Green IN Artificial Intelligence from a Software Perspective: State-of-the-Art and Green Decalogue

This work presents a structured view of the state-of-the-art research on Artificial Intelligence (AI), from the point of view of efficiency and reduction of the energy consumption of AI software. We analysed the current research on energy consumption of AI algorithms and its improvements, which gave us a starting literature corpus of 2,688 papers that we identified as Green AI with a software perspective. We structure this corpus into Green IN AI and Green BY AI, which led us to discover that only 36 of them could be considered Green IN AI. After some quick insights about Green BY AI, we then introduce our main contribution: a systematic mapping of Green IN AI. We provide an in-depth analysis of the AI models that we observed during the mapping, and what solutions have been proposed for improving their energy efficiency. We also analyse the energy evaluation methodologies employed in Green IN AI, discovering that most papers opt for a softwarebased energy estimation approach and 27% of all papers do not document their methodology. We finish by synthetising our insights from the mapping into a Decalogue of Good Practices for Green AI.