Focus on Carbon Dioxide Footprint of AI/ML Model Training

Abstract

In the last decade, AI/ML has accomplished impressive achievements in various fields. Although this increasing usage of AI/ML systems eases our daily life in different perspectives, it may have side effects. The latter includes the various consequences of high computational cost for training and testing such models. In some cases, these computations leave a considerable amount of carbon footprint without significantly improving the model accuracy over consecutive training iterations. As a result, negative impacts of AI/ML model training/testing should be taken into consideration when developing such systems. In this paper, we analyse the trade-off between carbon foot-print and AI/ML model accuracy using four different model classes: CNN, RNN, GRU and LSTM. Our results show that the accuracy achieved by training the model is not in line with the amount of carbon dioxide emitted. Besides, the carbon footprint patterns cannot be generalized for the four classes of tested models. We observe that the model with the best training accuracy does not generate the highest carbon footprint. We conclude with guidelines for training efficient AI/ML systems that also incorporate carbon footprint emission for a sustainable approach.

Ähnliche Arbeiten