3:30 p.m. - 4:30 p.m. Location: ECSS 3.910
Please join us for an Electrical and Computer Engineering Seminar featuring Kyuho Lee of the Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea. Thursday, March 23, 2017, 3:30pm-4:30pm, in the ECSS 3.910 conference room.
Mr. Lee will speak on the topic of Analog-Digital Hybrid Computing and Mixed-Mode Circuit Implementations of Neural Networks and Fuzzy Inference Systems, and the Artificial Intelligence Applications.
Abstract: As the revolutionary era of Deep Neural Network (DNN) has emerged, several DNN hardware accelerators have been recently researched with digital circuit implementations. However, they suffer from huge power consumption when it comes to mobile applications, where low-power consumption is the crucial requirement. In this talk, analog-digital hybrid computing as well as the detailed mixed-mode circuits for DNN and Neuro-Fuzzy Inference System (NFIS) will be introduced for efficient and effective hardware implementation. The benefits of mixed-mode implementation can be summarized as: 1) fast speed by parallelized computation, 2) less silicon area occupation, and 3) low-power consumption. The power can be further reduced by optimizing current level used within the network. In addition, digital implementation of controller provides accurate training of DNN and NFIS, as well as high programmability of the network parameters required for analog circuits. Therefore, the mixed-mode implementation becomes not only an efficient, but also an effective approach to DNN and NFIS hardware implementation.
Also, a mixed-mode System-on-Chip (SoC) for the automotive black box system that provides intelligent Advanced Driver Assistance System (ADAS) functions will be presented. The proposed system features the world’s first SoC implementation of Semi-Global Matching algorithm as well as the mixed-mode NFIS for intelligent decision making for the vehicle. By utilizing the mixed-mode computing, the SoC has dual-mode operations: 1) high-performance computation for ADAS and deep risk prediction in driving-mode and 2) ultra-low-power record triggering in parking-mode. As a result, the proposed ADAS SoC achieved the highest energy efficiency compared to the state-of-the-art ADAS SoCs, and was presented in ISSCC 2016, COOLCHIPS 2016, and HOTCHIPS 2016 followed by publication in JSSC 2017.
Biography: Kyuho Lee received B.S. and M.S. degrees in the school of Electrical Engineering from Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea in 2012 and 2014, respectively. Now he is currently pursuing the Ph.D. degree in the same school. Throughout his research, he has implemented 6 SoC’s including low-power mixed-mode neural network classifier, a high-throughput Network-on-Chip architecture for fast computer vision processing, a low-energy object matching processor for real-time object recognition on mobile platform, and real-time convolutional neural network processor. He recently developed an intelligent vision SoC for ADAS which includes high-performance stereo matching for precise depth extraction as well as intelligent decision making system for deep risk prediction, and presented the system at ISSCC 2016, COOLCHIPS 2016, and HOTCHIPS 2016.