Deeply Quantized Neural Networks: Enabling In Sensor Computing

AI on the edge as a paradigm promises unparalleled energy efficiency and data privacy among other benefits. This has resulted in a rising trend of approaches to bring AI computing near or embedded directly inside a sensing element. In this paper, we describe a workflow of executing contemporary deep learning applications along with schemes to overcome the challenges of implementing this class of applications in a resource-constrained computing environment. We demonstrate the performance on a state-of-the-art AI-capable MEMS IMU sensor product from STMicroelectronics called the Intelligent Sensor Processing Unit (ISPU)