PIP-PACA: An Interpretable Image Classification Framework via Prototype-Aware Clustering Attention

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) excel in image classification accuracy but lack transparency, limiting their use in critical applications.
• Prototype-based methods offer interpretability by linking inputs to semantic prototypes, but struggle with global context and prototype relationships.
• Existing methods like PIP-Net provide interpretability but require improvements in modeling complex relationships.

Purpose of the Study:

• To develop an interpretable image classification model, PIP-PACA, that addresses limitations in global context modeling and prototype relationship utilization.
• To enhance the transparency and reliability of deep learning models in image classification tasks.
• To improve the efficiency and effectiveness of prototype-based learning through novel attention mechanisms.

Main Methods:

• Proposed the prototype-aware clustering attention (PACA) module, featuring learnable cluster centers for projecting features into a prototype space.
• Implemented a bidirectional attention mechanism between features and prototypes for capturing global information with linear complexity.
• Incorporated a normalization operation during feature extraction to stabilize distributions and improve prototype matching.

Main Results:

• PIP-PACA preserves the interpretability of existing prototype-based methods.
• Achieved notable improvements in image classification accuracy, sparsity, and prototype purity compared to baseline methods.
• Demonstrated the effectiveness of the clustering-based attention mechanism in enhancing prototype learning.

Conclusions:

• The proposed PIP-PACA model offers a superior approach to interpretable image classification by effectively integrating clustering and attention mechanisms.
• The PACA module provides a computationally efficient and effective way to model global context and prototype relationships.
• This work advances the field of explainable AI in computer vision, paving the way for more reliable and transparent deep learning applications.