I’m speaking at ITA today about our new paper “Do ImageNet Classifiers Generalize to ImageNet?” http://people.csail.mit.edu/ludwigs/papers/imagenet.pdf … (1/10)
Prikaži ovu nit
-
These four images were selected by MTurkers as a *bow*: “a weapon for shooting arrows, composed of a curved piece of resilient wood with a taught cord to propel the arrow.” (6/10)pic.twitter.com/ykbhJrxNIO
Prikaži ovu nit
Takeway 1: It’s fine to build giant models that suck tons of information out of a holdout set. It’s really hard to overfit in our standard ML paradigm. (7/10)
-
-
Takeaway 2: Distribution shift is real and scary. Simply changing a pool of MTurk workers causes a huge drop in accuracy. Imagine what happens when you take these models and put them out in the wild. (8/10)
Prikaži ovu nit -
This is the result of tireless heroic work by
@lschmidt3 ,@BeccaRoelofs , and@Vaishaal, who had the patience and determination to see this through no matter how many obstacles they hit. It’s such a privilege to be part of this amazing project. (9/10)Prikaži ovu nit -
PS the paper is currently on hold by the arxiv. This is either because our final pdf file is slightly too large, or because arxiv is following
@OpenAI's lead and deeming all AI/ML research too dangerous to release. (10/10)Prikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.