Almost every modern neural network works the same way: a tidy assembly line of layers firing in lockstep, trained by gradient descent. NeuronSoup throws that out. Signals wander through a shared pool of neurons along individual paths, arriving at different times because of built-in delays. When two paths cross the same neuron, the later signal meets whatever the earlier one left behind, and they reinforce or cancel like ripples meeting on a pond.

There is no gradient here at all. The whole thing, its wiring, weights, and delays, is bred by a genetic algorithm tuning a genome of about 14,600 numbers. On MNIST digits (using pre-computed ResNet18 features as input), an evolved network hit 85.9% accuracy and fit in 115 KB. The authors argue this style adapts how much computation each input gets and finds cross-path interactions that ordinary networks must be hand-designed to have.

This is a read of the abstract, so the paper is where the mechanism is argued.