Research map / Label-efficient and robust learning / Recognition and long-tail

Deep Unsupervised Hashing via External Guidance

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DUH-EG, an unsupervised image hashing method that selects representative WordNet nouns via CLIP text features and clustering, builds external features for images, and trains with a bidirectional contrastive loss aligning internal and external hash codes.

Problem

Deep unsupervised hashing for image retrieval relies only on internal visual structure, which limits discriminability due to semantic ambiguity; the paper asks how external textual knowledge can provide semantic guidance for hash learning.

Approach

Select diverse, representative nouns from WordNet via K-means on CLIP text embeddings, form per-image external features by softmax-weighted matching with CLIP image features, and use a bidirectional contrastive loss over augmented-view and external hash codes, with external-feature similarity used to treat latent semantic positives as positives instead of negatives.

Key results

• On CIFAR-10 with VGG-16, DUH-EG reports 0.887-0.888 MAP at 16/32 bits, and averages a 3.39% improvement over the strongest CNN-based baselines.
• With CLIP ViT-B/16 backbone, DUH-EG reaches 0.931-0.934 MAP on CIFAR-10 and 0.854-0.878 on MSCOCO, averaging 2.95% and 2.95% gains over top ViT-based methods on CIFAR-10 and MSCOCO respectively.
• Ablation shows external knowledge (EK) gives an average 11.37% gain over image-only augmentation on MSCOCO.

Measured effects reported in the paper

Surprising findings

• Ablation on MSCOCO shows the SIM (false-negative removal) module gives a 5.30% gain at 16 bits when used alone.

Limitations

• Performance depends on the distribution alignment between training data and external textual knowledge (domain shift risk, per impact statement).
• Hyperparameters such as T2 are tuned per dataset with narrow ranges, with best values varying across datasets (0.95-0.99).

Experimental setup

Datasets: CIFAR-10 NUS-WIDE Flickr25k MSCOCO WordNet

Benchmarks: CIFAR-10 NUS-WIDE Flickr25k MSCOCO

Metrics: MAP@1000 MAP@5000 precision-recall curves TopN-precision

Baselines: CIBHash DATE MeCoQ DSCH NSH CGHash DDCH HAMAN HiHPQ HARR LGH HHCH

Builds on: CIBHash DATE CLIP WordNet

Closest papers by meaning

• Deep Unsupervised Hashing with Latent Semantic Components (AAAI 2022 · similarity 0.89)
• Creating Something From Nothing: Unsupervised Knowledge Distillation for Cross-Modal Hashing (CVPR 2020 · similarity 0.89)
• Effective Comparative Prototype Hashing for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.88)
• An Asymmetric Augmented Self-Supervised Learning Method for Unsupervised Fine-Grained Image Hashing (CVPR 2024 · similarity 0.87)
• Vision-guided Text Mining for Unsupervised Cross-modal Hashing with Community Similarity Quantization (AAAI 2025 · similarity 0.87)
• KALAHash: Knowledge-Anchored Low-Resource Adaptation for Deep Hashing (AAAI 2025 · similarity 0.87)