Towards FAIR for Machine Learning Models

Authors: Daniel Katz (University of Illinois), Chunhua "Leo" Liao (Lawrence Livermore National Laboratory), Gnana Bharathy (Australian Research Data Commons, University of Technology Sydney)

Abstract: Machine learning (ML) is increasingly used in HPC, both for training new models and testing/using existing models. To make efficient use of HPC and human resources, it's essential to effectively share developed ML models, and to think of them as products, like data and software; this can be done by applying the FAIR principles to them. This BoF will discuss FAIR for ML models, and aim to bring HPC ML researchers together with the larger ML and information systems communities. It will be 1/4 presentation and 3/4 discussion of bottlenecks and future community activities/organization.

Long Description: Machine learning is increasingly used in HPC, with work in both training new models and testing/using existing models. To make efficient use of HPC and human resources, it's essential to effectively share developed ML models. This means we need to think of them as products, just like we do with data and software, and this can be done by applying the FAIR (findable, accessible, interoperable, reusable) principles to them, as funding calls are starting to discuss. This BOF will discuss FAIR for ML models, and help bring SC community members interested in further work together in collaboration with the wider (non-HPC) ML and information sciences community. The BOF will also briefly address the interaction between data, infrastructure and ML, and acknowledge allied concerns such as ethics and privacy in this context. The BOF will be 1/4 presentation, and 3/4 discussion of bottlenecks and future community activities and organization.

This BOF is relevant to 1) SC researchers who build ML models and want to share them, but are not sure what metadata to use to describe them, or how to share them; 2) SC researchers who want to use ML models that others have developed, including to compare results with their own models; 3) institutions that provide computing resources that are used for ML and need policies on how to maximize the impact of these resources; 4) institutions that support ML researchers and need to understand factors to use in evaluating how their work is impacting the community; and 5) researchers who are developing techniques or frameworks aimed for improving FAIRness of ML models, training datasets, and workflows in HPC. It is new to SC.

The outcomes of this BOF will be 1) increased understanding of ML models as shareable products; 2) exchanging information about active projects related to making ML FAIR for HPC; and 3) sharing experiences and lessons when applying FAIR principles to ML in HPC.

URL:


Back to Birds of a Feather Archive Listing