Anandkumar: more audience ideas
- we need to encourage people to care about the science
- make the contribution the code, not the paper (papers as documentation)
- focus on the steps before you get to the code
- record all the little trivial tricks
#icml2018 #ReproducibleML
-
Show this thread
-
Anandkumar: more audience ideas - student perspective: I'd rather have the code than a really detailed paper because it's faster & I'm under a lot of time pressure
#icml2018 - increase the incentives for reproducing & publishing negative results#ReproducibleML1 reply 2 retweets 5 likesShow this thread -
Anandkumar: Different forms of reproduciblity - not reproducible in theory (conditions in paper cannot be recreated) - reproducible in theory but not in practice (require enormousee resources, "magic" hyperpameters which are not shared/hard to find)
#icml2018#ReproducibleML1 reply 3 retweets 5 likesShow this thread -
Anandkumar: And the ideal case, which is reproducible in both theory & practice
#icml2018#ReproducibleML1 reply 1 retweet 2 likesShow this thread -
Anandkumar: Lots of work needed to develop - best practices for describing the algorithm - bets practices for publishing code (Jupyter code? hard for complex conde) - unit tests (helpful for toy examples, but may not find all the bugs)
#icml2018#ReproducibleML1 reply 2 retweets 5 likesShow this thread -
Anandkumar: We could also choose to focus on code that can run out of the box? (Docker? Automatic validation?)
#icml2018#ReproducibleML1 reply 2 retweets 2 likesShow this thread -
Anandkumar: We already have all the technology we need to ensure reproduciblity. The problem is social engineering: how can we incentivize people to care about reproduciblity?
#icml2018#ReproducibleML1 reply 5 retweets 7 likesShow this thread -
Anandkumar: I've worked on a number of projects explaining concepts using Jupyter: - https://gluon.mxnet.io/ for machine learning - https://github.com/tensorly/tensorly … for descibing algorithm along with code
#icml2018#ReproducibleML1 reply 10 retweets 17 likesShow this thread -
Anandkumar: There's also Dockerizing, like AWS SageMaker/CodaLab, that lets reproducers just grab and go
#icml2018#ReproducibleML1 reply 1 retweet 3 likesShow this thread -
Anandkumar: The tools exist! The big question is, how do we get people to actually use them?
#icml2018#ReproducibleML1 reply 5 retweets 5 likesShow this thread
Nice work live tweeting all this! 
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.