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

1 Introduction

• Domain shift motivates adaptation, but collecting extensive labeled data in a new domain is costly, and conventional UDA and SSL methods can perform worse than training on labeled source and target samples in SSDA.The introduction notes that deep domain adaptation has largely focused on unsupervised methods, while deep semi-supervised learning has focused on identical-domain settings.Source: §1 · pp. 1–2
• In SSDA, labeled-target supervision attracts corresponding features and their neighborhoods toward source clusters, while less-correlated unlabeled target samples remain unaligned, creating intra-domain discrepancy.This divides the target distribution into aligned and unaligned subdistributions, helping explain why UDA and SSL methods can underperform training on labeled source and target samples alone.Source: §1 · pp. 2–3
• The proposed framework uses attraction to align target subdistributions, perturbation to propagate labels without disrupting already aligned features, and class-aware exploration to modulate prototypes.The perturbation scheme places target subdistributions in an intermediate region; exploration locally modulates class prototypes as a complement to attraction and perturbation.Source: §1 · p. 3
• The authors report state-of-the-art performance against vanilla deep neural networks, UDA, SSL, and SSDA methods on DomainNet, Office-Home, and Office.The introduction describes extensive experiments on these domain adaptation datasets.Source: §1 · p. 3

2.1 Unsupervised Domain Adaptation

• UDA methods reduce cross-domain divergence, translate domain appearance, or treat the source as partially labeled data for semi-supervised learning.Examples include MMD and adversarial confusion for divergence reduction, and Drop-to-Adapt with virtual adversarial perturbation.Source: §2.1 · p. 4

2.2 Semi-supervised Learning

• DA addresses data from distributions with domain discrepancy, whereas SSL assumes labeled and unlabeled data come from the same distribution.Deep SSL methods include augmentation consistency, adversarial perturbation, prediction or parameter ensembling, self-training, and augmentation distribution alignment.Source: §2.2 · p. 4

2.3 Semi-supervised Domain Adaptation

• Prior SSDA methods use labeled target data to address domain mismatch, but none discuss or mitigate intra-domain discrepancy, which this paper makes its focus.Prior approaches include auxiliary constraints, subspace learning, soft-label interpolation, and minimax training to bring unlabeled target samples closer to class prototypes. The visualization shows that three labeled target samples per class can attract their neighborhoods, separating target features into aligned and unaligned subdistributions.Source: §2.3 · pp. 4–5, Figure 2 · p. 5

• Intra-domain discrepancy is an internal distribution gap among subdistributions; unlike SSDA, UDA typically has insufficiently distinct target subdistributions because domain samples have unique correlations.Source: §3 · p. 5
• In SSDA, labeled target samples deterministically divide the target distribution into aligned and unaligned subdistributions, leaving some unlabeled samples distant from source clusters and inaccurately predicted.The paper reports this pattern even when only three target labels per class are given.Source: §3 · p. 6
• Inter-domain discrepancy minimization underestimates the gap and works less effectively in SSDA, and can negatively affect the class-aware aligned target subdistribution.Most target errors occur in the unaligned subdistribution, so SSDA depends on aligning it.Source: §3 · p. 6
• Conventional SSL assumptions fail in SSDA because target samples do not share one distribution, so label propagation can spread errors among wrongly predicted samples and is meaningless among correctly predicted ones.The authors propose a framework to address intra-domain discrepancy.Source: §3 · p. 6

4.1 Problem Formulation

• SSDA trains on labeled source, labeled target, and unlabeled target samples to improve target-domain discriminability.Source: §4.1 · pp. 6–7

4.2 Spherical Feature Space with Prototypes

• A similarity-based prototypical classifier trains normalized features and class prototypes in a spherical space using source and labeled target data.The classifier reduces intra-class variation; the proposed schemes align normalized feature distributions in that space. The classifier approach is adopted from.Source: §4.2 · p. 7

4.3 Attraction Scheme

• Attraction globally aligns unaligned and aligned target subdistributions by minimizing their estimated intra-domain discrepancy with multi-width Gaussian-kernel MMD.Because labeled target samples may be too few to represent the aligned subdistribution, the estimate uses the complex distribution of labeled source and target data.Source: §4.3 · p. 8

4.4 Perturbation scheme

• The perturbation scheme moves target subdistributions toward an intermediate region to propagate predictions and modulate prototypes without disrupting already aligned features.It perturbs prototypes in an entropy-maximizing direction, optimizes bounded feature perturbations toward them, and regularizes perturbed and original data with KL divergence; conventional perturbation can cause negative transfer in SSDA.Source: §4.4 · pp. 8–9

4.5 Exploration scheme

• Exploration locally modulates prototypes and aligns selected unlabeled target features whose element-wise entropy is below a threshold, using the nearest prototype’s class.This complements global attraction with class-aware alignment and enables aligned features to be perturbed farther toward the unaligned subdistribution.Source: §4.5 · pp. 9–10

4.6 Overall framework and training objective

• The overall objective is a weighted sum of supervision, attraction, perturbation, and exploration losses within the integrated framework.Figure 3 depicts the feature extractor, trainable class prototypes, supervision module, and modules for the proposed schemes, with normalized features and prototypes in the same spherical feature space.Source: Figure 3 · p. 7, §4.6 · p. 10

5.1 Experimental Setup

• Experiments covered DomainNet, Office-Home, and Office using one- or three-shot target labels, with additional 5-, 10-, and 20-label comparisons.The study included seven DomainNet scenarios, 12 Office-Home scenarios, and two Office scenarios; DomainNet has six domains and 345 classes, Office-Home four domains and 65 classes, and Office three domains and 31 classes.Source: §5.1 · pp. 10–11
• The comparison included SSDA, UDA, SSL, and non-adaptation baselines, with all methods using the same backbone architecture.Baselines included MME, SagNet, DANN, ADR, CDAN, ENT, and S+T; labeled target samples supervised the UDA methods during training.Source: §5.1 · p. 11

5.2 Results

• 2.8% (ResNet-34, three-shot), 1.2% (ResNet-34, one-shot), 0.7% (AlexNet, three-shot), and 0.4% (AlexNet, one-shot) were DomainNet gains over the best baseline.The method outperformed most individual DomainNet scenarios, with a few exceptions.Source: §5.2 · p. 12
• Our method outperformed all baselines on average on Office-Home across both backbones and outperformed the other baselines on Office.It performed best in most Office-Home scenarios. On Office, the comparison used a three-shot setting.Source: Table 2 · p. 11, §5.2 · pp. 12–13, Table 3 · p. 12

5.3 Analysis

• With more target labels, our method remained superior on ResNet-34 throughout the tested cases, while ENT surpassed state-of-the-art SSDA methods above 20 labels on AlexNet and five on ResNet-34.The comparison varied labeled target samples from 0 to 20 per class; the authors relate ENT’s gains at higher label ratios to the shift toward an SSL setting.Source: Figure 4 · p. 12, §5.3 · p. 13
• On DomainNet, attraction improved performance on average while DANN and MMD rarely helped or caused negative transfer; the proposed perturbation improved results while VAT degraded them.The authors attribute the perturbation scheme’s effect to directing perturbations toward the intermediate region of target subdistributions; exploration complemented the other schemes.Source: §5.3 · pp. 13–14, Table 4 · p. 13
• On Real-to-Sketch DomainNet features, visualizations show source-target clusters aligning over training and target clusters becoming well accumulated by the final stage.The visualization used 20 of the 126 classes for display, without selecting classes by top-20 accuracy.Source: §5.3 · p. 14, Figure 5 · p. 14

6 Conclusions

• The method outperformed conventional UDA and SSL methods on DomainNet, Office-Home, and Office.The experiments on these datasets validate the method’s effectiveness.Source: §6 · p. 14
• Attraction minimizes estimated target-domain intra-domain discrepancy, while perturbation moves aligned and unaligned target features into intermediate regions of target subdistributions.The perturbation scheme addresses both well-aligned and unaligned target features.Source: §6 · p. 14
• Exploration selectively aligns features locally and class-wise, complementing attraction and perturbation.Together, the three schemes address intra-domain discrepancy while aligning cross-domain feature distributions.Source: §6 · p. 14

Additional Performance Comparisons with Varying Number of Target Labels

• In 5-shot and 10-shot DomainNet settings, the method outperformed all baselines by a considerable average margin and led in every scenario but one.Experiments used ResNet-34 as the backbone.Source: §1 · pp. 17–18
• At five shots, ENT matched S+T, while at ten shots it outperformed the state-of-the-art SSDA method on average.These settings correspond to 1.8% ∼6.7% and 3.6% ∼13.5% of target samples, respectively; the paper attributes ENT’s results to SSDA increasingly resembling target-domain SSL as labeled target samples increase.Source: §1 · p. 18
• Conventional domain adaptation methods performed less effectively than S+T or even worse in the five-shot and ten-shot comparisons.The baselines included conventional domain adaptation, conventional semi-supervised learning, and a state-of-the-art SSDA method.Source: §1 · p. 18