Green AutoML

AutoML is a resource and time-consuming task resulting in a large environmental footprint. Therefore it is key to figure out smart ways to save resources and reduce the carbon emissions of AutoML, coined as Green AutoML. On one hand, one can make the AutoML method itself more efficient, e.g., using data compression, zero-cost AutoML, energy-aware objective functions, multi-fidelity evaluations, or intelligent early termination. On the other hand, one can search for energy-efficient pipelines, e.g., having a model size constraint, energy-efficient architectures, or model compression. Overall, an important question arises: what to do with the savings? Very straightforward, one could terminate early and therefore gift the saved time and resources somehow to the planet. On the other hand, one could use the (already budgeted) time to further search for better candidate pipelines. In any case, quantifying the environmental footprint is the foundation for any research within the field of Green AutoML. This non-trivial task can be approximated in terms of counting CPU / GPU hours, which can be easily obtained with the support of suitable tools.

AutoML for Sustainability

AutoML can also be used for sustainable applications, thereby searching for well-performing machine learning pipelines and also minimizing their carbon emissions. One example is the usage of AutoML for plastic waste detection [Theodorakopoulos et. al. 2023].