The ML research community has long been driven by the need to publish, which results in a stark, sometimes ridiculous bias towards complexity. Remember to ask: "can we do this with k-means and logistic regression?"
Show this thread
Can we do super-resolution with k-means & logreg? Probably Can we do generative models (structured latent space from which on can sample high-res images) with interpolated patches with k-means & logreg? Probably Pixel-to-actions on Atari games with k-means & logreg? Probably Etc
-
-
Can we do 83% accuracy on CIFAR10 with k-means and logistic regression? Yes (Coates 2011).
Show this thread -
I think we should have a "simple, strong baselines" workshop at major ML conferences. This would help focus ML research on real progress.
Show this thread
End of conversation
New conversation -
-
-
As always, there's a small finite set of concepts and techniques that can be extended and applied to a theoretically infinite set of use cases.
-
The data also shows. There's more use of logistic regression and similar simpler (not easier) techniques in the industry. ML researchers should keep it coming anyway.
End of conversation
New conversation -
-
-
Performance metrics abstracted from practical use contribute to that problem. DL is good at offering apparent improvement over k-means plus logistic regression with no real benefits.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
There's some recent work that showed kNN is enough for pixel-to-actions, but the reviewers hated it and it will prob get rejected ...https://twitter.com/hardmaru/status/932439978017136640 …
-
@srchvrs what are your comments? - Show replies
New conversation -
-
-
Why don't you demonstrate that? It should be easy to do and would be a bunch of great publications if it works!
- End of conversation
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.