InfoDisent: Explainability of image classification models by information disentanglement

In this work, we introduce InfoDisent, a hybrid approach that attaches a lightweight, interpretable head to any frozen pretrained backbone. Unlike fully post-hoc explainers, InfoDisent requires a single training pass of this head over the dataset. Therefore it occupies a middle ground between post-hoc and ante-hoc methods, as it merges the flexibility of post-hoc methods with the concept-level modeling capabilities of self-explainable neural networks, such as ProtoPNets. We demonstrate the effectiveness of InfoDisent through computational experiments and user studies across various datasets using modern backbones such as ViTs and convolutional networks. While InfoDisent achieves competitive performance within the class of interpretable models, we observe an accuracy-interpretability trade-off when compared to black-box counterparts, especially visible in CNNs. Notably, InfoDisent generalizes the prototypical parts approach to novel domains (ImageNet).