Green Artificial Intelligence Standard

The Green AI Standard aims to develop a standard and raise awareness for best environmental practices in AI research and development

The climate issue in AI

Developing machine learning models is extremely costly for the environment (Strubell et al. (2019))

• Training BERT on a GPU is roughly equivalent to a trans-American flight (650kg CO2)
• One (512px) BigGAN experiment is equivalent to a trans-Atlantic roundtrip (~1 to 2t of CO2)
• Neural architecture search experiments for Transformer is emitting as much as 50 years of an average human life (~280t of CO2)

Information and communications technology is on track to create 3.5% of global emissions by 2020–which is more than the aviation and shipping industries–and could hit 14% by 2040 (Guardian (2018)). We need to take a stand now!

Best practices in development

1. Report time to retrain machine learning models (e.g. GigaFLOPS till convergence, Strubell et al. (2019))
2. Report sensitivity of hyperparameters for machine learning models (e.g. variance with respect to Hyperparameters searched, Strubell et al. (2019))
3. Use more efficient alternatives to brute-force grid search for hyperparameter tuning (e.g. random or bayesian search, Strubell et al. (2019))

Best practices in infrastructure

Minimize costs and carbon emissions by sharing local infrastructure instead of relying on on-demand cloud computing resources (Strubell et al. (2019))

Best practices in deployment

The fossil fuel industry is responsible for most of the world's CO2 emission by a large margin. Artificial intelligence has been a driving force of optimizing gas and oil extraction processes. By following the Standard, we pledge to not make developed applications available for fossil fuel focused usage.

Offset your resulting emissions

We recommend offsetting your emissions to certified carbon neutrality projects if possible. Offsets can be calculated via MyClimate and purchased here:

• MyClimate
• Gold Standard
• ActForest

You can also measure how much power your deep learning model has consumed via Power Meter. Note that it only covers GPU consumption.

Show your commitment with a badge on your repository

👇 The Pledge Badge	👇 The Carbon Neutral Badge

The pledge badge shows your commitment to do the best to reduce the greenhouse gas emissions caused by your research by following the best practices developed by the Green AI Standard

The Carbon Neutral Badge shows that your greenhouse gas emissions caused by your code repository are offsetted. The badge should link to the offset certificate for verification

Acknowledgement

The green ring is inspired by the Climate Reality project