ProtoS-ViT: Visual foundation models for sparse self-explainable classifications

TL;DR

ProtoS-ViT introduces a new architecture that enhances explainability in visual classification by leveraging pre-trained Vision Transformers and a comprehensive set of evaluation metrics, advancing the development of transparent AI models.

Contribution

The paper proposes a novel architecture that improves explanation quality in prototypical networks using frozen ViT backbones and introduces extensive metrics for evaluating explanation effectiveness.

Findings

Outperforms existing prototypical models in explanation quality

Provides compact, human-recognizable explanations

Effective for both general and biomedical image classification

Abstract

Prototypical networks aim to build intrinsically explainable models based on the linear summation of concepts. Concepts are coherent entities that we, as humans, can recognize and associate with a certain object or entity. However, important challenges remain in the fair evaluation of explanation quality provided by these models. This work first proposes an extensive set of quantitative and qualitative metrics which allow to identify drawbacks in current prototypical networks. It then introduces a novel architecture which provides compact explanations, outperforming current prototypical models in terms of explanation quality. Overall, the proposed architecture demonstrates how frozen pre-trained ViT backbones can be effectively turned into prototypical models for both general and domain-specific tasks, in our case biomedical image classifiers. Code is available at \url{https://github.com/hturbe/protosvit}.