Visual Domain Adaptation with Manifold Embedded Distribution Alignment

Jindong Wang; Wenjie Feng; Yiqiang Chen; Han Yu; Meiyu Huang; Philip S. Yu

Source-linked AI summary

Visual Domain Adaptation with Manifold Embedded Distribution Alignment

Jindong Wang, Wenjie Feng, Yiqiang Chen, Han Yu, Meiyu Huang, Philip S. Yu

arXiv:1807.07258v2Jul 19, 2018cs.CVcs.LG

TL;DR

Visual domain adaptation must reduce cross-domain divergence despite distorted feature spaces and the unequal importance of marginal and conditional distributions. MEDA learns a domain-invariant classifier on the Grassmann manifold with dynamic distribution alignment, and experiments report improved classification accuracy over traditional and deep methods.

Problem
Existing adaptation methods either align distributions in distorted original feature spaces or learn subspaces without fully reducing divergence, while treating marginal and conditional distributions as equally important.
Method
MEDA learns a domain-invariant classifier in the Grassmann manifold with structural risk minimization and dynamically evaluates the relative importance of marginal and conditional distributions.
Results
3.5% average classification-accuracy improvement over state-of-the-art traditional and deep methods was reported across 7 real-world image datasets.
Takeaways & Limitations
MEDA provides manifold-based adaptation with quantitative weighting of marginal and conditional distribution alignment.
Takeaways & Limitations
The domain-adaptation formulation assumes shared feature and label spaces between source and target domains.

Abstract

from arXiv · show

Visual domain adaptation aims to learn robust classifiers for the target domain by leveraging knowledge from a source domain. Existing methods either attempt to align the cross-domain distributions, or perform manifold subspace learning. However, there are two significant challenges: (1) degenerated feature transformation, which means that distribution alignment is often performed in the original feature space, where feature distortions are hard to overcome. On the other hand, subspace learning is not sufficient to reduce the distribution divergence. (2) unevaluated distribution alignment, which means that existing distribution alignment methods only align the marginal and conditional distributions with equal importance, while they fail to evaluate the different importance of these two distributions in real applications. In this paper, we propose a Manifold Embedded Distribution Alignment (MEDA) approach to address these challenges. MEDA learns a domain-invariant classifier in Grassmann manifold with structural risk minimization, while performing dynamic distribution alignment to quantitatively account for the relative importance of marginal and conditional distributions. To the best of our knowledge, MEDA is the first attempt to perform dynamic distribution alignment for manifold domain adaptation. Extensive experiments demonstrate that MEDA shows significant improvements in classification accuracy compared to state-of-the-art traditional and deep methods.

1 INTRODUCTION

Visual domain adaptation addresses cross-domain recognition by reducing distribution divergence, but existing approaches face degenerated feature transformation and unevaluated weighting of marginal and conditional alignment. MEDA combines manifold learning with dynamic distribution alignment and improves average accuracy in experiments.

Domain adaptation leverages labeled source-domain samples because obtaining sufficient labeled target-domain data is expensive and time-consuming.
Existing approaches reduce divergence through instance reweighting, subspace learning, or marginal and conditional distribution alignment.
Subspace learning does not eliminate feature divergence, while alignment in the original feature space is hindered by distorted features.
MEDA learns a domain-invariant classifier in the Grassmann manifold and dynamically weights marginal and conditional distributions.
MEDA provides quantitative evaluation of the relative importance of marginal and conditional distributions in domain adaptation.
3.5% average classification-accuracy improvement was achieved over state-of-the-art traditional and deep methods on 7 real-world image datasets.

2 RELATED WORK

Related methods separately address subspace structure, distribution alignment, or domain-invariant classifier learning. MEDA differs by aligning distributions on a manifold while evaluating their relative importance.

Subspace learning: Subspace Alignment aligns domain base vectors but fails to adapt feature distributions, while SDA and CORAL add distribution or statistical alignment with remaining limitations.
Subspace learning: Geodesic-flow and Grassmann-based methods model domain relationships on manifolds, but prior Grassmann adaptation ignored conditional distribution alignment.
Distribution alignment: Joint distribution adaptation and its extensions align marginal and conditional distributions equally, preventing evaluation of their differing importance.
Distribution alignment: MEDA evaluates each distribution’s quantitative importance and aligns distributions in the manifold to overcome feature distortion in the original space.
Domain-invariant classifier learning: ARTL and DMM remain vulnerable to original-space feature distortion, while DIP focuses on feature transformation and aligns only marginal distributions.

3 MANIFOLD EMBEDDED DISTRIBUTION ALIGNMENT

MEDA combines manifold feature learning with dynamic distribution alignment to learn a domain-invariant classifier through structural risk minimization. It transforms features into the Grassmann manifold, weights marginal and conditional alignment according to an adaptive factor, and exploits manifold geometry.

Overall approach: MEDA uses structural risk minimization to learn a domain-invariant classifier while adding constraints that promote source–target distribution consistency.The objective includes sample loss, regularization, dynamic distribution alignment, and Laplacian regularization.
Manifold feature learning: Manifold feature learning maps source and target representations into the Grassmann manifold to reduce distortion associated with transformations in the original feature space.MEDA models domains as d-dimensional subspaces and uses geodesic flow between them; GFK is used for computational efficiency.
Dynamic distribution alignment: Dynamic distribution alignment combines marginal and class-conditional alignment through an adaptive factor µ rather than assigning them fixed equal weights.When µ approaches 0, marginal alignment is emphasized; when µ approaches 1, conditional alignment is emphasized; µ = 0.5 treats them equally.
Dynamic distribution alignment: The alignment term measures distribution divergence with maximum mean discrepancy, including projected MMD for alignment associated with the learned classifier.MMD is computed in an RKHS using differences between embedded sample means.
Dynamic distribution alignment: MEDA estimates µ quantitatively from global and local domain structure using the A-distance, and recomputes the estimate during each adaptation iteration.The estimate can change because the feature distribution may vary after conditional distribution evaluation.
Classifier learning: MEDA can learn a cross-domain function directly without explicit classifier training, unlike methods that require a separate classifier-learning stage.This property distinguishes MEDA from JGSA and CORAL as described by the authors.

4 EXPERIMENTS AND EVALUATIONS

MEDA is evaluated on seven public image datasets spanning multiple cross-domain recognition tasks, using the paper’s manifold-alignment algorithm and classifier-learning procedure.

Experiments use seven public image datasets, including Office+Caltech10, USPS+MNIST, ImageNet+VOC2007, and Office-31.
The MEDA algorithm learns manifold features, generates target soft labels, constructs a transformed-feature kernel, updates the classifier, and returns f.
Office+Caltech10 provides 12 transfer tasks among Amazon, Webcam, DSLR, and Caltech domains with 10 common classes.
USPS+MNIST contributes two digit-recognition tasks, U →M and M →U, across different image distributions.
ImageNet and VOC2007 contribute two five-class transfer tasks, I →V and V →I.

4.2 State-of-the-art Comparison Methods

MEDA is compared with traditional feature-learning and distribution-alignment methods, as well as deep domain-adaptation baselines, across benchmark accuracy tables.

The study compares MEDA with several state-of-the-art traditional and deep domain-adaptation approaches.
The comparison reports accuracy on Office+Caltech10 using SURF and DeCaf6 features, and on USPS+MNIST and ImageNet+VOC2007.
Traditional learning methods: Traditional baselines include 1NN, SVM, PCA, TCA, GFK, JDA, TJM, ARTL, CORAL, and related adaptation methods.
Deep domain adaptation methods: Deep baselines include AlexNet, DDC, DAN, Deep CORAL, and DUCDA.

4.3 Experimental Setup

The experiments use extracted features for traditional methods, target-domain accuracy as the metric, and parameter settings including manifold dimensions, iterations, kernels, and regularization.

Experimental setup: Traditional methods use 256 SURF features for USPS+MNIST, 800 SURF and 4,096 DeCaf6 features for Office+Caltech10, and 4,096 DeCaf6 features elsewhere.
Experimental setup: Deep methods operate on the original images rather than the extracted features used by traditional methods.
Parameter sensitivity: Figure 3 compares accuracy in original and manifold spaces across values of µ, with dashed lines marking the best baseline.
Parameter setting: MEDA uses manifold dimensions d = 20, 30, 40 across the Office+Caltech10, USPS+MNIST, and ImageNet+VOC datasets, respectively.
Parameter setting: The experiments run T = 10 iterations with an RBF kernel and regularization parameters p = 10, λ = 10, η = 0.1, and ρ = 1.
Evaluation metric: Classification Accuracy on the target domain is the evaluation metric, comparing predicted and true target labels.

4.4 Experimental Results and Analysis

Across 28 tasks, MEDA generally outperforms traditional, subspace, distribution-alignment, and deep baselines, while component studies evaluate its classifier-learning terms.

Overall comparison: MEDA outperformed competing traditional and deep methods on 21 of 28 tasks.
Overall comparison: 73.2% average accuracy exceeded JGSA’s 69.7% by 3.5%, corresponding to an 11.6% average error reduction.
Interpretation: The analysis attributes MEDA’s scope to reducing distribution divergence across a wide range of image datasets.
Traditional methods: MEDA generally surpassed distribution-alignment and subspace-learning methods, whose limitations correspond to residual domain shift and feature distortion.
Deep methods: MEDA outperformed AlexNet, DDC, DAN, DCORAL, and DUCDA on Office+Caltech10 datasets.

4.5 Effectiveness Analysis

MEDA’s ablations support manifold feature learning, dynamic distribution alignment, and estimation of the adaptive factor as effective components of domain adaptation.

Manifold feature learning: 25.8% mean-accuracy improvement was observed on task C →A when manifold feature learning was included.The comparison was computed as (56.5 −44.9)/44.9 × 100%.
Dynamic distribution alignment: Optimal µ values varied across tasks, taking values 0.2, 0, and 1, supporting task-dependent distribution alignment.Different µ values can produce the same classification result on some tasks.
Adaptive-factor estimation: Estimated ˆµ produced performance close to µopt and sometimes exceeded grid search, with a 0.9% performance variation on C →A.The reported variation was (57.0 −56.5)/57.0 × 100%.
Component ablation: SRM, dynamic distribution alignment, and Laplacian regularization were each reported as important and indispensable to MEDA.The experiments also found distribution alignment especially important and Laplacian regularization beneficial for capturing manifold structure.

4.6 Parameter Sensitivity

MEDA remained robust across broad parameter ranges, so several parameters could be selected without extensive fine-tuning in real applications.

Practical selection: The experiments concluded that MEDA parameters do not need to be fine-tuned in real applications.This conclusion follows the reported parameter-sensitivity experiments.
Manifold parameters: MEDA was robust across wide ranges of manifold subspace dimension d and neighbor count p.The tested ranges were d ∈{10, 20, · · · , 100} and p ∈{2, 4, · · · , 64}.
Regularization parameters: MEDA maintained robust performance across a wide range of regularization parameters λ, η, and ρ.Reported best-choice ranges included λ ∈[0.5, 1, 000] and η ∈[0.01, 1].

4.7 Convergence and Time Complexity

MEDA reached steady performance quickly and had running-time complexity comparable to strong traditional baselines, while deep-method timing was not directly comparable.

Convergence: MEDA reached steady performance in fewer than 10 iterations during empirical convergence analysis.The paper presents this as a training advantage in cross-domain tasks.
Time complexity: MEDA achieved running-time complexity comparable to ARTL and JGSA while also reporting superior classification accuracy.The timing comparison used an Intel Core i7-4790 CPU with 24 GB memory.

5 CONCLUSIONS

The paper concludes that MEDA addresses both identified domain-adaptation challenges through manifold-based classification and dynamic distribution alignment, with strong experimental performance.

Conclusion: MEDA addresses degenerated feature transformation and unevaluated distribution alignment in visual domain adaptation.The conclusion identifies these as the two challenges handled by the approach.
Conclusion: MEDA learns a domain-invariant classifier using structural risk minimization while performing dynamic distribution alignment.The method also provides a feasible solution for quantitatively calculating the adaptive factor.
Conclusion: Extensive experiments on large-scale public image-classification datasets reported superiority over state-of-the-art traditional and deep domain-adaptation methods.The conclusion states this as the result of the empirical evaluation.