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

KALAHash: Knowledge-Anchored Low-Resource Adaptation for Deep Hashing

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 1 citations (OpenAlex)

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.

Introduces low-resource adaptation for deep hashing and proposes KALAHash, combining CLoRA (LoRA whose down-projection anchors are class-name text embeddings from CLIP) with KIDDO (text-anchored alignment loss plus discrete cyclic coordinate descent for binary codes) to adapt CLIP ViT-B/32 with 1-8 shots per class.

Problem

Deep hashing for approximate nearest neighbor search degrades severely when a pre-trained model must be adapted with only a few labeled samples per class, due to distribution shift and embedding collapse.

Approach

Replace the random LoRA down-projection vectors with class-level textual knowledge embeddings (dynamically selected top-r per input via cosine similarity) so the low-rank update is anchored in class semantics, and add a text-to-code alignment loss optimized with discrete cyclic coordinate descent.

Key results

• On CIFAR-10 1-shot, KALAHash reaches 57.54 mAP versus 46.66 for the best listed baseline CSQ and 39.53 for DCH, with 8.91%-18.01% gains over baselines.
• Plugging CLoRA into eight deep hashing baselines improves mAP by 1.01%-4.18% on NUS-WIDE, 0.93%-2.80% on MS-COCO, and 0.96%-12.52% on CIFAR-10.
• Ablations show removing CLoRA drops CIFAR-10 mAP by 11.16 points and removing KIDDO drops it by 6.65 points; CLoRA adds only about 0.01-0.05 ms of inference overhead.

Measured effects reported in the paper

Surprising findings

• Full fine-tuning of the backbone collapses the embedding space in the 1-shot setting (Silhouette Score 50.0%, mAP 14.7%), performing worse than locking the backbone.
• Under extremely low resource settings, the state-of-the-art full-data hashing methods are not competitive, and the authors note that MDSH with full fine-tuning on full CIFAR-10 also suffers distribution shift.

Limitations

• Performance remains far below full-data SOTA in some settings; the method is not shown to be competitive when many shots are available without further tuning.
• Baseline comparisons rely on a locked-backbone protocol that the authors acknowledge may limit baseline performance.

Experimental setup

Datasets: NUS-WIDE MS-COCO CIFAR-10 ImageNet-100 CUB-200

Benchmarks: NUS-WIDE MS-COCO CIFAR-10 ImageNet-100 CUB-200

Metrics: mAP@K Silhouette Score inference time

Baselines: HashNet DSDH DCH GreedyHash CSQ OrthoHash HSWD MDSH LoRA Prompt Tuning Bias Tuning UGH

Builds on: LoRA CLIP DSDH discrete cyclic coordinate descent MDSH

Closest papers by meaning

• DistillHash: Unsupervised Deep Hashing by Distilling Data Pairs (CVPR 2019 · similarity 0.88)
• Robust Domain Adaptive Hashing via Structural Noise Modeling and Correction (AAAI 2026 · similarity 0.88)
• Effective Comparative Prototype Hashing for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.88)
• Deep Unsupervised Hashing via External Guidance (ICML 2025 · similarity 0.87)
• Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization (ICML 2025 · similarity 0.86)
• PromptHash:Affinity-Prompted Collaborative Cross-Modal Learning for Adaptive Hashing Retrieval (CVPR 2025 · similarity 0.85)