Non-determinism and Failure Modes in Machine Learning

摘要

Determinism is a key concern in the certification of software for safety-critical systems. In this paper, we evaluate the role of determinism in certification standards, using airborne software as example. We analyze and speculate how the requirements and underlying concepts related to determinism can be adapted for Machine Learning algorithms. In addition, we systematically identify and analyze a large set of factors that contribute to variations of behavior in machine learning systems across multiple levels. Our suggestion is that such variability factors are handled in a similar fashion to failure modes in current software and systems development. We propose that the method followed and the identified set of factors is taken as a step towards a global catalog that can assist both developers and assessors in attaining certifiable machine learning systems.