Neuron Abandoning Attention Flow: Visual Explanation of Dynamics Inside CNN Models.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Explaining the decision-making process of Convolutional Neural Networks (CNNs) remains a significant challenge.
• Visualizing attention dynamics within CNNs is crucial for understanding their classification behavior.

Purpose of the Study:

• To introduce a novel method, Neuron Abandoning Attention Flow (NAFlow), for visually explaining attention evolution in CNNs.
• To address the unsolved problem of understanding intermediate layer neuron contributions to CNN classification decisions.

Main Methods:

• Developed a cascading neuron abandoning back-propagation algorithm to exclude unused neurons in intermediate CNN layers.
• Proposed a Neuron Abandoning Back-Propagation module to generate Back-Propagation Feature Maps (BPFM) by inverting CNN layers.
• Introduced a channel contribution weights module using Jacobian Matrix for similarity metric-based CNN models.

Main Results:

• NAFlow effectively visualizes attention flow dynamics within CNNs.
• The method precisely excludes neurons not contributing to classification decisions.
• Demonstrated effectiveness across eleven CNN models for diverse tasks including general image classification, contrastive learning, few-shot learning, and image retrieval.

Conclusions:

• NAFlow provides a powerful tool for interpreting CNN attention mechanisms.
• The proposed method enhances the explainability of deep learning models in computer vision.
• This work offers significant advancements in understanding and debugging CNNs.