Dendritic Structure Enables Parameter-Efficient Neural Representations

Artificial neural networks generally rely on simplified point-neurons, abstracting away the branched, dendritic structure of biological neurons. In this work, we investigate whether incorporating explicit dendritic structure provides a useful computational constraint for artificial networks. Using DendriNet, a neural-network layer composed of sparsely connected excitatory and inhibitory units with defined dendritic tree topologies and divisive inhibition, we replace the fully connected layers of a pretrained vision backbone with DendriNet blocks and evaluate the representational and scaling properties of dendritic constraints on ImageNet-1k. Our results demonstrate that DendriNets match or exceed dense point-neuron accuracy with up to a $24$-fold reduction in effective parameters, while class selectivity increases systematically from distal dendritic branches to the soma.