Towards the systematic reporting of the energy and carbon footprints of machine learning
Source: ACM Digital Library
In this scientific article published in The Journal of Machine Learning Research, and written by researchers Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky and Joelle Pineau.
The primary aim of this research study is to put forward the hypothesis that the lack of data on the carbon emissions of AI models is due to the complexity of data collection. For the researchers, this complexity is due to the fact that certain factors that need to be taken into account have data that varies according to certain conditions.
The research team is then putting in place a framework to simplify data collection.
Figure 1 - A diagram demonstrating how the released version of the tool works.
Source: Towards the systematic reporting of the energy and carbon footprints of machine learning.
To conclude their research, and based on their findings, the researchers propose a series of 7 measures to be taken into account to improve the cost of carbon emissions downwards.
Encourage research into energy efficiency using dashboards;
Conduct experiments in environmentally-friendly regions;
Reduce overheads to use efficient algorithms and resources;
Consider energy-performance trade-offs before deploying energy-intensive models;
Selecting an efficient test environment, particularly in field research;
Ensuring reproducibility to reduce energy consumption due to replication difficulties;
Consistent reporting of energy and carbon measurements.