Scholarlink

<p>Label noise and class imbalance are two types of data bias that have attracted widespread attention in the past, but few methods can address both of them simultaneously. Recently, some works have begun to explore handling the two biases concurrently. In this article, we combine feature-level sample selection with logit-level knowledge distillation and logit adjustment to form a more complete collaborative training framework using two neural networks, which is termed <b>D</b>ynamic <b>N</b>oise and <b>I</b>mbalance <b>W</b>eighted <b>D</b>istillation (DNIWD). Firstly, we construct two types of sample sets, which are dynamic high-confidence set and basic confidence set. Based on the former, we estimate the centroids for each class in the latent space and select clean and easy examples for the peer network based on the uncertainty. Secondly, based on the latter, we perform knowledge distillation between the existing two networks to facilitate the learning of all classes, letting the network adaptively adjust the weight of distillation loss based on its own outputs. Meanwhile, we add an auxiliary classifier to each network and apply an improved balanced loss to train it, in order to boost the generalization performance of tail classes in more severe cases of class imbalance and provide balanced predictions for constructing confidence sample sets. Compared to state-of-the-art methods, <b>DNIWD</b> achieves significant improvement on synthetic and real-world datasets.</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

Mitigating noisy labels in long-tailed image classification via multi-level collaborative learning

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Xinyang Zhou,
  • Zhijie Wen,
  • Yuandi Zhao,
  • Jun Shi,
  • Shihui Ying

摘要

Label noise and class imbalance are two types of data bias that have attracted widespread attention in the past, but few methods can address both of them simultaneously. Recently, some works have begun to explore handling the two biases concurrently. In this article, we combine feature-level sample selection with logit-level knowledge distillation and logit adjustment to form a more complete collaborative training framework using two neural networks, which is termed Dynamic Noise and Imbalance Weighted Distillation (DNIWD). Firstly, we construct two types of sample sets, which are dynamic high-confidence set and basic confidence set. Based on the former, we estimate the centroids for each class in the latent space and select clean and easy examples for the peer network based on the uncertainty. Secondly, based on the latter, we perform knowledge distillation between the existing two networks to facilitate the learning of all classes, letting the network adaptively adjust the weight of distillation loss based on its own outputs. Meanwhile, we add an auxiliary classifier to each network and apply an improved balanced loss to train it, in order to boost the generalization performance of tail classes in more severe cases of class imbalance and provide balanced predictions for constructing confidence sample sets. Compared to state-of-the-art methods, DNIWD achieves significant improvement on synthetic and real-world datasets.