Metric Learning with Eigenvalue Optimization
Yiming Ying
Assistant Professor
Computer Science College of Engineering, Mathematics and Physical Sciences University of Exeter
Abstract: Distance metric is a fundamental concept in Machine Learning since a proper choice of a metric has crucial effects on the performance of both supervised and unsupervised learning algorithms.
In this talk I will present our recent work in this challenging research direction, starting with an introduction to metric learning problems. The main theme of this talk is to present a novel eigenvalue-optimization framework for learning a Mahalnobis metric from data. Within this context, we introduce a novel metric learning approach called DML-Eigen which is shown to be equivalent to a well-known eigenvalue optimization problem called minimizing the maximal eigenvalue of a symmetric matrix. Moreover, we show that similar ideas can be extended to large margin nearest classifiers
(LMNN) and maximum-margin matrix factorisation for collaborative filtering (e.g. predicting customers’ preference to products). This novel framework not only provides new insights into metric learning but also opens new avenues to the design of efficient metric learning algorithms. Indeed, first-order algorithms scalable to large datasets are developed and their convergence analysis will be discussed in detail. At last we show the competitiveness of our methods by various experiments on benchmark datasets. In particular, we report an encouraging result on a challenging face verification dataset called Labeled Faces in the Wild (LFW).
Short Bio: Dr. Yiming Ying received his B.S. degree in mathematics from Hangzhou University in 1997, Hangzhou, China and his PhD degree in mathematics from Zhejiang University in 2002, Hangzhou, China. Currently he is a Lecturer (Assistant Professor) in Computer Science in the School of Engineering, Computing and Mathematics at the University of Exeter, United Kingdom. His research interests include machine learning, learning theory, optimization, probabilistic graphical models and the applications to computer vision, bioinformatics and multimedia data analysis. http://emps.exeter.ac.uk/staff/yy267