Source-linked AI summary
Deep Hashing Network for Unsupervised Domain Adaptation
Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, Sethuraman Panchanathan
TL;DR
Deep domain adaptation needs labeled source data to recognize unlabeled target data, while hashing addresses efficient storage and retrieval. The paper introduces Office-Home and DAH, which learns binary codes from labeled source and unlabeled target data using supervised hashing, target entropy, and MK-MMD alignment. Experiments report efficient hash codes that outperform competitive unsupervised domain-adaptation baselines, while some results show that increasing code length does not necessarily improve mAP.
Problem
Deep domain-adaptation methods require suitable labeled-data benchmarks, while unsupervised target hashing must learn useful codes without target labels.
Method
The paper introduces Office-Home and a DAH framework that learns hash codes from labeled source data and unlabeled target data using supervised hash, target entropy, and MK-MMD losses.
Results
The framework learns efficient hash codes that outperform existing competitive baselines for unsupervised domain adaptation across multiple transfer tasks.
Takeaways & Limitations
DAH integrates category assignment under weak or insufficient labels with target hash-code estimation in one framework, and Office-Home supports further domain-adaptation research.
Takeaways & Limitations
The derivation assumes equal numbers of source and target points in each batch and an even batch size.
Abstract
from arXiv · showhide
In recent years, deep neural networks have emerged as a dominant machine learning tool for a wide variety of application domains. However, training a deep neural network requires a large amount of labeled data, which is an expensive process in terms of time, labor and human expertise. Domain adaptation or transfer learning algorithms address this challenge by leveraging labeled data in a different, but related source domain, to develop a model for the target domain. Further, the explosive growth of digital data has posed a fundamental challenge concerning its storage and retrieval. Due to its storage and retrieval efficiency, recent years have witnessed a wide application of hashing in a variety of computer vision applications. In this paper, we first introduce a new dataset, Office-Home, to evaluate domain adaptation algorithms. The dataset contains images of a variety of everyday objects from multiple domains. We then propose a novel deep learning framework that can exploit labeled source data and unlabeled target data to learn informative hash codes, to accurately classify unseen target data. To the best of our knowledge, this is the first research effort to exploit the feature learning capabilities of deep neural networks to learn representative hash codes to address the domain adaptation problem. Our extensive empirical studies on multiple transfer tasks corroborate the usefulness of the framework in learning efficient hash codes which outperform existing competitive baselines for unsupervised domain adaptation.
1. Introduction
The paper addresses unsupervised domain adaptation and storage-efficient recognition by combining labeled source data, unlabeled target data, and deep hash-code learning. It introduces Office-Home and a Domain Adaptive Hashing framework with supervised, unsupervised, and distribution-alignment losses.
- Motivation: Deep models need substantial labeled data, motivating domain adaptation using labeled data from a related source domain.Unsupervised adaptation assumes labels are available only in the source domain, while target labels are unavailable.
- Motivation: Hashing converts high-dimensional data into compact binary codes for fast retrieval and low memory cost.The proposed network outputs binary hash codes instead of probability values for classification.
- Contributions: Office-Home contains around 15,500 images across 4 domains and 65 object categories for evaluating domain adaptation.The dataset is introduced specifically to evaluate the proposed algorithm.
- Contributions: Domain Adaptive Hashing learns target hash codes using supervised source hashing, unsupervised target entropy, and MK-MMD losses.The supervised loss groups same-class source samples, the entropy loss aligns target samples with source categories, and MK-MMD reduces domain discrepancy.
2. Related Work
Prior work includes shallow and deep domain-adaptation methods, unsupervised hashing, and cross-modal hashing. The paper positions DAH as addressing the previously unstudied combination of deep hashing and unsupervised domain adaptation.
- Domain Adaptation: Shallow domain-adaptation methods modify source classifiers, learn transformations, reweight source data, or project data using predetermined features.These approaches address domain shift without jointly learning deep transferable representations.
- Domain Adaptation: Deep domain-adaptation methods learn transferable features, including representations trained to make source and target domains indistinguishable.Other methods jointly adapt features and classifiers using residual transfer networks.
- Hashing: Unsupervised hashing targets efficient storage and retrieval, while cross-modal hashing learns shared codes for related modalities but is not unsupervised domain adaptation.The cited cross-modal methods mainly support retrieval across text and image domains.
- Novelty: DAH is presented as the first deep-hashing approach for unsupervised domain adaptation.The paper combines deep hash-code learning with adaptation between related source and target domains.
3. Domain Adaptive Hashing Networks
The DAH network learns binary hash codes for source and target data by combining supervised source hashing, unsupervised target entropy, and multi-kernel domain alignment. Its losses encourage category-discriminative codes while reducing source–target feature discrepancy.
- Network architecture: DAH uses a deep CNN with a hashing layer to produce binary codes h_i ∈ {−1, +1}^d for source and target inputs.The network converts outputs ψ(x) into hash codes using the sign function.
- Domain alignment: The MK-MMD loss aligns source and target feature representations at multiple fully connected layers to address domain shift.The final layer outputs are U^s and U^t, and domain discrepancy is reduced across network layers.
- Supervised hashing: Supervised hashing uses source labels and pairwise similarity to make same-category hash codes similar and different-category codes dissimilar.Hash-code similarity is linked to Hamming distance through the dot product h_i^T h_j.
- Supervised hashing: The continuous relaxation replaces discrete h_i with real-valued u_i, while tanh activation and quantization loss reduce approximation and binarization errors.The relaxation bounds network outputs between −1 and +1 and penalizes distance from sgn(u_i).
- Unsupervised target hashing: For unlabeled target data, the entropy loss encourages each target probability vector to become one-hot and align with one source category.Similarity to multiple source examples from a category provides a more robust category assignment, while dissimilarity to other categories is encouraged.
- Joint objective: The overall DAH objective combines supervised source hashing, target entropy minimization, and domain adaptation, with γ and η controlling the latter two terms.Hash values are obtained from the network output U using H = sgn(U).
4. The Office-Home Dataset
Office-Home is introduced as a larger, multi-domain object-recognition dataset for evaluating deep domain adaptation methods. It contains roughly 15,500 images across 65 everyday-object categories and four distinct visual domains.
- Motivation: Existing vision domain-adaptation datasets are limited in size and unsuitable for validating deep learning algorithms.The paper notes that many standard datasets predate the widespread use of deep learning.
- Dataset composition: Office-Home contains around 15,500 images from 65 everyday-object categories distributed across four domains.Each domain contains images from the same 65 categories.
- Dataset composition: The four domains are Art, Clipart, Product, and Real-World, representing artistic depictions, clipart, background-free product images, and camera photographs.These domains provide varied visual forms of everyday objects for domain-adaptation evaluation.
- Data construction: Images were collected from public websites and manually filtered for quality, size, and content.Product images were collected from Amazon, while Art and Real-World images came from public-domain sources.
5. Experiments
The experiments evaluate DAH on Office and Office-Home domain-adaptation tasks, recognition accuracy, feature alignment, and hashing retrieval. DAH generally performs strongly across transfer tasks, with domain adaptation improving compact hash-code learning.
- Datasets: The Office-Home dataset contains everyday-object images organized into 4 domains and 65 categories, while Office contains 3 domains.Office-Home has around 15,500 images; Office has around 4,100 images.
- Unsupervised Domain Adaptation: Experiments evaluate DAH for unsupervised domain adaptation using labeled source data and unlabeled target data across Office and Office-Home.Comparisons include shallow and deep domain-adaptation baselines, with fc7 features used for shallow methods.
- Recognition Accuracy: DAH consistently outperforms baselines across Office-Home domains, although DANN marginally surpasses DAH on the Office dataset.The authors relate the Office result to its smaller number of categories and argue that DAH is more effective on the larger-category Office-Home setting.
- Ablation: DAH outperforms DAH-e, indicating that minimizing target entropy improves alignment between source and target samples and increases accuracy.DAH-e is the variant with the entropy-loss coefficient η set to zero.
- Feature Analysis: DAH features show the least source-target discrepancy and the strongest domain overlap and category clustering among Deep, DAN, and DAH representations.These patterns are assessed using A-distance and t-SNE embeddings for 10 categories from Art and Clipart.
- Unsupervised Domain Adaptive Hashing: At 64 bits, DAH delivers the best hashing performance relative to NoDA, shallow and deep unsupervised hashing methods, and supervised hashing.Evaluation uses precision-recall curves and mean average precision; NoDA performs poorest because of domain mismatch.
6. Conclusions
The paper presents DAH, a deep domain-adaptive hashing framework for unsupervised domain adaptation, and introduces Office-Home for further research.
- DAH uses deep neural networks to learn efficient hash codes for unsupervised domain adaptation.
- The framework addresses category assignment with weak or insufficient labels and hash-code estimation for unlabeled target data.
- The paper integrates hash-code learning with unsupervised domain adaptation in a single framework.
- Office-Home is introduced as a dataset for further domain-adaptation research.
7. Loss Function Derivative
This section derives backpropagation gradients for the DAH loss, covering domain discrepancy, supervised hashing, and target entropy components.
- The derivative of Equation 8 is formulated for backpropagation, with γ and η controlling domain-adaptation and target-entropy loss importance.
- The MK-MMD derivative is developed for source and target layer outputs using batches containing paired source and target data points.
- The supervised hash-loss derivative is obtained with sgn(.) treated as constant to avoid nondifferentiability at zero.
- The entropy-loss derivation defines target-category probabilities and derives gradients with respect to source and target outputs.
- The entropy loss is re-phrased using logarithmic terms before completing its derivative formulation.
- Indicator functions specify when derivative terms contribute, including conditions involving source categories and output relationships.
8. Unsupervised Domain Adaptation: Additional Results
Additional Office-Home experiments compare recognition accuracy across hash-code lengths and domain-adaptation methods.
- Increasing the hash-code length from 64 to 128 bits increases average recognition accuracy because of greater representation capacity.
- The 16-bit setting has lower recognition accuracy than the longer-code settings.
- Recognition accuracies are reported for Office-Home domain-adaptation experiments across Art, Clipart, Product, and Real-World domains.
9. Unsupervised Domain Adaptive Hashing: Additional Results
Additional hashing experiments evaluate 16- and 128-bit codes on Office-Home using precision-recall curves and mean average precision.
- 16- and 128-bit DAH results are presented in Figures 6 and 7, respectively, with corresponding mAP values in Tables 6 and 7.
- Increasing hash-code length does not necessarily increase mean average precision across hashing methods.
- For Real-World at 128 bits, DAH performs better than supervised hashing, indicating that domain adaptation can help learn a better generalized model in some cases.
- The precision-recall figures compare NoDA, shallow unsupervised hashing, deep unsupervised hashing, DAH, and supervised hashing.