Do We Need Exotic Models? Engineering Metrics to Enable Green Machine Learning from Tackling Accuracy-Energy Trade-offs

Abstract

Machine learning (ML) has been shown to bypass key limitations of traditional methods (i.e., physical testing and numerical simulations) and hence presents itself as an attractive technology in engineering. While the integration of ML brings exciting opportunities, it also introduces unique challenges. One such challenge is related to the heavy reliance of ML on large datasets and computing facilities – for algorithm development, training, deployment, and storage. To realize Green ML (GML), this paper argues that ML users are to be cognizant of the hidden costs of energy consumption and subsequent carbon emissions arising from ML modeling, and hence they are ethically bound to apply ML responsibly. In this pursuit, a series of simple and exotic ML algorithms are examined, and their performance on a relatively large dataset (∼8000 observations) is documented on five fronts; predictive performance, model size, training time, energy consumption, and generatedcarbon emissions. In addition, this work also examines the influence of algorithmic architecture, processing language, number of features, as well as dataset size on model predictivity and energy consumption. Findings from this investigation infer that a 23–99% reduction in energy consumption and carbon emissions can be attained (while maintaining a comparable level of accuracy) by adopting simple as opposed to exotic ML models. The same findings have also led to the development of two new metrics that can tie the predictivity (i.e., level of accuracy) to the amount of energy consumed per algorithm. These metrics can be used to compare model performance in a similar manner to that traditionally used to assess the accuracy of ML predictions, thereby integrating energy-based awareness as a key dimension for model comparison.