Investigating the Effect of Network Pruning on Performance and Interpretability
Researchers investigated how network pruning affects the performance and interpretability of GoogLeNet on ImageNet. They applied various pruning techniques and retraining strategies, finding that performance could be maintained or even improved with sufficient retraining. However, their experiments using the Mechanistic Interpretability Score (MIS) showed no clear link between pruning rate and interpretability, suggesting MIS may not always align with intuitive understanding of model decisions. AI
IMPACT Provides insights into optimizing deep learning models for efficiency and understanding their decision-making processes.