Welcome to ARChip Lab
清华大学 2025年有1-2名直博名额,欢迎报考!

ARChip Lab has been working on AI from theory to application. Interdisciplinary involvement including biology and neurology explains the basic theory of AI from the perspective of biological behavior, thus, it supports and provides a more scientific explanation to algorithm, which is the at the next level. Combining with mathematics, the algorithm offers solubility to neural networks. In terms of application, the Lab is developing a special chip architecture that is used on neural network algorithm and automobile pilot. With the help of hardware, a high performance of the chip design will be achieved.

ARChip Lab团队进行从基础理论到实际应用的人工智能研究。理论层有与脑科学、神经学交叉的课题组,通过进行对人工智能回归本源的研究,在细胞的生物学行为角度、以及脑区层次解析人工智能的基础理论,从而支持算法层的研究,以提供更为科学的解释;中间算法层将传统的算法研究与应用数学学科结合,利用数学这门工具以更科学的手段对神经网络等算法进行可解析性研究,关注算法可解释性、鲁棒性、小型化;在应用层,我们开发专用的芯片架构,主要面向神经网络计算与自动驾驶场景的处理器架构,并通过团队中算法的硬件优化支持,以软硬件协同设计思想进行芯片开发,以取得比纯做芯片设计更为高效的结果。

Kaisheng Ma is now an Assistant Professor in Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University. He got Ph.D. in Department of Computer Science and Engineering, The Pennsylvania State University. His research focuses on computer architecture, implanted devices, AI Algorithms Design, focusing on interpretation, robustness and compact model design. Dr. Ma has won many awards, including: 2015 HPCA Best Paper Award, 2016 IEEE MICRO Top Picks, 2017 ASP-DAC Best Paper Award. 2018 EDAA Best Dissertation Award. Dr. Ma has many honors, including 2016 Penn State CSE Department Best Graduate Research Award (Among ~170 Ph.D. students), 2016 Cover Feature of NSF ASSIST Engineering Research Center Newsletter (Among 40 graduate students across four participating universities.), 2011 Yang Fuqing & Wang Yangyuan Academician Scholarship (1/126, Peking University.).

AI Algorithms AI Chip Architecture, Design and Tapeout Implantable Chips
For algorithm designing, our research interests can be summarized as:(1) model miniaturization, and (2) model robustness. We focus on domain specific architectures for artificial intelligence, especially algorithm-inspired design. With the help of better semiconductor technology and AI technology, we can understand the nature of human intelligence better.
We have proposed a series of techniques by refining the training samples and training process. Experiments show that these methods can significantly improve model accuracy, robustness and efficiency. Fabricated in UMC 55nm SP process, our first tapeout was successfully accomplished in Nov, 2019. This chip contains 256 PEs and 369KB on-chip SRAM. For the evaluation setup, we design the mother board for our chip and interconnected it to Xilinx VC706 FPGA test board via FMC. Our work is focusing on miniaturize implanted device which can record neural signal and establish an enough reliable wireless communication link to record brain signal from different areas.
© Copyright 2020 ARChip Lab - All rights reserved