Exploring the Accuracy - Energy Trade-off in Machine Learning

Citation

Brownlee A, Adair J, Haraldsson S & Jabbo J (2021) Exploring the Accuracy - Energy Trade-off in Machine Learning. In: 2021 IEEE/ACM International Workshop on Genetic Improvement (GI). Genetic Improvement Workshop at 43rd International Conference on Software Engineering, Madrid, Spain, 30.05.2021-30.05.2021. Piscataway, NJ: IEEE. https://doi.org/10.1109/GI52543.2021.00011

Abstract
Machine learning accounts for considerable global electricity demand and resulting environmental impact, as training a large deep-learning model produces 284 000kgs of the greenhouse gas carbon dioxide. In recent years, search-based approaches have begun to explore improving software to consume less energy. Machine learning is a particularly strong candidate for this because it is possible to trade off functionality (accu-racy) against energy consumption, whereas with many programs functionality is simply a pass-or-fail constraint. We use a grid search to explore hyperparameter configurations for a multilayer perceptron on five classification data sets, considering trade-offs of classification accuracy against training or inference energy. On one data set, we show that 77% of energy consumption for inference can saved by reducing accuracy from 94.3% to 93.2%. Energy for training can also be reduced by 30-50% with minimal loss of accuracy. We also find that structural parameters like hidden layer size is a major driver of the energy-accuracy trade-off, but there is some evidence that non-structural hyperparameters influence the trade-off too. We also show that a search-based approach has the potential to identify these trade-offs more efficiently than the grid search.