PaperlayerView exact revision on arXiv

Source-linked AI summary

Attract, Perturb, and Explore: Learning a Feature Alignment Network for Semi-supervised Domain Adaptation

Taekyung Kim, Changick Kim

arXiv:2007.09375v1cs.CV

TL;DR

SSDA must exploit limited labeled target data while handling both domain shift and semi-supervised learning assumptions. The paper introduces intra-domain discrepancy and addresses it with attraction, perturbation, and exploration, achieving state-of-the-art performance across DomainNet, Office-Home, and Office.

  • Problem

    Conventional domain adaptation and semi-supervised learning methods often perform less effectively or negatively in SSDA because intra-domain discrepancy has not been addressed.

  • Method

    The framework alleviates intra-domain discrepancy through attraction, domain-adaptive perturbation, and class-aware exploration schemes.

  • Results

    The method achieves state-of-the-art performance on DomainNet, Office-Home, and Office.

  • Takeaways & Limitations

    Addressing intra-domain discrepancy provides the paper's route to aligning cross-domain features for SSDA.

Abstract

from arXiv · show

Although unsupervised domain adaptation methods have been widely adopted across several computer vision tasks, it is more desirable if we can exploit a few labeled data from new domains encountered in a real application. The novel setting of the semi-supervised domain adaptation (SSDA) problem shares the challenges with the domain adaptation problem and the semi-supervised learning problem. However, a recent study shows that conventional domain adaptation and semi-supervised learning methods often result in less effective or negative transfer in the SSDA problem. In order to interpret the observation and address the SSDA problem, in this paper, we raise the intra-domain discrepancy issue within the target domain, which has never been discussed so far. Then, we demonstrate that addressing the intra-domain discrepancy leads to the ultimate goal of the SSDA problem. We propose an SSDA framework that aims to align features via alleviation of the intra-domain discrepancy. Our framework mainly consists of three schemes, i.e., attraction, perturbation, and exploration. First, the attraction scheme globally minimizes the intra-domain discrepancy within the target domain. Second, we demonstrate the incompatibility of the conventional adversarial perturbation methods with SSDA. Then, we present a domain adaptive adversarial perturbation scheme, which perturbs the given target samples in a way that reduces the intra-domain discrepancy. Finally, the exploration scheme locally aligns features in a class-wise manner complementary to the attraction scheme by selectively aligning unlabeled target features complementary to the perturbation scheme. We conduct extensive experiments on domain adaptation benchmark datasets such as DomainNet, Office-Home, and Office. Our method achieves state-of-the-art performances on all datasets.

1 Introduction

The paper identifies intra-domain discrepancy as a central SSDA issue and proposes three complementary schemes to alleviate it. Experiments across DomainNet, Office-Home, and Office report state-of-the-art performance.

  • 1 Introduction: The paper introduces intra-domain discrepancy, arising when labeled-target supervision aligns some target features while leaving others unaligned.This separates the target domain into aligned and unaligned subdistributions.
  • 1 Introduction: The proposed framework combines attraction, perturbation, and exploration to align target features by reducing intra-domain discrepancy.Attraction aligns target subdistributions, perturbation regularizes intermediate target regions, and exploration modulates class prototypes selectively.
  • 1 Introduction: The perturbation scheme moves target subdistributions toward their intermediate region to propagate labels without damaging already aligned features.
  • 1 Introduction: The exploration scheme locally modulates class prototypes and selectively aligns unlabeled target features in complement to the other schemes.
  • 1 Introduction: The attraction scheme aligns the unaligned target subdistribution to the aligned subdistribution through intra-domain discrepancy minimization.
  • 1 Introduction: The method achieves state-of-the-art performance among deep neural network, UDA, SSL, and SSDA baselines.

2 Related Work

Related work distinguishes domain adaptation from semi-supervised learning and positions SSDA as a task bridging source and target distributions with limited target labels. The paper argues that prior methods do not address SSDA's intra-domain discrepancy.

  • 2 Related Work: Prior UDA methods reduce cross-domain divergence, translate domain appearance, or learn domain-invariant representations.
  • 2 Related Work: Domain adaptation addresses data from distinct distributions, whereas semi-supervised learning assumes labeled and unlabeled data come from an identical distribution.
  • 2 Related Work: SSDA bridges a source distribution toward a target distribution using partially labeled target samples, but relatively few works have studied it.
  • 2 Related Work: The paper states that existing methods do not discuss or mitigate intra-domain discrepancy in SSDA.

3 Intra-domain Discrepancy

Intra-domain discrepancy is an internal gap between target subdistributions that SSDA supervision can make deterministic. This gap limits conventional domain adaptation and semi-supervised learning methods because aligned and unaligned target samples require different treatment.

  • 3 Intra-domain Discrepancy: Intra-domain discrepancy is an internal distribution gap among subdistributions within a domain.
  • 3 Intra-domain Discrepancy: In SSDA, labeled-target supervision pulls samples and neighborhoods toward source clusters while less-correlated unlabeled samples remain distant, creating aligned and unaligned subdistributions.
  • 3 Intra-domain Discrepancy: Conventional domain adaptation methods are less suitable because reducing inter-domain discrepancy does not directly align the unaligned target subdistribution where most errors occur.
  • 3 Intra-domain Discrepancy: Conventional SSL methods can propagate errors within wrongly predicted target subdistributions because SSDA violates SSL's identical-distribution assumption.

4 Method

The method adapts SSDA in a spherical feature space by combining attraction, domain-adaptive perturbation, and class-aware exploration with supervised training.

  • Spherical Feature Space with Prototypes: The prototypical classifier trains normalized embeddings and class prototypes using source and labeled target samples, while the schemes align their spherical feature distributions.The classifier uses similarity-based prototypes to reduce intra-class variation.
  • Attraction scheme: Attraction minimizes the estimated discrepancy between aligned and unaligned target subdistributions to achieve global target alignment.The scheme estimates feature-distribution divergence between target subdistributions and directly minimizes it.
  • Perturbation scheme: Domain-adaptive perturbation moves target samples toward an intermediate region, propagates predictions, modulates prototypes, and regularizes perturbed and original data with KL divergence.It perturbs prototypes along an entropy-maximizing direction before optimizing bounded sample perturbations.
  • Exploration scheme: Exploration selectively aligns low-entropy unlabeled target samples using the nearest prototype’s class, complementing global attraction and perturbation.The selected set is defined by an entropy threshold and filtered with an indicator function.
  • Overall framework: The framework combines supervision, attraction, perturbation, and exploration losses into one weighted training objective.Its feature extractor, trainable class prototypes, and scheme-specific modules operate in a shared spherical feature space.

5 Experiments

Experiments evaluate the method across DomainNet, Office-Home, and Office using multiple adaptation scenarios, backbones, baselines, and target-label budgets. The method generally outperforms competing approaches, while ablations and feature visualizations analyze how its schemes contribute.

  • Experimental Setup: Experiments cover 7 DomainNet, 12 Office-Home, and 2 Office adaptation scenarios, using one or three labeled target samples per class and additional 5-, 10-, and 20-shot comparisons.The study uses standard domain-adaptation benchmarks and compares against SSDA, UDA, SSL, and no-adaptation baselines.
  • Experimental Setup: The experiments use AlexNet and ResNet-34 backbones with matched labeled source and target samples and twice as many unlabeled target samples per mini-batch.Optimization uses SGD with learning rate 0.01, momentum 0.9, and weight decay 0.0005.
  • Performance Comparison on DomainNet: The method outperformed the best baseline by 2.8% in three-shot and 1.2% in one-shot settings on ResNet-34, with smaller gains on AlexNet across DomainNet scenarios.The reported average gains were 0.7% in three-shot and 0.4% in one-shot settings on AlexNet.
  • Performance comparison with varying number of target labels: With more labeled target samples, the method maintained superior performance across all tested cases on ResNet-34, including scenarios where one- or three-shot performance was weaker.The comparison varied labeled target samples from 0 to 20 per class on the Real-to-Clipart DomainNet scenario.
  • Feature alignment analysis: The t-SNE visualization tracks feature-alignment progress through the proposed method during training.The visualization is presented as six panels, labeled (a) through (f).
  • Ablation and scheme analysis: Conventional adversarial perturbation degraded performance, whereas the proposed perturbation scheme improved it by directing target features toward the intermediate region between subdistributions.The exploration scheme also complemented the other schemes in the ablation analysis.

6 Conclusions

The paper identifies intra-domain discrepancy in SSDA and addresses it through attraction, perturbation, and exploration schemes. Experiments on three benchmarks validate the framework against conventional UDA and SSL methods.

  • 6 Conclusions: The framework aligns cross-domain feature distributions by alleviating target-domain intra-domain discrepancy through attraction, perturbation, and exploration.Attraction minimizes discrepancy, perturbation moves target features toward an intermediate region, and exploration performs selective class-wise alignment.
  • 6 Conclusions: The method outperformed conventional UDA and SSL methods on DomainNet, Office-Home, and Office.The conclusion reports validation across all three datasets.

Domain Adaptation (Supplementary Material)

The supplementary material reports additional DomainNet comparisons using five-shot and ten-shot settings with a ResNet-34 backbone.

  • Domain Adaptation (Supplementary Material): Additional comparisons evaluate DomainNet across five-shot and ten-shot settings using ResNet-34.The cited table covers classification accuracy across seven scenarios.

1 Additional Results and Analysis

Additional DomainNet experiments show that the proposed method remains strong as the number of labeled target samples increases, while conventional methods can become less effective and ENT can become competitive.

  • Additional Performance Comparisons with Varying Number of Target Labels: The method outperformed all baselines on average in both five-shot and ten-shot DomainNet settings, with superior results in all but one scenario.The comparison includes domain-adaptation, semi-supervised, and state-of-the-art SSDA baselines.
  • Additional Performance Comparisons with Varying Number of Target Labels: As target-label counts increase, conventional domain-adaptation methods can perform worse than S+T, while ENT becomes comparable at five shots and surpasses the SSDA method at ten shots on average.The five- and ten-shot settings correspond to 1.8%–6.7% and 3.6%–13.5% of target samples, respectively.
Loading 2007.09375v1…