Sustainability of Artificial Intelligence Systems
Description
This research area explores several dimensions of sustainability of Artificial Intelligence (AI) systems, with special emphasis on Machine Learning (ML) systems: environmental, economic, technical, social, and individual sustainability. The research aims to foster sustainability in all the stages of the MLOps/ML lifecycle.
Team
Projects Overview
Title | Type | Goal |
---|---|---|
Towards green AI-based software systems: an architecture-centric approach | Research | To provide data scientists and software engineers tool-supported, architecture-centric methods for the modeling and development of green AI-based systems. |
Generation of energy labels for certifying the efficiency of machine learning systems | Technology transfer | Studying the viability of tool-supported energy labels for certifying ML systems. |
Greening AI with Software Engineering | Research network | Proposing to connect current developments in computer science applications with the urgency of transitioning to (and maintaining) a sustainable society. |
Audiovisual and digital material for data engineering | Teaching innovation | Defining and improving university courses regarding the technical sustainability of ML systems in the MLOps lifecycle, including best software engineering practices for ML systems. |
Outcomes/Main Contributions
Research:
-
Environmental sustainability of ML systems (Green AI) aiming to reduce the environmental impact in the ML lifecycle by:
- Reporting guidelines of ML carbon emissions during training.
- Reducing ML model size and using lightweight architectures.
- Applying interactive mid-training tactics such as early stopping.
- Evidence-based decision making on where to execute (in terms of cloud-edge continuum and hardware).
- Optimizing the selection of serving infrastructure (e.g., runtime engines, execution providers).
- Technical Sustainability of ML Systems:
- Economic Sustainability of ML Systems:
Technology Transfer:
- Tools for environmental and economic sustainability of ML systems, such as the GAISSALabel tool for energy efficiency label for ML systems.
- Contributions to standards activities such as “ISO/IEC DTR 20226 Information technology — Artificial intelligence — Environmental sustainability aspects of AI systems” and the Spanish UNE working group CTN-UNE 71/SC 42/GT 1 "Evaluación de la eficiencia energética de los sistemas de inteligencia artificial."
Teaching:
- The state-of-the-art outcomes of this research area (and the SE4AI research area) are integrated into two subjects:
- “Machine Learning Systems in Production” (MLOps) of the Master of Data Science at UPC.
- “Advanced Topics of Data Engineering 2” (TAED2) of the Bachelor degree of Data Science and Engineering at UPC.
Community:
- Co-organizers of the 2024 and 2025 editions of the International Workshop on Green and Sustainable Software (GREENS).
Team Leader and Contact Person
Silverio Martínez-Fernández
Contact Silverio Martínez-Fernández
Share: