Scholarlink

<p>The exponential growth in deep learning model complexity, particularly in computer vision, has raised pressing concerns about energy consumption and environmental sustainability. This paper introduces the energy-conscious adaptive framework (ECAF), a novel approach to optimizing the trade-off between model performance and energy efficiency during training. ECAF integrates real-time energy profiling into the training pipeline, dynamically adjusting data augmentation intensity and selective fine-tuning strategies based on instantaneous energy metrics and model performance feedback. The proposed framework employs an adaptive optimization algorithm that prioritizes energy efficiency while maintaining or improving model accuracy and F1-scores. Extensive experiments conducted on benchmark datasets, including CIFAR-10, ImageNet, and a custom satellite imagery dataset, demonstrate that ECAF achieves significant reductions in energy consumption-up to 35%-compared to state-of-the-art techniques like AutoAugment and NASNet, without compromising predictive performance. ECAF’s contributions to sustainable AI practices underscore its potential for scaling energy-efficient deep learning across diverse domains.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient deep learning training: an energy-conscious adaptive framework (ECAF)

  • 期刊论文

  • 发表日期：

    2025年3月19日

查看全文
  • Pulkit Dwivedi,
  • Mansi Kajal

摘要

The exponential growth in deep learning model complexity, particularly in computer vision, has raised pressing concerns about energy consumption and environmental sustainability. This paper introduces the energy-conscious adaptive framework (ECAF), a novel approach to optimizing the trade-off between model performance and energy efficiency during training. ECAF integrates real-time energy profiling into the training pipeline, dynamically adjusting data augmentation intensity and selective fine-tuning strategies based on instantaneous energy metrics and model performance feedback. The proposed framework employs an adaptive optimization algorithm that prioritizes energy efficiency while maintaining or improving model accuracy and F1-scores. Extensive experiments conducted on benchmark datasets, including CIFAR-10, ImageNet, and a custom satellite imagery dataset, demonstrate that ECAF achieves significant reductions in energy consumption-up to 35%-compared to state-of-the-art techniques like AutoAugment and NASNet, without compromising predictive performance. ECAF’s contributions to sustainable AI practices underscore its potential for scaling energy-efficient deep learning across diverse domains.