InfoCAM: An information-weighted class activation mapping for explaining visual neural networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning models, particularly visual neural networks, exhibit black-box behavior due to complex nonlinear functions and hyperparameter tuning.
• Lack of transparency hinders understanding of decision-making processes in visual neural networks.

Purpose of the Study:

• To enhance the transparency and interpretability of deep visual neural networks.
• To propose a saliency visual explanation framework that provides faithful explanations.

Main Methods:

• Proposed Information-weighted Class Activation Mapping (InfoCAM), a novel saliency visual explanation framework.
• Introduced a Dual-Stream Information Bottleneck (DSIB) module using variational inference to decompose feature activations into discriminative and noise streams.
• Maximized mutual information between the discriminative stream and output prediction to assign faithful weights to activation maps.

Main Results:

• InfoCAM generates reliable visual explanations by faithfully assigning weights to feature activation maps.
• The framework is robust against the shattered gradient problem and integrates seamlessly with various network architectures and tasks.
• Improved faithfulness evaluation metrics (average drop and average increase) by clarifying theoretical bounds and reducing sensitivity to output score fluctuations.

Conclusions:

• InfoCAM offers a faithful explanation method for visual neural networks, enhancing transparency.
• Demonstrated superior performance in perturbation testing and energy-based pointing game evaluations.
• InfoCAM provides a significant advancement in creating interpretable deep learning models for computer vision.