Deep Hashing Network for Unsupervised Domain Adaptation

Hemanth VenkateswaraJose EusebioShayok ChakrabortySethuraman Panchanathan

article2017CVPR2,557 citations

Introduces the Office-Home benchmark dataset and presents a deep hashing framework that learns transferable binary representations for accurate and storage-efficient unsupervised domain adaptation.

Listen

The article addresses the challenge of training deep neural networks for object recognition when labeled data is scarce in a target domain but available in a related source domain. This unsupervised domain adaptation problem is compounded by the need for efficient data storage and retrieval amid growing digital image collections. Hashing techniques offer compact binary codes for fast similarity search, yet prior methods have not integrated deep feature learning with domain adaptation for this setting.

The work introduces the Office-Home dataset, containing roughly 15,500 images across 65 categories from four visual domains, and proposes the Domain Adaptive Hashing (DAH) network. DAH replaces the final classification layer of a pre-trained VGG-F convolutional network with a hashing layer trained by three loss terms: a supervised loss that encourages similar hash codes for same-category source images, an unsupervised entropy loss that aligns unlabeled target images to source categories, and a multi-kernel maximum mean discrepancy term that reduces distribution shift across multiple network layers.

Experiments on the Office and Office-Home benchmarks show that DAH achieves the highest average classification accuracy on Office-Home (45.54 percent) among compared methods and remains competitive on Office. The learned 64-bit hash codes also yield higher mean average precision for retrieval than other unsupervised hashing approaches, with domain alignment visibly reducing feature discrepancy in t-SNE visualizations. Performance gains are larger when the number of categories is high.

These results indicate that jointly learning transferable hash codes and reducing domain shift within a single deep network improves both classification and retrieval under label scarcity. The framework is particularly useful for large-scale vision applications where annotation cost is high and storage efficiency matters. The new Office-Home dataset provides a more challenging testbed than prior collections for evaluating such methods.

Further validation on additional domains, larger hash lengths, and real-world deployment scenarios would strengthen confidence in the approach. The authors note that domain-adversarial training can outperform DAH when category counts are small, suggesting hybrid strategies may be worth exploring.

Arizona State University computer vision computer vision applications deep learning deep hashing network unsupervised domain adaptation hash code learning

arXiv: 1706.07522 hemanthdv/da-hash

Paper: Learning Transferable Features with Deep Adaptation Networks, Mingsheng Long et al. (2015). It introduces deep adaptation networks using multi-kernel maximum mean discrepancy to align distributions across network layers, providing the foundational domain alignment mechanism adapted into DAH.
Paper: Deep Domain Confusion: Maximizing for Domain Invariance, Eric Tzeng et al. (2014). It establishes the core methodology of inserting bottleneck adaptation layers and optimizing maximum mean discrepancy loss directly inside deep CNNs to achieve domain confusion.
Paper: Deep Transfer Learning with Joint Adaptation Networks, Mingsheng Long et al. (2016). It extends deep domain adaptation by aligning joint distributions of features and predictions across task-specific layers, directly informing multi-layer adaptation schemes.
Paper: Deep CORAL: Correlation Alignment for Deep Domain Adaptation, Baochen Sun et al. (2016). It provides the framework for embedding differentiable statistical alignment objectives directly within deep convolutional neural networks for unsupervised domain adaptation.
Paper: Adapting Visual Category Models to New Domains, Kate Saenko et al. (2010). It introduced the classic Office dataset benchmark and metric-learning paradigms that form the foundational evaluation setting for unsupervised visual domain adaptation.
Paper: A theory of learning from different domains, Shai Ben-David et al. (2010). It provides the foundational theoretical bounds and divergence measures governing error reduction when transferring classifiers across shifted data distributions.
Book: Domain-Adversarial Training of Neural Networks, Yaroslav Ganin et al. (2016). It details domain-adversarial training and gradient reversal for extracting invariant features, serving as a primary benchmark and contrasting adaptation strategy.
Paper: Similarity Search in High Dimensions via Hashing, A. Gionis et al. (1999). It introduces the fundamental concepts of locality-sensitive hashing for high-dimensional similarity search that underpin modern deep hashing layers.

Paper: Conditional Adversarial Domain Adaptation, Mingsheng Long et al. (2017). It advances unsupervised domain adaptation on benchmarks like Office-Home by conditioning adversarial adaptation directly on multilinear classifier predictions and prediction entropy.
Paper: Moment Matching for Multi-Source Domain Adaptation, Xingchao Peng et al. (2018). It scales moment-matching domain adaptation from single-source settings to complex multi-source visual recognition challenges.
Paper: CyCADA: Cycle-Consistent Adversarial Domain Adaptation, Judy Hoffman et al. (2018). It expands deep domain adaptation beyond feature alignment by integrating pixel-level cycle-consistent generative translation with semantic consistency constraints.
Paper: Deep Visual Domain Adaptation: A Survey, Mei Wang et al. (2018). It provides a comprehensive taxonomy and comparative survey of deep visual domain adaptation paradigms, synthesizing discrepancy-based and adversarial methods.
Paper: A Comprehensive Survey on Transfer Learning, Fuzhen Zhuang et al. (2019). It presents a broad survey of transfer learning mechanisms and benchmarks, placing deep domain adaptation within the wider taxonomy of transfer methods.

Cover for Deep Hashing Network for Unsupervised Domain Adaptation

Abstract

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.

Topics

computer vision computer vision applications deep learning hashing representation learning retrieval efficiency transfer learning

Keywords

deep hashing network unsupervised domain adaptation hash code learning joint distribution adaptation unsupervised entropy loss transferable features cross-modal hashing transfer learning

1 Introduction
2 Related Work
3 Domain Adaptive Hashing Networks
3.1 Reducing Domain Disparity
3.2 Supervised Hashing for Source Data
3.3 Unsupervised Hashing for Target Data
3.4 Domain Adaptive Hash Network
4 The Office-Home Dataset
5 Experiments
5.1 Datasets
5.2 Implementation Details
5.3 Unsupervised Domain Adaptation
5.4 Unsupervised Domain Adaptive Hashing
6 Conclusions
References
7 Loss Function Derivative
7.1 Derivative for MK-MMD
7.2 Derivative for Supervised Hash Loss
7.3 Derivative for Unsupervised Entropy Loss
8 Unsupervised Domain Adaptation: Additional Results
9 Unsupervised Domain Adaptive Hashing: Additional Results

Knowls

Knowl 1 — Domain Adaptive Hashing (DAH) Framework
model/method
Domain Adaptive Hashing (DAH) is a deep neural network framework for unsupervised domain adaptation that learns compact binary hash codes to classify unlabeled target data. Built on a convolutional neural network backbone (specifically, a VGG-F network pre-trained on ImageNet), the network replaces the standard final classification layer with a specialized hashing layer $\text{hash-fc8}$ structured as $\text{fc8} \to \text{batch-norm} \to \tanh(\cdot)$ . This layer maps high-dimensional image representations to continuous vectors $u_i \in [-1, +1]^d$ , which are subsequently binarized into $d$ -bit hash codes $h_i = \text{sgn}(u_i) \in \{-1, +1\}^d$ .

Given labeled source data $\mathcal{D}_s = \{(x_i^s, y_i^s)\}_{i=1}^{n_s}$ with labels $y_i^s \in \{1, \dots, C\}$ and unlabeled target data $\mathcal{D}_t = \{x_i^t\}_{i=1}^{n_t}$ , DAH trains the network parameters end-to-end to minimize the joint objective function: $J(U) = \mathcal{L}(U_s) + \gamma \mathcal{M}(U_s, U_t) + \eta \mathcal{H}(U_s, U_t)$ where $U_s = \{u_i^s\}_{i=1}^{n_s}$ , $U_t = \{u_i^t\}_{i=1}^{n_t}$ , $U = U_s \cup U_t$ , and:
- $\mathcal{L}(U_s)$ is a supervised pairwise hashing loss on labeled source data with continuous relaxation and quantization regularization.
- $\mathcal{M}(U_s, U_t)$ is the multi-kernel Maximum Mean Discrepancy (MK-MMD) loss evaluated across fully connected layers $\mathcal{F} = \{\text{fc6}, \text{fc7}, \text{fc8}\}$ to reduce distribution disparity between domains.
- $\mathcal{H}(U_s, U_t)$ is an unsupervised entropy loss on unlabeled target data that drives each target representation to align closely with exactly one source category.
- $\gamma > 0$ and $\eta > 0$ are hyperparameter weights balancing domain alignment and target entropy regularization.
Knowl 2 — Supervised Source Hashing Loss with Quantization Regularization
model/method
In Domain Adaptive Hashing (DAH), the supervised hash loss $\mathcal{L}(U_s)$ ensures that source instances from the same class produce similar binary hash codes and those from different classes produce dissimilar codes.

For source data points $x_i^s, x_j^s$ with continuous network outputs $u_i, u_j \in [-1, +1]^d$ , a pairwise similarity label is defined as $s_{ij} = 1$ if $y_i^s = y_j^s$ and $s_{ij} = 0$ if $y_i^s \neq y_j^s$ . The probability of similarity given the continuous representations is modeled using the sigmoid function $\sigma(z) = \frac{1}{1 + \exp(-z)}$ : $p(s_{ij} \mid u_i, u_j) = \begin{cases} \sigma(u_i^\top u_j), & s_{ij} = 1 \\ 1 - \sigma(u_i^\top u_j), & s_{ij} = 0 \end{cases}$

Assuming the elements of the similarity matrix $S = \{s_{ij}\}$ are independent, minimizing the negative log-likelihood with a continuous relaxation quantization penalty yields the source hashing loss: $\min_{U_s} \mathcal{L}(U_s) = - \sum_{s_{ij} \in S} \left( s_{ij} u_i^\top u_j - \log\left(1 + \exp(u_i^\top u_j)\right) \right) + \sum_{i=1}^{n_s} \|u_i - \text{sgn}(u_i)\|_2^2$ where $\text{sgn}(\cdot)$ is the element-wise sign function. To mitigate class imbalance between similar ( $s_{ij} = 1$ ) and dissimilar ( $s_{ij} = 0$ ) pairs, the similarity values can be reweighted (e.g., $s_{ij} \in \{0, 10\}$ ).
Knowl 3 — Unsupervised Target Entropy Loss
model/method
Because unlabeled target instances in unsupervised domain adaptation lack category annotations, Domain Adaptive Hashing (DAH) uses an unsupervised entropy minimization loss based on soft assignment probabilities to source classes.

Assuming $K$ source samples per class $j \in \{1, \dots, C\}$ , let $u_k^{sj} \in [-1, +1]^d$ denote the continuous output for the $k$ -th source sample belonging to category $j$ , and let $u_i^t \in [-1, +1]^d$ be the continuous output for target sample $x_i^t$ . The probability $p_{ij}$ that target sample $x_i^t$ belongs to category $j$ is computed as the normalized exponential inner product across all $K$ source exemplars of that class: $p_{ij} = \frac{\sum_{k=1}^K \exp({u_i^t}^\top u_k^{sj})}{\sum_{l=1}^C \sum_{k=1}^K \exp({u_i^t}^\top u_k^{sl})}$ satisfying $\sum_{j=1}^C p_{ij} = 1$ .

The entropy loss $\mathcal{H}(U_s, U_t)$ across all target instances is defined as: $\mathcal{H}(U_s, U_t) = -\frac{1}{n_t} \sum_{i=1}^{n_t} \sum_{j=1}^C p_{ij} \log(p_{ij})$

Minimizing this entropy forces the probability vector $p_i = [p_{i1}, \dots, p_{iC}]^\top$ to approach a low-entropy one-hot distribution, ensuring that each target data point is strongly aligned with source samples of exactly one category and dissimilar to source samples of all other categories.
Knowl 4 — Multi-Layer Multi-Kernel Maximum Mean Discrepancy Loss
equation
In Domain Adaptive Hashing (DAH), domain discrepancy between source and target feature representations is reduced across fully connected layers $\mathcal{F} = \{\text{fc6}, \text{fc7}, \text{fc8}\}$ using multi-kernel Maximum Mean Discrepancy (MK-MMD): $\mathcal{M}(U_s, U_t) = \sum_{l \in \mathcal{F}} d_k^2(U_s^l, U_t^l)$ where $U_s^l = \{u_i^{s,l}\}_{i=1}^{n_s}$ and $U_t^l = \{u_i^{t,l}\}_{i=1}^{n_t}$ are the activations of layer $l$ for source and target samples, respectively.

The squared MK-MMD distance in reproducing kernel Hilbert space $\mathcal{H}_k$ with characteristic multi-kernel $k(\cdot, \cdot)$ is: $d_k^2(U_s^l, U_t^l) = \|\mathbb{E}[\phi(u^{s,l})] - \mathbb{E}[\phi(u^{t,l})]\|_{\mathcal{H}_k}^2$ where $\phi(\cdot)$ is the mapping associated with kernel $k(x, y) = \langle \phi(x), \phi(y) \rangle$ . The characteristic kernel is defined as a convex combination of $\kappa$ positive semi-definite Gaussian kernels $\{k_m\}_{m=1}^\kappa$ : $k(x, y) = \sum_{m=1}^\kappa \beta_m k_m(x, y), \quad k_m(x, y) = \exp\left(-\frac{\|x - y\|_2^2}{\sigma_m}\right)$ with weights $\beta_m = 1/\kappa$ , base bandwidth $\sigma$ set to the median pairwise distance in the training data, and multi-kernel bandwidths $\sigma_m \in [2^{-8}\sigma, 2^8\sigma]$ spaced by factors of 2.

In mini-batch training with an even batch size $n$ of source representations $U = \{u_i\}_{i=1}^n$ and target representations $V = \{v_i\}_{i=1}^n$ , the unbiased linear-time estimator is computed over quadruplets $w_i = [u_{2i-1}, u_{2i}, v_{2i-1}, v_{2i}]$ for $i \in \{1, \dots, n/2\}$ : $\mathcal{M}(U, V) = \frac{2}{n \kappa} \sum_{m=1}^\kappa \sum_{i=1}^{n/2} \left[ k_m(u_{2i-1}, u_{2i}) + k_m(v_{2i-1}, v_{2i}) - k_m(u_{2i-1}, v_{2i}) - k_m(u_{2i}, v_{2i-1}) \right]$
Knowl 5 — Office-Home Benchmark Dataset
definition
The Office-Home dataset is an object recognition benchmark designed to evaluate domain adaptation algorithms under significant domain shift. It contains approximately 15,500 images across 65 everyday object categories organized into 4 distinct domains:
- Art (Ar): Artistic depictions in sketches, paintings, drawings, and ornamentation (minimum 15 images/category, image resolutions from $117 \times 85$ to $4384 \times 2686$ pixels, baseline LIBLINEAR SVM accuracy $44.99 \pm 1.85\%$ ).
- Clipart (Cl): Clipart images collected from web sources (minimum 39 images/category, resolutions from $18 \times 18$ to $2400 \times 2400$ pixels, baseline accuracy $53.95 \pm 1.45\%$ ).
- Product (Pr): Commercial product images without background, crawled from Amazon (minimum 38 images/category, resolutions from $75 \times 63$ to $2560 \times 2560$ pixels, baseline accuracy $66.41 \pm 1.18\%$ ).
- Real-World (Rw): Photographs of objects in natural office and home environments captured with regular cameras (minimum 23 images/category, resolutions from $88 \times 80$ to $6500 \times 4900$ pixels, baseline accuracy $59.70 \pm 1.04\%$ ).
The dataset has an average of approximately 70 images per category (maximum 99 images). Deep feature baselines are computed using 5-fold cross-validation with a LIBLINEAR SVM on fc7 activations extracted from a pre-trained VGG-F network.
Knowl 6 — Target Category Assignment from Continuous and Hash Representations
model/method
Following training of the Domain Adaptive Hashing (DAH) network on labeled source data $\mathcal{D}_s$ and unlabeled target data $\mathcal{D}_t$ , target predictions are obtained using the learned continuous outputs or binary hash codes.

For a target instance $x_i^t$ , the network outputs continuous representation $u_i^t \in [-1, +1]^d$ and binary hash code $h_i^t = \text{sgn}(u_i^t) \in \{-1, +1\}^d$ . Given $K$ exemplar representations $\{u_k^{sj}\}_{k=1}^K$ for each source class $j \in \{1, \dots, C\}$ , the posterior category assignment probability is computed as: $f(x_i^t) = p(y = j \mid u_i^t) = \frac{\sum_{k=1}^K \exp({u_i^t}^\top u_k^{sj})}{\sum_{l=1}^C \sum_{k=1}^K \exp({u_i^t}^\top u_k^{sl})}$

The predicted category $\hat{y}_i^t$ is assigned via: $\hat{y}_i^t = \arg\max_{j \in \{1, \dots, C\}} p(y = j \mid u_i^t)$

During retrieval tasks, test binary hash codes $h_i^t$ can be compared directly against stored database codes $h_j$ using Hamming distance: $\text{dist}_H(h_i, h_j) = \frac{1}{2}(d - h_i^\top h_j)$

Knowl 7 — Unsupervised Domain Adaptation Recognition Accuracies on Office-Home

data/table

Target classification accuracies (%) across the 12 transfer tasks on the Office-Home dataset (65 classes) comparing DAH (with $d = 64$ bits) against shallow transfer learning baselines and deep domain adaptation architectures:

Expt.	Ar $\to$ Cl	Ar $\to$ Pr	Ar $\to$ Rw	Cl $\to$ Ar	Cl $\to$ Pr	Cl $\to$ Rw	Pr $\to$ Ar	Pr $\to$ Cl	Pr $\to$ Rw	Rw $\to$ Ar	Rw $\to$ Cl	Rw $\to$ Pr	Avg.
GFK	21.60	31.72	38.83	21.63	34.94	34.20	24.52	25.73	42.92	32.88	28.96	50.89	32.40
TCA	19.93	32.08	35.71	19.00	31.36	31.74	21.92	23.64	42.12	30.74	27.15	48.68	30.34
CORAL	27.10	36.16	44.32	26.08	40.03	40.33	27.77	30.54	50.61	38.48	36.36	57.11	37.91
JDA	25.34	35.98	42.94	24.52	40.19	40.90	25.96	32.72	49.25	35.10	35.35	55.35	36.97
DAN	30.66	42.17	54.13	32.83	47.59	49.78	29.07	34.05	56.70	43.58	38.25	62.73	43.46
DANN	33.33	42.96	54.42	32.26	49.13	49.76	30.49	38.14	56.76	44.71	42.66	64.65	44.94
DAH-e	29.23	35.71	48.29	33.79	48.23	47.49	29.87	38.76	55.63	41.16	44.99	59.07	42.69
DAH	31.64	40.75	51.73	34.69	51.93	52.79	29.91	39.63	60.71	44.99	45.13	62.54	45.54

All shallow baselines (GFK, TCA, CORAL, JDA) use deep features extracted from the fc7 layer of a pre-trained VGG-F network. DAH-e is an ablation of DAH with target entropy loss disabled ( $\eta = 0$ ). DAH achieves the highest average classification accuracy ( $45.54\%$ ), outperforming DANN ( $44.94\%$ ), DAN ( $43.46\%$ ), and DAH-e ( $42.69\%$ ). When evaluating across code lengths, DAH achieves an average accuracy of $31.36\%$ at $d = 16$ bits and $46.26\%$ at $d = 128$ bits.

Knowl 8 — Unsupervised Domain Adaptation Recognition Accuracies on Office

data/table

Target classification accuracies (%) across the 6 transfer tasks on the Office dataset (31 classes across Amazon (A), Dslr (D), and Webcam (W)) comparing DAH against shallow and deep domain adaptation baselines:

Expt.	A $\to$ D	A $\to$ W	D $\to$ A	D $\to$ W	W $\to$ A	W $\to$ D	Avg.
GFK	48.59	52.08	41.83	89.18	49.04	93.17	62.32
TCA	51.00	49.43	48.12	93.08	48.83	96.79	64.54
CORAL	54.42	51.70	48.26	95.97	47.27	98.59	66.04
JDA	59.24	58.62	51.35	96.86	52.34	97.79	69.37
DAN	67.04	67.80	50.36	95.85	52.33	99.40	72.13
DANN	72.89	72.70	56.25	96.48	53.20	99.40	75.15
DAH-e	66.27	66.16	55.97	94.59	53.91	96.99	72.31
DAH	66.47	68.30	55.54	96.10	53.02	98.80	73.04

DAH achieves an average accuracy of $73.04\%$ , exceeding DAN ( $72.13\%$ ) and DAH-e ( $72.31\%$ ), while DANN attains $75.15\%$ . Domain-adversarial training is competitive when the number of categories is small (31 in Office), whereas DAH shows superior performance when the category count is large (65 in Office-Home), because category-agnostic adversarial alignment can produce feature distributions that are less classification-friendly across many categories.

Knowl 9 — Unsupervised Domain Adaptive Hashing Image Retrieval Performance

data/table

Mean Average Precision (mAP) for image retrieval across different target domains evaluated at a hash code length of 64 bits:

Expt.	NoDA	ITQ	KMeans	BA	BDNN	DAH	SuH
Amazon	0.324	0.465	0.403	0.367	0.491	0.582	0.830
Webcam	0.511	0.652	0.558	0.480	0.656	0.717	0.939
Art	0.155	0.191	0.170	0.156	0.193	0.302	0.492
Clipart	0.160	0.195	0.178	0.179	0.206	0.333	0.622
Product	0.239	0.393	0.341	0.349	0.407	0.414	0.774
Real-World	0.281	0.323	0.279	0.273	0.336	0.533	0.586
Avg.	0.278	0.370	0.322	0.301	0.382	0.480	0.707

For the Office-Home tasks, Art is the source domain for Clipart, Product, and Real-World, and Clipart is the source domain for Art. For Office, Amazon and Webcam form source-target pairs.

NoDA (Deep Pairwise-Supervised Hashing trained on source without adaptation) achieves an average mAP of $0.278$ .
Unsupervised hashing methods (ITQ, KMeans, BA, BDNN) ignore source labels and achieve average mAPs between $0.301$ and $0.382$ .
DAH integrates supervised source hashing, domain alignment, and target entropy minimization, achieving an average mAP of $0.480$ .
SuH (Supervised Hashing trained directly with target labels) provides an empirical upper bound of $0.707$ .

Knowl 10 — Exact Gradients of the Domain Adaptive Hashing Loss Function
equation
In Domain Adaptive Hashing (DAH), backpropagation updates network parameters via partial derivatives of the overall loss $J = \mathcal{L}(U_s) + \gamma \mathcal{M}(U_s, U_t) + \eta \mathcal{H}(U_s, U_t)$ with respect to source representations $u_q \in U_s$ and target representations $v_q \in U_t$ .

The derivative of the supervised source hashing loss $\mathcal{L}(U_s)$ with respect to source continuous vector $u_q$ (treating $\text{sgn}(u_q)$ as constant) is: $\frac{\partial \mathcal{L}}{\partial u_q} = \sum_{j=1}^{n_s} 2\left(\sigma(u_q^\top u_j) - s_{qj}\right) u_j + 2\left(u_q - \text{sgn}(u_q)\right)$ where $\sigma(z) = \frac{1}{1 + \exp(-z)}$ and $s_{qj}$ is the symmetric pairwise similarity indicator.

The derivative of the target entropy loss $\mathcal{H}(U_s, U_t)$ with respect to target output $v_q$ is: $\frac{\partial \mathcal{H}}{\partial v_q} = -\frac{1}{n_t} \sum_{j=1}^C \left(\log(p_{qj}) + 1\right) \left[ \sum_{k=1}^K \bar{p}_{qjk} u_k^j - p_{qj} \sum_{l=1}^C \sum_{k'=1}^K \bar{p}_{qlk'} u_{k'}^l \right]$ and with respect to source exemplar $u_q^p$ (the $q$ -th exemplar of class $p$ ) is: $\frac{\partial \mathcal{H}}{\partial u_q^p} = -\frac{1}{n_t} \sum_{i=1}^{n_t} \sum_{j=1}^C v_i \left(\log(p_{ij}) + 1\right) \left[ \sum_{k=1}^K \mathbb{I}\{j=p, k=q\} \bar{p}_{ijk} - p_{ij} \bar{p}_{ipq} \right]$ where $\bar{p}_{ijk} = \frac{\exp(v_i^\top u_k^j)}{\sum_{l=1}^C \sum_{k'=1}^K \exp(v_i^\top u_{k'}^l)}$ , $p_{ij} = \sum_{k=1}^K \bar{p}_{ijk}$ , and $\mathbb{I}\{\cdot\}$ is the indicator function.

Coverage note — None was omitted; all primary contributed methods, dataset specifications, loss functions, analytic gradients, and empirical benchmarks on domain adaptation and hashing retrieval are captured.

References

1.Y. Aytar and A. Zisserman. Tabula rasa: Model transfer for object category detection. In IEEE ICCV, 2011. 2
2.S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan. A theory of learning from different domains. Machine learning, 79(1-2):151–175, 2010. 6
3.Y. Bengio, A. Courville, and P. Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828, 2013. 2
4.L. Bruzzone and M. Marconcini. Domain adaptation problems: A dasvm classification technique and a circular validation strategy. IEEE, PAMI, 32(5):770–787, 2010. 2
5.Y. Cao, M. Long, J. Wang, Q. Yang, and P. S. Yu. Deep visual-semantic hashing for cross-modal retrieval. In ACM-SIGKDD, 2016. 2
6.Z. Cao, M. Long, and Q. Yang. Transitive hashing network for heterogeneous multimedia retrieval. In AAAI, 2016. 2
7.M. A. Carreira-Perpin´an and R. Raziperchikolaei. Hashing with binary autoencoders. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 557–566, 2015. 2, 7
8.K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman. Return of the devil in the details: Delving deep into convolutional nets. In BMVC, 2014. 2, 4
9.R. Chattopadhyay, Q. Sun, W. Fan, I. Davidson, S. Panchanathan, and J. Ye. Multisource domain adaptation and its application to early detection of fatigue. ACM Transactions on Knowledge Discovery from Data (TKDD), 6(4):18, 2012. 2
10.W.-S. Chu, F. De la Torre, and J. F. Cohn. Selective transfer machine for personalized facial action unit detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3515–3522, 2013. 2
11.T.-T. Do, A.-D. Doan, and N.-M. Cheung. Learning to hash with binary deep neural network. In European Conference on Computer Vision, pages 219–234. Springer, 2016. 2, 7
12.J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng, and T. Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In ICML, pages 647–655, 2014. 2, 6
13.L. Duan, I. W. Tsang, and D. Xu. Domain transfer multiple kernel learning. IEEE PAMI, 34(3):465–479, 2012. 2
14.V. Erin Liong, J. Lu, G. Wang, P. Moulin, and J. Zhou. Deep hashing for compact binary codes learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2475–2483, 2015. 2
15.R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. Liblinear: A library for large linear classification. Journal of machine learning research, 9(Aug):1871–1874, 2008. 5, 6
16.B. Fernando, A. Habrard, M. Sebban, and T. Tuytelaars. Unsupervised visual domain adaptation using subspace alignment. In CVPR, pages 2960–2967, 2013. 1, 2
17.Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V. Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17(59):1–35, 2016. 1, 2, 3, 6
18.X. Glorot, A. Bordes, and Y. Bengio. Domain adaptation for large-scale sentiment classification: A deep learning approach. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 513–520, 2011. 1, 2
19.B. Gong, K. Grauman, and F. Sha. Connecting the dots with landmarks: Discriminatively learning domain-invariant features for unsupervised domain adaptation. In ICML (1), pages 222–230, 2013. 2
20.B. Gong, Y. Shi, F. Sha, and K. Grauman. Geodesic flow kernel for unsupervised domain adaptation. In IEEE CVPR, 2012. 1, 2, 5, 6
21.Y. Gong and S. Lazebnik. Iterative quantization: A procrustean approach to learning binary codes. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 817–824. IEEE, 2011. 7
22.Y. Gong, S. Lazebnik, A. Gordo, and F. Perronnin. Iterative quantization: A procrustean approach to learning binary codes for large-scale image retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2916–2929, 2013. 2, 4, 7
23.R. Gopalan, R. Li, and R. Chellappa. Domain adaptation for object recognition: An unsupervised approach. In 2011 international conference on computer vision, pages 999–1006. IEEE, 2011. 2
24.A. Gretton, D. Sejdinovic, H. Strathmann, S. Balakrishnan, M. Pontil, K. Fukumizu, and B. K. Sriperumbudur. Optimal kernel choice for large-scale two-sample tests. In Advances in neural information processing systems, pages 1205–1213, 2012. 3, 4, 5, 11
25.K. He, F. Wen, and J. Sun. K-means hashing: An affinity-preserving quantization method for learning binary compact codes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2938–2945, 2013. 2, 7
26.S. Hochreiter, Y. Bengio, P. Frasconi, and J. Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001. 4
27.J. Hoffman, E. Rodner, J. Donahue, K. Saenko, and T. Darrell. Efficient learning of domain-invariant image representations. In ICLR, 2013. 2
28.K. Jarrett, K. Kavukcuoglu, Y. Lecun, et al. What is the best multi-stage architecture for object recognition? In 2009 IEEE 12th International Conference on Computer Vision, pages 2146–2153. IEEE, 2009. 5
29.Q.-Y. Jiang and W.-J. Li. Deep cross-modal hashing. arXiv preprint arXiv:1602.02255, 2016. 2
30.W.-J. Li, S. Wang, and W.-C. Kang. Feature learning based deep supervised hashing with pairwise labels. In IJCAI, 2016, 2016. 4, 7
31.M. Long, Y. Cao, J. Wang, and M. Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97–105, 2015. 1, 2, 3, 4, 5, 6
32.M. Long, J. Wang, G. Ding, J. Sun, and P. Yu. Transfer joint matching for unsupervised domain adaptation. In CVPR, pages 1410–1417, 2014. 2
33.M. Long, J. Wang, G. Ding, J. Sun, and P. S. Yu. Transfer feature learning with joint distribution adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2200–2207, 2013. 1, 2, 5, 6
34.M. Long, H. Zhu, J. Wang, and M. I. Jordan. Unsupervised domain adaptation with residual transfer networks. In NIPS, 2016. 1, 2, 3, 5
35.P. Lucey, J. F. Cohn, T. Kanade, J. Saragih, Z. Ambadar, and I. Matthews. The extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified expression. In CVPR, pages 94–101. IEEE, 2010. 5
36.Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5
37.M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1717–1724, 2014. 2
38.S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang. Domain adaptation via transfer component analysis. Neural Networks, IEEE Trans. on, 22(2):199–210, 2011. 1, 2, 6
39.S. J. Pan and Q. Yang. A survey on transfer learning. IEEE TKDE, 22(10):1345–1359, 2010. 2
40.M. Pantic, M. Valstar, R. Rademaker, and L. Maat. Web-based database for facial expression analysis. In ICME. IEEE, 2005. 5
41.V. M. Patel, R. Gopalan, R. Li, and R. Chellappa. Visual domain adaptation: A survey of recent advances. IEEE signal processing magazine, 32(3):53–69, 2015. 2
42.K. Saenko, B. Kulis, M. Fritz, and T. Darrell. Adapting visual category models to new domains. In ECCV, 2010. 1, 2, 5
43.S. Shekhar, V. M. Patel, H. V. Nguyen, and R. Chellappa. Generalized domain-adaptive dictionaries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 361–368, 2013. 1
44.B. Sun, J. Feng, and K. Saenko. Return of frustratingly easy domain adaptation. In ICCV, TASK-CV, 2015. 1, 2, 6
45.A. Torralba and A. A. Efros. Unbiased look at dataset bias. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 1521–1528. IEEE, 2011. 5
46.E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko. Simultaneous deep transfer across domains and tasks. In Proceedings of the IEEE International Conference on Computer Vision, pages 4068–4076, 2015. 1, 2, 3
47.A. Vedaldi and K. Lenc. Matconvnet – convolutional neural networks for matlab. In Proceeding of the ACM Int. Conf. on Multimedia, 2015. 5
48.J. Wang, H. T. Shen, J. Song, and J. Ji. Hashing for similarity search: A survey. arXiv preprint arXiv:1408.2927, 2014. 1
49.J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural networks? In Advances in neural information processing systems, pages 3320–3328, 2014. 2, 3
50.H. Zhu, M. Long, J. Wang, and Y. Cao. Deep hashing network for efficient similarity retrieval. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 4

Citation

MLA

Venkateswara, H., et al. “Deep Hashing Network for Unsupervised Domain Adaptation”. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 5385–94, https://doi.org/10.1109/CVPR.2017.572.

APA

Venkateswara, H., Eusebio, J., Chakraborty, S., & Panchanathan, S. (2017). Deep Hashing Network for Unsupervised Domain Adaptation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 5385–5394. https://doi.org/10.1109/CVPR.2017.572

Chicago

Venkateswara, H., J. Eusebio, S. Chakraborty, and S. Panchanathan. 2017. “Deep Hashing Network for Unsupervised Domain Adaptation”. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 5385–94. https://doi.org/10.1109/CVPR.2017.572.

Harvard

Venkateswara, H. et al. (2017) “Deep Hashing Network for Unsupervised Domain Adaptation”, 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 5385–5394. Available at: https://doi.org/10.1109/CVPR.2017.572.

Vancouver

1. Venkateswara H, Eusebio J, Chakraborty S, Panchanathan S (2017) Deep Hashing Network for Unsupervised Domain Adaptation. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp 5385–5394

BibTeX

@inproceedings{Venkateswara_2017, title={Deep Hashing Network for Unsupervised Domain Adaptation}, url={http://dx.doi.org/10.1109/CVPR.2017.572}, DOI={10.1109/cvpr.2017.572}, booktitle={2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}, publisher={IEEE}, author={Venkateswara, Hemanth and Eusebio, Jose and Chakraborty, Shayok and Panchanathan, Sethuraman}, year={2017}, month=July, pages={5385–5394} }

Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: IEEE

Deep Hashing Network for Unsupervised Domain Adaptation

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Domain Adaptive Hashing (DAH) Framework

Knowl 2 — Supervised Source Hashing Loss with Quantization Regularization

Knowl 3 — Unsupervised Target Entropy Loss

Knowl 4 — Multi-Layer Multi-Kernel Maximum Mean Discrepancy Loss

Knowl 5 — Office-Home Benchmark Dataset

Knowl 6 — Target Category Assignment from Continuous and Hash Representations

Knowl 7 — Unsupervised Domain Adaptation Recognition Accuracies on Office-Home

Knowl 8 — Unsupervised Domain Adaptation Recognition Accuracies on Office

Knowl 9 — Unsupervised Domain Adaptive Hashing Image Retrieval Performance

Knowl 10 — Exact Gradients of the Domain Adaptive Hashing Loss Function

References

Citation

BibTeX

Source Code

Access the Paper