You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
I debated at #DALI2019 about enforcing code release and #reproducibility in #AI #ML #conferences. This is the right thing to do: it is fair and it will ensure honest progress. Otherwise we are headed to next #AIwinter. https://anima-ai.org/2019/01/04/reproducibility-debate-should-code-release-be-compulsory-for-conference-publications/ …
Update on #reproducibilty in #AI #ML #conferences I added Twitter discussions at end of the original blog post. I hope we can enforce reproducibility for upcoming #NeurIPS End of vapor-ware #mediahype Let true science begin https://anima-ai.org/2019/01/04/reproducibility-debate-should-code-release-be-compulsory-for-conference-publications/ …
Hi Anima, can you do something about NVIDIA offering zero dollars to sponsor the top architecture conferences? Thanks for everything you do!
I can't understand if a researcher publishes a paper what's wrong with providing the code? Thank you for the blog-post
YES PRECISELY! And if we release our code, we should imho also at least point to any public datasets that some of the results are based on. I know this can sometimes be tricky with privacy issues, but what good is public code that depends on other restrictions?
It seems the counter arguments read more as excuses than arguments. That being said, I think there's a very important scientific argument for arguing against code release: *independent* replication of the underlying scientific idea is crucial for progress.
If some algorithm is so intensely sensitive to hyperparameters or bugs in the original implementation, then the scientific idea will die away as its replication fails. You can argue this is wasted time for those who tried to replicate, but it's of great value to science.
IIRC, rapid progress in Robocup came because each year people published their code. This acted as a ratchet: the code from the best team in one year became the basis for the next year. So even "leaderboard chasing" improves with open code.
It's analogous to natural selection. How can methods improve if nobody really knows which methods have true fitness and which don't? We are now in the realm that we just trust authority, if DeepMind, MIT publish then we treat it as gospel.
@AnimaAnandkumar It was an honor debating alongside you for a cause I too personally believe in!
I made an edit to my post on #reproducibility at #DALI2019. I checked with @ShalitUri that he was fine with being named. It was a pleasure to debate along with him and we came up with the points collaboratively.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
