Image Comprehension on Mobile and Cloud Platforms

Dr. Soheil Ghiasi
Professor Electrical and Computer Engineering, UC Davis, Davis, CA

Cerent Engineering Science Complex, Salazar Hall 2009A
4:00 PM

Abstract – Convolutional Neural Networks (CNNs) have led to remarkable breakthroughs in a number of important pattern recognition tasks, such as image and voice recognition. Modern CNNs are composed of many layers, have millions (sometimes billions) of parameters, and require massive amount of computation to train them. The training is typically performed using large amount of data on massively-parallel servers equipped with graphics processing units (GPUs). Subsequently, the trained models are utilized in many application scenarios to run 'inference' on user's data, often on resourceconstrained mobile devices. The computational demand of running modern CNNs creates interesting technical challenges, in particular as it relates to deployment of trained models to resourceconstrained devices, such as smart phones and mobile robots. This talk presents an overview of this emerging area. We will also discuss some ongoing work in our group, where we focus on optimizing CNN inference via custom FPGA-based accelerators or mobile graphics processing units (mobile GPUs).

Dr. Soheil Ghiasi is a professor of electrical and computer engineering at the University of California, Davis. His research interests are design and optimization of embedded systems for signal processing, machine learning and multimedia applications. He received his B.S. degree from Sharif University of Technology, Tehran, Iran in 1998, and his M.S. and Ph.D. in Computer Science from University of California, Los Angeles in 2002 and 2004, respectively. He has served on the organizing and technical program committees of numerous technical conferences, and several journals in the general area of computing systems. He is a senior member of IEEE and ACM.