On July 5 2017, Prof. Weibo Hu from Nankai University visited the Institute of Microelectronics of Chinese Academy of Sciences (IMECAS). He delivered a lecture entitled “Memory-Assisted Analog Signal Processing in Deep Learning Chips” in IMECAS hosted by Prof. Qi Liu. More than 30 local professionals and graduate students attended the meeting.
The concept of machine learning appeared in the computer programing as early as 1959. These years because of unprecedented processing power, huge available data, appealing market value and other factors, deep learning as a kind of machine learning has become a hot research topic. However, implementing and running deep learning in digital circuits are very power-hungry and hardware-demanding. Because the neuromorphic network which is normally used in the deep learning has millions or even billions of multipliers and accumulators (MAC) which are very expensive in digital circuits. Memory-assisted analog signal processing provides a fundamentally different way to implement MAC with acceptable precisions. Analog circuits can use the Ohm law to simply implement the multiplication, and use the Kirchoff Current Law to implement the summation. The analog way seems to dramatically be energy-efficient, but it has many design challenges and some application bottlenecks. In this lecture, Prof. Hu introduced the reasons behind memory-assisted analog signal processing in the perspective of a mixed-signal circuit designer, and previous and current designs, and explained the future research direction. The new points of his research on deep learning chips made a big splash and reached an academic discussion.
After the lecture, a symposium was held with the participating researchers from the IMECAS, both sides introduced their recent works and had a heated discussion.
Prof. Hu was giving a talk.
Biography
Dr. Hu has got Bachelor and Master from Harbin Institute of Technology and Peking University in 2005 and 2008, respectively. He got his Ph. D. from Texas Tech. University in 2014. He had worked at the Mixed-Signal Department in Qualcomm in San Diego for about 4 years before he got a faculty position in Nankai University. He has published about 20 journal or conference papers, got about 200 Google Scholar citations and hold 3 patents. During his time in Qualcomm, as a main circuit designer he took part in the development of 4 audio CODEC chips which have been sold more than half billion dies.In 2016, he led a team to build up an integrated circuit lab in Nankai University. All team members are either from top fabless companies or top US universities. A major research area is the application of the deep learning on integrated circuits, including the implementation of the deep learning chip via memory-assisted analog circuits and the application of the deep learning algorithm on circuit verification or CAD.
