Neurogram

Find in this repository an implementation of the CPPN-NEAT algorithm to create networks that represent images. I tried to keep things simple and lightweight, using only matplotlib as a dependency for drawing the images. The dependency networkx was added for drawing the networks for debugging purposes. The networks themselves are simple Python objects. 6 networks (and hence images) are generated at random, and from then on two images can crossover in various ways into 6 new images, and so on.

I was heavily inspired by this blogpost. They are able to generate way prettier images and explain how they did it in the post. I did not bother to fine-tune my parameters, this was merely a fun exercise to try and implement a neural network that evolves using a genetic algorithm. The topological sorting of the nodes was an interesting sub-problem to solve, because the nodes are not arranged in layers like in “normal” neural networks. According to Wikipedia, listing the nodes in the opposite order of their last visit of a depth-first search produces a topological sorting. So that is what I did.

Also, I used uv for the first time (late to the party, I know), was impressed by the speed and convenience. I especially enjoyed the inline metadata format which allows the dependencies to be declared in the .py file itself. I will definitely use this more often for small scripts like this. Some output screenshots: