What is it about?
The paper presents a new extension specification to describe Machine Learning Models (MLM) within the context of Spatio-Temporal Asset Catalogs (STAC) to help FAIR (Findable, Accessible, Interoperable, Reusable) AI within the geospatial domain. https://github.com/stac-extensions/mlm
Featured Image
Photo by Kevin Stadnyk on Unsplash
Why is it important?
Current Model Card definitions and published AI/ML models on hubs like HuggingFace do not capture sufficient metadata about geospatial and spatio-temporal properties to replicate experiments. Furthermore, the scale of datasets involved in training such models, and their retrieval, is insufficiently described to properly understand how to reuse the models and understand their application domain. Using the proposed specification, interoperability, provenance and trust of decision-making pipelines involving such models can provide better understanding of the process with standardized references and metadata.
Perspectives
The specification aggregates multiple previous efforts attempting similar solutions, but each lacking some considerations over certain aspects (application domain, runtime requirements, cross-referencing sources, unifying the ecosystem with other existing metadata extensions). This new specification brought together many of the involved parties and agencies involved in the previous efforts, to unify all aspects under a common reference. The specification also comes with a set of tools that facilitate its use and integration with existing code bases and common development packages. This was a lacking aspect of previous efforts, which did not help in their adoption. Finally, by regrouping multiple agencies and companies working together on this new specification, sustainability of the solution is improved. This was also an issue with previous references, which quickly lost traction due to insufficient or undefined maintenance stakeholders.
Francis Charette-Migneault
Centre de recherche informatique de Montreal
Read the Original
This page is a summary of: Machine Learning Model Specification for Cataloging Spatio-Temporal Models (Demo Paper), October 2024, ACM (Association for Computing Machinery),
DOI: 10.1145/3681769.3698586.
You can read the full text:
Resources
Contributors
The following have contributed to this page
