Browse State-of-the-Art Machine Learning Papers with Code
paperswithcode.com
paperswithcode.com
https://join.slack.com/t/paperswithcode/shared_invite/enQtNT...
Previously a ShowHN: https://news.ycombinator.com/item?id=18496347
Also would be great to include papers with SOTA results on “tabular” Multivariate datasets, the kind that arise in numerous applications, e.g. EHR/MHR, advertising, finance, etc. In other words, something like the UCI ML Repository datasets (which are mostly “small” but still would be great to know the SOTA models for those), and much larger versions of such datasets — I often see papers applying ML to tabular healthcare datasets but the datasets are often not available.
I think it should be mandatory to publish your code, dataset (or at least a sample of the verification data), and then the trained model weights if you are publishing a neural network "application" paper.
It's very simple to put it on github etc, and then link it in the paper, disappointing when it doesn't happen.
This problem doesn't just exist for neural networks. There are lots of papers published presenting some algorithm, but they never provide the source you can quickly check or experiment. Certainly now the technology exists to allow things to be more open (ie github).
Is there any way to add some kind of flag/sorting mechanism on this for code that requires massive GPUs/ computing power to reproduce?