Today we are launching new sites for computer science, physics, mathematics, astronomy and statistics. Partnering with arXiv, you can use these sites to sync code to show on arXiv paper pages. These sites are live today, and the code tab is now enabled for arXiv papers from all fields! Explore (and add code to) our new portal here: https://portal.paperswithcode.com
As a result of this expansion, we are now tracking artifacts for over 600k research papers. This is just the beginning, and we are deepening our coverage in the weeks and months ahead. We hope our efforts increase code availability for all fields of science — making it a research norm — so the entire research community can progress more quickly together!
Building the New Open Science
Accelerate scientific progress by making research easier to grasp, use and extend
Papers with Code began by targeting problems experienced by the machine learning community. We set out to tackle the problems of accessibility, reproducibility and information overload. We tackled these problems with products such as code matching to papers, scientific results extraction, a task taxonomy, and a methods corpus. We are grateful to the 6000+ individual editors who are contributing to this free resource.
But these problems are not exclusive to machine learning: the same problems exist in broader computational science. That is the context behind today’s expansion of Papers with Code. We are expanding in response to two trends that we observe within the scientific community.
The first trend is the increasing demand for reproducibility within science.
In October we launched our collaboration with arXiv to support code on paper pages. The message resonated with the broader scientific community and we saw a surge in edits with people submitting code to their papers. We also saw lots of physics, astronomy and other domain papers being added to PwC — before we’d even had a chance to consider serving more sciences!
So the direct reason for our expansion today is to serve the broader scientific community. Looking over the last 5 years, code is available for 25% of ML papers. This contrasts with a code availability of 2.3% of papers in other fields. So we will help more researchers tackle this common problem, and help drive up code availability in all fields.
The second trend is the growth of deep learning as applied to science.
One of the new trends we see in our data is the rise of ML being used in scientific tasks. Neural networks are now a critical part of the toolkit for scientists, from protein folding to galaxy detection. So we are also expanding to serve the “machine learning in science” community. For now we are targeting code availability as a first step, which is the same approach we took for the main ML resource.
These two trends are not destined to be entwined. ML-powered science does not have to be open science. But we believe that combining the force of ML-powered scientific discovery with the force of open research collaboration is multiplicative. This compounds scientific progress in an open way to everyone’s benefit.
At Papers with Code, we want to work towards this goal of Open Science: where research is increasingly open and extensible, and where the method base is increasingly computational and machine learning driven. If you think this is where science should go, then join us by contributing to the resource today. Add code, add results — help science progress!
We hope you have a happy holiday season, and we look forward to what we can build together as a community in 2021!
The Papers with Code team
(Robert, Ross, Viktor, Marcin, Ludovic, Elvis, Guillem)