Source-linked AI summary
Domain Adaptation with Auxiliary Target Domain-Oriented Classifier
Jian Liang, Dapeng Hu, Jiashi Feng
TL;DR
Domain adaptation must transfer knowledge across shifted domains, while source-biased pseudo-labeling can produce low-quality target labels. ATDOC introduces a target-only auxiliary classifier using memory-based nearest-centroid and neighborhood-aggregation methods, and it achieves strong results across diverse adaptation and scarce-labeled SSL settings.
Problem
Source-trained pseudo-labeling ignores domain shift, biases predictions toward labeled source data, and can propagate errors in the target domain.
Method
ATDOC introduces a target-only auxiliary classifier with memory-based nearest-centroid and neighborhood-aggregation classifiers, without additional network parameters.
Results
ATDOC-NA consistently outperforms feature-level alignment and other regularization methods across UDA, SSDA, PDA, and scarce-labeled SSL tasks; with MixMatch it reaches 86.3% mean accuracy on VisDA-C.
Takeaways & Limitations
ATDOC can be integrated with existing domain-invariant feature-learning methods and further boost adaptation performance.
Takeaways & Limitations
The paper focuses on the vanilla closed-set setting, where source and target domains share the same categories.
Abstract
from arXiv · showhide
Domain adaptation (DA) aims to transfer knowledge from a label-rich but heterogeneous domain to a label-scare domain, which alleviates the labeling efforts and attracts considerable attention. Different from previous methods focusing on learning domain-invariant feature representations, some recent methods present generic semi-supervised learning (SSL) techniques and directly apply them to DA tasks, even achieving competitive performance. One of the most popular SSL techniques is pseudo-labeling that assigns pseudo labels for each unlabeled data via the classifier trained by labeled data. However, it ignores the distribution shift in DA problems and is inevitably biased to source data. To address this issue, we propose a new pseudo-labeling framework called Auxiliary Target Domain-Oriented Classifier (ATDOC). ATDOC alleviates the classifier bias by introducing an auxiliary classifier for target data only, to improve the quality of pseudo labels. Specifically, we employ the memory mechanism and develop two types of non-parametric classifiers, i.e. the nearest centroid classifier and neighborhood aggregation, without introducing any additional network parameters. Despite its simplicity in a pseudo classification objective, ATDOC with neighborhood aggregation significantly outperforms domain alignment techniques and prior SSL techniques on a large variety of DA benchmarks and even scare-labeled SSL tasks.
1. Introduction
Domain adaptation transfers knowledge across shifted domains, but source-trained classifiers can produce biased target pseudo labels. ATDOC addresses this bias with a target-only auxiliary classifier built using memory-based non-parametric methods.
- Domain adaptation transfers knowledge from label-rich source data to label-scarce target scenarios affected by dataset shift.
- Source-trained classifiers may classify target samples poorly when target data lies far from the source domain.
- Pseudo-labeling can propagate target-domain errors because predictions are biased toward labeled source data.
- ATDOC introduces a target-only auxiliary classifier to generate more target-oriented pseudo labels without additional network parameters.
- ATDOC-NA uses a memory bank and neighborhood aggregation to build a domain-specific classifier and produce pseudo labels with confidence weights.
- ATDOC reports competitive or better results across unsupervised, semi-supervised, partial-set adaptation, and scarce-labeled SSL settings.
2. Related Work
Related work addresses adaptation through domain-invariant representations, domain-specific modeling, SSL regularization, pseudo-labeling, and memory mechanisms. ATDOC combines target-oriented pseudo-labeling with unsupervised memory updates to address limitations in these approaches.
- Deep domain adaptation commonly aligns feature distributions to learn transferable representations shared across source and target domains.
- Other adaptation methods specialize domain-specific encoders, private components, or batch-normalization layers instead of fully sharing model parameters.
- SSL methods exploit unlabeled data through assumptions or regularizers such as entropy minimization, consistency, graph propagation, and augmentation.
- Pseudo-labeling treats high-confidence predictions as labels and retrains with cross-entropy, while later work addresses noise and class imbalance.
- Memory banks support global structure learning under mini-batch optimization by retaining information from earlier samples.
- ATDOC differs from closely related memory-based SSL by updating its class-centroid memory bank unsupervisedly and using self-learning on unlabeled data.
3. Methodology
ATDOC addresses source-classifier bias in unsupervised and semi-supervised domain adaptation by learning target-oriented pseudo labels from unlabeled-data structure. It uses a memory module with non-parametric nearest-centroid and neighborhood-aggregation classifiers that can complement existing adaptation methods.
- Problem setting: Domain adaptation here targets labeling unlabeled target samples from source and target data, mainly under the shared-category setting.The framework covers unsupervised adaptation without labeled target data and semi-supervised adaptation with a labeled target subset.
- Pseudo-labeling motivation: Standard pseudo-labeling selects maximum-probability predictions but can remain biased toward labeled source data and propagate target-domain errors.A gradually increasing loss coefficient can mitigate early pseudo-label noise, but conventional regularization does not fully use unlabeled-data structure.
- ATDOC framework: ATDOC introduces an auxiliary target-domain classifier to improve pseudo labels under domain shift without adding parametric classifier networks.The method uses a memory mechanism to exploit information across target samples while avoiding trivial sample selection and alternating training.
- Nearest centroid classifier: The nearest-centroid classifier stores target-oriented class centroids in a memory bank, updates them with pseudo labels using exponential moving averages, and assigns labels using centroid distance.Cosine distance is used by default, and the resulting pseudo labels train a standard cross-entropy objective.
- Neighborhood aggregation: Neighborhood aggregation retrieves nearest target neighbors from a feature-and-prediction memory bank and averages their soft predictions to capture global target structure.Prediction sharpening and class balancing prepare stored predictions, while confidence-weighted cross-entropy trains on the aggregated pseudo labels.
- Training objective: ATDOC losses can be combined with labeled-data training and integrated into existing domain-alignment or semi-supervised learning methods.For unlabeled target data, the objective combines supervised label-smoothed cross-entropy with either nearest-centroid or neighborhood-aggregation pseudo-label loss.
4. Experiments
Experiments evaluate ATDOC across closed-set, semi-supervised, partial-set, and scarce-labeled settings using several domain-adaptation benchmarks. ATDOC-NA generally achieves the strongest results, benefits from integration with existing methods, and improves as its neighborhood and training strategies are tuned.
- Experimental setup: Experiments cover Office-31, Office-Home, VisDA-C, and DomainNet-126 across closed-set UDA, SSDA, PDA, and scarce-labeled SSL tasks.Office-31 has six transfer tasks, Office-Home has twelve, VisDA-C evaluates synthetic-to-real transfer, and DomainNet-126 uses four domains and 126 classes.
- Closed-set UDA: ATDOC-NA obtains the best average accuracy on Office-31 and outperforms BNM in four of six tasks.ATDOC-NC remains competitive, while combining ATDOC-NA with CDAN+E further increases average accuracy.
- Closed-set UDA: 86.3% mean accuracy is achieved by ATDOC-NA combined with MixMatch on VisDA-C, outperforming recent UDA methods.ATDOC-NA also beats BNM and MCC, and on Office-Home it outperforms ATDOC-NC and BNM in average accuracy.
- Semi-supervised DA: ATDOC-NA outperforms BNM and MCC under both one-shot and three-shot SSDA protocols on DomainNet-126.The difference between ATDOC-NA's one-shot and three-shot results is relatively small, and its variants outperform prior APE results.
- Partial-set UDA: ATDOC-NA achieves better average accuracy than RTNetadv on Office-Home partial-set UDA, while ATDOC-NC remains competitive.The authors attribute NC's competitiveness partly to class centroids providing a global-level representation that resists noise.
- Analysis: Linear ramp-up is effective because early pseudo labels or classifier outputs are unreliable, while ATDOC-NA accuracy grows and eventually converges.The convergence and ramp-up behavior is examined for ATDOC-NA and compared with BNM.
- Ablation study: Ablations find weighting and class-balancing sharpening effective, larger neighborhood size m beneficial, and λ = 0.2 suitable for both Office-31 and VisDA-C.Using m = 1 is risky on the small Office-31 dataset and produces worse results.
5. Conclusion
The paper presents ATDOC as a simple regularization approach for dataset shift in domain adaptation. ATDOC-NA consistently outperforms feature-level alignment and other regularization methods across several adaptation and scarce-label settings.
- ATDOC addresses dataset shift in domain adaptation through a new regularization framework.
- ATDOC-NA outperforms feature-level domain alignment and other regularization methods with consistent margins across UDA, SSDA, PDA, and scarce-labeled SSL tasks.
- The authors identify universal domain adaptation and semantic segmentation as future extensions.
6. Correction
The camera-ready version corrects a typo in Eq. (6), while retaining all previously reported experimental results.
- A typo in Eq. (6) was corrected without requiring changes to the previous experimental results.