Structure aware graph community cluster pruning for efficient neural network compression in Parkinson's disease...

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Medical Informatics

Background:

• Deep neural networks (DNNs) require significant computational and memory resources, limiting their application in medical and edge environments.
• Existing neural network pruning methods often lack functional group preservation, leading to performance degradation.

Purpose of the Study:

• To introduce Graph-Community Cluster Pruning (GCCP), a novel structured pruning framework for efficient DNN compression.
• To evaluate GCCP's effectiveness in preserving model performance while reducing computational and memory footprints.

Main Methods:

• Constructing neuron activation similarity graphs to identify functionally cohesive neuron groups.
• Utilizing Louvain community detection to cluster neurons based on activation patterns.
• Employing gradient-weighted saliency for ranking and pruning within identified communities.

Main Results:

• Achieved a 37.23% parameter reduction and 37.84% FLOPs reduction in Parkinson's disease detection.
• Maintained high test accuracy (94.87%) and AUC (0.969) post-pruning.
• GCCP outperformed conventional pruning methods, especially at high compression ratios.

Conclusions:

• GCCP is a robust and interpretable method for compressing DNNs, suitable for resource-constrained medical applications.
• The framework enables the development of efficient, clinically reliable diagnostic models.
• Community visualization offers insights into neural specialization and functional diversity preservation.