PaperlayerView exact revision on arXiv

Source-linked AI summary

Aligning Domain-specific Distribution and Classifier for Cross-domain Classification from Multiple Sources

Yongchun Zhu, Fuzhen Zhuang, Deqing Wang

arXiv:2201.01003v1cs.LGcs.AIcs.CV

TL;DR

The paper targets MUDA settings where diverse source domains make shared domain-invariant representations difficult and leave domain-specific class boundaries underused. It proposes separate source-target feature alignments followed by classifier-output alignment, and reports strong benchmark performance across multi-source transfer tasks. The method assumes unlabeled target samples and uses domain-specific feature extractors and classifiers.

  • Problem

    MUDA must handle multiple diverse source domains, but prior methods commonly seek shared representations without considering domain-specific class boundaries.

  • Method

    The framework learns separate domain-invariant representations for each source-target pair, trains domain-specific classifiers, and aligns their target outputs using decision boundaries.

  • Results

    MFSAN outperforms compared baseline methods on most multi-source transfer tasks, while its representation and classifier-alignment variants support the proposed design.

  • Takeaways & Limitations

    The results indicate that combining pair-specific representations with domain-specific class-boundary consideration is important for MUDA.

  • Takeaways & Limitations

    The target data are sampled without observed labels.

Abstract

from arXiv · show

While Unsupervised Domain Adaptation (UDA) algorithms, i.e., there are only labeled data from source domains, have been actively studied in recent years, most algorithms and theoretical results focus on Single-source Unsupervised Domain Adaptation (SUDA). However, in the practical scenario, labeled data can be typically collected from multiple diverse sources, and they might be different not only from the target domain but also from each other. Thus, domain adapters from multiple sources should not be modeled in the same way. Recent deep learning based Multi-source Unsupervised Domain Adaptation (MUDA) algorithms focus on extracting common domain-invariant representations for all domains by aligning distribution of all pairs of source and target domains in a common feature space. However, it is often very hard to extract the same domain-invariant representations for all domains in MUDA. In addition, these methods match distributions without considering domain-specific decision boundaries between classes. To solve these problems, we propose a new framework with two alignment stages for MUDA which not only respectively aligns the distributions of each pair of source and target domains in multiple specific feature spaces, but also aligns the outputs of classifiers by utilizing the domain-specific decision boundaries. Extensive experiments demonstrate that our method can achieve remarkable results on popular benchmark datasets for image classification.

Introduction

The paper addresses MUDA, where multiple diverse source domains make distribution alignment harder than in SUDA and common domain-invariant representations may be inadequate. It proposes two-stage alignment of domain-specific distributions and classifier outputs, supported by experiments on benchmark datasets.

  • MUDA is practically important because labeled data are often available from multiple source domains.
  • Combining multiple sources into one domain may improve performance through data expansion, but the improvement may be limited.
  • Common feature-space alignment is difficult because shifts among multiple sources increase the mismatch between source and target domains.
  • The proposed framework first aligns each source-target pair in a separate feature space to learn multiple domain-invariant representations.
  • It then aligns classifier outputs for target samples using domain-specific decision boundaries.
  • The authors conduct comprehensive experiments on three well-known benchmarks to validate the proposed model.

Related Work

Related work covers SUDA methods for reducing source-target discrepancy and MUDA methods that use multiple source-specific models or shared representations. The paper distinguishes its approach by learning separate feature representations for source-target pairs and aligning classifier outputs using domain-specific boundaries.

  • Single-source Unsupervised Domain Adaptation (SUDA): SUDA methods include reweighting training data and transforming source and target subspaces into closer lower-dimensional manifolds.
  • Single-source Unsupervised Domain Adaptation (SUDA): Deep SUDA methods commonly match distribution mean embeddings or use domain discriminators to learn domain-invariant features.
  • Single-source Unsupervised Domain Adaptation (SUDA): Other SUDA approaches reconstruct features into images or transform whitened source features using target covariance.
  • Multi-source Unsupervised Domain Adaptation (MUDA): MUDA is needed because dataset shifts among multiple source domains prevent straightforward combination into one source domain.
  • Multi-source Unsupervised Domain Adaptation (MUDA): Prior deep MUDA methods use source-target classifiers and discriminators or combine gradients from multiple domain discriminators while learning common representations.
  • Multi-source Unsupervised Domain Adaptation (MUDA): The proposed approach instead maps each source-target pair into multiple feature spaces and aligns classifier outputs using domain-specific decision boundaries.

Method

MFSAN uses shared and source-specific network components to align each source–target pair in separate feature spaces, then reduces disagreement among source-specific classifiers on target samples.

  • Domain-specific distribution alignment: The first alignment stage matches each source–target pair’s distributions in a specific feature space to learn multiple domain-invariant representations.The framework uses MMD to estimate discrepancy between each source domain and the target domain.
  • Architecture: MFSAN combines a common feature extractor with N unshared domain-specific extractors and N source-specific classifiers.The common extractor maps all domains into a shared feature space, while each domain-specific extractor maps a source–target pair into its own space.
  • Source classification: Each domain-specific classifier is a softmax predictor trained with cross-entropy on labeled source-domain samples.The classifier receives the domain-specific representation produced after the common and source-specific feature extractors.
  • Domain-specific classifier alignment: The second alignment stage minimizes disagreement among classifiers’ probabilistic outputs for target samples, particularly near class boundaries.At prediction time, the framework averages the outputs of all classifiers rather than using a separate voting operator.
  • Optimization: Training updates the common extractor, all domain-specific extractors, and all classifiers using classification, MMD, and classifier-discrepancy losses.The procedure samples labeled source images and unlabeled target images, computes the respective representations and losses, and minimizes the total objective.

Experiments

Experiments evaluate MFSAN against domain-adaptation baselines across ImageCLEF-DA, Office-31, and Office-Home, with ablations, visualizations, convergence, and sensitivity analyses. Results show benefits from combining sources, using multiple feature spaces, and aligning domain-specific classifiers.

  • Benchmark evaluation: MFSAN is evaluated on three benchmark datasets: ImageCLEF-DA, Office-31, and Office-Home.The evaluation covers multiple transfer-task combinations across these datasets.
  • Benchmark evaluation: Source Combine performs better than Single Best in most transfer tasks, indicating that combining source domains is helpful.The authors attribute this improvement to data enrichment.
  • Benchmark evaluation: MFSAN outperforms all compared baseline methods on most multi-source transfer tasks.The comparison includes deep SUDA and MUDA baselines under source-combine, single-best, and multi-source standards.
  • Ablation analysis: MFSANmmd outperforms DAN with source combination, contrasting multiple domain-invariant representations in multiple feature spaces with common representations in one feature space.The authors interpret this result as evidence that extracting common representations for all domains is difficult.
  • Ablation analysis: MFSAN with disc loss produces classifiers with smaller performance gaps than MFSANmmd, supporting the effectiveness of classifier-output alignment.The ablation compares variants with mmd loss, disc loss, or both.
  • Feature visualization: In t-SNE visualizations for task D →A and D, W →A, MFSAN representations are reported as better than those from DAN comparisons.The comparison is used to validate aligning both domain-specific distributions and classifiers.
  • Convergence and sensitivity: All algorithms almost converge after 1.5×10^4 iterations, while MFSAN with disc loss has smaller classifier gaps and higher accuracy.The convergence comparison is conducted on task D, W →A.

Conclusion

The paper proposes MFSAN, a general MUDA framework that jointly aligns source-target distributions in multiple feature spaces and aligns outputs from multiple source classifiers. Experiments on three image datasets demonstrate the framework’s effectiveness.

  • MFSAN simultaneously learns multiple domain-invariant representations by aligning each source-target pair in a separate feature space.The framework addresses domain-specific distributions rather than forcing all domains into one common representation.
  • MFSAN also aligns classifier outputs from multiple sources while learning domain-specific representations.Its classifier-alignment component incorporates domain-specific decision-boundary information.
  • Experiments on three image datasets demonstrate the effectiveness of the proposed framework.
  • MFSAN can integrate different kinds of MMD and discrepancy-loss functions.
Loading 2201.01003v1…