Structured Pruning Strategy Based on Interpretable Machine Learning

Structured Pruning is an important strategy for neural network optimization. However, conventional pruning methods often focus solely on filter parameter magnitudes or their spatial distribution, disregarding the semantic information encoded in parameters and the close relationship between filter parameters and data. To address this issue, we propose a neural network pruning strategy based on interpretable machine learning. This strategy combines interpretable machine learning techniques to analyze the importance of filters in understanding the content of data. The importance derived from this analysis is then used to rank and prune filters. Experimental results demonstrate that our proposed strategy achieves significant pruning benefits across VGG, ResNet models, and the CIFAR-10 dataset. Furthermore, our approach offers interpretability, allowing for a thorough exploration of the mutual information between filter parameters and image data.