Perspectives on AI-ML Safety Assurance
Emmanuel Ledinot, Philippe Quéré, Philippe Baufreton, Jean Gassino, Franck Serratrice, Hugues Bonnin, Damien Chabrol, Amina Mekki-Mokhtar, Olivier Appere, Joseph Machrouh · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
AI-ML suffers from a reliability glass-ceiling phenomenon (e.g.~10 -3 error/inference), making it incompatible with safety-criticality.Several orders of magnitude are missing.We explain why, we point to the characteristics of ML that conflict with the assurance objectives assigned to safety-critical developments.Could encapsulation of ML constituents into fault-tolerant architectures, ML development assurance, and software/hardware development assurance, altogether mitigate the gap?We argue that in spite of impressive progress of ML state-of-the-art, the answer is negative.Drawing from Topological Data Analysis (TDA) and set-based non-linear control, we propose to supplement ML point-based specification and verification with volume-based specification and verification to meet 10 -5 err./ inf.levels, as a minimum.We outline the rationale of a new research field we name (Ultra) Reliable Machine Learning, at the confluence of TDA, statistics on manifolds, and ML safety assurance.Some cross-domain safety regulation principles guide the underlying rationale.We illustrate the methodology on image classification.