Interpret Neural Networks by Extracting Critical Subnetworks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) demonstrate high performance across various AI fields.
• Increasing demand for interpretability and robustness in DNNs.
• Current methods struggle to fully explain DNN decision-making processes.

Purpose of the Study:

• To develop a method for understanding DNN functional mechanisms.
• To enhance model interpretability and robustness.
• To propose novel adversarial example detection techniques.

Main Methods:

• Extraction of critical subnetworks, defined as important channels whose suppression severely degrades performance.
• Analysis of layerwise semantic behavior using extracted subnetworks.
• Development of adversarial example detection methods based on subnetwork properties.

Main Results:

• Critical subnetworks provide insights into layerwise model behavior.
• Subnetworks enable more accurate visual explanations via attribution methods.
• Proposed detection methods show potential for increasing model robustness against adversarial attacks.

Conclusions:

• Extracting critical subnetworks is a viable approach for DNN interpretability.
• This method offers a novel perspective for understanding internal model workings.
• The findings contribute to developing more robust and trustworthy AI systems.