This isn't necessarily surprising. Maybe clinicians can be trained to get good at it, but, maybe they can't. Back before computers beat Go, the conventional wi

This isn't necessarily surprising. Maybe clinicians can be trained to get good at it, but, maybe they can't. Back before computers beat Go, the conventional wisdom was that it was a hard target because humans can leverage their pattern-finding systems to prune the state space much more efficiently than computers can. Which got me thinking: what about a game that's the opposite? Like the game's state space is embedded in the timbre of a complex tone, and you make moves by twiddling three or four knobs. It doesn't have to be a very complex game for computers to do a lot better at this, extracting meaning out of white-ish noise is not something we're good at. Assuming this isn't a data artifact, it might be one of those cases. Some pattern in the brachiation of the veins or the placement of the rod and cone cells, which just looks like noise to us, but which an ML algorithm can find and use to separate male and female retinas. There's no reason to think that humans would ever get very good at that, although I never count humans out given examples like chicken sexing[0], a notoriously opaque skill which people can nonetheless acquire, despite having little facility in explaining to others how they do it. [0]: https://en.wikipedia.org/wiki/Chick_sexing#Cultural_references https://en.wikipedia.org/wiki/Chick_sexing#Cultural_referenc...