Introduction
Efficient image retrieval from large-scale datasets relies on hashing, which maps high-dimensional data into compact binary codes. While unsupervised contrastive learning has advanced this field, current methods often overlook the inherent hierarchical structure of semantic data and the geometric limitations of Euclidean space. Hyperbolic Hierarchical Contrastive Hashing (HHCH) addresses these gaps by leveraging hyperbolic geometry and a novel hierarchical learning framework to produce more discriminative hash codes.
The Hierarchical Semantic Challenge
Standard contrastive learning treats all non-identical images in a batch as equally negative. However, real-world semantics are naturally organized into hierarchies. For instance, while a "Husky" and a "Poodle" are distinct entities, they share a "Dog" ancestor. Treating them as unrelated negatives can confuse the model and lead to suboptimal feature representations.
Furthermore, Euclidean space is mathematically ill-suited for representing such tree-like structures due to its polynomial volume growth. In contrast, hyperbolic space—specifically the Poincaré ball model—exhibits exponential volume growth, allowing for the embedding of complex hierarchies with minimal distortion.
The HHCH Framework
The HHCH method introduces three primary innovations to bridge the gap between geometry and semantics:
- Hyperbolic Embedding: Data is projected into a Poincaré ball instead of Euclidean space. This provides the necessary "room" at the manifold's edges to preserve hierarchical relationships in low dimensions.
- Hierarchical Hyperbolic K-Means: The authors developed a bottom-up clustering approach adapted for hyperbolic space using Einstein midpoints. This discovers latent "taxons" or prototypes at multiple levels of granularity (coarse to fine).
- Hierarchical Contrastive Learning: The model employs a dual-level strategy. At the instance level, it uses the discovered hierarchy to avoid selecting negatives that share common ancestors, preventing "false negative" confusion. At the prototype level, it aligns image embeddings with their respective cluster centers across the hierarchy, ensuring hash codes capture both fine-grained and coarse semantic features.
Experimental Evaluation
The authors benchmarked HHCH against state-of-the-art (SOTA) methods on four standard datasets: ImageNet, CIFAR-10, NUS-WIDE, and FLICKR25K.
As shown in the results, HHCH consistently outperforms existing unsupervised hashing techniques. On the ImageNet dataset, HHCH achieved a Mean Average Precision (mAP) of 0.826 at 64 bits, representing a significant 4.5% improvement over previous benchmarks. The method also demonstrates superior performance at lower bit counts (e.g., 16 bits), highlighting the efficiency of hyperbolic representations for capturing dense semantic information.
Conclusion and Future Directions
HHCH demonstrates that geometric priors and hierarchical awareness are crucial for effective unsupervised hashing. By moving beyond Euclidean constraints, the model produces more semantically consistent hash codes. While the hierarchical K-Means step introduces some computational overhead, the gains in retrieval accuracy and feature disentanglement are substantial. Future work may extend this hyperbolic approach to multi-modal retrieval and even more complex hierarchical construction schemes.
Summary updated