François CholletVerified account

supervised and unsupervised models are not necessarily probabilistic models by default.

seems like for non probabilistic models it's usually more clear if supervised or unsupervised. I'm thinking like kNN or k-means

i know its weird, but it can't just be conditional, ex ppl consider text prediction unsupervised, although its conditional

I think because of old models such as pLSA which model the joint probability?

(just speculating). Also seems very blurry when eg you have a softmax over 1M words

back in the days, in the intersection of cv/neuronsci/ml unsupervised learning was about projection pursuit

but the quality of the projections was measured as how useful they were for SL tasks. So UL always lacked a good def.

isn't the textbook def that UL is about modeling joint distributions? Just in practice not the most useful distinction