AI for sustainability should be sustainable in its own: results from project FAIR-Spoke 9

Current Machine Learning technology can help address sustainability challenges in various sectors and better assess and mitigate natural risks. However, this technology has a substantial ecological footprint, and its accelerating adoption across industries risks turning it into a threat to sustainability in itself. This paper summarizes the results of research activities of project FAIR-Spoke 9 (Green-aware AI) that concern developing foundational methods for improving the eco-efficiency and sustainability of data-driven AI pipelines. This goal has been pursued by two research directions: reducing the amount of training data in a task-aware way; and leveraging background information. The first research direction has led to methods for selecting training data in a task-oriented adaptive way. The second line has yielded training methods exploiting users’ feedback and prior models or devoted to training multi-task modular models. A high-level technical description of two representative methods of both types is offered in the paper, along with a discussion of research implications, limitations, and future research directions.