A Patch-Masking Approach to Explain CNNs and Vision Transformers through Class-Specific Impacts
A new method is presented for explaining CNNs and ViTs image classification through a set of interpretable rules composed of one or more antecedents, which combine pixel-level properties and patch-level features, analyzing the impact of each region on the model’s classification.
Jean-Marc Boutay, Damian Boquete, Deniz Köprülü et al.
· 0 citations