My research goal is to realize Intermittent TinyML, enabling battery-less tiny devices to intermittently execute deep neural networks (DNN) via ambient power. It is a novel research direction at the intersection of intermittent computing and deep learning, and once realized, would create innovative applications.

My research team has developed a suite of system runtimes and libraries that help AI and IoT application developers easily build low-cost, intermittent-aware inference systems. Notably, our intermittent operating system [TCAD’20] was the first to enable multitasking concurrency on intermittent system, making complicated intermittent applications increasingly possible. HAWAII [TCAD’20], which comprises an inference engine and API library, enables hardware accelerated intermittent DNN inference. iNAS [TECS’21] was the first framework that introduces intermittent execution behavior into neural architecture search to automatically find intermittently-executable DNN models. DynBal [TECS’23] introduces runtime reconfiguration to intermittent inference engines, mitigating the risk of non-termination under extremely weak power. NodPA [TECS’25] captures the non-determinism of dynamic neural networks, allowing them to be executed on intermittent systems for the first time.

HAWAII and iNAS respectively received the Best Paper Award at IEEE/ACM CODES+ISSS 2020 and 2021, whereas DynBal and NodPA were Best Paper Award nominees in 2023 and 2025.