Comment by tednoob

Is this method used during training? Seems to me there could be a point to only backpropagate when the model is wrong?