0.1% accuracy increments correspond to 10 images in the testing set; they should be reporting standard error bars with those numbers.