Source-linked AI summary
Cross-Domain Adaptive Clustering for Semi-Supervised Domain Adaptation
Jichang Li, Guanbin Li, Yemin Shi, Yizhou Yu
TL;DR
SSDA is challenged by limited labeled target data and source-dominated training, which can weaken target-domain feature representations. CDAC addresses this with adversarial adaptive clustering, cluster-wise cross-domain alignment, and high-confidence pseudo-labeling. Experiments report state-of-the-art performance across benchmark SSDA datasets.
Problem
Few labeled target samples and source-dominated training limit discriminative target-domain representations and can disconnect or misalign target features.
Method
CDAC groups unlabeled target features into clusters, aligns clusters across domains, and uses high-confidence pseudo-labels to strengthen target cluster cores.
Results
CDAC achieves state-of-the-art semi-supervised domain adaptation performance on DomainNet, Office-Home, and Office.
Takeaways & Limitations
Cluster-wise alignment jointly addresses inter-domain and intra-domain adaptation while stronger target cluster cores facilitate adversarial learning.
Takeaways & Limitations
The SSDA setting assumes only one-shot or few-shot labeled target data, with N_l much smaller than N_s and N_u.
Abstract
from arXiv · showhide
In semi-supervised domain adaptation, a few labeled samples per class in the target domain guide features of the remaining target samples to aggregate around them. However, the trained model cannot produce a highly discriminative feature representation for the target domain because the training data is dominated by labeled samples from the source domain. This could lead to disconnection between the labeled and unlabeled target samples as well as misalignment between unlabeled target samples and the source domain. In this paper, we propose a novel approach called Cross-domain Adaptive Clustering to address this problem. To achieve both inter-domain and intra-domain adaptation, we first introduce an adversarial adaptive clustering loss to group features of unlabeled target data into clusters and perform cluster-wise feature alignment across the source and target domains. We further apply pseudo labeling to unlabeled samples in the target domain and retain pseudo-labels with high confidence. Pseudo labeling expands the number of ``labeled" samples in each class in the target domain, and thus produces a more robust and powerful cluster core for each class to facilitate adversarial learning. Extensive experiments on benchmark datasets, including DomainNet, Office-Home and Office, demonstrate that our proposed approach achieves the state-of-the-art performance in semi-supervised domain adaptation.
1. Introduction
SSDA must address both inter-domain and intra-domain gaps, but few labeled target samples and source-dominated training limit target feature discrimination. CDAC combines cluster-wise alignment, adversarial adaptive clustering, and high-confidence pseudo-labeling to improve target cluster cores and reported performance.
- SSDA uses a small number of labeled target samples to complement labeled source data, while adaptation must reduce both inter-domain and intra-domain gaps.
- Source-dominated training can leave unlabeled target features disconnected from labeled target samples and mismatched with source features.
- CDAC groups unlabeled target features into clusters and performs cluster-wise feature alignment across source and target domains.This is intended to achieve inter-domain and intra-domain adaptation simultaneously.
- High-confidence pseudo-labeling expands labeled target data and forms more robust, powerful class cluster cores for adversarial learning.
- Extensive experiments on DomainNet, Office-Home, and Office report state-of-the-art SSDA performance for CDAC.
2. Related Work
Prior work includes adversarial feature alignment, pseudo-labeling, and target-domain intra-domain adaptation. CDAC differs by using adaptive cluster-wise alignment to address both cross-domain alignment and target-domain feature cohesion.
- Adversarial Learning for UDA: Adversarial domain adaptation methods commonly pursue sample-wise or distribution-wise feature alignment across domains.
- Adversarial Learning for UDA: CDAC introduces adaptive cluster-wise feature alignment to achieve inter-domain and intra-domain adaptation together.
- Pseudo Labeling on UDA: Pseudo-labeling methods assign reliable labels to unlabeled data using prediction information to support semi-supervised training.
- Pseudo Labeling on UDA: CDAC uses high-confidence pseudo-labels to expand labeled target samples and strengthen class cluster cores for adversarial learning.
- Semi-supervised Domain Adaptation: SSDA leverages one or a few labeled target samples per class to improve domain adaptation through target supervision.
- Semi-supervised Domain Adaptation: Earlier SSDA approaches combine adversarial cross-domain alignment with methods that reduce target-domain intra-domain discrepancy.
3. Methodology
CDAC combines adversarial adaptive clustering, pseudo labeling, and consistency training to jointly address intra-domain clustering and cross-domain feature alignment in SSDA.
- Overall Loss: Standard cross-entropy supervision is applied to labeled source and target data alongside the proposed unsupervised objectives.The network uses a feature extractor and normalized linear classifier to produce probabilistic predictions.
- Adversarial Adaptive Clustering: CDAC introduces adversarial adaptive clustering to group unlabeled target features and perform cluster-wise alignment across source and target domains.The approach targets inter-domain and intra-domain adaptation simultaneously rather than relying only on sample-wise or distribution-wise alignment.
- Adversarial Adaptive Clustering: Pairwise similarities among unlabeled target features become binary targets for a cross-entropy loss that encourages similar samples to share classifier predictions.Similarity is derived from shared top-k ranked feature indices, with k set to 5.
- Adversarial Adaptive Clustering: Gradient reversal adversarially maximizes the clustering loss for the feature extractor while classification supervision remains enforced through the classifier.This design reduces source-dominated bias and encourages more domain-invariant target representations.
- Pseudo Labeling: High-confidence pseudo-labels expand the labeled target set and establish more robust class cluster cores for cross-domain cluster alignment.The method uses progressive pseudo labeling and retains labels according to a confidence threshold τ.
- Overall Loss: The model trains with predictions from differently transformed unlabeled images and applies consistency loss to keep those predictions consistent.Two transformed versions support the adversarial clustering and pseudo-labeling losses, while consistency training increases input diversity.
4. Experiments
Experiments evaluate CDAC on DomainNet, Office-Home, and Office under few-shot settings, comparing it with established SSDA and UDA methods. CDAC achieves strong benchmark performance, while ablations and analyses associate its gains with adaptive clustering, pseudo labeling, and improved cross-domain feature alignment.
- Experimental setup: CDAC is evaluated on DomainNet, Office-Home, and Office using standard 1-shot or 3-shot semi-supervised domain adaptation settings.DomainNet uses Real, Clipart, Painting, and Sketch domains; Office-Home and Office are also evaluated under established benchmark protocols.
- Benchmark comparisons: CDAC significantly outperforms competing methods across experiments with Alexnet and Resnet34 backbones.Results are reported in Tables 1–3 for 1-shot and 3-shot settings, with the strongest performance observed throughout the reported experiments.
- Benchmark comparisons: 6.6% and 6.8% are CDAC’s average gains over the previous best method on DomainNet with Alexnet under 1-shot and 3-shot settings, respectively.With Resnet34, CDAC exceeds the current best results by 6% and 4.3% under 1-shot and 3-shot settings.
- Ablation studies: 25.3% higher average performance results when adaptive clustering and pseudo-labeling losses are combined with cross-entropy than with the SSDA baseline.Using only adaptive clustering or pseudo-labeling yields average gains of 17.6% and 23.4%, respectively.
- Adaptive clustering analysis: Cluster Core Distance decreases during training, with CDAC reaching the minimum final value and better feature alignment than S+T and CDAC without pseudo labeling.The analysis measures distances between same-class source and target feature clusters to assess cluster-wise alignment.
- Pseudo-labeling analysis: 59.9% and 63.8% of total training examples receive correct pseudo-labels at best performance in the R→S and P→C DomainNet scenarios, respectively.The scenarios use Resnet34 with 3-shot and 1-shot settings, respectively.
5. Conclusions
The paper presents CDAC for SSDA, combining adversarial adaptive clustering, cluster-wise cross-domain alignment, and adapted pseudo labeling. Experiments and ablations validate the method's effectiveness.
- CDAC addresses SSDA by grouping unlabeled target features into clusters and aligning those clusters across domains.
- Its adapted pseudo-labeling component strengthens target-domain cluster cores to facilitate adversarial learning.
- Extensive experiments and ablation studies validate the proposed method.