Visual Foundation Models: A New Approach for Explainable AI

Artificial intelligence (AI) is increasingly permeating all areas of our lives. Visual Foundation Models (VFMs) in particular have caused a stir in recent years due to their impressive performance in image recognition and processing. However, as the complexity of these models increases, so does the need for transparency and explainability. How can we understand and comprehend the decisions of these "black boxes"? A promising approach lies in the development of so-called self-explainable models (SEMs).

The Challenge of Explainability

VFMs are often based on deep neural networks, whose workings are difficult to understand even for experts. This poses a problem, especially in sensitive application areas such as medicine or autonomous vehicle control. Here, it is essential to be able to understand the reasons for the AI's decisions in order to build trust and identify potential sources of error.

SEMs attempt to solve this problem by decomposing predictions into a weighted sum of interpretable concepts. For example, an SEM that classifies images of animals can justify its decision based on features such as "fur," "beak," or "wings." This approach promises more transparency and allows users to better understand the AI's decisions.

ProtoFM: An Efficient Approach for Explainable VFMs

A new approach in this area is ProtoFM, a method that combines VFMs with a prototypical architecture and specialized training objectives. Instead of retraining the entire VFM, only a lightweight "head" (approximately 1 million parameters) is trained on the existing, frozen VFM. This enables efficient and resource-saving implementation.

ProtoFM uses prototypes to explain the decision-making process. These prototypes represent characteristic features of the respective classes. The similarity of a new image to these prototypes determines the classification and simultaneously provides an understandable explanation for the decision.

Performance and Interpretability in Harmony

Evaluations of ProtoFM show promising results. The method not only achieves competitive classification performance but also surpasses existing models in terms of various interpretability metrics. This suggests that ProtoFM represents an important step towards explainable AI.

The Future of Explainable AI

The development of explainable AI models is an active research area. ProtoFM is an example of a promising approach that combines the power of VFMs with the need for transparency and interpretability. Future research will focus on further improving explainability and expanding applicability to various fields. This will help to strengthen trust in AI systems and enable their use in critical applications.

Bibliography: - https://arxiv.org/abs/2502.19577 - https://zenodo.org/records/14778494 - https://paperreading.club/page?id=287685 - https://arxiv.org/list/cs.CV/new - https://papers.ssrn.com/sol3/Delivery.cfm/323289e9-03c5-4683-9eac-4df7a6384096-MECA.pdf?abstractid=5106267&mirid=1 - https://www.researchgate.net/publication/388262906_Explainability_for_Vision_Foundation_Models_A_Survey - https://openreview.net/forum?id=xP1ROUJoyt&referrer=%5Bthe%20profile%20of%20Gianmarco%20Mengaldo%5D(%2Fprofile%3Fid%3D~Gianmarco_Mengaldo1) - https://ceur-ws.org/Vol-2444/ialatecml_paper1.pdf - https://www.sciencedirect.com/science/article/pii/S277266222300070X - https://github.com/samzabdiel/XAI