Reliability analysis on case-study traffic sign convolutional neural network on APSoC

Abstract

Deep learning has been widely used to solve computer vision applications such as autonomous cars and Advanced Driver Assistance Systems (ADAS). One of these computer vision applications is traffic sign recognition that plays an essential role in autonomous cars and ADAS. All-Programmable System-on-Chip (APSoC) is an excellent target platform for implementing ADAS applications due to its flexibility. However, as its Programmable Logic (PL) is implemented with SRAM cells and traffic sign recognition is a safety critical application, the reliability under radiation must be analyzed and hardening or mitigation techniques may be required. Thus, random and piled-up fault injections in the APSoC configuration bits were carried out as well as neutron-radiation experiments in order to evaluate the reliability of a Convolutional Neural Network (CNN). Comparison with related work shows that timing-multiplexing neural network architectures present an insignificantly higher Architecture Vulnerability Factor (AVF) than parallel pipelined architectures.