In-Situ Self-Powered Intelligent Vision System with Inference-Adaptive Energy Scheduling for BNN-Based Always-On Perception
TimeThursday, July 14th10:53am - 11:15am PDT
Location3000, Level 3
Event Type
Research Manuscript
Design of Cyber-physical Systems, Cloud Computing and IoT
DescriptionThis paper proposes an in-situ self-powered intelligent visual perception system that harvests light energy utilizing the indispensable image sensor. Binary neural network (BNN) processing architecture combining in-sensor-processing and computing-in-memory is adopted to achieve low-power inference. The harvested energy is allocated to the computation circuits layer by layer by a light-weight duty-cycling based energy scheduler. A software-hardware co-design method, which exploits the layer-wise error tolerance of BNN as well as the computing-error and energy consumption characteristics of the computation circuit, is proposed to determine the computation turn-on threshold energy level of each layer, achieving high energy efficiency for self-powered BNN inference.