Institute of Information Science, Academia Sinica

Energy harvesting technology and ultra-low-power microcontrollers have led to the advent of tiny battery-less devices capable of sustainable, maintenance-free operation. These devices can now run deep neural networks (DNNs) locally, shifting intelligent decision-making from the cloud to beyond the edge. However, because harvestable ambient energy is weak and unstable, battery-free devices operate intermittently and must preserve progress to withstand frequent power failures. This talk will first introduce the fundamental challenges of enabling deep learning on battery-less devices. Next, recent advancements in intermittent-aware runtime and design-time software will be presented. The system runtimes ensure that intermittency management does not offset the benefits of accumulative execution for accelerated DNN inference. In contrast, the design-time tools optimize DNN models for safe and efficient deployment on intermittently powered systems.

Hashan Roshantha Mendis is a Senior R&D Scientist at the Research Center for Information Technology Innovation (CITI), Academia Sinica, and Co-Director of the Embedded and Mobile Computing Lab at CITI. He received his Eng.D. degree from the Department of Computer Science at the University of York, UK, in 2017. His research focuses on intermittent systems and TinyML. His work has received best paper awards at leading international conferences, as well as the NSTC Postdoctoral Academic Research Award. He serves as an Associate Editor for IEEE ESL and has served on the Technical Program Committees of IEEE RTSS and ACM SAC, as well as as a reviewer for IEEE TCAD, ACM TECS, IEEE TMC, ACM TCPS, and IEEE TCASAI.