Improving Performance of the PRYSTINE Traffic Sign Classification by Using a Perturbation-Based Explainability...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Model interpretability is crucial for high-stakes applications like autonomous driving.
• Explainable AI (XAI) methods are vital for understanding black-box models, including Convolutional Neural Networks (CNNs).

Purpose of the Study:

• To evaluate the explainability of a traffic sign classifier developed by the PRYSTINE project.
• To use XAI insights for compressing the CNN classifier by removing less impactful kernels.
• To assess the impact of kernel compression on classification precision.

Main Methods:

• Implemented a post-hoc, local, perturbation-based forward explainability method to identify high- and low-impact kernels in the CNN.
• Excluded 'vague' kernels from the layer preceding the fully connected layer.
• Evaluated classifier precision across various kernel compression levels.

Main Results:

• Successfully distinguished between high- and low-impact kernels using the XAI method.
• Pruning 5% of kernels, guided by XAI, resulted in a 2% decrease in traffic sign and traffic light classification precision.
• Demonstrated a trade-off between model compression and classification accuracy.

Conclusions:

• The proposed methodology effectively uses XAI for CNN compression in traffic sign classification.
• This approach is valuable for applications where processing power and execution time are critical constraints.
• XAI-driven compression offers a viable strategy for optimizing deep learning models in resource-constrained environments.