Evolving Neural Networks.

Twelve tiny neural networks (two inputs, two hidden neurons, one output) evolve to compute XOR, with no training, only selection. Each generation every network is tried on the four input pairs and scored out of 4; the better half survive, and their offspring mix the parents' weights and mutate. Each row is one network, drawn once for each input pair, best network first; a neuron is green when it fires (activation above 0.5).