Exploring the Costs of Automation: A Comparative Study on the Energy Consumption and Performance of Open-Source Automated Machine and Deep Learning Libraries

<p dir="ltr"><i>Machine learning (ML) solutions have surged across various domains, requiring significant specialized expertise to execute. To make ML more accessible, automated machine learning (AutoML) and deep learning (AutoDL) have emerged, simplifying model development without extensive user intervention. While AutoML and DL offer more accessibility, their environmental impact given their increased computational demand, remains unexplored. This study addresses this gap by comparing the energy consumption and performance of five popular open-source AutoML libraries: FLAML, H2O AutoML, AutoGluon, TPOT, and AutoKeras. The experiment results of the study demonstrate a statistically significant difference with 99% confidence using the ANOVA and T-test methods in CPU and GPU energy consumption between these libraries. The study also provides discussion on the importance of considering both performance and sustainability when selecting libraries through a weighted scoring algorithm. The study's key findings show that Autogluon and FLAML offer a balanced approach, achieving good performance while minimizing energy consumption; H2O AutoML excels in model versatility; AutoKeras emphasizes performance over energy reduction; and TPOT excels more for tree-based algorithms, rather than general ML tasks. Future work may include investigating the impact of parameters like early stopping, training-test splits, and hyperparameter selection on energy consumption and exploring these libraries with various datasets to increase the generalizability of results.</i></p>