NeuronSoup evolves neural networks without backpropagation, hitting 85.9% accuracy

NeuronSoup, a new neural architecture from Subodh Kalia, abandons traditional deep learning's reliance on backpropagation. Instead, it uses a genetic algorithm to evolve asynchronous, delay-mediated signal propagation through a pool of shared neurons. Signals travel from input to output via variable-length paths, with hidden neurons reused across multiple routes. When two signals pass through the same neuron, their arrival timing and polarity create constructive or destructive interference, enabling complex temporal dynamics without a differentiable graph. The entire system — topology, weights, delays, and connectivity — is encoded in a flat 14,602-gene genome and co-evolved over generations.

Tested on 10-class MNIST digit classification using frozen ResNet18 features, NeuronSoup evolved 204 active paths through 266 hidden neurons (156 shared, one participating in 11 distinct paths) and reached 85.9% test accuracy after 10,000 generations. The trained model is only 115 KB. Kalia argues this addresses key deep learning limitations: no need for backpropagation, per-sample adaptive depth, and emergent lateral interactions. He notes that CMA-ES fails at this scale, making genetic algorithms uniquely suited. The architecture could generalize to other domains by swapping the encoder and output structure.