Source-linked AI summary
Cross-Domain Gradient Discrepancy Minimization for Unsupervised Domain Adaptation
Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, Ke Lu
TL;DR
Bi-classifier adversarial adaptation can produce small classifier discrepancies while assigning target samples to wrong categories, leaving target accuracy insufficiently constrained. CGDM minimizes cross-domain gradient discrepancy using clustering-based pseudo labels and self-supervised learning, and experiments report an advantage over prior methods across large-scale UDA datasets.
Problem
Bi-classifier adversarial methods can agree on wrong target categories because output discrepancy alone does not guarantee target accuracy.
Method
CGDM uses source-target gradient discrepancy as surrogate supervision, with clustering-based pseudo labels and self-supervised learning for target samples.
Results
The method's experiments on three large-scale UDA datasets demonstrate its advantage over previous methods.
Takeaways & Limitations
Gradient discrepancy is presented as a way to improve category-level distribution alignment while preserving target-sample classification accuracy.
Takeaways & Limitations
The method assumes labeled source data and unlabeled target data drawn from source and target distributions.
Abstract
from arXiv · showhide
Unsupervised Domain Adaptation (UDA) aims to generalize the knowledge learned from a well-labeled source domain to an unlabeled target domain. Recently, adversarial domain adaptation with two distinct classifiers (bi-classifier) has been introduced into UDA which is effective to align distributions between different domains. Previous bi-classifier adversarial learning methods only focus on the similarity between the outputs of two distinct classifiers. However, the similarity of the outputs cannot guarantee the accuracy of target samples, i.e., target samples may match to wrong categories even if the discrepancy between two classifiers is small. To challenge this issue, in this paper, we propose a cross-domain gradient discrepancy minimization (CGDM) method which explicitly minimizes the discrepancy of gradients generated by source samples and target samples. Specifically, the gradient gives a cue for the semantic information of target samples so it can be used as a good supervision to improve the accuracy of target samples. In order to compute the gradient signal of target samples, we further obtain target pseudo labels through a clustering-based self-supervised learning. Extensive experiments on three widely used UDA datasets show that our method surpasses many previous state-of-the-arts. Codes are available at https://github.com/lijin118/CGDM.
1. Introduction
UDA transfers knowledge from labeled source data to unlabeled target data, but bi-classifier methods can align target samples incorrectly because classifier agreement does not ensure accuracy. CGDM addresses this by using cross-domain gradient discrepancy as supervision, with clustering-based pseudo labels and self-supervised learning.
- Motivation: UDA transfers models trained on labeled source data to similar target domains without labels, reducing reliance on expensive annotation.The paper motivates UDA with costly large-scale labeling, including in medical image analysis.
- Existing approaches: Adversarial domain adaptation learns domain-invariant representations, but domain discriminators can neglect target category information and reduce feature discriminability.Bi-classifier adversarial learning is presented as another adversarial paradigm alongside domain-discriminator methods.
- Problem: Small discrepancy between two classifiers cannot guarantee accurate target predictions because both classifiers may assign the same wrong category.Ambiguous target samples can therefore produce inaccurate class-wise distribution alignment in later adversarial optimization.
- Proposed method: CGDM explicitly minimizes the discrepancy between gradient vectors produced by source and target samples, using gradient signals as surrogate supervision for target accuracy.The method aims to align the optimization paths of the two domains, not merely their final classifier outputs.
- Proposed method: Clustering-based pseudo labels support target-gradient computation, while self-supervised learning fine-tunes the model with source and target data to reduce ambiguous target samples.The contribution describes this as a generalized learning loss that can be applied to other UDA paradigms.
2. Related Work
UDA research includes sample-weighted, metric-learning, and adversarial approaches, with pseudo-labeling increasingly used to exploit target semantics. CGDM combines clustering-based pseudo labels with a bi-classifier adversarial framework and gradient-discrepancy alignment.
- UDA paradigms: Existing UDA methods are broadly grouped into sample-weighted adaptation, metric-learning adaptation, and adversarial-learning adaptation.Metric-learning examples include MMD, DAN, JAN, CMD, and MDD.
- Adversarial adaptation: Adversarial methods learn domain-invariant representations using either an additional domain discriminator or conditional-distribution alignment strategies.The first strategy distinguishes domain-specific features while a feature learner attempts to fool the discriminator.
- Pseudo-label adaptation: Pseudo-label methods exploit target semantics for regularization, class-centroid matching, conditional alignment, or self-training.These methods differ in how pseudo labels guide target-domain adaptation.
- CGDM framework: CGDM's framework passes source and target samples through a generator and two classifiers, combining supervised, self-supervised, adversarial, and gradient-discrepancy objectives.The adversarial process detects target samples outside the source support, while gradient discrepancy aligns distributions and improves target accuracy.
- Clustering-based pseudo labels: Clustering-based methods obtain target pseudo labels in feature space, including k-means for contrastive learning and weighted k-means for source-free self-supervision.CGDM follows this clustering strategy but uses pseudo labels to minimize cross-domain gradient discrepancy rather than directly self-train.
3. Method
CGDM augments bi-classifier adversarial learning by aligning source and target gradient signals, using clustering-based pseudo labels and self-supervised learning to preserve target semantics during adaptation.
- Bi-classifier adversarial learning: Bi-classifier adversarial learning trains a generator and two task-specific classifiers, then alternates source classification, target discrepancy maximization, and generator alignment.The classifiers maximize target-output divergence while preserving source accuracy; the generator subsequently minimizes that divergence.
- Motivation: Small classifier-output discrepancy cannot ensure accurate target predictions because both classifiers may assign the same wrong category.The paper gives nearly identical predictions [0.95, 0.03, 0.02] and [0.96, 0.02, 0.02] as an example.
- Cross-domain gradient discrepancy minimization: CGDM explicitly minimizes the discrepancy between gradient vectors generated from source and target samples when updating the generator.The gradient discrepancy loss uses cosine similarity and is intended to align distributions while considering sample semantics.
- Target pseudo labels: Because target labels are unavailable, CGDM assigns pseudo labels and uses entropy-weighted classification loss to reduce the influence of ambiguous target samples.Target pseudo labels are obtained through a clustering-based strategy, with nearest-centroid assignment and cosine distance used in this paper.
- Self-supervised learning: Self-supervised learning fine-tunes the model with source and target data to improve target-distribution discriminability and support category-level alignment.The self-supervised loss reinforces the source-training step using weighted pseudo-label classification.
- Overall objective: The final optimization adds trade-off-weighted self-supervision and gradient discrepancy constraints to the alternating training procedure.After generator updates, the target feature manifold is described as closer to the source manifold while target classification accuracy is preserved.
4. Theoretical Analysis
The analysis frames UDA error through source risk, domain divergence, and joint-hypothesis error, motivating CGDM as a way to address category-level misalignment while retaining adversarial alignment. Its optimization also encourages source and target learning procedures to follow similar paths.
- Domain-adaptation bound: The domain-adaptation bound separates target error into source error, H∆H-domain divergence, and the ideal joint-hypothesis error λ.Source error can be minimized using labeled source data, while λ reflects the error of the ideal joint hypothesis.
- Motivation from theory: MCD reduces H∆H-domain divergence with two hypotheses but treats λ as negligible, leaving category-level misalignment insufficiently addressed.The paper identifies this treatment as problematic when the optimal source and target hypotheses differ.
- Joint-hypothesis gap: When domains are category-level misaligned, no joint hypothesis may simultaneously minimize source and target errors, producing a large λ.CGDM therefore combines pseudo labels with gradient-discrepancy minimization rather than using source-model pseudo labels as direct supervision.
- Training procedure: The training algorithm obtains pseudo target labels, trains on both domains, maximizes classifier-output divergence, and updates the generator with a gradient-similarity constraint.These three steps are repeated across epochs with trade-off hyperparameters α and β.
- Optimization-path interpretation: Minimizing cross-domain gradient discrepancy encourages source and target learning procedures to follow a similar optimization path toward a shared hypothesis.The authors argue that this enables finer-grained optimization and a more accurate joint hypothesis, improving UDA accuracy.
5. Experiments
Experiments on DomainNet, VisDA-2017, and ImageCLEF show that CGDM improves unsupervised domain adaptation across many-category, large-domain-gap, and standard transfer settings. Model analyses further examine feature alignment, convergence, hyper-parameter sensitivity, ablations, and gradient-discrepancy variants.
- Results on DomainNet: CGDM surpasses mainstream domain adaptation methods in DomainNet mean accuracy across pairwise transfers among six domains.MCD and SWD have 6.7% and 3.6% lower mean accuracy than CGDM, respectively.
- Results on VisDA-2017: CGDM improves mean accuracy on VisDA-2017 by up to 29.9% over source-only training and by 10.4% and 5.9% over MCD and SWD.It performs best in five categories and reports 94.5%, 82.5%, and 49.2% accuracy for knife, sktbrd, and truck.
- Results on ImageCLEF: CGDM achieves the best average accuracy of 89.5% on ImageCLEF across six adaptation scenarios.The method also improves target-domain accuracy in difficult scenarios such as C →P.
- Model Analysis: CGDM produces category-level target-feature alignment with the source, while t-SNE comparisons show weaker target discriminability for domain-adversarial methods.MCD improves feature discriminability, but CGDM aligns target samples well at category level.
- Model Analysis: Training reduces source classification loss and increases target accuracy smoothly as iterations increase.The convergence example uses the C →I task on ImageCLEF.
- Model Analysis: Both self-supervised learning and gradient discrepancy minimization significantly improve ImageCLEF mean accuracy, with their combination performing best.The gradient discrepancy loss is the main new methodological contribution, while self-supervised learning is auxiliary.
6. Conclusion
The paper addresses inaccurate bi-classifier adversarial learning in domain adaptation by minimizing gradient discrepancy between source and target domains. It also uses self-supervised learning to obtain more reliable target pseudo labels, with experiments demonstrating the method’s advantage.
- The method minimizes gradient discrepancy between two domains to improve category-level distribution alignment.
- Self-supervised learning obtains more reliable pseudo labels for target samples.
- Experiments on large-scale datasets demonstrate the advantage of the proposed method.