Surprisingly, AI edge devices produce high carbon dioxide. A study revealed that training off-the-shelf language processing system emits 1400 pounds of CO2. Training full processing AI from scratch can cause 78000 pounds of emission. Han Cai and his team designed an efficient algorithm for training networks once for all, namely the Once-for-all model. In simpler words, conventional AI systems train each data set according to a set of rules each time they appear. Han Cai suggested that if we train a full system once and then that training set follows the whole approach without being trained from the beginning. This can save the AI environment cost. The whole initiative is called “GreenAI,” in which scientists are working to reduce carbon emission to run AI. Once-for-all model can reduce carbon emission 1300 times more than conventional methods. Read more

April, 2020