When a paper claims 0.1 top-5 on ImageNet with a feature extractor trained on 10% of the data w/o looking at labelspic.twitter.com/Qp5Nnp7pEb
-
Not that I think it's a bad idea. I've been trying to get stacked matrix-factorization-based conv feature extractors to work for years
If claims are true it would be v exciting. But I cannot believe it at this point. Too outlandish especially in light of my own experience
-
-
did anything from the paper stand out as potentially significant in comparison to your previous work?
-
the residual-like shortcut. There's a need to prevent information from disappearing from layer to layer in a stacked setting
End of conversation
New conversation -
-
-
If anyone has independent repro, hit me up. Meanwhile, I'll be at the bar
-
Which bar?
- Show replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.