Can we rely on AI? Reliability Issues in Artificial Neural Networks and Potential Solutions for Autonomous Vehicles

by Prof. Paolo Rech (University of Trento)

Tuesday 16 Apr 2024, 09:00 → 10:00 Europe/Zurich

13/2-005 (CERN)

HYBRID SEMINAR IN 13-2-005 AND ZOOM

Abstract:

Driverless cars are the new trend in the automotive market and, to boost deep space exploration, NASA and ESA are willing to add self-driving capabilities to their rovers. Ingenuity, landed in Mars in 2021, is the first autonomous vehicle to move outside of the Earth. To be implemented, a self-driving system needs to analyze a huge amount of images and signals in real time. This is achieved thanks to Convolutional Neural Networks (CNNs) executed on Graphics Processing Units (GPUs) or dedicated accelerators implemented in Field Programmable Gate Arrays (FPGAs) or in Application Specific Integrated Circuits (ASICs), such as the Google’s Tensor Processing Unit (TPU).

In this talk, after a brief description of radiation effects at physical level, we will investigate the reliability of GPUs, FPGAs, and TPUs executing neural networks, we will show if and why a neutron-induced corruption can modify the autonomous vehicles behaviors, and discuss the implications of these corruptions for the adoption of self-driving vehicles in large scale.

The evaluation, to be accurate and precise, is based on the combination of beam experiments and fault injection at different levels of abstractions (RTL, microarchitectural, and software). This combination allows us to have a realistic evaluation of the error rate, distinguish between tolerable errors and critical errors, and to design efficient and effective hardening solutions for neural networks. Exploiting the potential of machine learning and taking full advantage of the computing resources in modern accelerators it is possible to significantly improve the neural network reliability with nearly-zero overhead.

Zoom details are also in the invitation email.

Make sure you have the latest Zoom client app to be able to connect https://cern.zoom.us/ -> Download client

Paolo_CERN_2024_final.pdf

Paolo_CERN_2024_final.pptx