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

ICLR 2026 Conference Submission16405 Authors

Keywords: Artificial Intelligence, Dynamic Sparse Training, Brain Network Science, Network Topology

TL;DR: We created a new method to build sparse artificial neural networks inspired by biological dendrites, which boosts performance and efficiency in image classification and machine translation tasks.

Abstract: Artificial neural networks (ANNs) achieve remarkable performance but at the unsustainable cost of extreme parameter density. In contrast, biological networks operate with ultra-sparse, highly organized structures, where dendrites play a central role in shaping information integration. Here we introduce the Dendritic Network Model (DNM), a generative framework that bridges this gap by embedding dendritic-inspired connectivity principles into sparse artificial networks. Unlike conventional random initialization, DNM defines connectivity through parametric distributions of dendrites, receptive fields, and synapses, enabling precise control of modularity, hierarchy, and degree heterogeneity. This parametric flexibility allows DNM to generate a wide spectrum of network topologies, from clustered modular architectures to scale-free hierarchies, whose geometry can be characterized and optimized with network-science metrics. Across image classification benchmarks (MNIST, Fashion-MNIST, EMNIST, CIFAR-10), DNM consistently outperforms classical sparse initializations at extreme sparsity (99\%), in both static and dynamic sparse training regimes. Moreover, when integrated into state-of-the-art dynamic sparse training frameworks and applied to Transformer architectures for machine translation, DNM enhances accuracy while preserving efficiency. By aligning neural network initialization with dendritic design principles, DNM demonstrates that sparse bio-inspired network science modelling is a structural advantage in deep learning, offering a principled initialization frame to train scalable and energy-efficient machine intelligence.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 16405