PaperlayerView exact revision on arXiv

Source-linked AI summary

Model Adaptation: Unsupervised Domain Adaptation without Source Data

Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, Si Wu

arXiv:2502.19316v1cs.CV

TL;DR

The paper studies unsupervised model adaptation, where an existing source model must improve on a new domain using only unlabeled target data. It proposes 3C-GAN and regularization to enable source-free adaptation, and reports superior performance across multiple benchmarks, including 98.5% accuracy on MNIST→MNIST-M. A stated limitation is that the adversarial loss alone cannot guarantee semantic similarity to the input label.

  • Problem

    Unsupervised model adaptation seeks to improve an existing source prediction model on a target domain using only unlabeled target data because source data may be unavailable for privacy or practical reasons.

  • Method

    3C-GAN generates target-style training samples for collaborative improvement of the generator and prediction model, supplemented by weight and clustering-based regularization.

  • Results

    98.5% accuracy on MNIST→MNIST-M, and the method outperforms recent state-of-the-art results across multiple adaptation benchmarks without source data.

  • Takeaways & Limitations

    The results support adapting pretrained prediction models across domains without retaining or accessing source datasets.

  • Takeaways & Limitations

    The adversarial loss simulates the target distribution but cannot guarantee semantic similarity to the input label y.

Abstract

from arXiv · show

In this paper, we investigate a challenging unsupervised domain adaptation setting -- unsupervised model adaptation. We aim to explore how to rely only on unlabeled target data to improve performance of an existing source prediction model on the target domain, since labeled source data may not be available in some real-world scenarios due to data privacy issues. For this purpose, we propose a new framework, which is referred to as collaborative class conditional generative adversarial net to bypass the dependence on the source data. Specifically, the prediction model is to be improved through generated target-style data, which provides more accurate guidance for the generator. As a result, the generator and the prediction model can collaborate with each other without source data. Furthermore, due to the lack of supervision from source data, we propose a weight constraint that encourages similarity to the source model. A clustering-based regularization is also introduced to produce more discriminative features in the target domain. Compared to conventional domain adaptation methods, our model achieves superior performance on multiple adaptation tasks with only unlabeled target data, which verifies its effectiveness in this challenging setting.

1. Introduction

The paper addresses domain shift by adapting an existing source prediction model using only unlabeled target data, avoiding source-data dependence. It introduces 3C-GAN with regularization and reports strong performance across adaptation benchmarks.

  • Domain shift can seriously degrade visual recognition when test environments differ from the source domain.
  • Source-data adaptation is often impractical because privacy, security, storage, and transfer constraints may prevent access to large source datasets.
  • Unsupervised model adaptation transfers a pretrained source model across domains using only unlabeled target data under a shared label-space assumption.
  • Weight and clustering-based regularization stabilize adaptation, preserve source-model knowledge, and improve target-domain generalization.
  • 3C-GAN generates target-style training samples while the generator and prediction model collaboratively improve one another without source data.
  • Experiments on multiple benchmarks show that the method is effective and outperforms recent state-of-the-art results without source data.

2. Related Work

Related work primarily aligns source and target distributions or translates source data, while other methods use target data through co-training, pseudo-labeling, or entropy regularization. These approaches generally remain inapplicable when source data is unavailable during adaptation.

  • Most domain adaptation methods reduce source-target distribution discrepancy using source labels and domain-alignment objectives.
  • Generative approaches align feature or image distributions by generating source-like or target-like data with adversarial models.
  • Unsupervised image translation adds reconstruction, class-label, or semantic-feature constraints to preserve content without paired domain data.
  • Co-training, pseudo-labeling, and entropy regularization exploit unlabeled target data, but these methods still require source data during adaptation.
  • Incremental learning methods are related but require labeled target data for new tasks, unlike the proposed setting.

3. Proposed Method

The proposed method adapts a source prediction model using only unlabeled target data by alternating target-style generation and model adaptation. It combines collaborative class-conditional generation with source-weight preservation and clustering-based regularization.

  • Collaborative Class Conditional GAN: 3C-GAN adapts a pre-trained source prediction model using an unlabeled target dataset without accessing source data.The framework uses a discriminator and a label-conditioned generator alongside the existing prediction model.
  • Collaborative Class Conditional GAN: The generator produces target-style samples conditioned on labels and noise, while the discriminator distinguishes generated samples from real target samples.The generator is updated to match the target distribution, and the prediction model provides semantic guidance.
  • Collaborative Class Conditional GAN: The generator and prediction model collaborate iteratively: generated instances improve the model, whose enhanced predictions provide more accurate guidance for generation.The two procedures are repeated during training.
  • Weight Regularization: Weight regularization keeps the adapted model close to the fixed source model, stabilizing adaptation and preserving source knowledge.The authors report that this regularization improves adaptation in most cases.
  • Clustering-based Regularization: Clustering-based regularization minimizes target conditional entropy and adds local smoothness so decision boundaries avoid dense target regions.The local smoothness constraint encourages similar predictions for a target sample and its adversarially perturbed version.

4. Experiments

Experiments evaluate source-free model adaptation across digit, sign, Office-31, and VisDA17 benchmarks, with visualizations and ablations examining generation, feature clustering, and regularization. The method improves target-domain performance across these settings while using source data only to obtain the pretrained model.

  • Experimental setup: Experiments use digit, sign, Office-31, and VisDA17 adaptation benchmarks, with source data excluded during adaptation.Office-31 evaluations use ResNet50, while VisDA17 evaluations use ResNet101.
  • Digit and sign results: 98.5% accuracy on MNIST→MNIST-M improves the Source-Only baseline by around 40%, while the model increases target-domain performance on all digit and sign tasks.The method achieves over 99% accuracy on SVHN→MNIST, USPS→MNIST, and Syn.Sign→GTSRB.
  • Office-31 results: Around 3% and 4% average gains over GenToAdapt and MADA, respectively, are reported across six Office-31 adaptation tasks.On the four A⇆D and A⇆W tasks, the method exceeds the second-best method by 4.5% on average.
  • VisDA17 results: 81.6% class mean accuracy is achieved on VisDA17 with vanilla ResNet101, increasing to 83.3% with an enhanced backbone.The enhanced result surpasses SimDA with ResNet152 by 10.4%, and the method exceeds self-ensembling by 9.1% without data augmentation.
  • Visualization analysis: Late-stage generated samples have better quality than early-stage samples, indicating collaboration between the prediction model and generator during adaptation.The visualizations condition samples on class labels while sharing noise vectors across rows, and the generator learns class-conditional target distributions.
  • Ablation study: Target features form stronger class clusters after adaptation, while removing generation causes non-convergence and adding clustering regularization improves accuracy by around 1 to 3 percentage points.On MNIST→MNIST-M, clustering raises accuracy from 94.2% to 97.0%; on MNIST→USPS, it raises accuracy from 95.4% to 97.0%. Weight regularization further stabilizes performance, and removing the smoothness constraint reduces accuracy on A⇆W and A⇆D.

5. Conclusion

The paper proposes 3C-GAN for source-free unsupervised model adaptation, combining collaborative learning with weight and clustering regularization. Across multiple benchmarks, it achieves the best or comparable performance without source data.

  • 3C-GAN bypasses dependence on source data by incorporating generated images into adaptation.The prediction model and generator are mutually enhanced through collaborative learning.
  • Weight regularization stabilizes training by encouraging similarity to the source model.
  • Clustering-based regularization further improves target-domain generalization by stabilizing training.The paper introduces it alongside weight regularization to improve adaptation performance.
  • The method achieves the best or comparable performance against recent data-based adaptation methods without source data.Experiments span multiple domain adaptation benchmarks and support effectiveness across a broad class of scenarios.
Loading 2502.19316v1…