Decomposing neural networks as mappings of correlation functions