Ultra-Lightweight Edge Intelligence Using Vector Symbolic Architecture

Edge intelligence greatly benefits Internet-of-Things devices by providing insights at the sensor node without transmitting the data to and from the cloud. Although recent work has scaled deep neural networks down to microcontrollers, these solutions remain impractical for devices operating under strict energy budgets—such as battery-powered camera traps in remote forests or micro-robots in disaster zones. My work addresses that gap by exploring a brain-inspired Vector Symbolic Architecture (VSA) over conventional neural networks. I introduce HyperCam for low-power image classification, NavHD for micro-robot navigation, and Omen for dynamic optimization during inference. These methods each optimize VSA's encoding and training algorithms to achieve meaningful speedups and energy savings. Ongoing work seeks to demonstrate these advances as real-world systems, where VSA can extend the lifespan of intelligent systems in remote or energy-constrained environments.