Hiroki Matsutani

Keio University, Japan

An Ultra-Fast On-Device CNN Finetuning Approach for Resource-Limited Edge Devices

Abstract

Neural network retraining is known as a costly operation on resource-limited IoT devices, limiting runtime model optimization opportunities to address performance degradation of pretrained models. I this talk we will review lightweight on-device learning solutions of neural networks. Then we will present our FPGA-based sub-second CNN finetuning approach for resource-limited IoT devices by optimizing forward and backward passes of a parameter-efficient finetuning method. We will demonstrate our prototype systems and their applications.

Biography

Hiroki Matsutani received the BA, ME, and PhD degrees from Keio University, Yokohama, Japan, in 2004, 2006, and 2008, respectively. He is currently a Professor in the Department of Information and Computer Science, Keio University. His research interests are related to computer architecture, interconnection networks, hardware accelerators, and machine learning algorithms.