On Symmetry and Initialization for Neural Networks