Events Calendar

Computer Engineering Seminar

Speaker: Jae-sun Seo, Arizona State University

Talk Title: Efficient Digital Hardware Design for Machine Learning and Neuromorphic Algorithms

Abstract: In recent years, machine learning algorithms (e.g. convolutional neural networks, deformable parts model) have been widespread across a broad range of image, video, speech, and biomedical applications. For similar applications, there also has been a surge of interest in neuromorphic computing and spiking neural networks (e.g. TrueNorth), which more closely follow biological nervous systems. In this talk, we present our exemplary research work on efficient digital hardware design for both machine learning and neuromorphic algorithms.
On the machine learning side, algorithms trained by offline back propagation works well on pre-defined datasets, but state-of-the-art algorithms are compute-/memory-intensive, making it difficult to perform low-power real-time classification. Our prototype designs in FPGA and ASIC frameworks are presented that improve the energy-efficiency (GOPS/W) by optimizing computation, memory, and communication for representative large-scale networks.
On the neuromorphic side, the classification accuracies on MNIST or ImageNet datasets has not yet reached those of machine learning counterparts, but we find it suitable for unsupervised continuous online learning applications (e.g. defense, robotics, biomedical) aiming low power consumption. Building up on earlier work on on-chip STDP (spike-timing dependent plasticity) learning for pattern recognition (45nm) and spiking clustering for deep-brain sensing (65nm), we propose a versatile neuromorphic processor that can support various STDP learning and inhibition rules with large fan-in/out per neuron. Preliminary implementation results and future research directions will be discussed.

Biography: Jae-sun Seo received his Ph.D. degree from the University of Michigan in 2010 in electrical engineering. From 2010 to 2013, he was with IBM T. J. Watson Research Center, where he worked on energy-efficient circuits for high-performance processors and neuromorphic chip design for the DARPA SyNAPSE project. In January 2014, he joined Arizona State University as an assistant professor in the School of ECEE. During the summer of 2015, he was a visiting faculty at Intel Circuit Research Labs. His research interests include efficient hardware design of learning algorithms and integrated power management. He received the IBM outstanding technical achievement award in 2012, and serves on the technical program committee for ISLPED.

Host: Prof. Massoud Pedram

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248

Audiences: Everyone Is Invited

Contact: Annie Yu