Source-linked AI summary
Progressive Feature Alignment for Unsupervised Domain Adaptation
Chaoqi Chen, Weiping Xie, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, Tingyang Xu, Junzhou Huang
TL;DR
UDA must transfer knowledge from labeled source data to unlabeled target data while avoiding pseudo-label error accumulation and preserving category consistency. PFAN progressively aligns discriminative features using EHTS, APA, iterative alternative updates, and a temperature variate that slows source-loss convergence. The reported experiments show that PFAN outperforms state-of-the-art UDA approaches on three datasets.
Problem
Pseudo-label-based discriminative UDA methods are vulnerable to error accumulation because pseudo-labeling accuracy is not explicitly guaranteed.
Method
PFAN progressively aligns categories using EHTS-selected target samples, APA prototype alignment, alternating updates, and a temperature variate for the source classifier.
Results
PFAN outperforms state-of-the-art UDA approaches on three domain adaptation datasets.
Takeaways & Limitations
The approach uses target intra-class variance and cross-domain category consistency to address UDA problems.
Takeaways & Limitations
Pseudo-label-based target alignment remains constrained by falsely labeled samples, especially when domain discrepancy is large or mini-batches contain insufficient categorical information.
Abstract
from arXiv · showhide
Unsupervised domain adaptation (UDA) transfers knowledge from a label-rich source domain to a fully-unlabeled target domain. To tackle this task, recent approaches resort to discriminative domain transfer in virtue of pseudo-labels to enforce the class-level distribution alignment across the source and target domains. These methods, however, are vulnerable to the error accumulation and thus incapable of preserving cross-domain category consistency, as the pseudo-labeling accuracy is not guaranteed explicitly. In this paper, we propose the Progressive Feature Alignment Network (PFAN) to align the discriminative features across domains progressively and effectively, via exploiting the intra-class variation in the target domain. To be specific, we first develop an Easy-to-Hard Transfer Strategy (EHTS) and an Adaptive Prototype Alignment (APA) step to train our model iteratively and alternatively. Moreover, upon observing that a good domain adaptation usually requires a non-saturated source classifier, we consider a simple yet efficient way to retard the convergence speed of the source classification loss by further involving a temperature variate into the soft-max function. The extensive experimental results reveal that the proposed PFAN exceeds the state-of-the-art performance on three UDA datasets.
1. Introduction
UDA transfers knowledge from labeled source domains to unlabeled target domains, but domain shifts and pseudo-label errors complicate category-level alignment. PFAN addresses these challenges with progressive alignment using target intra-class variation, complementary pseudo-label selection and prototype alignment, and a temperature-adjusted source classifier.
- UDA transfers knowledge from a label-rich source domain to a fully unlabeled target domain despite data bias or shift.
- Global distribution matching can mix samples from different classes, motivating discriminative category-level alignment in the target domain.
- Pseudo-label-based methods risk error accumulation because target category losses depend on pseudo-labels whose accuracy is not explicitly guaranteed.
- Target samples exhibit easy, hard, and false-easy patterns, making progressive handling of intra-class variation necessary for reliable pseudo-labeling.Easy samples are close to the source domain, hard samples are ambiguous, and false-easy samples can receive confident but incorrect labels.
- PFAN progressively enforces category alignment through EHTS-selected pseudo-labeled samples and APA alignment of source and target class prototypes.APA statistically aligns cross-domain class distributions rather than backpropagating target category loss directly from pseudo-labels.
- PFAN further retards source classification-loss convergence with a temperature variate and exceeds state-of-the-art UDA performance on three datasets.
2. Related Work
Related UDA work includes adversarial feature alignment, distance-based distribution matching, and pseudo-label-driven discriminative learning. These approaches motivate explicit category-level alignment because global domain alignment alone does not ensure class-to-class overlap.
- Deep UDA methods use adversarial learning to seek feature spaces where source and target distributions are maximally confused.Gradient reversal is one example for training domain-invariant features.
- Distance-based methods measure domain discrepancy with MMD, KL-divergence, or Wasserstein distance while commonly matching marginal distributions P(Xs) and P(Xt).
- Exact domain-level alignment does not imply fine-grained class-to-class overlap, so category-level alignment is important without target labels.
- Pseudo-label methods compensate for missing target categorical information through conditional matching, asymmetric tri-training, iterative sample selection, or labeling all target samples.
3. Progressive Feature Alignment Network
PFAN combines EHTS, APA, and temperature-scaled source classification to progressively align domain features and category distributions. Its training alternates pseudo-label selection and prototype alignment while jointly addressing domain-level and category-level alignment.
- Architecture and training: PFAN consists of EHTS, APA, and a soft-max function with a temperature variate, trained through alternating iterative updates.EHTS provides pseudo-labeled target samples, while APA enforces cross-domain category alignment.
- Adaptive Prototype Alignment: The EHTS and APA updates work alternatively and iteratively, with EHTS supplying pseudo-labeled samples and APA mitigating errors introduced by false labels.The training procedure initializes the model on labeled source data, repeatedly runs EHTS, computes prototypes, and fine-tunes the model.
- Easy-to-Hard Transfer Strategy: EHTS computes source class prototypes and uses cosine similarity to select target samples whose scores exceed a progressively adjusted threshold.The threshold controls the growth rate of easy samples so harder samples can be selected in later training steps.
- Adaptive Prototype Alignment: APA aligns source and selected-target prototypes for each category to suppress false-easy samples and enforce cross-domain category consistency.It uses squared Euclidean prototype distance and adaptively updates global prototypes from accumulated local prototypes.
- Training Losses: A temperature T > 1 softens source-class probabilities and retards source classification-loss convergence, supporting a non-saturated source classifier during adaptation.The objective jointly incorporates source classification, domain confusion, and APA losses, while adversarial alignment addresses domain-level discrepancy.
- Theoretical Analysis: The theoretical analysis links cross-domain category alignment to removal of the optimality gap between source and target labeling functions.The paper frames this result through a domain-adaptation bound on expected target error.
4. Experiments
PFAN is evaluated across Office-31, ImageCLEF-DA, and digit-transfer tasks, where it generally outperforms comparison methods. Ablations and analyses associate its gains with progressive target-sample selection, adaptive prototype alignment, temperature-based source training, and more discriminative target features.
- Datasets and setup: Office-31 contains 4,110 images across Amazon, Webcam, and DSLR, and PFAN is evaluated on all six ordered domain-transfer combinations.The domains contain 2,817, 795, and 498 images, respectively.
- Comparisons with State-of-the-Arts: PFAN improves the Office-31 state-of-the-art average from 79.1% to 80.4% and performs especially strongly on hard tasks such as A →W and A →D.The comparison includes AlexNet, DDC, DAN, RTN, RevGrad, ADDA, JAN, ATT, MADA, and MSTN.
- Comparisons with State-of-the-Arts: PFAN outperforms comparison methods on most ImageCLEF-DA transfer tasks, indicating performance across the dataset’s six domain combinations.ImageCLEF-DA contains twelve classes from ImageNet, Pascal VOC, and Caltech-256, with 600 images per domain.
- Comparisons with State-of-the-Arts: On MNIST→SVHN, PFAN exceeds the reproduced suboptimal MSTN result by +4.8%, while also outperforming alternatives on SVHN→MNIST and MNIST→USPS.The authors attribute the challenging-task result to reliable pseudo-label selection and reduced influence from falsely labeled samples.
- Ablation Study: Removing any ablated component degrades performance, while EHTS outperforms random and full target-sample selection.The variants remove or alter EHTS, APA, accumulated prototype alignment, or temperature scaling.
- Pseudo-labeling Accuracy: EHTS keeps pseudo-labeling accuracy higher and stable during training, and test accuracy increases as more labeled target samples are incorporated.The analysis states that PFAN alleviates the negative influence of falsely labeled samples.
- Further Empirical Analysis: Higher temperature retards source-loss convergence, while PFAN reduces distribution discrepancy relative to AlexNet and learns target features with greater inter-class dispersion and lower intra-class variation than RevGrad.The feature comparison uses t-SNE visualizations for task A →W, with eight randomly selected classes.
5. Conclusion
PFAN addresses UDA by exploiting target-domain intra-class variance and cross-domain category consistency. EHTS and APA support reliable pseudo-label selection and reduce false-label bias, while retarding source-loss convergence further improves performance.
- PFAN exploits target-domain intra-class variance and cross-domain category consistency for UDA.
- EHTS selects reliable pseudo-labeled samples, while APA alleviates bias caused by falsely labeled samples.
- Retarding source classification-loss convergence further improves PFAN performance.
- PFAN outperforms state-of-the-art UDA approaches on three domain adaptation datasets.