Deep Learning Optimizers: A Sustainability Perspective on Energy and Emissions
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Optimizers play a pivotal role in the efficacy of deep neural network models, adjusting a model's parameters and reducing loss function. However, sustainability-related challenges emerge due to their associated computational costs and energy consumption. Despite numerous extant literature on the comparative accuracy of various deep learning optimizers, empirical research and theoretical guidance regarding their environmental footprint is still lacking. Thus, it is crucial to identify the optimal optimizer that maximizes accuracy while also prioritizing sustainability. This research, therefore, conducts rigorous comparative analyses of the energy consumption and associated environmental impacts of nine contemporary deep learning optimizers, with the goal of uncovering their ecological friendliness. Our research findings indicate that Stochastic Gradient Descent (SGD) markedly surpasses other optimizers in terms of energy efficiency. Additionally, commonly used optimizers such as Adam, Adagrad, and AdaMax emerge as viable choices in the construction of deep learning architectures. This research work will also provide invaluable insights for future endeavors in the conceptualization of energy-efficient deep learning implementations for sustainable AI development while reducing the ecological impact of advanced machine learning models.
Fields of Study & Subjects
Green IT and Sustainability
Stochastic Gradient Optimization Techniques
Energy Efficiency in Computing
Publication details
- Journal / source
- CRC Press eBooks
- Pages
- 167–179
- Publication date
- 27 March 2026
- Language
- EN
- Document type
- Book chapter