Papers with Code is an excellent resource for anyone doing research related to or involving machine learning as well as researchers interested in open science and reproducibility. As the name indicates, every paper included in the Papers with Code database includes the associated code in a GitHub repository. These associated GitHub repositories allow users to examine the code, discover contributors, and make a copy (known as a “fork”). In addition to GitHub integration, papers that use Python can link to Google Colaboratory, a resource that allows users to execute Python code in their browser. The combination of GitHub repositories and Colaboratory make Papers with Code a powerful platform for those interested in reproducing machine learning research.

The collaborative component of Papers with Code is not limited to forking GitHub repositories. Researchers can also add directly to a paper’s record. Users can add new code implementations via forked GitHub repository, evaluation tables, and tasks directly to a paper’s record. This allows the community around Papers with Code to build and share knowledge beyond the research paper. Users can share their own results and expertise, amplifying the reproducible component of the platform.

Getting started may seem daunting but Papers with Code is created with researchers that use machine learning in mind. Users can search and browse by method or model. For example, those interested in computer vision can browse papers that use specific computer vision methods. This browsing is possible because of the extensive keywords attached to each paper. Any method used within a paper is tagged and searchable, allowing easy cross-referencing for methods across papers.

Papers are also divided by academic discipline into Portals. Papers with Code has 6 portals: Machine Learning (main portal that includes all other portals), Computer Science, Physics, Mathematics, Astronomy, and Statistics. Selecting a portal allows researchers in those fields explore the specific machine learning applications for their discipline.

Users can also search by dataset. Currently, Papers with Code has identified over 4000 datasets that were used in their papers. Users can search by dataset or share their new dataset with the community. Each dataset contains metadata on the licensing, modalities, and languages as well. Papers with Code has also recently partnered with arxiv to make it easier for arxiv datasets to made available in the Papers with Code database.

Papers with Code also offers an active Slack channel that folks can use to ask questions about the site or troubleshoot problems. The moderators and community around the Slack channel are very responsive, often answering questions on the same day. The platform was also recently acquired by Facebook AI, another indicator of the success of Papers with Code.

by Samuel Putnam, Director of Bern Dibner Library, New York University

We welcome your comments and suggestions. If you have a resource that you would like to see highlighted please leave us a comment.