We ask whether the wiring topology of a real biological connectome, imposed as a fixed binary mask on a gradient-trained network, changes what that network can learn. The connectome is the complete Caenorhabditis elegans chemical synapse map: 448 nodes, 4,681 synapses, density 2.33%. We use the full adjacency matrix as-is, without cropping or tiling, so all topological properties of the connectome are preserved. Four conditions are compared across four tasks (digit classification, diabetes regression, two-moons, sequential digit recognition) and five random seeds: fully dense, randomly sparse at matched density, bio-topological, and magnitude-pruned (a sparse topology extracted from a trained dense network). On classification tasks, biological and random-sparse masks reach the same accuracy within statistical noise and both overfit less than the dense baseline. On regression and the recurrent task, the dense network wins, and again the two sparse conditions are indistinguishable from each other. The specific pattern of C. elegans connections adds nothing beyond matched random sparsity. We report this as a negative result with some care: it rules out wiring topology alone as a mechanism for computational advantage, but says nothing about what happens when topology is combined with biologically plausible learning rules, spiking dynamics, or dendritic computation. These are the subjects of the next stages of this project.
Maxence Arella (Fri,) studied this question.