A Dendritic-Inspired Network Science Generative Model for Topological Initialization of Connectivity in Sparse Artificial Neural Networks

Artificial neural networks (ANNs) achieve strong performance but at the cost of extreme parameter density. In contrast, biological networks operate with sparse, highly organized structures in which dendrites play a central role in information integration. We introduce the Dendritic Topological Network Model (DTNM), a generative framework that embeds dendritic-inspired connectivity principles into sparse artificial networks. Unlike conventional random initialization, DTNM defines connectivity through parametric distributions of dendrites, receptive fields, and synapses, providing explicit control over modularity, hierarchy, and degree heterogeneity. This parametric flexibility allows DTNM to produce diverse network topologies, from clustered modular architectures to hierarchical, hub-dominated structures, whose geometry can be characterized and optimized with network-science metrics. Across image classification benchmarks, DTNM consistently outperforms data-agnostic sparse initialization baselines at extreme sparsity (99%), in both static and dynamic sparse training regimes. When integrated into state-of-the-art dynamic sparse training frameworks and applied to Transformer architectures for machine translation, DTNM likewise outperforms the other evaluated methods. Crucially, in cross-task parameter-transfer evaluations with no target-task hyperparameter tuning, fixed DTNM configurations still outperform the evaluated data-agnostic baselines. We further introduce a data-informed extension, iDTNM, that incorporates input-correlation statistics into the same dendritic structure and narrows the performance gap with data-informed sparse initialization methods. These results show that bio-inspired network topology can provide a transferable structural prior for scalable and sparse models.