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

Xiyu Meng, Tan Tang, Yuhua Zhou, Zihan Yan, Dazhen Deng, Yongheng Wang, Yuhan Wu, Yingcai Wu

Abstract: Highlights•VIS4SL: Visual analytics tool for diagnosing shortcut learning in neural networks.•Identifies shortcuts via multi-faceted visual analytics and robust proxy comparisons.•Case studies in image classification demonstrate practical efficacy.•Multi-level feature analysis (model, category, instance) using perturbations.•Enhances model transparency via interactive visual feedback and explanations.