NeuronSoup: Evolving Asynchronous, Shared-Neuron Temporal Graphs without Backpropagation

TL;DR - NeuronSoup is a backprop-free neural architecture that co-evolves topology, weights, delays, and connectivity via a genetic algorithm, using asynchronous signal propagation through a pool of shared neurons. It matters as a proof-of-concept that non-differentiable, gradient-free methods can produce compact, per-sample-adaptive computation.

• Replaces synchronous layer-by-layer processing with asynchronous, delay-mediated signals; paths route continuous values through shared hidden neurons, where arrival timing/polarity causes constructive or destructive interference.
• A genetic algorithm optimizes a flat 14,602-gene real-valued genome; the authors claim GAs are the right tool here and report that CMA-ES fails at this scale.
• On 10-class MNIST using frozen ResNet18 features, it evolved 204 active paths through 266 hidden neurons (156 shared) and reached 85.9% test accuracy after 10,000 generations, in a 115 KB model.
• Claimed advantages: no differentiable computation graph, per-sample adaptive depth, and emergent lateral pathway interactions; results are limited to one benchmark with pre-extracted features, so generalization claims remain unproven.