different stages. We’ll show that different types of examples are important in different learning stages, allowing the use of curriculum learning approaches. We’ll continue by discussing the resulting qualitative differences in active learning strategies for the high and low-budget regimes. We’ll conclude by analyzing the order in which neural networks learn new data points, finding vast similarities across different architectures and in several data domains.