Green AI
In plain wordsA movement to make AI research and products use less energy and computing power, and to report those costs openly alongside accuracy.
AI research and practice that counts the computation, energy and carbon behind a result as part of its quality, weighing efficiency together with accuracy when methods are compared, and that discloses those costs. Its originators set it against Red AI: work that improves accuracy mainly by spending far more computation, for steeply diminishing returns, and that therefore favors well-funded labs. Proposed practices include reporting floating-point operations, energy and emissions for training and inference; publishing efficiency–accuracy curves; preferring efficient architectures and smaller or distilledknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run. models; scheduling work for low-carbon times and places; and reusing rather than retraining models. The broader label sustainable AI adds hardware lifecycles, water and the use of AI for environmental goals.
The workshop's Green AI checklist asks authors to report the GPU hours behind every number in the results table.
Origin
Introduced by Roy Schwartz, Jesse Dodge, Noah A. Smith and Oren Etzioni of the Allen Institute for AI in 'Green AI' (arXiv July 2019; Communications of the ACM, 2020), which also coined 'Red AI'.
Related
- environmental impact of AIenvironmental impact of AI, n.: The effects of AI on the planet: the electricity, water, materials and land used to make chips, build data centers and train and run models, and the pollution that results.
- Jevons paradoxJevons paradox, n.: The observation that making something more efficient can make people use more of it overall because it gets cheaper — as with coal in the 1800s, and with AI computing today.
- AI data centerAI data center, n.: A building or campus packed with AI chips, power equipment and cooling, built to train and run AI models. The biggest use as much electricity as a city.
- computecompute, n.: The raw computing power used to train and run AI, usually measured by how many specialized chips are used and for how long. With data and algorithms, it is a main driver of AI progress.
- knowledge distillationknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run.
- scaling lawsscaling laws, n.: Measured patterns showing that an AI model's error shrinks smoothly and predictably as the model is made bigger and given more data and computing power, so labs can forecast a giant model's performance from small, cheap experiments.
Sources
- Schwartz, R., Dodge, J., Smith, N. A. & Etzioni, O. (2020). Green AI. Communications of the ACM 63(12), 54–63.
Mentioned in
environmental impact of AIenvironmental impact of AI, n.: The effects of AI on the planet: the electricity, water, materials and land used to make chips, build data centers and train and run models, and the pollution that results.