Scholarlink

<p>Learning robust feature representations from long-tail distributed data is essential. Recently, contrastive learning has shown impressive progress in addressing long-tail learning challenges. While contrastive learning aims to optimize the lower bound of mutual information between feature distribution and label distribution, the previous approaches often substantially rely on less accurate and unrealistic assumptions about model distribution and overlook the long-tail nature of the instance space. Consequently, these methods fail to achieve a sufficiently tight lower bound. To address these concerns, we first propose a loss function derived from mini-Batch instance Features and Class Prototypes to construct a Conditional Gaussian mixture distribution (CGM-BF-CP), and prove its generalization ability from the perspective of generalization error upper bound. Then we create high quality and high relevance KNN graph to model relation between features. And propose a corresponding loss function, i.e., Graph based Contrast Learning Loss (GCLL). The feature information can be transferred between classes through this graph, so that the tail class features can be better learned. The experimental results on Cifar10/100-LT and ImageNet-LT show that our proposed model is competitive with the latest state-of-the-art methods. Our code is available at <a href="https://github.com/error030/CGM-BP-CP/tree/main">https://github.com/error030/CGM-BP-CP/tree/main</a>.</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

Long-tailed classification by efficient contrast learning with high quality and high relevance latent features

  • 期刊论文

  • 发表日期：

    2025年2月3日

查看全文
  • Hong-li Yuan,
  • Jian-wei Liu

摘要

Learning robust feature representations from long-tail distributed data is essential. Recently, contrastive learning has shown impressive progress in addressing long-tail learning challenges. While contrastive learning aims to optimize the lower bound of mutual information between feature distribution and label distribution, the previous approaches often substantially rely on less accurate and unrealistic assumptions about model distribution and overlook the long-tail nature of the instance space. Consequently, these methods fail to achieve a sufficiently tight lower bound. To address these concerns, we first propose a loss function derived from mini-Batch instance Features and Class Prototypes to construct a Conditional Gaussian mixture distribution (CGM-BF-CP), and prove its generalization ability from the perspective of generalization error upper bound. Then we create high quality and high relevance KNN graph to model relation between features. And propose a corresponding loss function, i.e., Graph based Contrast Learning Loss (GCLL). The feature information can be transferred between classes through this graph, so that the tail class features can be better learned. The experimental results on Cifar10/100-LT and ImageNet-LT show that our proposed model is competitive with the latest state-of-the-art methods. Our code is available at https://github.com/error030/CGM-BP-CP/tree/main.