Reliability modeling for three-version machine learning systems through Bayesian networks

Abstract

Machine learning (ML) is extensively employed in AI-powered systems including safety-critical applications such as autonomous vehicles. The outputs from ML models are sensitive to real-world input data and error-prone, thereby improving the reliability of ML systems’ outputs has become a critical challenge in ML system design. In this paper, we introduce N-version ML architectures to enhance the ML system reliability and propose Bayesian Networks (BNs) models to evaluate the reliability of system outputs targeting three-version ML systems. The proposed BN reliability models allow us to formulate five distinct types of three-version ML architectures that are composed of diverse models and diverse input data sources. To validate the BN reliability models with real samples from ML systems, we conduct empirical studies on traffic sign recognition tasks and evaluate prediction performance. As a result, we find the prediction residuals between the observed reliability and the predicted reliability by the BN reliability models are less than 0.015 across all data sets, which is much better than the prediction performance by the baseline model. In addition, in comparison to the previous reliability models without exploiting BNs, the proposed models exhibit an advantage in reliability prediction, except for the triple model with single input architecture.