发表机构
Universidade da Coruña; CITIC（科鲁尼亚大学; 中信）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
针对深度学习模型部署难题，提出轻量级训练策略，解耦特征提取与分类器优化，预计算特征减少开销，重新设计分类器头。在多种架构和数据集上评估，显著减少训练时间，兼顾精度，还能大幅减少二氧化碳排放，提供可持续方案。
AI 中文摘要
深度学习模型在图像分类方面取得了领先成果，但因计算成本和能源需求面临部署挑战。我们提出一种轻量级训练策略，使模型的归一化层适应新领域，并将特征提取与分类器优化解耦，通过仅预计算一次特征来减少开销。重新设计的带有基于边缘加权损失的分类器头，无需端到端反向传播即可进一步减少模糊性。在四种卷积神经网络架构（ResNet18、ResNet50、MobileNet、DenseNet121）、三种Transformer模型（ViT、Swin和DeiT）以及三个医学数据集（脑癌MRI、BreakHis和PatchCamelyon）上进行评估，我们的方法显著减少了所需的训练时间，仅在精度上有微小权衡，常常匹配或超越基线性能。这种效率转化为将二氧化碳排放量减少几个数量级，为资源受限的临床或原型制作环境提供了实用且环境可持续的解决方案。
英文摘要
Deep learning models achieve state-of-the-art image classification but face deployment challenges due to computational costs and energy demands. We propose a lightweight training strategy that adapts normalization layers of the model to the new domain and decouples feature extraction from classifier optimization, reducing overhead by precomputing features only once. A redesigned classifier head with margin-based weighted loss further minimizes ambiguity without end-to-end backpropagation. Evaluated across four CNN architectures (ResNet18, ResNet50, MobileNet, DenseNet121), three Transformer models (ViT, Swin and DeiT) and three medical datasets (Brain Cancer MRI, BreakHis and PatchCamelyon), our approach significantly reduces the required training time with only a marginal accuracy trade-off, often matching or surpassing baseline performance. This efficiency translates to reducing CO2 by orders of magnitude, offering a practical and environmentally sustainable solution for resource-constrained clinical or prototyping environments.
