arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2305.03378cs.CVcs.LG
收藏

面向长尾识别中的有效协作学习

Towards Effective Collaborative Learning in Long-Tailed Recognition

Tsinghua Shenzhen International Graduate School, Tsinghua University(清华大学深圳国际研究生院);International Digital Economy Academy(国际数字经济研究院)

Zhengzhuo Xu, Zenghao Chai, Chengyin Xu, Chun Yuan, Haiqin Yang

AI总结:

针对长尾识别中多专家协作学习的知识迁移类别不平衡问题,提出含重加权蒸馏损失、特征级蒸馏和对比代理任务的ECL框架,在四个数据集上达到SOTA性能。

摘要:

真实世界数据通常存在严重的类别不平衡和长尾分布问题,其中少数类的样本占比远低于多数类。现有研究多采用多专家架构来缓解模型在少数类上的不确定性,这类方法通过协作学习(即在线蒸馏)来聚合多个专家的知识。本文发现,专家之间的知识迁移在类别分布上是不平衡的,这导致少数类的性能提升十分有限。为解决该问题,我们通过比较分别受在线蒸馏监督和标签标注监督的两个分类器的预测结果,提出了一种重加权蒸馏损失。我们还指出,特征级蒸馏能显著提升模型性能并增强特征鲁棒性。最后,我们提出了有效协作学习(Effective Collaborative Learning, ECL)框架,该框架整合了一个对比代理任务分支以进一步提升特征质量。在四个标准数据集上的定量和定性实验表明,ECL取得了当前最优的性能,详细的消融实验也验证了ECL各组件的有效性。

英文摘要:

Real-world data usually suffers from severe class imbalance and long-tailed distributions, where minority classes are significantly underrepresented compared to the majority ones. Recent research prefers to utilize multi-expert architectures to mitigate the model uncertainty on the minority, where collaborative learning is employed to aggregate the knowledge of experts, i.e., online distillation. In this paper, we observe that the knowledge transfer between experts is imbalanced in terms of class distribution, which results in limited performance improvement of the minority classes. To address it, we propose a re-weighted distillation loss by comparing two classifiers' predictions, which are supervised by online distillation and label annotations, respectively. We also emphasize that feature-level distillation will significantly improve model performance and increase feature robustness. Finally, we propose an Effective Collaborative Learning (ECL) framework that integrates a contrastive proxy task branch to further improve feature quality. Quantitative and qualitative experiments on four standard datasets demonstrate that ECL achieves state-of-the-art performance and the detailed ablation studies manifest the effectiveness of each component in ECL.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑