Energy-aware and dynamic training of deep neural networks (EADTrain) for sustainable AI

Abstract

The growing complexity of deep neural networks, particularly in the domain of computer vision, has led to increasing concerns regarding their energy consumption and environmental impact. To tackle these issues, we propose EADTrain, an innovative training framework that emphasizes energy-conscious learning. EADTrain integrates live energy monitoring within the training cycle, enabling dynamic adjustments to data augmentation strategies and selective fine-tuning based on ongoing energy consumption and model performance feedback. This responsive training mechanism helps achieve an optimal trade-off between computational efficiency and predictive accuracy. We assess EADTrain across several visual recognition tasks using benchmark datasets including CIFAR-10, ImageNet, and a custom satellite imagery dataset. The experimental results show that EADTrain reduces energy usage by up to 35% compared to leading methods, without compromising classification accuracy or F1-score. These findings position EADTrain as a scalable and environmentally efficient framework for training deep learning models in energy-constrained settings.