NeuronSoup evolves a 115 KB asynchronous shared-neuron network with no backpropagation, hitting 85.9% on MNIST features

arXiv·low signal

Submitted July 16 by Subodh Kalia, the architecture replaces synchronous layer-by-layer processing with delay-mediated asynchronous signal propagation through a shared neuron pool, co-evolving topology, weights, delays, and connectivity via a genetic algorithm across a 14,602-gene genome. The evolved network uses 204 active paths through 266 hidden neurons, with 156 neurons shared across paths and one participating in 11. Accuracy is modest and the task is easy, so treat this as an architecture-exploration result rather than a performance claim — the notable properties are per-sample adaptive compute depth and a 115 KB trained model.