I think a key difference is the physics renormalization structures use fairly regular or uniform weights and the deep learning plays a lot with the weights. So there are going to be pretty big differences in behavior.