EE Seminar: Prof. Lior Wolf, CS TAU

~~(The talk will be given in English)

Prof. Lior Wolf
Tel Aviv University, School of Computer Science
Monday, November 24, 2014
15:00 - 16:00
Room 011, Kitot Bldg.

Web-Scale Training for Face Recognition
Abstract
Scaling machine learning methods to massive datasets has attracted considerable attention in recent years, thanks to easy access to ubiquitous sensing and data from the web. Face recognition is a task of great practical interest for which (i) very large labeled datasets exist, containing billions of images; (ii) the number of classes can reach tens of millions or more; and (iii) complex features are necessary in order to encode subtle differences between subjects, while maintaining invariance to factors such as pose, illumination, and aging. We present an elaborate pipeline that consists of a crucial network compression step followed by a new bootstrapping scheme for selecting a challenging subset of the dataset for efficient training of a higher capacity network. We show that two distinct properties have surprising effects on the transferability of deep convolutional networks: (1) The bottleneck of the network serves as an important regularization, and that (2) in contrast to conventional wisdom, performance saturation exist in CNN's (when the number of training samples grows) unless the samples are chosen judiciously. By addressing these issues, we are able to greatly improve face recognition accuracy on the widely used LFW benchmark, both in the verification (1:1) and identification (1:N) protocols, and to bridge between the literature and the industry, for the first time, by directly comparing with the state of the art Commercially-Off-The-Shelf system and show a sizable leap in performance.