Artificial Intelligence plays a dual role with respect to climate change. First as a powerful tool to predict events, model many complex phenomena and to optimize many processes. There is, however, a second area that has been neglected by researchers and industry: the ecological impact of artificial intelligence itself. Only recently some light has been cast in this direction.
On one hand, it has been forecasted that by 2030 half of the world’s electric energy consumption will be attributed to computing facilities. On the other hand, recent studies show the design and training of state of the art machine learning models produced the same amount of CO2 as six medium cars during their lifespan. This raises concerns on how to make an ecologically-viable artificial intelligence. In this talk I will go through recent advances on the combination of cloud computing, transfer and active learning, model reuse, and evolutionary computing, among others, that could lead to an eco-savvy AI.
However, this needs yet to be properly understood as it calls for a coordinated effort involving different areas of AI, ML, HPC and theoretical computer science.