Source-linked AI summary
DistillHash: Unsupervised Deep Hashing by Distilling Data Pairs
Erkun Yang, Tongliang Liu, Cheng Deng, Wei Liu, Dacheng Tao
TL;DR
Unsupervised deep hashing is limited by unreliable semantic similarity signals, despite the efficiency of hashing for large-scale search. DistillHash distills confident data pairs from noisy local-structure labels using Bayes-optimal analysis and learns hash functions through Bayesian learning; it consistently outperforms state-of-the-art methods on three benchmark datasets.
Problem
Unsupervised deep hashing lacks reliable supervisory similarity signals, limiting its ability to achieve satisfactory search performance.
Method
DistillHash analyzes noisy local-structure similarity labels against Bayes-optimal labels, automatically distills confident data pairs, and learns representations and hash codes with a Bayesian framework.
Results
DistillHash consistently obtains the best MAP results across hash lengths on FLICKR25K, NUSWIDE, and CIFAR10, improving average MAP over SSDH by 3.08%, 4.01%, and 2.86%, respectively.
Takeaways & Limitations
Distilling data pairs with more accurate similarity relationships supports state-of-the-art unsupervised deep hashing performance across three benchmark datasets.
Takeaways & Limitations
Initial similarity labels constructed from local structures are unreliable and may contain label noise; the theoretical selection also assumes locally invariant flip rates.
Abstract
from arXiv · showhide
Due to the high storage and search efficiency, hashing has become prevalent for large-scale similarity search. Particularly, deep hashing methods have greatly improved the search performance under supervised scenarios. In contrast, unsupervised deep hashing models can hardly achieve satisfactory performance due to the lack of reliable supervisory similarity signals. To address this issue, we propose a novel deep unsupervised hashing model, dubbed DistillHash, which can learn a distilled data set consisted of data pairs, which have confidence similarity signals. Specifically, we investigate the relationship between the initial noisy similarity signals learned from local structures and the semantic similarity labels assigned by a Bayes optimal classifier. We show that under a mild assumption, some data pairs, of which labels are consistent with those assigned by the Bayes optimal classifier, can be potentially distilled. Inspired by this fact, we design a simple yet effective strategy to distill data pairs automatically and further adopt a Bayesian learning framework to learn hash functions from the distilled data set. Extensive experimental results on three widely used benchmark datasets show that the proposed DistillHash consistently accomplishes the state-of-the-art search performance.
1. Introduction
The paper addresses unreliable supervisory signals in unsupervised deep hashing by distilling confident data pairs from noisy local-structure similarities, then learns hash functions from them. DistillHash is reported to outperform current state-of-the-art unsupervised hashing methods on three benchmark datasets.
- Unsupervised deep hashing lacks reliable semantic similarity signals because quantization and reconstruction objectives do not exploit similarities between data points.
- DistillHash assigns initial pairwise similarity labels from local data structures and treats them as noisy because they may not represent semantic similarities.
- Under a mild assumption, pairs whose noisy labels agree with Bayes-optimal semantic labels can be distilled.
- The method provides criteria for selecting distilled pairs and automatically collects them for hash learning.
- Experiments on three popular benchmark datasets show that DistillHash outperforms current state-of-the-art unsupervised hashing methods.
2. Related Work
Prior work spans supervised and unsupervised hashing, including deep models, while noisy-label research provides related strategies for handling corrupted supervision. DistillHash connects these areas by treating initial similarity relationships as noisy labels.
- Hashing methods are commonly divided into supervised models that use semantic labels and unsupervised models trained without supervised information.
- Supervised hashing methods preserve semantic similarity in Hamming space using approaches such as inner-product approximation, label regression, and pairwise optimization.
- Unsupervised methods learn binary codes from unlabeled data through techniques including PCA with iterative quantization and discrete graph optimization.
- Unsupervised deep hashing uses neural architectures for feature extraction and hash mapping, including auto-encoding and transformation-based similarity schemes.
- Noisy-label research distinguishes label-noise-tolerant classification from noise-cleansing methods that filter corrupted labels using training-sample information.
- DistillHash treats initial similarity relationships as noisy labels and models their relationship to Bayes-optimal classifier labels.
3. Approach
DistillHash addresses unsupervised hashing by estimating noisy pairwise similarity from local structure, identifying reliable distilled pairs under a mild assumption, and learning hash functions from them with Bayesian learning.
- DistillHash relates noisy-label probabilities to Bayes-optimal semantic labels to determine when reliable pair labels can be recovered.The Bayes classifier assigns similarity according to whether η(xi, xj) is at least 0.5, while unsupervised learning estimates the noisy probability ˜η.
- Unsupervised similarity labels are initialized from feature distances using thresholds, but local-structure labels may contain noise.Pairs are labeled similar when d(i, j) ≤ t1 and dissimilar when d(i, j) > t2.
- 3.1. Collecting Distilled Data Pairs Automatically: The assumption limits label noise locally, and experiments on three benchmark datasets empirically verify that it applies well to the constructed noisy labels.The paper describes this assumption as weaker than requiring the flip rate to be bounded by 0.5.
- 3.1. Collecting Distilled Data Pairs Automatically: Under Assumption 1, the method bounds label-flip rates using local neighborhoods and selects pairs whose estimated noisy probabilities cross reliability thresholds.Selected pairs receive Sij = +1 or −1 according to upper-bound-adjusted thresholds, forming the distilled data-pair set.
- 3.2. Bayesian Learning Framework: After distillation, Bayesian learning estimates hash codes from pairwise similarity likelihoods and also estimates the noisy conditional probability.A pairwise logistic model uses hash-code inner products, which reflect Hamming distance for binary codes.
4. Experiments
Experiments on FLICKR25K, NUSWIDE, and CIFAR10 compare DistillHash with unsupervised hashing baselines using MAP, TopN-precision, and precision-recall. DistillHash consistently achieves the strongest reported retrieval performance, while its distilled-pair variant improves over training on initial similarity labels.
- MAP results: DistillHash consistently obtains the best MAP results across hash bit lengths on all three datasets.Compared with ITQ, average MAP improves by 6.89%, 13.97%, and 7.73% on FLICKR25K, NUSWIDE, and CIFAR10; compared with SSDH, improvements are 3.08%, 4.01%, and 2.86%.
- Retrieval curves: DistillHash achieves the best TopN-precision and precision-recall performance among the evaluated approaches on the three datasets with 16- and 32-bit codes.These results cover both Hamming ranking-based evaluation and hash lookup evaluation.
- Training analysis: DistillHash converges within 1,000 iterations in all reported cases, and its performance is not sensitive to hyper-parameter o over [2, 20].The experiments set o to 4 after the sensitivity analysis.
- Ablation study: DistillHash improves over DistillHash* trained on initial similarity labels by 2.11%–4.23% across datasets and hash lengths.The comparison isolates distilled data-pair learning because the two variants share the same Bayesian learning framework.
5. Conclusions
DistillHash is a new unsupervised deep hashing approach that combines theoretical analysis of noisy local-structure labels, automatic data-pair distillation, and Bayesian hash learning. Experiments on three benchmark datasets show that it surpasses other state-of-the-art methods.
- DistillHash investigates how noisy labels from local structures relate to Bayes optimal classifier labels.
- The method provides a simple scheme to automatically distill data pairs with potentially reliable similarity labels.
- DistillHash uses a Bayesian learning framework to learn hash codes from the distilled data set.
- Experiments on three benchmark datasets show that DistillHash surpasses other state-of-the-art methods.