A Counterfactual Account of Algorithmic Robustness

Abstract Accuracy plays an important role in the deployment of machine learning algorithms. But accuracy is not the only epistemic property that matters. For instance, it is well-known that algorithms may perform accurately during their training phase but experience a significant drop in performance when deployed in real-world conditions. To address this gap, people have turned to the concept of algorithmic robustness. Roughly, robustness refers to an algorithm’s ability to maintain its performance across a range of real-world and hypothetical conditions. In this paper, we develop a rigorous account of algorithmic robustness grounded in Robert Nozick’s counterfactual sensitivity and adherence conditions for knowledge. By bridging insights from epistemology and machine learning, we offer a novel conceptualization of robustness that captures key instances of algorithmic brittleness while advancing discussions on reliable AI deployment. We also show how a sensitivity-based account of robustness provides notable advantages over related approaches to algorithmic brittleness, including causal and safety-based ones.