- This event has passed.
Exploiting metric structure for more accurate classification (Prof. Mike Izbicki)
February 2, 2022 @ 4:15 pm - 5:30 pm
Title: Exploiting metric structure for more accurate classification
Speaker: Mike Izbicki, Department of Mathematical Sciences, Claremont McKenna College
Abstract: Classification problems often have many semantically similar classes. For example, the famous ImageNet dataset contains classes for 80 different dog breeds, 40 different bird species, and 25 types of vehicles. This semantic structure can be formalized using a metric space, with semantic similarity of classes encoded by the distance function. In this talk, I’ll describe the “tree loss”, which is the first technique with provable performance guarantees for exploiting this metric structure. I’ll also show that the tree loss has better empirical performance than competing algorithms on image, text, and vector data.