Neural networks (NNs) achieve breakthroughs in computer vision and natural language processing, yet their “black box” nature persists. Traditional methods prioritise parameter optimisation and loss design, overlooking NNs’ fundamental structure as topologically organised nonlinear computational systems. This work proposes a complex network theory framework decoding structure–function coupling by mapping convolutional layers, fully connected layers, and Dropout modules into graph representations. To overcome limitations of heuristic compression techniques, we develop a topology-sensitive adaptive pruning algorithm that evaluates critical paths via node strength centrality, preserving structural–functional integrity. On CIFAR-10, our method achieves 55.5% parameter reduction with only 7.8% accuracy degradation—significantly outperforming traditional approaches. Crucially, retrained pruned networks exceed original model accuracy by up to 2.63%, demonstrating that topology optimisation unlocks latent model potential. This research establishes a paradigm shift from empirical to topologically rationalised neural architecture design, providing theoretical foundations for deep learning optimisation dynamics.
Zhang et al. (Fri,) studied this question.