Energy-efficient circuits and system architectures to enable intelligence at the edge of the cloud

Abstract

Internet of Things (IoT) devices are collecting a large amount of data for video processing, monitoring health, etc. Transmitting the data from the sensor to the cloud requires a large aggregate bandwidth. The objective of the proposed research is to leverage advances in machine learning to perform in-sensor computation, thus reducing the transmission bandwidth, preserving data privacy and enabling low-power operation. The proposed research demonstrates a system design and IC designs to achieve energy efficiency. As a system prototype, we demonstrate a light-powered always ON gesture recognition system. As circuit innovations, we show voltage and time-based matrix multiplying ADCs (MMADCs), compressive sensing ADCs (CS-ADCs) along with measurement results. The proposed time-based MMADC is digitally synthesizable, can operate at supply as low supply as 0.4V and demonstrates higher energy efficiency compared to the state of the art designs. As SoC innovation, we propose time-based reinforcement learning for the edge computing implemented along with sensors and actuators to demonstrate autonomous obstacle avoidance robot.