Last week we looked at the information bottleneck theory of DNNs. This ICLR'18 submission provides a critical analysis - turns out the choice of activation function has/had a lot to do with it!http://blog.acolyer.org/2017/11/24/on-the-information-bottleneck-theory-of-deep-learning …
-
-
There's also a really interesting discuss there between the authors!
-
I guess if anything this demonstrates linking back to original pdf is useful and removes misunderstandings like here.

-
Although I guess they did link to the pdf. But I meant to the thread, although, yeah, I've learned something.

-
I guess the only thing I don't understand is if it's open and still anonymous/double blind how can you give credit if you wanted to use the ideas? Why not wait till review is done so you can unanonymise it and make reviews open and pdf available?
-
They are citeable already (there's a bibtex link) and I guess the idea is that more people can contribute to the review straight away. I'm still processing but I think I kind of like it.
-
Yeah, I get the open review at the time of review as opposed to making the whole thing available after. I alse get the double blind thing. But together? It seems incompatible to me. Also citing it as anonymous seems like not citing it? Maybe it gets updated afterwards?
-
Maybe if there was a DOI it would make more sense? To link the citations once the names are out?
-
yeah, I wonder if there are any barriers to that. If not it would seem like a neat solution.
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.