现代信息科技2025，Vol.9Issue(14)：42-48,7.DOI:10.19850/j.cnki.2096-4706.2025.14.009
基于多层次对比学习的细粒度表示学习
Fine-grained Representation Learning Based on Multi-level Contrast Learning
摘要
Abstract
In recent years,advances in unsupervised representation learning have often relied on a predetermined number of classes to enhance feature extraction and clustering performance.However,this has sparked a debate regarding two core issues.The first issue is the necessity of predefining the number of classes and the second issue is whether class labels can adequately capture the fine-grained semantic features within the data.In real scenarios,the true number of classes in a dataset is often unknown,and the limitation casts doubt on the effectiveness of class labels in representation learning.To address these issues,this paper proposes a Contrastive Disentangling(CD)framework that learns data representations without relying on category priors.The framework employs a multi-level Contrastive Learning strategy,including an instance-level contrastive loss that distinguishes the representations of different samples,a feature-level contrastive loss that enhances the independence of outputs from different feature extraction heads,and a normalized entropy loss that ensures feature diversity and prevents representation collapse.Extensive experimental results on the CIFAR-10,CIFAR-100,STL-10,and ImageNet-10 datasets demonstrate that CD outperforms existing methods when category information is missing or ambiguous.关键词
表征解耦/聚类/对比学习/表征学习/无监督学习Key words
Representation Disentangling/clustering/Contrastive Learning/Representation Learning/Unsupervised Learning分类
信息技术与安全科学引用本文复制引用
江厚望,詹仕华..基于多层次对比学习的细粒度表示学习[J].现代信息科技,2025,9(14):42-48,7.基金项目
福建省自然科学基金项目(2021J01129) （2021J01129）
福建省高等学校教育技术研究会基金项目(H2000134A) （H2000134A）
福建农林大学横向科技创新基金项目(KHF190015) （KHF190015）
