Is the size alone enough to do that with a "fairly good accuracy"? I can't help thinking of all the deep learning tests that inadvertently found unwanted shortcuts.