Operating DNNs on mobile devices is critical for mitigating privacy concerns, reducing network dependency, and enabling real-time inference in ML-based applications. We develop resource-efficient on-device inference techniques with adaptive DNN architectures that dynamically adjust to available resources. Our work spans the full Android software stack — from kernel-level scheduling to framework and application-level optimization — to deliver targeted, practical solutions for mobile deployment.
Energy storage systems power a vast range of devices, from smartphones to electric vehicles. We apply machine learning to the challenges of remaining capacity estimation, hybrid energy storage system design, and the development of novel sensing techniques. Our research integrates battery experiments, hardware design, and software development to build intelligent management systems that improve reliability, longevity, and efficiency across diverse platforms.
IoT sensor nodes are foundational to emerging applications such as smart farms and environmental monitoring. Maintenance constraints make battery-free, light energy-harvesting nodes a compelling solution. We focus on energy optimization across both energy-consuming devices and energy-supplying harvesters, developing low-power hardware and reinforcement learning-based adaptation mechanisms that ensure stable, autonomous operation across varying environmental conditions.