Are you implying models that classify hand-written digits don’t generalize and only work on training data?