Feature Learning by Large-Scale Heterogeneous Networks with Application to Face Verification

Thomas S. Huang, University of Illinois at Urbana-Champaign

Usage Details

Feature learning lies at the core of all visual recognition problems, which are as diverse as object categorization, scene understanding, video event recognition, face verification, and soft- biometrics. To date, one of the most successful approaches for feature leaning is Deep Neural Networks (DNN) (aka deep learning). However existing deep learning methods are limited; they are not able to actively model semantics present at many levels in the image. This project focuses on the Heterogeneous Network (HN) for learning image features. As a natural extension to the recently popular Deep Neural Network, HN perfectly accommodates the structural image semantics via a novel layer-wise training strategy. Existing research indicates that the learning of large-scale, many-layer neural networks is computationally expensive. Here we propose a distributed learning framework for HNs using Asynchronous Stochastic Gradient Descent (ASGD), which scales up to thousands of computers.