Seminar in Computing + Mathematical Sciences

Goal of metric learning is to learn a notion of distance---or a metric---in the representation space that yields good prediction performance on data. In this tutorial we explore some classic ways one can efficiently find good metrics. Starting from the basics, we'll cover classic techniques like Mahalanobis Metric for Clustering (MMC) and Large Margin Nearest Neighbor (LMNN) and discuss key principles that make these techniques effective in improving prediction performance. We will also study some extensions and see how metric learning has helped in ranking problems (information retrieval) and large scale classification.