From the record
Real cases, written from the published record. Predict first; the record and its sources come after.
1907 · Berlin
Clever Hans
A horse named Hans taps out answers to sums with his hoof, and he is usually right. A 1904 panel of experts watched him and found no trick: no signals, and Hans answered strangers too.
The psychologist Oskar Pfungst then ran his own tests, changing one thing at a time.
Which change made Hans start getting the answers wrong?
What happened. When the person asking knew the answer, Hans was right about nine times in ten. When nobody near him knew it, he was right about one time in sixteen. Hans was reading tiny, unconscious movements of the questioner, who relaxed when he reached the right number. Pfungst found he could steer Hans on purpose with the same small head movements, and that he also cued him without meaning to.
The horse was smart, at reading people. The test was measuring the wrong thing.
Pause to think. A model scores 100% on a test. What two questions do you ask before believing it?
2019 · Berkeley
ImageNet, again
ImageNet is a famous test of image recognition, and for years models were tuned to score well on its test set.
Researchers built a new test set the same way as the original: same sources, same labelling steps. Then they ran the top models on it.
How much accuracy did the top models lose on the new set?
What happened. Accuracy fell by about 11 to 14 percentage points. The order of the models mostly held: a model that was better on the old set was usually better on the new one.
What we don’t know. The authors could not fully explain the drop; small differences in how the new images were chosen may account for part of it.
Pause to think. Can a test be used up by the people who keep tuning to it?