Does CLIP's generalization performance mainly stem from high train-test similarity?

Prasanna Mayilvahanan, Thaddus Wiedemer, Evgenia Rusak, Matthias Bethge, Wieland Brendel

Browse the full ICLR paper archive.