A look at the modularity of neural networks. They develop a notion of modularity. Specifically, modules in neural nets are groups of neurons that are highly connected internally and weakly connected externally.
First experiment is to train a network on a classification task, and then try to cluster modules. Then they reshuffle the weights of the network and again re-cluster. They find that the generally trained networks are more modular than the randomly shuffled ones. Interestingly, though, CIFAR is not very modular by this metric (i.e. the trained network tends to be LESS modular than the shuffled one).
They then try this experiment with dropout. Dropout makes the networks much more modular.
They then examine whether this modularity is a result of the optimization algorithm (ADAM). They show that running ADAM on an unlearnable dataset does not result in a particularly modular network (although they had some outliers on some of the longer training runs that were very modular).
Similarly, they tried to memorize a small, random noise dataset. They show that generally the networks that memorize the data are more clusterable (although dropout seems to prevent this but also does not learn the data).
They then do lesioning experiments where they “kill” certain modules. They show that some of these modules end up being highly impactful in the classification accuracy of the network.