VIS4SL: A visual analytic approach for interpreting and diagnosing shortcut learning

Abstract

Shortcut learning, a phenomenon where deep neural networks inadvertently learn irrelevant features, has been extensively discussed due to its impact on model generalization and unexpected failures. Interpreting and diagnosing shortcut learning is challenging due to its diverse manifestations and multiple influencing factors. To assist data scientists in these tasks, we introduce VIS4SL, an interactive visual analytics approach that harnesses both human intelligence and computational power. VIS4SL integrates a perturbation-based method with comprehensive visualizations to facilitate an understandable analysis of learned features. We also present a set of comparative visualizations that allow for the evaluation of model explanations against robust proxies, particularly human explanations, to quantify the degree of shortcut learning and assess model components. Two case studies, involving natural image classification and visualization classification, demonstrate the efficacy of VIS4SL in practical applications. Our findings reveal that the model uses the orientation of bars to differentiate between bar charts and Pareto charts. Furthermore, we explore how interactive visualizations enhance data scientists’ understanding of shortcut learning, enabling the development of more precise deep learning models.