Scholarlink

<p>Training deep neural networks (DNNs) is increasingly recognized as a major contributor to the energy footprint of Artificial Intelligence (AI). Existing dataset pruning (DP) and active learning (AL) techniques reduce training data volumes but often introduce costly computations that undermine their energy-saving potential. This paper introduces <i>Play it Straight</i> and its enhanced variant <i>Re-Play it Straight</i>, two adaptive training algorithms that combine random subset sampling with lightweight AL-inspired instance selection. The proposed framework achieves a better balance between accuracy and energy efficiency by incrementally fine-tuning models on small, informative subsets, while controlling computational overhead. Experiments on multiple benchmark datasets demonstrate substantial reductions in training energy compared to state-of-the-art DP and AL methods, with <i>Re-Play it Straight</i> consistently delivering superior performance. These results highlight the potential of our approach to support more sustainable deep learning practices, contributing to the broader goals of Green AI.</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

An efficient model training framework for green AI

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月11日

查看全文
  • Francesco Scala,
  • Sergio Flesca,
  • Luigi Pontieri

摘要

Training deep neural networks (DNNs) is increasingly recognized as a major contributor to the energy footprint of Artificial Intelligence (AI). Existing dataset pruning (DP) and active learning (AL) techniques reduce training data volumes but often introduce costly computations that undermine their energy-saving potential. This paper introduces Play it Straight and its enhanced variant Re-Play it Straight, two adaptive training algorithms that combine random subset sampling with lightweight AL-inspired instance selection. The proposed framework achieves a better balance between accuracy and energy efficiency by incrementally fine-tuning models on small, informative subsets, while controlling computational overhead. Experiments on multiple benchmark datasets demonstrate substantial reductions in training energy compared to state-of-the-art DP and AL methods, with Re-Play it Straight consistently delivering superior performance. These results highlight the potential of our approach to support more sustainable deep learning practices, contributing to the broader goals of Green AI.