Keypoint Counting Classifiers: Turning Vision Transformers into Self-Explainable Models Without Training

2025-12-19

AI Summary

To address the lack of interpretability in Vision Transformers (ViTs), this paper proposes Keypoint Counting Classifiers (KCCs), the first zero-training self-explaining framework for ViTs: it requires no fine-tuning, introduces no additional parameters, and instead extracts semantically meaningful keypoints solely by analyzing internal attention maps of pre-trained ViTs. Classification is performed via cross-image keypoint matching and counting. The method inherently supports intuitive, human-readable visualizations; its decision process is semantically aligned and fully traceable, significantly enhancing human-AI collaboration trustworthiness. Extensive experiments across multiple benchmark datasets demonstrate that KCCs consistently outperform existing self-explaining baselines in both explanation quality and classification accuracy. By offering an efficient, lightweight, and plug-and-play interpretability enhancement paradigm, KCCs provide a practical solution for general-purpose vision foundation models.

Technology Category

• Computer Vision: Interpretability, Explainability, and Transparency
• Machine Learning: Transparent, Interpretable, Explainable ML
• Humans and AI: Explainable AI (XAI) for Human Understanding

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Current approaches for designing self-explainable models (SEMs) require complicated training procedures and specific architectures which makes them impractical. With the advance of general purpose foundation models based on Vision Transformers (ViTs), this impracticability becomes even more problematic. Therefore, new methods are necessary to provide transparency and reliability to ViT-based foundation models. In this work, we present a new method for turning any well-trained ViT-based model into a SEM without retraining, which we call Keypoint Counting Classifiers (KCCs). Recent works have shown that ViTs can automatically identify matching keypoints between images with high precision, and we build on these results to create an easily interpretable decision process that is inherently visualizable in the input. We perform an extensive evaluation which show that KCCs improve the human-machine communication compared to recent baselines. We believe that KCCs constitute an important step towards making ViT-based foundation models more transparent and reliable.

Problem

• Turning Vision Transformers into self-explainable models without retraining.
• Providing transparency and reliability to ViT-based foundation models.
• Creating an interpretable, visualizable decision process for ViTs.

Innovation

• Transforms Vision Transformers into self-explainable models without retraining
• Uses keypoint matching between images for interpretable decision processes
• Provides visualizable explanations directly in the input image space