Novel Arithmetics in Deep Neural Networks for Autonomous Driving

This paper is a technical report that focuses on the challenges and opportunities presented by novel arithmetics in deep neural networks (DNNs) for signal processing, specifically in the context of autonomous driving applications. It outlines the crucial need for real-time processing on embedded systems due to the safety-critical nature of autonomous driving, which cannot rely on cloud computing due to latency and security concerns. The report reviews various computing arithmetic options and their implications for the implementation of DNN accelerators, aiming for accurate and efficient processing of automotive sensor data. It discusses the rise of alternative arithmetic styles like BFLOAT16 and flexpoint, as well as the Posit format, which seeks to balance precision and resource efficiency. Sections of the report address the specific signal processing functions applicable to DNNs, such as activation functions and the use of lookup tables. The report concludes with a discussion of trade-offs between accuracy and computational complexity in DNN signal processing for autonomous driving.