Back

Transfer Feature Learning with Joint Distribution Adaptation

Mingsheng LongJianmin WangGuiguang DingJiaguang SunPhilip S. Yu

article2013ICCV1,843 citations

Proposes Joint Distribution Adaptation, a principled dimensionality reduction framework that iteratively aligns both marginal and class-conditional distributions using pseudo labels to boost cross-domain image classification accuracy without requiring target labels.

Listen

In real-world computer vision applications, acquiring large amounts of accurately labeled data for every new environment or camera setup is expensive and time-consuming. While transfer learning allows organizations to train models using abundant labeled data from an existing source domain and apply them to an unlabeled target domain, standard techniques often fail when the two environments differ significantly in both overall data patterns (marginal distributions) and class-specific relationships (conditional distributions).

The article introduces and evaluates Joint Distribution Adaptation (JDA), a transfer learning approach designed to adapt both marginal and conditional distributions simultaneously without requiring any labeled samples in the target environment.

To achieve this, the method combines standard data dimensionality reduction with a statistical distance metric that measures domain differences across feature distributions and predicted class-level distributions. Because target domain labels are unavailable, the approach generates initial pseudo labels using a base classifier and iteratively refines these labels and the shared feature representation until the system stabilizes. The authors evaluated this technique across 36 cross-domain image classification tasks covering digits, faces across multiple camera poses, and diverse object recognition datasets, comparing it against standard non-transfer baselines and four state-of-the-art domain adaptation algorithms.

The key findings demonstrate that Joint Distribution Adaptation delivers substantial performance gains. First, it achieved an average classification accuracy of 57.37% across all 36 evaluation sets, outperforming the strongest transfer baseline by 7.57% in absolute accuracy—amounting to a 15.07% relative reduction in classification error. Second, it markedly outperformed standard non-transfer classification, which averaged only 37.46% accuracy and suffered major failures in difficult visual tasks. Third, the iterative pseudo label refinement process converged reliably within just 10 iterations across test conditions. Fourth, the approach demonstrated high operational stability over a broad range of core parameter settings, avoiding brittleness.

These results show that explicitly aligning both marginal and conditional distributions is critical for robust domain adaptation. By relying on sufficient distribution statistics rather than exact density estimations, the approach successfully tolerates noisy initial pseudo labels and extracts features that enhance between-domain similarity while preserving category distinctions. Operationally, this capability allows organizations to reuse existing image libraries across different operational environments without the cost, timeline delays, or manual effort associated with annotating new target datasets.

Organizations developing computer vision systems across varying operational settings should consider adopting joint distribution matching pipelines to minimize manual data annotation costs. Computational trade-offs should be weighed during deployment: while this method requires slightly more computation than single-distribution baselines, it processes data orders of magnitude faster than complex density-estimation methods. As a future step, the underlying alignment framework can be explored in conjunction with other advanced representation learning models, such as sparse coding.

Confidence in these findings is supported by consistent outperformance across 36 diverse cross-domain image datasets. However, practitioners should recognize the boundary conditions: the approach assumes that the source and target domains share identical feature sets and class categories. Additionally, because the method relies on a closed-form dimensionality reduction step, performance in production environments will depend on the representativeness of the unlabeled target samples provided during training.

Tsinghua UniversityUniversity of Illinois Chicagoclassifierscomputer visionconvolutional neural networksTransfer feature learningJoint distribution adaptationcross-domain classification
Long et al (2013).pdf
  • Paper: A theory of learning from different domains, Shai Ben-David et al. (2010). This paper establishes the foundational generalization bounds and distribution discrepancy metrics that motivate adapting both marginal and conditional feature distributions.
  • Paper: Correcting Sample Selection Bias by Unlabeled Data, Jiayuan Huang et al. (2006). It introduces non-parametric distribution matching via Maximum Mean Discrepancy in reproducing kernel Hilbert spaces, which serves as the mathematical core for JDA's alignment objective.
  • Paper: A Survey on Transfer Learning, Sinno Jialin Pan et al. (2010). This comprehensive survey outlines the core taxonomy and formal definitions of transfer learning that frame JDA's dimensionality reduction approach.
  • Paper: Adapting Visual Category Models to New Domains, Kate Saenko et al. (2010). It introduces the standard cross-domain visual benchmark (Office) and metric-learning paradigms on which JDA evaluates domain transfer performance.
  • Paper: Analysis of Representations for Domain Adaptation, Shai Ben-David et al. (2006). It provides theoretical analysis demonstrating that effective domain adaptation requires learning representations that minimize cross-domain distribution divergence.
Cover for Transfer Feature Learning with Joint Distribution Adaptation

Abstract

Transfer learning is established as an effective technology in computer vision for leveraging rich labeled data in the source domain to build an accurate classifier for the target domain. However, most prior methods have not simultaneously reduced the difference in both the marginal distribution and conditional distribution between domains. In this paper, we put forward a novel transfer learning approach, referred to as Joint Distribution Adaptation (JDA). Specifically, JDA aims to jointly adapt both the marginal distribution and conditional distribution in a principled dimensionality reduction procedure, and construct new feature representation that is effective and robust for substantial distribution difference. Extensive experiments verify that JDA can significantly outperform several state-of-the-art methods on four types of cross-domain image classification problems.

Topics

classifierscomputer visionconvolutional neural networksdimensionality reductionimage classificationmarginalsrepresentation learningtransfer learningattention mechanism

Keywords

Transfer feature learningJoint distribution adaptationcross-domain classificationtransductive transfersubspace learningunlabeled target dataimage classification

Table of Contents

  • 1. Introduction
  • 2. Related Work
  • 3. Joint Distribution Adaptation
  • 3.1. Problem Definition
  • 3.2. Proposed Approach
  • 3.2.1 Feature Transformation
  • 3.2.2 Marginal Distribution Adaptation
  • 3.2.3 Conditional Distribution Adaptation
  • 3.2.4 Optimization Problem
  • 3.2.5 Iterative Refinement
  • 3.3. Learning Algorithm
  • 3.4. Computational Complexity
  • 4. Experiments
  • 4.1. Data Preparation
  • 4.2. Baseline Methods
  • 4.3. Implementation Details
  • 4.4. Experimental Results
  • 4.5. Effectiveness Verification
  • 4.6. Parameter Sensitivity
  • 4.7. Convergence and Time Complexity
  • 5. Conclusion and Future Work
  • 6. Acknowledgments
  • References

Knowls

  1. Knowl 1 — Joint Distribution Adaptation Optimization Formulation

    model/method

    Joint Distribution Adaptation (JDA) learns a shared low-dimensional representation Z=ATXZ = A^T X for cross-domain transfer learning where the source domain Ds={(x1,y1),…,(xns,yns)}\mathcal{D}_s = \{(x_1, y_1), \dots, (x_{n_s}, y_{n_s})\} is labeled and the target domain Dt={xns+1,…,xns+nt}\mathcal{D}_t = \{x_{n_s+1}, \dots, x_{n_s+n_t}\} is completely unlabeled. Let X=[x1,…,xn]∈Rm×nX = [x_1, \dots, x_n] \in \mathbb{R}^{m \times n} be the input data matrix packing all n=ns+ntn = n_s + n_t examples across mm shared features, and let H=In−1n11TH = I_n - \frac{1}{n} \mathbf{1}\mathbf{1}^T be the centering matrix, where InI_n is the n×nn \times n identity matrix and 1\mathbf{1} is the n×1n \times 1 column vector of all ones.

    To simultaneously maximize the variance of the projected data while minimizing the Maximum Mean Discrepancy (MMD) across domains for both the marginal distributions (c=0c=0) and class-conditional distributions for each class c∈{1,…,C}c \in \{1, \dots, C\}, JDA optimizes the transformation matrix A∈Rm×kA \in \mathbb{R}^{m \times k} (k≪mk \ll m) via the constrained objective:

    min⁡ATXHXTA=Ik∑c=0Ctr(ATXMcXTA)+λ∥A∥F2\min_{A^T X H X^T A = I_k} \sum_{c=0}^C \text{tr}\left( A^T X M_c X^T A \right) + \lambda \|A\|_F^2

    where M0∈Rn×nM_0 \in \mathbb{R}^{n \times n} is the marginal MMD matrix, Mc∈Rn×nM_c \in \mathbb{R}^{n \times n} (c=1,…,Cc=1, \dots, C) are class-conditional MMD matrices constructed using source ground-truth labels and target pseudo labels, λ>0\lambda > 0 is a regularization parameter ensuring well-posedness, and ∥⋅∥F\|\cdot\|_F denotes the Frobenius norm. Based on the generalized Rayleigh quotient, fixing the constraint ATXHXTA=IkA^T X H X^T A = I_k corresponds to maximizing data reconstruction variance as in Principal Component Analysis (PCA).

  2. Knowl 2 — MMD Coefficient Matrices for Marginal and Class-Conditional Adaptation

    definition

    In Joint Distribution Adaptation, distribution differences are quantified nonparametrically via empirical Maximum Mean Discrepancy (MMD) using coefficient matrices M0M_0 and McM_c for c∈{1,…,C}c \in \{1, \dots, C\}.

    For the marginal distribution discrepancy (c=0c=0), the matrix M0∈Rn×nM_0 \in \mathbb{R}^{n \times n} comparing nsn_s source instances Ds\mathcal{D}_s and ntn_t target instances Dt\mathcal{D}_t (n=ns+ntn = n_s + n_t) is defined elementwise by:

    (M0)ij={1ns2,xi,xj∈Ds1nt2,xi,xj∈Dt−1nsnt,otherwise(M_0)_{ij} = \begin{cases} \frac{1}{n_s^2}, & x_i, x_j \in \mathcal{D}_s \\ \frac{1}{n_t^2}, & x_i, x_j \in \mathcal{D}_t \\ -\frac{1}{n_s n_t}, & \text{otherwise} \end{cases}

    For the class-conditional distribution discrepancy of class c∈{1,…,C}c \in \{1, \dots, C\}, let Ds(c)={xi∈Ds:y(xi)=c}\mathcal{D}_s^{(c)} = \{x_i \in \mathcal{D}_s : y(x_i) = c\} with cardinality ns(c)=∣Ds(c)∣n_s^{(c)} = |\mathcal{D}_s^{(c)}| denote the source examples of class cc with true labels y(xi)y(x_i), and let Dt(c)={xj∈Dt:y^(xj)=c}\mathcal{D}_t^{(c)} = \{x_j \in \mathcal{D}_t : \hat{y}(x_j) = c\} with cardinality nt(c)=∣Dt(c)∣n_t^{(c)} = |\mathcal{D}_t^{(c)}| denote the target examples assigned to class cc via pseudo labels y^(xj)\hat{y}(x_j). The class-conditional MMD matrix Mc∈Rn×nM_c \in \mathbb{R}^{n \times n} is defined by:

    (Mc)ij={1(ns(c))2,xi,xj∈Ds(c)1(nt(c))2,xi,xj∈Dt(c)−1ns(c)nt(c),(xi∈Ds(c)∧xj∈Dt(c))∨(xj∈Ds(c)∧xi∈Dt(c))0,otherwise(M_c)_{ij} = \begin{cases} \frac{1}{(n_s^{(c)})^2}, & x_i, x_j \in \mathcal{D}_s^{(c)} \\ \frac{1}{(n_t^{(c)})^2}, & x_i, x_j \in \mathcal{D}_t^{(c)} \\ -\frac{1}{n_s^{(c)} n_t^{(c)}}, & \left(x_i \in \mathcal{D}_s^{(c)} \wedge x_j \in \mathcal{D}_t^{(c)}\right) \lor \left(x_j \in \mathcal{D}_s^{(c)} \wedge x_i \in \mathcal{D}_t^{(c)}\right) \\ 0, & \text{otherwise} \end{cases}

  3. Knowl 3 — Generalized Eigendecomposition Solution for JDA

    theoretical result

    Using Lagrange multipliers, the constrained optimization problem of Joint Distribution Adaptation (JDA):

    min⁡ATXHXTA=Iktr(AT(X∑c=0CMcXT+λIm)A)\min_{A^T X H X^T A = I_k} \text{tr}\left( A^T \left( X \sum_{c=0}^C M_c X^T + \lambda I_m \right) A \right)

    is formulated via the Lagrangian function:

    L(A,Φ)=tr(AT(X∑c=0CMcXT+λIm)A)+tr((Ik−ATXHXTA)Φ)L(A, \Phi) = \text{tr}\left( A^T \left( X \sum_{c=0}^C M_c X^T + \lambda I_m \right) A \right) + \text{tr}\left( (I_k - A^T X H X^T A) \Phi \right)

    where Φ=diag(φ1,…,φk)∈Rk×k\Phi = \text{diag}(\varphi_1, \dots, \varphi_k) \in \mathbb{R}^{k \times k} contains the Lagrange multipliers. Setting ∂L∂A=0\frac{\partial L}{\partial A} = 0 yields the generalized eigendecomposition problem:

    (X∑c=0CMcXT+λIm)A=XHXTAΦ\left( X \sum_{c=0}^C M_c X^T + \lambda I_m \right) A = X H X^T A \Phi

    The optimal transformation matrix A∈Rm×kA \in \mathbb{R}^{m \times k} is formed by the kk smallest eigenvectors corresponding to the kk smallest eigenvalues in Φ\Phi. The optimal low-dimensional representations of the data are obtained as Z=ATXZ = A^T X.

  4. Knowl 4 — Joint Distribution Adaptation Algorithm

    algorithm

    The Joint Distribution Adaptation (JDA) algorithm iteratively solves a generalized eigenvalue problem to extract domain-invariant features and updates target pseudo labels with a base classifier.

    Input: Input matrix X∈Rm×nX \in \mathbb{R}^{m \times n}, source labels ys∈{1,…,C}nsy_s \in \{1, \dots, C\}^{n_s}, subspace dimension kk, regularization parameter λ\lambda.
    Output: Transformation matrix A∈Rm×kA \in \mathbb{R}^{m \times k}, embeddings Z∈Rk×nZ \in \mathbb{R}^{k \times n}, adaptive classifier ff.
    Construct marginal MMD matrix M0M_0 using source and target index sets
    Initialize conditional MMD matrices Mc:=0M_c := 0 for $c \in \{1, \dots, C\}
    repeat
        Solve generalized eigendecomposition:
            (X∑c=0CMcXT+λIm)A=XHXTAΦ\left( X \sum_{c=0}^C M_c X^T + \lambda I_m \right) A = X H X^T A \Phi
        Select the kk smallest eigenvectors to form transformation matrix AA
        Compute embedded representations Z:=ATXZ := A^T X
        Train base classifier ff on source embeddings {(ATxi,yi)}i=1ns\{(A^T x_i, y_i)\}_{i=1}^{n_s}
        Predict target pseudo labels y^j:=f(ATxj)\hat{y}_j := f(A^T x_j) for all $j \in \{n_s + 1, \dots, n_s + n_t\}
        Construct conditional MMD matrices {Mc}c=1C\{M_c\}_{c=1}^C using updated pseudo labels y^\hat{y}
    until Convergence
    Return classifier ff trained on {(ATxi,yi)}i=1ns\{(A^T x_i, y_i)\}_{i=1}^{n_s}

    In standard implementations, the base classifier ff is a 1-Nearest Neighbor (1-NN) classifier. The iteration loop terminates upon convergence, which empirically occurs within T=10T = 10 iterations. Subspace dimension kk is set around k=100k=100, and regularization λ\lambda is set to 0.10.1 for digit/face tasks and 1.01.0 for object tasks.

  5. Knowl 5 — Kernel Joint Distribution Adaptation (Kernel-JDA)

    model/method

    For nonlinear transfer learning problems, input examples X=[x1,…,xn]X = [x_1, \dots, x_n] are implicitly mapped via ψ(x)\psi(x) into a reproducing kernel Hilbert space (RKHS) with kernel matrix K=ψ(X)Tψ(X)∈Rn×nK = \psi(X)^T \psi(X) \in \mathbb{R}^{n \times n}. According to the Representer theorem, the adaptation matrix in RKHS can be expressed as an expansion over training samples A∈Rn×kA \in \mathbb{R}^{n \times k}. The Kernel-JDA optimization problem is formulated as:

    min⁡ATKHKTA=Ik∑c=0Ctr(ATKMcKTA)+λ∥A∥F2\min_{A^T K H K^T A = I_k} \sum_{c=0}^C \text{tr}\left( A^T K M_c K^T A \right) + \lambda \|A\|_F^2

    where H=In−1n11TH = I_n - \frac{1}{n} \mathbf{1}\mathbf{1}^T is the centering matrix, M0M_0 is the marginal MMD matrix, and McM_c (c=1,…,Cc=1, \dots, C) are the class-conditional MMD matrices. The optimal matrix A∈Rn×kA \in \mathbb{R}^{n \times k} is obtained by solving the generalized eigendecomposition:

    (K∑c=0CMcKT+λIn)A=KHKTAΦ\left( K \sum_{c=0}^C M_c K^T + \lambda I_n \right) A = K H K^T A \Phi

    and taking the kk smallest eigenvectors.

  6. Knowl 6 — Computational Complexity of JDA

    theoretical result

    For an input dataset with mm features, n=ns+ntn = n_s + n_t samples across source and target domains, CC classes, kk subspace bases, and TT pseudo-label refinement iterations (under conditions where k≪min⁡(m,n)k \ll \min(m, n) and T≪min⁡(m,n)T \ll \min(m, n)):

    1. Solving the dense m×mm \times m generalized eigendecomposition takes O(Tkm2)O(T k m^2) operations across TT iterations.
    2. Constructing the C+1C+1 MMD matrices of size n×nn \times n requires O(TCn2)O(T C n^2) operations.
    3. Embedding projection and base classifier prediction take O(Tmn)O(T m n) operations.

    Summing these components, the overall computational complexity of the linear JDA algorithm is:

    O(Tkm2+TCn2+Tmn)O\left(T k m^2 + T C n^2 + T m n\right)

  7. Knowl 7 — Benchmark Datasets and Setup for Cross-Domain Image Classification

    experimental setup

    Joint Distribution Adaptation is evaluated on 36 cross-domain transfer tasks constructed from four visual benchmarks:

    1. USPS + MNIST (Digit): USPS (1,800 images) and MNIST (2,000 images) share 10 digit classes (C=10C=10), standardized to 16×1616 \times 16 pixel grayscale features (m=256m=256), forming 2 tasks: USPS →\to MNIST and MNIST →\to USPS.
    2. COIL20 (Object): 1,440 images across 20 objects (m=1,024m=1,024 grayscale pixels, C=20C=20), partitioned by view angles into COIL1 (quadrants 1 and 3: [0∘,85∘]∪[180∘,265∘][0^\circ, 85^\circ] \cup [180^\circ, 265^\circ], 720 images) and COIL2 (quadrants 2 and 4: [90∘,175∘]∪[270∘,355∘][90^\circ, 175^\circ] \cup [270^\circ, 355^\circ], 720 images), forming 2 tasks: COIL1 →\to COIL2 and COIL2 →\to COIL1.
    3. PIE (Face): 11,554 images across 68 individuals (m=1,024m=1,024, C=68C=68) split across 5 poses: PIE1 (C05, left), PIE2 (C07, upward), PIE3 (C09, downward), PIE4 (C27, frontal), PIE5 (C29, right). Pairing each pose as source and target yields 5×4=205 \times 4 = 20 cross-domain tasks.
    4. Office + Caltech-256 (Object): Four domains: Amazon (A, 958 images), Webcam (W, 295 images), DSLR (D, 157 images), and Caltech-256 (C, 1,123 images), sharing 10 classes (C=10C=10) with 800-bin z-score standardized SURF histograms (m=800m=800). Pairing each yields 4×3=124 \times 3 = 12 tasks.

    Evaluation is reported in classification Accuracy on unlabeled target data:

    Accuracy=∣{x:x∈Dt∧y^(x)=y(x)}∣∣{x:x∈Dt}∣\text{Accuracy} = \frac{|\{x : x \in \mathcal{D}_t \wedge \hat{y}(x) = y(x)\}|}{|\{x : x \in \mathcal{D}_t\}|}

  8. Knowl 8 — Cross-Domain Classification Accuracy Comparison Across 36 Datasets

    data/table

    Classification accuracy (%) of Joint Distribution Adaptation (JDA) compared to 1-Nearest Neighbor (NN), Principal Component Analysis (PCA + NN), Geodesic Flow Kernel (GFK + NN), Transfer Component Analysis (TCA + NN), and Transfer Subspace Learning (TSL + NN) across 36 cross-domain visual recognition tasks.

    Dataset NN PCA GFK TCA TSL JDA
    USPS vs MNIST 44.70 44.95 46.45 51.05 53.75 59.65
    MNIST vs USPS 65.94 66.22 67.22 56.28 66.06 67.28
    COIL1 vs COIL2 83.61 84.72 72.50 88.47 88.06 89.31
    COIL2 vs COIL1 82.78 84.03 74.17 85.83 87.92 88.47
    PIE1 vs PIE2 26.09 24.80 26.15 40.76 44.08 58.81
    PIE1 vs PIE3 26.59 25.18 27.27 41.79 47.49 54.23
    PIE1 vs PIE4 30.67 29.26 31.15 59.63 62.78 84.50
    PIE1 vs PIE5 16.67 16.30 17.59 29.35 36.15 49.75
    PIE2 vs PIE1 24.49 24.22 25.24 41.81 46.28 57.62
    PIE2 vs PIE3 46.63 45.53 47.37 51.47 57.60 62.93
    PIE2 vs PIE4 54.07 53.35 54.25 64.73 71.43 75.82
    PIE2 vs PIE5 26.53 25.43 27.08 33.70 35.66 39.89
    PIE3 vs PIE1 21.37 20.95 21.82 34.69 36.94 50.96
    PIE3 vs PIE2 41.01 40.45 43.16 47.70 47.02 57.95
    PIE3 vs PIE4 46.53 46.14 46.41 56.23 59.45 68.45
    PIE3 vs PIE5 26.23 25.31 26.78 33.15 36.34 39.95
    PIE4 vs PIE1 32.95 31.96 34.24 55.64 63.66 80.58
    PIE4 vs PIE2 62.68 60.96 62.92 67.83 72.68 82.63
    PIE4 vs PIE3 73.22 72.18 73.35 75.86 83.52 87.25
    PIE4 vs PIE5 37.19 35.11 37.38 40.26 44.79 54.66
    PIE5 vs PIE1 18.49 18.85 20.35 26.98 33.28 46.46
    PIE5 vs PIE2 24.19 23.39 24.62 29.90 34.13 42.05
    PIE5 vs PIE3 28.31 27.21 28.49 29.90 36.58 53.31
    PIE5 vs PIE4 31.24 30.34 31.33 33.64 38.75 57.01
    C →\to A 23.70 36.95 41.02 38.20 44.47 44.78
    C →\to W 25.76 32.54 40.68 38.64 34.24 41.69
    C →\to D 25.48 38.22 38.85 41.40 43.31 45.22
    A →\to C 26.00 34.73 40.25 37.76 37.58 39.36
    A →\to W 29.83 35.59 38.98 37.63 33.90 37.97
    A →\to D 25.48 27.39 36.31 33.12 26.11 39.49
    W →\to C 19.86 26.36 30.72 29.30 29.83 31.17
    W →\to A 22.96 31.00 29.75 30.06 30.27 32.78
    W →\to D 59.24 77.07 80.89 87.26 87.26 89.17
    D →\to C 26.27 29.65 30.28 31.70 28.50 31.52
    D →\to A 28.50 32.05 32.05 32.15 27.56 33.09
    D →\to W 63.39 75.93 75.59 86.10 85.42 89.49
    Average 37.46 39.84 41.19 47.22 49.80 57.37

    JDA outperforms all baselines on every dataset, achieving an overall average classification accuracy of 57.37%, representing an absolute gain of 7.57% (15.07% error reduction) over the best performing baseline (TSL at 49.80%).

  9. Knowl 9 — Empirical Verification of Embedding Properties, Convergence, and Runtime

    empirical result

    Empirical validation of Joint Distribution Adaptation (JDA) on benchmark datasets demonstrates the following properties:

    1. Distance Reduction: On PIE1 vs PIE2, evaluating true aggregate MMD distance using ground-truth labels shows that JDA achieves a substantially lower distribution distance than NN, PCA, and TCA. TCA only minimizes marginal MMD, whereas JDA reduces both marginal and conditional distribution discrepancies at each iteration.
    2. Embedding Structure: In the 20-nearest-neighbor similarity matrix of learned embeddings for the first 5 classes of PIE1 vs PIE2, JDA establishes high within-class similarity across domains (top-right and bottom-left diagonal blocks) while preserving low between-class similarity (off-diagonal blocks). By contrast, TCA exhibits substantial cross-domain between-class similarity, reducing discrimination capability.
    3. Convergence: Classification accuracy increases monotonically and MMD distance decreases monotonically with iterations, converging reliably within T=10T = 10 iterations.
    4. Runtime: On the PIE1 vs PIE2 dataset (m=1,024m = 1,024, n=4,961n = 4,961), execution time is: NN (4.85 s), PCA (2.63 s), GFK (4.58 s), TCA (3.80 s), TSL (1789 s), and JDA (46.32 s). JDA runtime scales as approximately TT times that of TCA and is orders of magnitude faster than TSL.

Coverage note — None was omitted.

References

  1. 1.A. Arnold, R. Nallapati, and W. W. Cohen. A comparative study of methods for transductive transfer learning. In Proceedings of ICDMW, 2007.
  2. 2.Y. Aytar and A. Zisserman. Tabula rasa: Model transfer for object category detection. In Proceedings of ICCV, 2011.
  3. 3.L. Bruzzone and M. Marconcini. Domain adaptation problems: A dasvm classification technique and a circular validation strategy. IEEE TPAMI, 32(5), 2010.
  4. 4.M. Chen, K. Q. Weinberger, and J. C. Blitzer. Co-training for domain adaptation. In Proceedings of NIPS, 2011.
  5. 5.N. FarajiDavar, T. de Campos, J. Kittler, and F. Yan. Transductive transfer learning for action recognition in tennis games. In Proceedings of ICCVW, 2011.
  6. 6.B. Gong, Y. Shi, F. Sha, and K. Grauman. Geodesic flow kernel for unsupervised domain adaptation. In Proceedings of CVPR, 2012.
  7. 7.R. Gopalan, R. Li, and R. Chellappa. Domain adaptation for object recognition: An unsupervised approach. In Proceedings of ICCV, 2011.
  8. 8.A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Scholkopf, and A. J. Smola. A kernel method for the two-sample problem. In Proceedings of NIPS, 2006.
  9. 9.G. Griffin, A. Holub, and P. Perona. Caltech-256 object category dataset. Technical report, Caltech, 2007.
  10. 10.M. Guillaumin and V. Ferrari. Large-scale knowledge transfer for object localization in imagenet. In Proceedings of CVPR, 2012.
  11. 11.I.-H. Jhuo, D. Liu, D.-T. Lee, and S.-F. Chang. Robust visual domain adaptation with low-rank reconstruction. In Proceedings of CVPR, 2012.
  12. 12.L. Jie, T. Tommasi, and B. Caputo. Multiclass transfer learning from unconstrained priors. In Proc. of ICCV, 2011.
  13. 13.C. H. Lampert and O. Kr¨omer. Weakly-paired maximum covariance analysis for multimodal dimensionality reduction and transfer learning. In Proceedings of ECCV, 2010.
  14. 14.C. H. Lampert, H. Nickisch, and S. Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In Proceedings of CVPR, 2009.
  15. 15.S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang. Domain adaptation via transfer component analysis. IEEE TNN, 22(2):199–210, 2011.
  16. 16.S. J. Pan and Q. Yang. A survey on transfer learning. IEEE TKDE, 22:1345–1359, 2010.
  17. 17.Q. Qiu, V. M. Patel, P. Turaga, and R. Chellappa. Domain adaptive dictionary learning. In Proceedings of ECCV, 2012.
  18. 18.B. Quanz, J. Huan, and M. Mishra. Knowledge transfer with low-quality data: A feature extraction issue. IEEE TKDE, 24(10), 2012.
  19. 19.S. D. Roy, T. Mei, W. Zeng, and S. Li. Socialtransfer: Crossdomain transfer learning from social streams for media applications. In Proceedings of ACM MM, 2012.
  20. 20.K. Saenko, B. Kulis, M. Fritz, and T. Darrell. Adapting visual category models to new domains. In Proceedings of ECCV, 2010.
  21. 21.S. Satpal and S. Sarawagi. Domain adaptation of conditional probability models via feature subsetting. In Proceedings of PKDD, 2007.
  22. 22.S. Si, D. Tao, and B. Geng. Bregman divergence-based regularization for transfer subspace learning. IEEE TKDE, 2010.
  23. 23.Q. Sun, R. Chattopadhyay, S. Panchanathan, and J. Ye. A two-stage weighting framework for multi-source domain adaptation. In Proceedings of NIPS, 2011.
  24. 24.H. Wang, F. Nie, H. Huang, and C. Ding. Dyadic transfer learning for cross-domain image classification. In Proceedings of ICCV, 2011.
  25. 25.S. Wang, S. Jiang, Q. Huang, and Q. Tian. Multi-feature metric learning with knowledge transfer among semantics and social tagging. In Proceedings of CVPR, 2012.
  26. 26.E. Zhong, W. Fan, J. Peng, K. Zhang, J. Ren, D. Turaga, and O. Verscheure. Cross domain distribution adaptation via kernel mapping. In Proceedings of KDD, 2009.

Citation

MLA
Long, M., et al. “Transfer Feature Learning with Joint Distribution Adaptation”. 2013 IEEE International Conference on Computer Vision, 2013, pp. 2200–07, https://doi.org/10.1109/ICCV.2013.274.
APA
Long, M., Wang, J., Ding, G., Sun, J., & Yu, P. S. (2013). Transfer Feature Learning with Joint Distribution Adaptation. 2013 IEEE International Conference on Computer Vision, 2200–2207. https://doi.org/10.1109/ICCV.2013.274
Chicago
Long, M., J. Wang, G. Ding, J. Sun, and P. S. Yu. 2013. “Transfer Feature Learning with Joint Distribution Adaptation”. 2013 IEEE International Conference on Computer Vision, 2200–2207. https://doi.org/10.1109/ICCV.2013.274.
Harvard
Long, M. et al. (2013) “Transfer Feature Learning with Joint Distribution Adaptation”, 2013 IEEE International Conference on Computer Vision. IEEE, pp. 2200–2207. Available at: https://doi.org/10.1109/ICCV.2013.274.
Vancouver
1. Long M, Wang J, Ding G, Sun J, Yu PS (2013) Transfer Feature Learning with Joint Distribution Adaptation. In: 2013 IEEE International Conference on Computer Vision. IEEE, pp 2200–2207

BibTeX

@inproceedings{Long_2013, title={Transfer Feature Learning with Joint Distribution Adaptation}, url={http://dx.doi.org/10.1109/ICCV.2013.274}, DOI={10.1109/iccv.2013.274}, booktitle={2013 IEEE International Conference on Computer Vision}, publisher={IEEE}, author={Long, Mingsheng and Wang, Jianmin and Ding, Guiguang and Sun, Jiaguang and Yu, Philip S.}, year={2013}, month=Dec, pages={2200–2207} }
Metadata:Crossref

Access the Paper

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

Open PDF

License: IEEE