<p>The growing sophistication and resource-intensive nature of deep learning algorithms, especially in the realm of computer vision, have raised significant ecological concerns owing to the high energy demands involved in model training. To address this, we propose the efficient adaptive transformation learning with augmentation framework, which is specifically designed to enhance sustainability within deep learning pipelines. The framework incorporates ongoing power consumption tracking throughout the training process to responsively calibrate the strength of data augmentation and to selectively update layers depending on real-time feedback from both energy and model performance. This adaptive mechanism enables a well-regulated compromise between power efficiency, model precision, and training duration. The framework undergoes extensive benchmarking on three diverse datasets–CIFAR-10, ImageNet, and a tailored satellite imagery corpus–utilizing both legacy convolutional models (ResNet, VGG) and contemporary scalable backbones (ViT-B/16, ConvNeXt-T). EATL-A achieves top-tier accuracy scores of 94.5%, 88.8%, and 90.6%, respectively, surpassing current energy-aware benchmarks like PINN-DT, ssProp, and <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7610_Article_IEq1.gif" Format="GIF" Height="20" Rendition="HTML" Resolution="72" Type="Linedraw" Width="56" /> </InlineMediaObject> <EquationSource Format="TEX">$\hbox {TinyM}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>TinyM</mtext> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>Net-V3 across both performance and F1-score. It further reduces energy consumption by up to 27% and shortens training time by up to 40% relative to standard transfer learning. Additionally, by translating energy metrics into carbon-equivalent emissions, EATL-A is shown to yield the lowest <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7610_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="32" /> </InlineMediaObject> <EquationSource Format="TEX">$\hbox {CO}_{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>CO</mtext> <mn>2</mn> </msub> </math></EquationSource> </InlineEquation> output, up to 4.6 mg less than the next-best alternative on CIFAR-10. These results underscore EATL-A’s scalability, robustness, and environmental relevance, offering a practical and generalizable framework for sustainable AI across both cloud-based and edge computing platforms.</p>

Balancing performance and energy efficiency: the method for sustainable deep learning

期刊论文
发表日期：

2025年7月9日

查看全文

Pulkit Dwivedi,
Benazir Islam

摘要

The growing sophistication and resource-intensive nature of deep learning algorithms, especially in the realm of computer vision, have raised significant ecological concerns owing to the high energy demands involved in model training. To address this, we propose the efficient adaptive transformation learning with augmentation framework, which is specifically designed to enhance sustainability within deep learning pipelines. The framework incorporates ongoing power consumption tracking throughout the training process to responsively calibrate the strength of data augmentation and to selectively update layers depending on real-time feedback from both energy and model performance. This adaptive mechanism enables a well-regulated compromise between power efficiency, model precision, and training duration. The framework undergoes extensive benchmarking on three diverse datasets–CIFAR-10, ImageNet, and a tailored satellite imagery corpus–utilizing both legacy convolutional models (ResNet, VGG) and contemporary scalable backbones (ViT-B/16, ConvNeXt-T). EATL-A achieves top-tier accuracy scores of 94.5%, 88.8%, and 90.6%, respectively, surpassing current energy-aware benchmarks like PINN-DT, ssProp, and $\hbox {TinyM}^{2}$ ${TinyM}^{2}$ Net-V3 across both performance and F1-score. It further reduces energy consumption by up to 27% and shortens training time by up to 40% relative to standard transfer learning. Additionally, by translating energy metrics into carbon-equivalent emissions, EATL-A is shown to yield the lowest $\hbox {CO}_{2}$ ${CO}_{2}$ output, up to 4.6 mg less than the next-best alternative on CIFAR-10. These results underscore EATL-A’s scalability, robustness, and environmental relevance, offering a practical and generalizable framework for sustainable AI across both cloud-based and edge computing platforms.