The paper “Efficient Diverse Redundant DNNs for Autonomous Driving“, coauthored by BSC authors Martí Caro, Jordi Fornt and Jaume Abella, has been accepted for publication in the 47th IEEE International Conference on Computers, Software & Applications (COMPSAC). The conference will be held in Torino (Italy) between the 26–30 of June 2023.
The submission has been included as part of the Autonomous Systems Symposium (ASYS) within COMPSAC. The ASYS offers a forum to discuss wide-ranging topics in autonomy in technical systems. It offers a venue for discussions between top-level experts from academia and industry about the current state and future perspectives of autonomous systems.
The SAFEXPLAIN paper on “Efficient Diverse Redundant DNNs for Autonomous Driving” presents initial research contribution towards making acceleration solutions based on Artificial Intelligence (AI), and particularly those based on Deep Neural Networks (DNNs), amenable for their use in safety-critical systems in general, and in automotive systems in particular.
Our work provides a solution for preventing random hardware faults in DNN accelerators from leading to a system failure by making redundant accelerators diverse using a much cheaper solution than usual full duplication with staggering.
Instead, we exploit the fact that DNNs provide stochastic results whose correctness is assessed semantically rather than at bit level, and combines time and space approximate – and low cost – redundancy to detect errors.
Part of our ongoing work includes extending such solutions to detect also DNN model mispredictions by smartly designing redundant and diverse DNN models and their hardware and software realizations
Figure 1 shows an algorithm combining DNN diverse redundant outcomes into a single (and correct) outcome.
Stay tuned to hear about SAFEXPLAIN participation in the ASYS.