Interactive exploration of CNN interpretability via coalitional game theory.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are crucial for image classification.
• Neuron feature visualization aids in understanding CNN mechanisms.
• Current methods for CNN interpretability are often inefficient due to the large number of neurons.

Purpose of the Study:

• To develop efficient metrics for assessing neuron and layer interpretability in CNNs.
• To enhance the exploration of CNN mechanisms and identify important neurons.
• To create interactive interfaces for improved CNN interpretability analysis.

Main Methods:

• Proposed Neuron Interpretive Metric (NeuronIM) inspired by SHapley Additive exPlanation (SHAP).
• NeuronIM quantifies feature expression ability by comparing neuron visualizations with SHAP images.
• Introduced Layer Interpretive Metric (LayerIM) by averaging NeuronIM values for convolution layers.
• Developed two interactive interfaces for multi-view display and user interaction.

Main Results:

• NeuronIM effectively assesses individual neuron feature expression.
• LayerIM provides a quantitative measure for convolution layer interpretability.
• Interactive interfaces facilitate efficient exploration of CNN interpretability.
• Effectiveness demonstrated through model pruning experiments and use cases.

Conclusions:

• NeuronIM and LayerIM significantly improve the efficiency of CNN interpretability exploration.
• The proposed metrics and interfaces enable rapid identification of important neurons and layers.
• This work offers a practical approach to understanding and optimizing CNN models.