Adversarial Reweighting with ɑ-Power Maximization for Domain Adaptation

Xiang Gu, Xi Yu, Yan Yang, Jian Sun, Zongben Xu

TL;DR

This paper introduces Adversarial Reweighting with $\alpha$-Power Maximization (ARPM), a novel framework designed for Partial Domain Adaptation (PDA) where the source domain contains extra "private" classes not present in the target domain. ARPM mitigates the harmful effects of "negative transfer" by adversarially down-weighting irrelevant source samples and introduces a powerful $\alpha$-power maximization loss to sharply reduce target prediction uncertainty. The method sets new performance benchmarks across multiple domain adaptation settings, including partial, open-set, universal, and test-time adaptation.

What Problem Does This Paper Address?

Standard Unsupervised Domain Adaptation (UDA) assumes that the source and target domains share the exact same set of classes. However, real-world scenarios are rarely so clean. In Partial Domain Adaptation (PDA), the source domain is large and diverse, containing "private" classes that do not exist in the target domain.

When traditional DA methods try to align the feature distributions of the two domains globally, they suffer from negative transfer: target domain samples are mistakenly aligned with these source-private classes, severely degrading classification accuracy.

Illustration of negative transfer caused by source-private classes in PDA. Target samples (blue) are incorrectly aligned with source-private classes (red squares).

The challenge lies in automatically identifying and filtering out these irrelevant source classes without access to any target labels.

Methods

The authors propose ARPM, which tackles PDA through three main components:

  1. Adversarial Reweighting Model: To eliminate the influence of source-private classes, the model learns an individual weight for each source sample. It minimizes the Wasserstein distance between the target feature distribution and the reweighted source feature distribution. Through an adversarial game, a discriminator tries to distinguish the domains, while the reweighting module learns to assign smaller weights to source-private samples to "fool" the discriminator.
The framework minimizes the Wasserstein distance between the reweighted source feature distribution and the target feature distribution to learn weights.
  1. $\alpha$-Power Maximization: Instead of using standard entropy minimization to reduce prediction uncertainty on the target domain, ARPM maximizes the sum of the $\alpha$-power of the classification scores: $$H_\alpha(\mathbf{p}) = \sum_k p_k^\alpha$$ For larger values of $\alpha$ (the authors use $\alpha=6$), this loss effectively ignores highly uncertain samples near the decision boundary during training, focusing instead on pushing ambiguous samples toward confident, one-hot predictions.
Intuitive motivation of ARPM: Source data is reweighted, and target samples are pushed toward confident source common class regions using $\alpha$-power maximization.
  1. Neighborhood Reciprocity Clustering: To ensure local consistency, this module enforces that target samples that are reciprocal nearest neighbors in feature space produce similar classification outputs, boosting overall robustness.

Datasets

The framework was evaluated on five standard PDA benchmarks:

  • Office-31: 4,652 images across 31 categories from three domains (Amazon, DSLR, Webcam).
  • Office-Home: A challenging dataset with 65 classes across four domains (Art, Clip Art, Product, Real-World).
  • VisDA-2017: A large-scale synthetic-to-real dataset containing 12 classes.
  • ImageNet-Caltech: Built from the shared classes of ImageNet and Caltech-256.
  • DomainNet: A massive, noisy benchmark spanning six domains and 345 classes.

Key Insights and Ideas

  • Sample-level vs. Class-level Reweighting: Unlike previous methods that assign a single weight to an entire class, ARPM learns individual weights per sample. This allows it to down-weight specific source samples that are outliers or less relevant to the target domain, even within shared classes.
  • Why $\alpha$-Power Beats Entropy: Standard entropy minimization can be unstable because its gradient remains large even for highly uncertain samples. By contrast, $\alpha$-power maximization (with $\alpha > 2$) yields near-zero gradients for highly ambiguous samples (the center of the probability simplex). This prevents noisy, incorrect target predictions from corrupting the feature extractor during early training phases.
  • Versatility Beyond PDA: The core mechanism of managing uncertainty via learned weights naturally extends to other complex "open-world" settings, such as Open-Set DA, Universal DA, and Test-Time Adaptation (TTA).

Key Results

ARPM consistently outperformed state-of-the-art PDA methods across all benchmarks:

  • Office-Home: Achieved a top average accuracy of 81.8%, outperforming the next best method (MOT) by 1.2%.
  • VisDA-2017: Reached 93.2% accuracy, a solid improvement over MOT (92.4%) and SLM (91.7%).
  • DomainNet: Showed a massive jump to 75.0% average accuracy compared to the conference version's 65.5%.
  • Test-Time Adaptation (TTA): On ImageNet-R, their Test $\alpha$-Power Maximization (TPM) variant achieved 38.3% accuracy, outperforming the popular TENT method (36.8%).

Limitations and Future Work

The authors note that both the adversarial reweighting model and the $\alpha$-power maximization mechanism fundamentally rely on having access to target domain data during the adaptation phase. Consequently, extending this approach to Domain Generalization (DG)—where the target domain is completely unseen during training—remains a non-trivial challenge. Future work will explore adapting these uncertainty-driven reweighting mechanisms to target-free scenarios.

Summary updated

Abstract

The practical Domain Adaptation (DA) tasks, e.g., Partial DA (PDA), open-set DA, universal DA, and test-time adaptation, have gained increasing attention in the machine learning community. In this paper, we propose a novel approach, dubbed Adversarial Reweighting with $\alpha$-Power Maximization (ARPM), for PDA where the source domain contains private classes absent in target domain. In ARPM, we propose a novel adversarial reweighting model that adversarially learns to reweight source domain data to identify source-private class samples by assigning smaller weights to them, for mitigating potential negative transfer. Based on the adversarial reweighting, we train the transferable recognition model on the reweighted source distribution to be able to classify common class data. To reduce the prediction uncertainty of the recognition model on the target domain for PDA, we present an $\alpha$-power maximization mechanism in ARPM, which enriches the family of losses for reducing the prediction uncertainty for PDA. Extensive experimental results on five PDA benchmarks, i.e., Office-31, Office-Home, VisDA-2017, ImageNet-Caltech, and DomainNet, show that our method is superior to recent PDA methods. Ablation studies also confirm the effectiveness of components in our approach. To theoretically analyze our method, we deduce an upper bound of target domain expected error for PDA, which is approximately minimized in our approach. We further extend ARPM to open-set DA, universal DA, and test time adaptation, and verify the usefulness through experiments.
BibTeX
@misc{gu2024adversarial,
  author = {Xiang Gu and Xi Yu and Yan Yang and Jian Sun and Zongben Xu},
  title = {{Adversarial Reweighting with ɑ-Power Maximization for Domain Adaptation}},
  year = {2024},
  eprint = {2404.17275},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.17275},
}
Figure