As AI adoption becomes more widespread, so do concerns about its energy consumption. Our latest paper proposes a method to train AI models more efficiently. The method reduces energy use by more than half when training a ResNet-18 model on ImageNet. arxiv.org/pdf/2405.02385
This is achieved by decorrelating layer inputs using an iterative decorrelation method initially described in Ahmad et. al. (2022). arxiv.org/abs/2203.13203
We are excited to explore further how other AI models, like LLMs, can be trained more efficiently to achieve substantial reductions in AI-related carbon emissions.