Engineering Trustworthy AI : A Developer Guide for Empirical Risk Minimization

Abstract

AI systems are increasingly used for critical decisions that transcend all important fields of private and public life. These systems often use empirical risk minimisation (ERM) to train powerful prediction models such as deep neural networks. The output of the predictive model runs the risk of unintentional bias, opacity, and other adverse effects. To ensure the safety of these systems, it is vital to consider these risks already in the design stage of the model. The EU acknowledged the potential sensitivity of the predictions/decisions made about persons which led to the formulation of the Ethics Guidelines for Trustworthy AI laying down seven key requirements for trustworthy AI. So far, the design of ERM-based methods prioritises accuracy over trustworthiness. This paper discusses how key requirements for trustworthy AI can be translated into design choices for ERM components. We map the design space of ML systems to the core objectives of trustworthy AI: fairness, privacy, robustness, and explainability. Our theory is instrumental in the design of trustworthy ML systems that minimise privacy leakage and are robust against (intentional) perturbations during their operation, such as disseminating fake news. The operation of trustworthy ML systems should also be transparent or explainable to its users. Finally, ML systems must be fair and not discriminate against specific user groups. There is an urgent need for a more holistic approach to ML that includes key requirements for trustworthy AI.