Energy-Aware Model Training: A Power-Efficient Iterative Learning Approach

The increasing computational demands of machine learning models have raised concerns regarding energy efficiency, particularly in large-scale AI deployments, cloud environments, and edge computing. Traditional training algorithms focus on accuracy optimization while neglecting energy consumption, leading to high power usage and operational costs. This paper introduces the Power-Efficient Iterative Learning (PEIL) framework, which integrates power monitoring into the training process to reduce energy consumption while maintaining competitive accuracy. PEIL incorporates an energy-aware regularization term into the loss function and dynamically adjusts learning rates based on real-time power measurements, balancing computational efficiency and power optimization. To evaluate its effectiveness, PEIL is benchmarked against a baseline supervised learning model trained on the MNIST dataset using stochastic gradient descent. Experimental results demonstrate that PEIL significantly reduces energy consumption and power usage without compromising classification performance. The most crucial advantage of PEIL is its ability to decrease total energy usage by up to 63%, making it an excellent choice for power-aware machine learning applications. These findings highlight the importance of power-efficient AI training methodologies, paving the way for more sustainable and scalable deep learning applications.