Why Neural Networks Can Discover Symbolic Structures