Can we reconcile safety objectives with machine learning performances?
Abstract
The strong demand for more automated transport systems with enhanced safety, in conjunction with the explosion of technologies and products implementing machine learning (ML) techniques, has led to a fundamental questioning of the trust placed in machine learning. In particular, do state-of-the-art ML models allow us to reach such safety objectives? We explore this question through two practical examples from the railway and automotive industries, showing that ML performances are currently far from those required by safety objectives. We then describe and question several techniques aimed at reducing the error rate of ML components: model diversification, monitoring, classification with a reject option, conformal prediction, and temporal redundancy. Taking inspiration from a historical example, we finally discuss when and how new ML-based technologies could be introduced.
Origin : Files produced by the author(s)