The Interpretability Analysis of DCNN Models Based on Structured Pruning Compression

Structured pruning systematically removes redundant structural components (such as convolution kernels and filters) to compress deep convolutional neural network (DCNN) models and accelerate inference, with its efficiency optimization value being extensively studied. However, the potential of structured pruning in enhancing model interpretability remains underexplored. In this study, leveraging the advantage that structured pruning simplifies networks, we propose a progressive multi-scale sparse pruning algorithm and conduct an interpretability analysis on DCNN models. First, we quantify the intersection of union (IoU) using network dissection methods to compare the semantic purity characterized by filters (neurons) before and after pruning. Subsequently, we design a heuristic key neuron scoring strategy that focuses on contextual interaction information among neurons, thereby obtaining the global-level key decision pathway of the model for a specific category of image input, which provides a more intuitive semantic understanding of model decisions. Experimental results demonstrate that structured pruning can effectively strip away chaotic parts from model neurons, thereby improving the semantic purity of neurons. Moreover, based on the sparse connections of the pruned model, we are able to extract a clear key decision pathway with high semantic purity, revealing the hierarchical knowledge structure during decision-making.