devol

Why we included this project

devol is a small experiment that treats neural architecture search as a genetic algorithm problem. It encodes the usual network design choices, like layer counts, feature maps, activations, dropout, and optimizer, into a fixed-width genome. Then it breeds a population of Keras models through crossover and mutation, keeping the top performers each generation. The code is compact enough to read end to end, and the notebooks walk through MNIST and CIFAR examples so you can watch the search in action. Just set expectations: this was an early proof of concept rather than a maintained library, so it works best as a clear teaching reference for wiring evolutionary search to a deep learning framework, not something to install and trust in production.