NTU Team Receives Best Paper Award at Top Embedded Systems Conference

Date: 2019/11/4

Image1:The Best Paper Award is presented during the CODES+ISSS conference.Image2:Certificate of award for the CODES+ISSS Best Paper Award.

The Best Paper Award is presented during the CODES+ISSS conference.

Certificate of award for the CODES+ISSS Best Paper Award.

During the Embedded Systems Week (ESWEEK) held in New York this year, the research conducted by the industry-academia team led by NTU Professor Tei-Wei Kuo (郭大維) was selected for the Best Paper of the Year award by the program committee of ACM/IEEE International Conference on Hardware/Software Codesign and System Synthesis (CODES+ISSS), a conference co-organized by ACM (Association for Computing Machinery) and IEEE (Institute of Electrical and Electronics Engineers). This marked the first time a Taiwanese research team received such a great honor.

The award-winning paper, “Achieving Lossless Accuracy with Lossy Programming for Efficient Neural-Network Training on NVM-Based Systems,” was the brainchild of the team comprised of members from NTU, Academia Sinica, and Macronix International. Besides receiving this recognition, this paper was accepted by ACM TECS (ACM Transactions on Embedded Computing Systems), the world’s leading academic journal in the field of embedded systems. This year marked the 28th edition of ACM/IEEE CODES+ISSS. Since 2005, it has been joined by ACM EMSOFT (International Conference on Embedded Software) and CASES (International Conference on Compilers, Architecture, and Synthesis for Embedded Systems) to form ESWEEK. This grand event in the world of embedded systems attracts a lot of attention from both academia and industry, which makes the win highly commendable!

This study indicated that the training of neural networks involves a high demand for RAM capacity, which causes significant energy consumption and cost. To overcome this obstacle, the NTU team proposed the use of non-volatile memory, more precisely, phase change memory. To improve training efficiency without reducing the accuracy of the neural network, a number of detailed and in-depth observations were made on the “data stream” and “data content” of the neural network. This effectively writes data in the memory while extending memory life, solving the problems faced by today’s neural network training without changing the existing architecture and technology. Besides inspiring innovative research of large neural network architecture, it can also be applied to the inference phase of neural networks.

Winning the best paper award at a prestigious international conference is an example of successful collaboration among academia, government, and industry. The R&D capacity of the team reached its full potential thanks to the long-term support from NTU, the Ministry of Science and Technology, and Macronix. Not only did winning this award enhance the international visibility of Taiwan, it will certainly inspire greater future research and breakthroughs.

This article is also featured in No. 75 of NTU Highlights (December, 2019).

Scroll to Top button