The idea is simple. When is optimation too much optimation? If machine learning is suppose to be as efficient as the brain. It should not relearn all the weights if you want to increase the total number of weights. It will take to long to reoptimize. So the idea is for a system that can be multiplied with as little recalculation as possible. I don’t know how neural networks could cooperate. Maybe the answer from one network or part of it can be improved or rejected by another set of networks.