AI for Automotive and Aerospace Applications: Reliability Challenges and Solutions
Paolo Rech - UFRGS, BR/Università di Trento, IT
Abstract
The main goal of this lecture is to provide students with an overview of the challenges associated with the hardware and software necessary for executing AI applications, such as object detection, that represents one of the major advances in the technology for computing devices. All the major cars builder and chip designers are targeting self-driven vehicles and the NASA’s JPL Perseverance mission lunched at the end of July 2020, for instance, includes the first autonomous vehicle used for space exploration. The course proposes a revision of basic concepts of real-time systems, parallel or programmable architectures, safety-critical systems, and approximate computing. These concepts are used and applied to deeply understand the AI frameworks based on neural networks and their application in automotive and aerospace markets. A study of the limitations in terms of reliability and of the problems that can affect the correct execution of software and hardware will be presented. The focus will be on the study of both the hardware and the software necessary to detect object in a scene in real time. The problems and the constraints related to the security and reliability that can influence a safety-critical system will be considered.
Syllabus
The main topics covered during the course are:
- Introduction
- AI for autonomous vehicles: state of the art
- Neural networks based object detection
- Safety-critical applications concepts
- AI for automotive and aerospace applications
- Parallel and Programmable processors
- CNNs in GPUs and FPGAs. What else?
- Standard ISO 26262. Faults in hardware, errors in software
- Hardening techniques for AI.
- Energy consumption, performances, accuracy, fault tolerance: can we have it all?
Keywords: AI, reliability, autonomous vehicles, GPUs, accelerators, automotive, aerospace, fault tolerance