Abu Bakar
Current Institution: Northwestern University
Email: abubakar2023@u.northwestern.edu
Bio: Abu Bakar is a PhD candidate in the Department of Computer Science and Northwestern University. His research focuses on making batteryless systems a reality by making them efficient, robust, and resilient to dynamic energy-harvesting conditions. He explores new hardware designs and runtime systems and develops interactive tools to create functional and intelligent applications capable of real-time inference and self-adaptation in extreme energy harvesting conditions. His work has appeared in top systems conferences like ACM ASPLOS, UbiComp, HotMobile, and BuildSys and has attracted attention from prominent media outlets including Forbes, Scientific American, ACM Tech News, Changing America series of The Hill, Daily Mail, The Independent, and many others.
Abstract: Building Intelligent Batteryless Systems for Large-Scale Deployments
Today, most IoT devices are battery-powered and require regular battery replacement. With a trillion devices by 2035, we will be replacing 274 million batteries every day, which will be detrimental to the environment. To address this, my research enables the adoption of batteryless sensors at a large scale by making them intelligent, efficient, and robust to dynamic energy harvesting conditions.
Batteryless sensors traditionally perform sense and send operations, which is impractical for large-scale sensing due to the high energy requirements for communication. Therefore, to be feasible for real-world applications, these sensors must be made inference-capable so that data is processed locally. Though traditional deep neural networks have been implemented on batteryless sensors, their program state and memory management costs are huge. My recent work is the first to explore an alternate low-overhead inference engine, Tsetlin Machine (TM), for batteryless sensors. In contrast to neural networks, which use arithmetic operations, TMs use propositional logic to predict outcomes based on input data. My work also adds intelligence to runtime systems responsible for maintaining the forward progress and memory management of the applications while computing intermittently. I explore heuristic-based adaptation techniques to analyze and predict energy harvesting conditions in order to dynamically modulate the compute complexity of applications at run-time so that useful outcomes can be achieved in extreme energy harvesting conditions. When paired with low-overhead inference engines at the application level, this low-level adaptation paves the path for sustainable large-scale deployments.