
Episode 15: Martín Arjovsky, INRIA, on benchmarks for robustness and geometric information theory
Generally Intelligent
00:00
Using a Dataset to Find Markers for Canser in Space
Sevan: I think in general, there are two or three good bench marks. They show very clearly that just training on your pull data fails because of police correlations. But i think her coler is something that you can do with picks and indensities alone. In fact, if you pretrain on image net, will still come to the experiment of life,. no image and do nothing, just doin a liniar model at picks a level. At least the problem is real. You throw a comnetal fat sona. There's all these wilds meta benchmark the people from stamfordt har creative, but they're all super simplistic problems. Basily
Transcript
Play full episode