NeuronSoup: Evolving Asynchronous, Shared-Neuron Temporal Graphs without Backpropagation
NeuronSoup reports 85.9% MNIST test accuracy with a 115 KB evolved model and no backpropagation.
The paper proposes asynchronous signal propagation through shared hidden neurons, with timing and polarity producing interference between paths. Its genetic algorithm co-evolves topology, weights, delays, and connectivity across a 14,602-gene genome. In the MNIST setup, it uses frozen ResNet18 features and evolves 204 active paths through 266 hidden neurons. The author argues the design avoids differentiable graph requirements and can vary computation depth per sample. ArXiv · AI/CL/LG's note
The paper proposes asynchronous signal propagation through shared hidden neurons, with timing and polarity producing interference between paths. Its genetic algorithm co-evolves topology, weights, delays, and connectivity across a 14,602-gene genome. In the MNIST setup, it uses frozen ResNet18 features and evolves 204 active paths through 266 hidden neurons. The author argues the design avoids differentiable graph requirements and can vary computation depth per sample. ArXiv · AI/CL/LG's note
score 4