PaperlayerView exact revision on arXiv

Source-linked AI summary

Adversarial-Learned Loss for Domain Adaptation

Minghao Chen, Shuai Zhao, Haifeng Liu, Deng Cai

arXiv:2001.01046v1cs.CVcs.LG

TL;DR

Domain adaptation must transfer knowledge across domains when target labels are unavailable, while existing approaches separately emphasize distribution alignment or target-feature discrimination. ALDA learns a confusion matrix adversarially to correct pseudo-labels and construct a target loss, achieving state-of-the-art performance on four standard datasets.

  • Problem

    Unsupervised domain adaptation lacks target ground truth and must address domain shift while obtaining both aligned and discriminative target features.

  • Method

    ALDA uses adversarial noise-correcting domain discrimination to estimate a confusion matrix, correct pseudo-labels, and optimize the target classifier with the corrected loss.

  • Results

    ALDA achieves state-of-the-art performance on four standard unsupervised domain adaptation datasets.

  • Takeaways & Limitations

    ALDA combines domain-adversarial distribution alignment with self-training-based target-feature discrimination and theoretically aligns distributions while correcting target predictions.

  • Takeaways & Limitations

    Self-training methods can experience performance drops with unmatched target features, and pseudo-label assumptions may not hold.

Abstract

from arXiv · show

Recently, remarkable progress has been made in learning transferable representation across domains. Previous works in domain adaptation are majorly based on two techniques: domain-adversarial learning and self-training. However, domain-adversarial learning only aligns feature distributions between domains but does not consider whether the target features are discriminative. On the other hand, self-training utilizes the model predictions to enhance the discrimination of target features, but it is unable to explicitly align domain distributions. In order to combine the strengths of these two methods, we propose a novel method called Adversarial-Learned Loss for Domain Adaptation (ALDA). We first analyze the pseudo-label method, a typical self-training method. Nevertheless, there is a gap between pseudo-labels and the ground truth, which can cause incorrect training. Thus we introduce the confusion matrix, which is learned through an adversarial manner in ALDA, to reduce the gap and align the feature distributions. Finally, a new loss function is auto-constructed from the learned confusion matrix, which serves as the loss for unlabeled target samples. Our ALDA outperforms state-of-the-art approaches in four standard domain adaptation datasets. Our code is available at https://github.com/ZJULearning/ALDA.

Introduction

Unsupervised domain adaptation must transfer knowledge from labeled source data to unlabeled target data despite domain shift. ALDA combines domain-distribution alignment with target-feature discrimination by learning a confusion matrix that corrects pseudo-labels and constructs the target loss.

  • Domain shift makes models trained on existing labeled datasets difficult to generalize to new data with different appearance characteristics.
  • A discriminator network adversarially learns the confusion matrix, which corrects pseudo-label vectors and supplies training labels for target samples.
  • Domain-adversarial learning aligns source and target feature distributions, whereas self-training enhances target classification features but aligns distributions only implicitly.
  • Self-training can suffer performance drops with unmatched target features because its implicit feature alignment has no theoretical guarantee.
  • ALDA models pseudo-label noise with a confusion matrix whose components represent the conditional relationship between ground-truth labels and model predictions.
  • ALDA theoretically aligns feature distributions and corrects target classifier predictions while combining domain-adversarial learning with self-training.

Related Work

UDA transfers knowledge from labeled source domains to unlabeled target domains, with distribution alignment as a central strategy. Related self-training methods use predictions on unlabeled data to shape decision boundaries and implicitly encourage class-wise feature alignment.

  • Unsupervised domain adaptation transfers knowledge between labeled source and unlabeled target domains under a domain-discrepancy constraint.
  • A major UDA research direction aligns source and target distributions because target classifier accuracy is bounded by source accuracy and domain discrepancy.
  • Self-training methods use model predictions on unlabeled samples to smooth decision boundaries, minimize prediction entropy, or provide pseudo-labels.
  • Applied to UDA, self-training methods implicitly encourage class-wise feature alignment and achieve strong results across multiple UDA tasks.

Methods

ALDA combines pseudo-label correction with adversarial domain discrimination by learning a class-wise noise model and using it to construct a target-domain loss. Its discriminator and generator jointly align class-wise feature distributions while incorporating classifier predictions and source labels.

  • Adversarial Learning: The discriminator minimizes source corrected-label discrepancy to ground truth and pushes target corrected labels toward the opposite distribution.The generator maximizes the adversarial loss through a gradient-reversal minimax game, confusing the discriminator.
  • Target Loss Construction: Pseudo-labels replace unavailable target ground truth, but their discrepancy from the ideal loss motivates correcting them with an estimated confusion matrix.The corrected label vector is obtained from the confusion matrix and pseudo-label vector, enabling approximate optimization of the ideal target loss.
  • Noise Modeling: ALDA analyzes pseudo-label noise with a confusion matrix and estimates it using an additional noise-correcting domain discriminator.The discriminator takes deep features as input and produces a noise vector whose components represent the probability that a pseudo-label matches the correct label.
  • Adversarial Learning: The adversarial objective uses classifier predictions and label information to achieve class-wise feature alignment between source and target domains.This differs from common domain-adversarial learning by conditioning the discrimination process on class-related information.
  • Optimization: ALDA uses the unhinged loss for target samples because it is robust to the uniform component of label noise, alongside supervised source-domain loss.A source classification regularizer is added to stabilize adversarial discriminator training, requiring correct source classification in addition to domain discrimination.

Theoretical Insight

ALDA’s theory shows that its noise-correcting domain discrimination aligns source and target feature distributions, while the corrected target loss reduces pseudo-label noise under stated assumptions.

  • Distribution alignment: At the optimal discriminator and generator, ALDA aligns the source and target feature distributions: Ps = Pt.This result is established for the noise-correcting domain discrimination.
  • Distribution alignment: The aligned feature distributions yield a theoretical bound on target expected error through source error and feature discrepancy.The argument follows the domain-adaptation error bound relating target error to source error and domain discrepancy.
  • Pseudo-label correction: Optimizing the corrected target loss reduces noise in pseudo-labels.Theorem 2 analyzes the corrected loss under an optimal labeling function in the feature space.
  • Pseudo-label correction: The corrected target training retains or suppresses pseudo-label contributions according to the discriminator’s correction of target predictions.The resulting target classifier training is described as more efficient than the original pseudo-label method.

Experiments

ALDA is evaluated across four standard unsupervised domain adaptation datasets and compared with state-of-the-art methods, component ablations, and feature visualizations. Results consistently favor combining domain-adversarial learning with self-training and correcting pseudo-label noise.

  • Datasets and setup: ALDA is evaluated on digits, Office-31, Office-Home, and VisDA-2017 using standard unsupervised domain adaptation settings.The experiments include three digit adaptation scenarios, six Office-31 tasks, twelve Office-Home scenarios, and VisDA validation-set evaluation.
  • Image Results: ALDA significantly outperforms state-of-the-art methods on Office-31, especially on hard transfers A →W, A →D, D →A, and W →A.The comparison includes domain-adversarial methods such as DANN, JAN, and MADA, as well as CDAN+E.
  • Image Results: ALDA exceeds advanced methods on Office-Home and performs better than other domain adaptation methods on VisDA-2017 with ResNet-50.Office-Home is described as more difficult because it has more categories and larger appearance gaps between domains.
  • Digits Results: ALDA outperforms distribution-alignment and self-training baselines on digit datasets and substantially reduces the gap between unsupervised adaptation and supervised target training.Each digit experiment is run three times, with average results and variance reported.
  • Digits Results: Lowering the pseudo-label threshold δ from 0.9 to 0.6 improves digit adaptation, with δ = 0.6 achieving the best result.The paper attributes this to including more low-confidence target samples in training on relatively easy digit transfers.
  • Ablation study: Ablations show that direct DANN-plus-self-training is inferior to ALDA, removing regularization sharply reduces performance, and corrected pseudo-label losses outperform uncorrected losses.The unhinged loss also performs better than cross-entropy because it removes the uniform part of label noise and facilitates noise correction.
  • Feature visualization: For A →W adaptation, ALDA produces target clusters closely matched to corresponding source clusters, unlike ResNet-50, self-training, and DANN.The visualization indicates that ALDA’s target features are both aligned and discriminative.

Conclusion

The paper introduces ALDA to combine domain-adversarial learning and self-training for unsupervised domain adaptation. It learns a confusion matrix to correct pseudo-label noise, and reports theoretical and experimental effectiveness across four standard datasets.

  • Conclusion: ALDA models pseudo-label noise with a confusion matrix learned through noise-correcting domain discrimination.The target classifier is then optimized using the resulting corrected loss function.
  • Conclusion: ALDA is theoretically and experimentally proven effective for unsupervised domain adaptation and achieves state-of-the-art performance on four standard datasets.

Theoretical Proof

The theoretical analysis defines source and target feature distributions, then shows that ALDA’s optimal discriminator and generator align those distributions. Under an optimal labeling function, the discriminator output recovers the ideal target labeling structure.

  • Definitions: ALDA represents source and target features as distributions Ps and Pt generated by G in feature space F.The analysis assumes an ideal labeling function that correctly labels source features.
  • Theoretical Proof: The proof derives the adversarial objective by expanding its feature-space contribution and optimizing the noise vector ξ.A constant term C depends only on the number of classes K.
  • Theoretical Proof: The discriminator’s optimal point is expressed using source and target densities together with ideal and predicted label functions.The derivation uses Py∗ for ideal labels and Pŷ for classifier predictions.
  • Theoretical Proof: Ps(f) = Pt(f) for all f in Ps ∪ Pt, so the optimal discriminator and generator align the source and target feature distributions.This is the conclusion of Theorem 1.
  • Theoretical Proof: At the optimum, c(xt) assigns 1/2 to the predicted class and 1/(2K−2) to every other class.Theorem 2 states that this output equals the ideal labeling structure h y∗(ft).

Experimental Analysis

The experiment tracks pseudo-label weights during ALDA training on A →W with ResNet50. Early training assigns higher weights to correct pseudo-labels, consistent with the theoretical corrected-target-loss analysis.

  • Experimental Analysis: The A →W experiment with ResNet50 records pseudo-label weights c(xt)ŷt separately for correct and incorrect samples.The experiment is designed to examine the mechanism of the corrected target loss.
  • Experimental Analysis: Theorem 2 predicts large weights for correct pseudo-labels and small weights for incorrect pseudo-labels.The weight estimates p(yt = ŷt|ŷt, xt).
  • Experimental Analysis: At the beginning of training, c(xt)ŷt is high for correct samples, matching the predicted weighting behavior in Fig. 1.The passage attributes this early pattern to correct samples being easier to transfer and match with the source distribution.
Loading 2001.01046v1…