Scope
In (Luccioni et al, 2023) the authors provide a comprehensive survey of the factors that contribute to the carbon emissions of machine learning (ML) models. The authors highlight the energy consumption required to develop and deploy ML models, focusing on how different models and tasks generate varying levels of greenhouse gas emissions. The study analyses 95 models across different ML tasks such as image classification, machine translation, and object detection, quantifying their carbon footprint. The goal is to raise awareness of the environmental impact of ML models and to encourage the creation of a centralised repository for reporting and tracking carbon emissions in ML research.
Summary
Several key insights are drawn from the paper. The authors acknowledge that carbon emissions reporting is inconsistent across the ML field, leading to potential underestimation of emissions. Further, the paper mostly focuses on models from well-known tasks (e.g., NLP, computer vision), which may not reflect emissions from less-studied models or domains. Additionally, while the study accounts for carbon intensity using yearly averages, it does not leverage real-time data that could capture fluctuations in energy grid composition.
Relevance for EXIGENCE
This work is highly relevant to EXIGENCE project for energy metrics, energy measurements and energy optimisation, aimed at incentivising users to minimise energy consumption. By providing insights into how different tasks, models, and energy sources influence emissions, the paper supports the development of strategies to encourage the use of energy-efficient ML models. Additionally, it advocates for standardised reporting, which aligns well with the project’s goals of tracking and rewarding reduced carbon emissions.
- Luccioni, and A. Hernández-García, “Counting Carbon: A Survey of Factors Influencing the Emissions of Machine Learning,” arXiv:2302.08476, Feb. 2023.
