HCCL: Hierarchical Consistency-Driven Contrastive Learning for Semi-Supervised Learning

Abstract

While Semi-Supervised Learning (SSL) has witnessed substantial advancements, prevailing methods primarily operate at a single semantic granularity, overlooking the multi-scale semantic structures intrinsic to visual data. This limitation restricts the exploitation of rich feature constraints, often yielding representations that lack hierarchical coherence. To bridge this gap, we propose HCCL, a Hierarchical Consistency-driven Contrastive Learning framework. Unlike previous approaches that rely on fixed taxonomies, HCCL introduces an optimal transport-based online clustering mechanism to automatically discover latent coarse-grained and fine-grained semantic priors without external supervision. Specifically, we implement a resource-efficient prototype allocation strategy to capture salient intra-class variations and global semantic groupings. We then introduce a cross-granularity alignment module that enforces consistency between these discovered latent structures and the target task, thereby regularizing the feature space against noise. Furthermore, we construct adaptive affinity graphs based on these multi-granularity predictions to guide contrastive learning. Extensive experiments on CIFAR-10, CIFAR-100, STL-10, and Mini-ImageNet demonstrate that HCCL achieves highly competitive results against state-of-the-art competitors, particularly in challenging low-label regimes.