N-Version Machine Learning Models for Safety Critical Systems
Fumio Machida · 2019
Quality control of machine learning systems is a fundamental challenge in industries to provide intelligent services or products using machine learning. While recent advances in machine learning algorithms substantially improve the performance of intelligent tasks such as object recognition, their outputs are essentially stochastic and very sensitive to input data. Such an output uncertainty is a big obstacle to ensure the quality of safety critical applications like autonomous vehicle and hence architectural design to mitigate the impact of error output becomes a great importance. In this paper, we propose N-version machine learning architecture that aims to improve system reliability against probabilistic outputs of individual machine learning modules. The key idea of this architecture is exploiting two kinds of diversities; input diversity and model diversity. Our study first formally defines these diversity metrics and analytically shows the improved reliability by N-version machine learning architecture. Since we treat a machine learning module as a black-box, the proposed architecture and the reliability property are generally applicable to any machine learning algorithms and applications.