Raw LLM Responses

> Are you saying any supervised learning problem is trivial once we have labelled data? That seems like quite a stretch to me. not all supervised learning problems are trivial (... obviously). I think my argument -- particularly as it pertains to the case of using radiographic images to identify pre-cancer -- is that it's a seemingly straightforward task within a standardized environment. by this I mean: any machine that is being trained to identify cancer from radiographic images is single-purpose. there's no need to be concerned about unseen data -- this isn't a self-driving car situation where any number of potentially new, unseen variables can be introduced at any time. human cells are human cells, and, although there is definitely some variation, they're largely the same and share the same characteristics (I recognize I'm possibly conflating histological samples and radiographic data, but I believe my argument holds). my understanding of image recognition -- and I admit I almost exclusively work in NLP, so my knowledge of the history might be a little fuzzy -- is that the vast majority of the "problems" have to do with the fact that the tests are based on highly diverse images, i.e. trying to get a machine to differentiate between grouses and flamingos, each with their own unique environments surrounding them, while also including pictures of other random animals. in cancer screening, I imagine this issue is basically nonexistent. we're looking for a simple "cancer" or "not cancer," in a fairly constrained environment. of course I could be completely wrong, but I hope I'm not, because if I'm not: 1) that means cancer screening will effectively get democratized and any sort of bottleneck caused primarily by practitioner scarcity will be diminished if not removed entirely and, 2) I won't have made an ass out of myself on the internet (though I'd argue this has happened so many times before that who's counting?)