2019年10月28日，来自美国亚利桑那州坦佩市亚利桑那州立大学Yu (Kevin) Cao教授应邀来访我基地，并在微纳电子大厦205会议室做了题为《On-line Adaptation for Intelligent Dynamic Systems》的学术报告。微纳电子学系多位老师以及研究生、本科生参加了报告会，并与报告人进行了热烈的讨论。

On-line Adaptation for Intelligent Dynamic Systems

On-line adaptation is increasingly needed in the dynamic system, such as a self-driving vehicle and a smart drone. Constrained by the computing resource at the edge, such adaptation should be accurate and robust in algorithms, but also efficient in computation. This talk presents a joint algorithm-hardware exploration on this topic. We will start from the latest advances in cumulative learning algorithms, which significantly reduce memory usage and computation cost in model update. Next, we will map these algorithms to a hybrid architecture with NVMs and SRAMs, leveraging their diverse properties in storage density and access time. By exploiting the on-line adaptation scheme, we demonstrate that a marginal overhead in SRAM is sufficient to overcome severe accuracy loss with unreliable NVMs. Finally, we develop a hierarchical benchmarking tool to assess and optimize various device, circuit, architecture and algorithm choices in on-chip learning, with the focus on in-memory computing and interconnect fabrics.