Pareto-Optimal Trade-off Analysis of Accuracy vs. Energy in Sustainable Large-scale Deep Learning

Abstract

The rapid growth of machine learning and deep neural networks has led to unprecedented computational demands, raising concerns over energy consumption and environmental impact. This paper provides a comprehensive review of methods and best practices for designing, training, and deploying energy-efficient neural models in both natural language processing (NLP) and computer vision applications. We examine model optimization strategies, including pruning, quantization, knowledge distillation, and efficient architecture design (MobileNets, EfficientNet, DistilBERT), as well as the impact of scaling laws on computational cost and performance. The study further highlights the importance of energy and carbon footprint reporting in research, advocating systematic tracking via tools such as Weights & Biases and standardized preprints. Empirical insights from recent studies demonstrate the trade-offs between accuracy, model size, and energy efficiency, while emphasizing the role of edge and low-resource deployment in sustainable AI. Finally, we propose practical recommendations for adopting green AI practices in both academia and industry, including renewable-energy-aware scheduling, energy-profiling automation, and curriculum integration for sustainable AI education. This synthesis aims to inform researchers, practitioners, and policymakers on balancing model performance with environmental responsibility, contributing to the broader goal of sustainable AI development.