Sustainable Machine Learning: Evaluating the Environmental Cost of AutoML Algorithms in AI Development
Marc Schmitt · 2024
This study evaluates the carbon footprint (CF) of Automated Machine Learning (AutoML) algorithms in AI development, examining three datasets to assess emissions across various run-times and countries. It is shown that the carbon intensity (CI) of these systems is significantly influenced by the energy sources powering the computational infrastructure. A correlation between run-time and model accuracy is observed, showing diminishing returns in accuracy with increased run-time and its environmental cost. The findings highlight the crucial role of geographic location and regional energy mix in determining the carbon footprint of AI operations. In areas with low-carbon or renewable energy sources, AI systems exhibit a reduced carbon footprint, underscoring the importance of infrastructural and environmental context in AI’s ecological impact. This study calls for adopting energy-efficient locations and optimizing ML algorithms to achieve a balance between model accuracy and environmental costs.