3 ms·
> Perhaps an unpopular opinion, but isn't the model just representing the world as it is, rather than what we want it to be? Probably not, becayse its trainung
by dragonwriter 4y ago
> Perhaps an unpopular opinion, but isn't the model just representing the world as it is, rather than what we want it to be?
Probably not, becayse its trainung data almost certainly has biases compared to “the world as it is”, and the model itself has some error with regard to the training data.
It's representing one view of the world seen through multiple levels of filters each with their own set of (systematic or random) biases. (Which is not too dissimilar to any individual human, but any widely used model's biases are likely to have much wider impact than any single human’s.)
- colordrops 4y agoSure, it absolutely makes sense to adjust models to account for biased inputs. Say for instance the pictures you train on are from tundra but you want to use the model in a jungle region. But you should use vetted stats to determine the nature of the landscape to feed them model (e.g. 95% of land is jungle, so feed 95% jungle images), rather than feeding it tundra images because you think the jungle is too hot.
- dragonwriter 4y agoIf its a model for art, I’ll absolutely feed it tundra images if I think the jungle is (or just looks) too hot. Of course, I’ll also embrace the projected cool vibe of the resulting imagery as a feature, and won’t object to people pointing out that the model is biased toward tundra over jungle and that is something users who aren’t actively looking for that bias might want to be aware of.