¿What is the state-of-the-art research on Green Artificial Intelligence?

What is it about?

In this work, we reviewed 2.688 papers that study the energy consumption of Artificial Intelligence (AI). In particular, we were looking for research that was published between 2016 and 2021, and that focused on the energy efficiency of AI software, rather than the hardware it runs on. We observed that, out of the 2.688 papers, only 36 were looking into the actual consumption of AI software, with the other 2.659 focused instead on using AI to make other systems more energy-efficient (for example, using AI to reduce the energy consumption of a building's heating). Since our focus was the energy efficiency of AI software itself, we then performed an analysis of those last 36 papers, which resulted in a series of insights regarding the methodologies that researchers employ to evaluate the energy consumption of AI software (including, for example, that 27% of those 36 papers didn't report their methodology), a summary of the currently existing solutions for improving the energy efficiency of AI software, and what unique challenges stand in the way of AI energy efficiency. Finally, we propose the Green AI Decalogue, with 10 guidelines that are based on the insights gained during the analysis process of this work.

Featured Image

Photo by Pietro Jeng on Unsplash

Photo by Pietro Jeng on Unsplash

Why is it important?

The research published about the energy consumption of Artificial Intelligence has increased a lot during the last few years. Because parsing through this much information is a challenging task, systematic mappings like this can help its readers to understand what are the current research trends in the area of Green Artificial Intelligence, what energy evaluation methodologies are currently being used by researchers, which energy-saving solutions have been applied in which AI models, and which AI techniques are the most (an least) studied in the field.