Abstract

Reconfigurable technology has become popular for high performance and low power implementations of deep learning networks, such as Convolutional Neural Network (CNN) designs. Most CNN applications on FPGA are domain-specific involving, for example, detecting specific types of objects. This paper presents an end-to-end approach for efficient development of domain-specific applications on CNNs based on transfer learning techniques, targeting reconfigurable devices. Such techniques enable adaptation of pre-trained models to specific domains by replacing standard convolution layers with efficient convolution blocks, and by applying layer fusion to enhance hardware performance. The proposed approach is illustrated by optimising a pre-trained VGG-16 model for an image recognition task targeting a Stratix V device, with promising improvement in accuracy and performance.

Speaker's Biography

Wayne Luk is Professor of Computer Engineering in Department of Computing at Imperial College London. He was Visiting Professor at Stanford University from 2006 to 2009. He founded and leads the Computer Systems Section at Imperial College. His research interests include reconfigurable computing, field-programmable technology, and design automation. He is a fellow of the Royal Academy of Engineering, a fellow of the IEEE and a fellow of the BCS.