Publication

Abstract : Understanding the carbon footprint is crucial for sustainable AI development. This work presents a framework called “carbon aware optimistic machine learning recommendation framework” (CAOMLF) that aims to balance machine learning performance with environmental sustainability by analyzing carbon emissions and energy consumption. This framework applies a multi-objective approach to balance between the predictive performance and corresponding environmental impact of carbon footprint using a trade-off score. This work evaluates basic machine learning tasks (classification, regression, and clustering) across various data types (text, image, and tabular) in a variety of scenarios and gives recommendations based on the trade-off score. For each of these tasks, a suite of five distinct model architectures have been systematically evaluated. This proposed framework consists of a training phase and a recommendation phase. In the training phase, the system captures data including accuracy and carbon emission for machine learning tasks with varying size of datasets in dimension. In the recommendation phase, the train once approach is used to suggest a better machine learning model with optimal hyperparameters for balancing performance and environmental impact.