Deep learning has facilitated a host of recent advancements within automation and across computer science, and offers significant potential in delivering further computational innovations across science and industry. These cutting-edge systems often use advanced techniques such as deep neural networks to improve classification and prediction accuracy; however, much of this work strives for competitive performance on global leaderboards without considering the consequences of this persuit further afield. For this reason, today’s deep models are incredibly computationally intensive, consuming considerable energy budgets to store and process vast parameter and data sets, inflicting a significant (but largely hidden) environmental, financial, and social cost. The subsequent review consolidates these impacts, describing the current issues raised by the energy-intensive nature of modern deep learning models and their training algorithms. Recent work within the field of Green AI is then explored, and proposed hardware and software solutions aiming to minimise the computational and energetic cost of deep learning training algorithms are examined in an attempt to demonstrate how the energy efficiency and sustainability of advanced models can be improved.
Keywords: Deep learning, energy efficiency, data efficiency, green AI, sustainable deep learning