Understanding Convolutional Neural Networks From Excitations.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Saliency maps explain Convolutional Neural Network (CNN) decisions.
• Current methods use gradients, limiting complex model explanations and negative gradient use.
• This restricts interpretive veracity and full layer information utilization.

Purpose of the Study:

• Introduce a novel method, Positive and Negative Excitation (PANE), for CNN interpretability.
• Enable direct, gradient-free extraction of PANE for each layer.
• Improve saliency map generation by leveraging complete layer information.

Main Methods:

• Directly extract Positive and Negative Excitation (PANE) for each layer, bypassing gradient reliance.
• Introduce a double-chain backpropagation procedure to organize excitations into saliency maps.
• Conduct comprehensive experiments on binary and multiclassification tasks.

Main Results:

• The proposed PANE method significantly improves salient and minor pixel removal.
• PANE offers superior guidance for generating inconspicuous adversarial perturbations.
• Experimental results demonstrate the effectiveness of layer-by-layer PANE extraction and correlation verification.

Conclusions:

• PANE provides a gradient-free approach for enhanced CNN interpretability.
• The method achieves state-of-the-art performance in improving saliency map quality.
• PANE offers a more comprehensive understanding of CNN decision-making processes.