Mixture of Gaussian-Distributed Prototypes With Generative Modelling for Interpretable and Trustworthy Image...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Prototypical-part methods improve image recognition interpretability by linking predictions to training prototypes.
• Existing point-based prototype learning methods suffer from limited representation power and performance degradation due to prototype projection.
• Current methods overlook sub-salient object regions, limiting their ability to capture crucial classification information.

Purpose of the Study:

• To introduce a new generative paradigm, Mixture of Gaussian-distributed Prototypes (MGProto), for learning prototype distributions.
• To enhance the representation power of prototypes for trustworthy out-of-distribution (OoD) detection and improve predictive performance.
• To develop a prototype mining strategy that considers both active and sub-salient object parts.

Main Methods:

• Proposed a generative paradigm, MGProto, to learn prototype distributions using Gaussian distributions.
• Developed a prototype mining strategy to incorporate sub-salient object regions alongside active ones.
• Implemented a pruning strategy to enhance model compactness by removing low-importance prototypes.

Main Results:

• MGProto achieves state-of-the-art performance in image recognition across multiple benchmark datasets.
• Demonstrated superior out-of-distribution (OoD) detection capabilities compared to existing methods.
• Provided encouraging interpretability results, showcasing the model's decision-making process.

Conclusions:

• MGProto offers a robust solution for interpretable image classification and trustworthy OoD detection.
• The generative approach and enhanced prototype mining strategy overcome limitations of previous methods.
• The proposed method achieves competitive performance while maintaining model compactness and interpretability.