Multi-Granularity Alignment-aware Contrastive learning for open-world semi-supervised learning

doi:10.1016/j.patcog.2026.114946

Multi-Granularity Alignment-aware Contrastive learning for open-world semi-supervised learning

2026-09-25

0

PRE

AI

DOI：10.1016/j.patcog.2026.114946

Original

Original request for help

Save

Abstract

En 中文

• A novel open-world SSL framework MGrAC is proposed to synergistically integrate pseudo-labeling and contrastive learning through multi-granularity alignment. • Multi-Granularity Classification constructs hierarchical prototypes to mitigate the known-class overfitting and enabling the structured representation learning. • Weakly Supervised Contrastive Learning guided by multi-level pseudo-labels enhances the feature discriminability without the semantic misalignment. • Target Granularity Alignment enforces cross-level consistency to resolve optimization conflicts and ensure coherent hierarchical semantics. • Experiments show that MGrAC achieves the state-of-the-art performance, and demonstrates the robustness in both generic and fine-grained open-world scenarios.

Keywords:

Semi-supervised learning

Open-world semi-supervised learning

Novel class discovery

Multi-granularity learning

Multi-Granularity Alignment-aware Contrastive learning for open-world semi-supervised learning

Abstract

Journal

Organization

Cited Papers