A modular neural network works by dividing the input domain into segments, assigning a separate neural network to each sub-domain. This paper introduces the self-splitting modular neural network, in which the partitioning of the input domain occurs during training. It works by first attempting to solve a problem with a single network. If that fails, it finds the largest chunk of the input domain that was successfully solved, and sets that aside. The remaining unsolved portion(s) of the input domain are then recursively solved according to the same strategy. Using standard backpropagation, several large problems are shown to be solved quickly and with excellent generalization, with very little tuning, using this divide-and-conquer approach.