> Interestingly, that's the "trick" behind a lot of the seemingly magic skill of geo guessers. The best players have played so much, that they now "see" things that a regular person wouldn't even consider to look for, like the camera quality, what year the car was from, and so they narrow down the possible countries by those aspects, before even looking at the "picture".
Slightly OT, but this happens constantly with ML classifiers on any highly multi-dimensional problem. At first it seems like magic, and then someone digs into the principal components of the prediction, and finds a mixture of a few highly specific factors that -- in the worst case -- is an artifact of the dataset itself (image blur or color bias, for example).
Also common is that the predictive factors aren't pathological -- they're just "boring" -- and therefore the performance of the model is dismissed by the practitioner ("oh, I'd have thought of that, since it's only using a few common traits that are well-understood.")
Slightly OT, but this happens constantly with ML classifiers on any highly multi-dimensional problem. At first it seems like magic, and then someone digs into the principal components of the prediction, and finds a mixture of a few highly specific factors that -- in the worst case -- is an artifact of the dataset itself (image blur or color bias, for example).
Also common is that the predictive factors aren't pathological -- they're just "boring" -- and therefore the performance of the model is dismissed by the practitioner ("oh, I'd have thought of that, since it's only using a few common traits that are well-understood.")