FAIR for Machine Learning (FAIR4ML) IG rev-002

The idea of FAIR (findable, accessible, interoperable, and reusable) in the context of scientific data management and stewardship was developed in 2014 and turned into specific principles in 20161 . Along the way, the idea was generalized in concept to apply to both data and other digital scholarly objects, but it has become clear in practice that what works for data may not directly work for all other digital objects. For example, both previous and ongoing work show that many of the guiding FAIR principles need to either be re-written or reinterpreted for software, resulting in the FAIR principles for Research Software2 , already with an adoption commitment from different communities and institutes3 . The FAIR principles also can apply to machine learning tools and models, though a direct application is not always possible as machine learning combines aspects of data, software and computational workflows.

There is a large amount of work around FAIR, both in RDA and elsewhere, initially focused on data and now software and other products but generally not ML models. There has already been some community work in this direction, both within RDA as well as beyond. This subject was first discussed at RDA during the RDA Virtual Plenary 16 as poster 31b (FAIR principles for ML models – https://doi.org/10.5281/zenodo.4271995). Additional activities include two dedicated BoF sessions at RDA VP17 (https://www.rd-alliance.org/defining-fair-machinelearning-ml) and RDA VP18 that aimed to capture the overall perspective on the topic. The discussion around FAIR for Machine Learning continued in further events under different domains; during the FAIR Festival, the efforts of FAIR4ML were presented together with similar initiatives for Software and Workflows. A position article was presented at the 2nd Workshop on Data and research objects management for Linked Open Science (DAMALOS 2021)4 . Additionally, a dedicated BoF session was held in the context of SC21 (“Towards FAIR for Machine Learning Models”, November 18th 2021), with a focus on FAIR for ML models, bringing together HPC ML researchers and the larger ML and information systems communities. Moreover the initiative was presented in the context of “Best Practices for Reusability of Machine Learning Models: Guideline and Specification” during ESIP 2021 in July 2021, as well as in the “Improving ‘FAIRness’ and ‘Fairness’ of AI/ML in Geoscience” session in January 2022. Finally, two informal Community Calls were organized, a kickoff in July 2021 and a follow-up in January 2022, aiming to organize and structure the efforts towards an RDA Interest Group.

Please see attached document for full Charter.

FAIR4ML-RDA-IG-Charter.pdf