Structural Entropy Dynamics in CNN Training: A Three-Phase Guided Framework with Applications in Training Optimization

Recent advances in convolutional neural networks CNNs have achieved re-markable success across computer vision domains, yet the inherent complexity and opaque nature of their training processes continue to impede further effi-ciency improvements. As a quantitative indicator of graph structural complexi-ty, structural entropy offers a novel perspective for analyzing the training dy-namics of neural networks. This work proposes a graph structure abstraction-based representation method for CNNs, establishing a quantitative framework for training complexity assessment through the transformation of computation-al graphs into weighted directed graphs followed by structural entropy calcula-tion. Through systematic monitoring of classical CNN architectures, we identi-fy a three-phase evolution pattern of complexity dynamics: Adjustment Phase, Convergence Phase, and Specialization Phase, thereby formulating a structural entropy-guided characterization framework for CNN training processes. Fur-thermore, by establishing the correlation between dynamic structural entropy features and model performance, we develop optimization strategies including entropy-aware early stopping criteria and adaptive learning rate scheduling. Experimental results demonstrate that the proposed methodology achieves 27 training acceleration without sacrificing model accuracy, providing a principled approach to enhance CNN training efficiency through complexity-aware opti-mization.