PaperlayerView exact revision on arXiv

Source-linked AI summary

Discriminative Adversarial Domain Adaptation

Hui Tang, Kui Jia

arXiv:1911.12036v2cs.CVcs.LG

TL;DR

Unsupervised domain adaptation must classify unlabeled target instances despite distribution shifts and limited target labels, while separate task and domain classifiers struggle to align joint feature-category distributions. DADA integrates these classifiers with a mutually inhibitory adversarial objective, extending it to partial and open set settings. The paper reports new state-of-the-art results across all three adaptation settings on benchmark datasets.

  • Problem

    Separate task and domain classifiers have limited ability to align joint feature-category distributions across domains, especially for complex multimodal distributions.

  • Method

    DADA uses an integrated category-domain classifier with a mutually inhibitory adversarial objective, extended as DADA-P and DADA-O for partial and open set adaptation.

  • Results

    The proposed methods achieve new state-of-the-art performance for closed, partial, and open set domain adaptation on benchmark datasets.

  • Takeaways & Limitations

    DADA explicitly promotes joint distribution alignment while supporting adaptation settings with mismatched source and target label spaces.

  • Takeaways & Limitations

    Open set performance depends on choosing an appropriate q: values that are too small miss unknown targets, while values that are too large misclassify known targets.

Abstract

from arXiv · show

Given labeled instances on a source domain and unlabeled ones on a target domain, unsupervised domain adaptation aims to learn a task classifier that can well classify target instances. Recent advances rely on domain-adversarial training of deep networks to learn domain-invariant features. However, due to an issue of mode collapse induced by the separate design of task and domain classifiers, these methods are limited in aligning the joint distributions of feature and category across domains. To overcome it, we propose a novel adversarial learning method termed Discriminative Adversarial Domain Adaptation (DADA). Based on an integrated category and domain classifier, DADA has a novel adversarial objective that encourages a mutually inhibitory relation between category and domain predictions for any input instance. We show that under practical conditions, it defines a minimax game that can promote the joint distribution alignment. Except for the traditional closed set domain adaptation, we also extend DADA for extremely challenging problem settings of partial and open set domain adaptation. Experiments show the efficacy of our proposed methods and we achieve the new state of the art for all the three settings on benchmark datasets.

Introduction

Unsupervised domain adaptation transfers labeled source-domain knowledge to unlabeled target domains despite distribution shifts. DADA addresses limitations of separate task and domain classifiers with integrated adversarial designs for closed, partial, and open set adaptation, achieving state-of-the-art results across all three settings.

  • Motivation: Unsupervised domain adaptation learns target classifiers from labeled related source data when target-domain labels are unavailable or costly to obtain.The source and target domains may have shifted distributions even when they share the same label space.
  • Motivation: Separate task and domain classifiers make fine-grained category-level alignment difficult, particularly for complex multimodal distributions.Their feature extractor may compensate for insufficiently aligned feature distributions, producing mode-collapse-related limitations.
  • DADA: DADA integrates category and domain classification and uses a mutually inhibitory adversarial objective to promote joint feature-category distribution alignment.The method explicitly couples domain and category predictions for each input instance.
  • Extensions: DADA-P extends the method to partial domain adaptation with reliable category-level weighting that reduces the influence of outlier source instances.Partial adaptation has a target label space subsumed by the source label space.
  • Extensions: DADA-O extends the method to open set adaptation by balancing shared-label joint alignment with classification of outlier target instances.Open set adaptation has a source label space subsumed by the target label space.
  • Results: Experiments report new state-of-the-art performance for closed, partial, and open set domain adaptation on benchmark datasets.The reported result covers all three adaptation settings.

Related Works

Prior work develops domain-adversarial, category-aware, and confidence-weighted approaches for closed, partial, and open set domain adaptation. These methods progressively incorporate category information or instance weighting to address distribution and label-space differences.

  • Closed Set Domain Adaptation: Closed set methods include untied feature alignment, similarity-based classification, category-aware adversarial training, and cluster-assumption constraints.Examples include ADDA, SimNet, MADA, CDAN, and VADA.
  • Partial Domain Adaptation: Partial domain adaptation methods reduce outlier-source influence using domain-based importance weights or category predictions.DADA-P instead weights its source discriminative adversarial loss by reliable category confidence.
  • Open Set Domain Adaptation: Open set domain adaptation addresses settings where the target label space contains labels absent from the source label space.The supplied related-work passage introduces this problem setting but does not detail a specific prior method.

Method

DADA uses an integrated category-and-domain classifier with adversarial objectives that explicitly couple category and domain predictions. Its extensions target joint-distribution alignment in closed-set, partial, and open-set adaptation.

  • Method: DADA integrates the task and domain classifiers into F(·), rather than using separate classifiers stacked on G(·).The integrated classifier outputs K category probabilities plus one domain-related output.
  • Discriminative Adversarial Learning: The proposed adversarial objective creates a mutually inhibitory relation between category and domain predictions for each input.This interaction is designed to make category discrimination and domain distinction compete during optimization.
  • Discriminative Adversarial Learning: Under practical conditions, minimizing the source loss over F(·) discriminates task categories and domains, while maximizing it over G(·) aligns the source domain to the target while retaining discriminability.The resulting classifier-feature-extractor optimization forms the paper’s adversarial training mechanism.
  • Discriminative Adversarial Learning: For target instances, DADA uses conditional category probabilities to weight domain predictions and explicitly create category-domain interplay for joint-distribution alignment.This design is motivated by reducing mode collapse through alignment with multiple related categories.
  • Extension for Partial Domain Adaptation: DADA-P addresses partial adaptation by reducing source-outlier influence, promoting joint alignment in the shared label space while alleviating negative transfer.The partial setting has a target label space subsumed by the source label space.
  • Extension for Open Set Domain Adaptation: DADA-O balances adaptation and outlier rejection by selecting q to separate unknown target instances from known ones without excessively misclassifying known instances.Too-small q harms unknown classification, whereas too-large q can misclassify known targets as unknown.

Experiments

Experiments evaluate DADA across closed set, partial, and open set domain adaptation benchmarks, including ablations and quantitative comparisons. DADA and its extensions outperform comparative methods across these settings.

  • Datasets and settings: Experiments use Office-31 and Syn2Real benchmarks across closed set, partial, and open set domain adaptation settings.Office-31 uses ResNet-50; Syn2Real-C closed set uses ResNet-101, partial adaptation uses ResNet-50, and Syn2Real-O open set uses ResNet-152.
  • Ablation study: DADA significantly outperforms DANN and DANN-CA in Office-31 ablations, supporting its mutually inhibitory category-domain design.Removing entropy minimization or target discrimination reduces performance relative to the complete DADA configuration.
  • Quantitative comparison: Figure 3 compares average true-category probabilities across task classifiers for target instances on the A → W transfer.The comparison includes No Adaptation, DANN, DANN-CA, and DADA, using target labels only to quantify category-level domain discrepancy.
  • Partial domain adaptation: +15.53% is DADA-P's improvement over all comparative methods on Syn2Real-C partial domain adaptation.The authors attribute this result to reducing the negative influence of source outliers while promoting joint distribution alignment in the shared label space.
  • Open set domain adaptation: DADA-O outperforms all comparative methods on both Known and Mean metrics for Syn2Real-O open set domain adaptation.It improves over AODA by a large margin when the target known-to-unknown ratio is much smaller than 1.

Conclusion

The paper concludes that DADA uses an integrated classifier and mutually inhibitory category-domain predictions to align feature-category joint distributions. Extensions address partial and open set adaptation, with benchmark experiments supporting efficacy in all three settings.

  • Conclusion: DADA addresses joint-distribution alignment limits caused by mode collapse from separately designed task and domain classifiers.Its integrated classifier explicitly couples category and domain predictions.
  • Conclusion: DADA's adversarial objective encourages mutually inhibitory category and domain predictions for each input instance.The method explicitly uses target discriminative information during adversarial training.
  • Conclusion: DADA is extended beyond closed set adaptation to partial and open set domain adaptation.The paper reports benchmark evidence for all three problem settings.

B Generalization Error Analysis for Our Proposed DADA

The analysis derives a hierarchy of distance measures that upper-bound source-target domain discrepancy and relates DADA to the strongest computable bound available under unlabeled target data. It argues that DADA's integrated, category-regularized adversarial training can better approximate target error from source error.

  • Generalization error analysis: DADA is proved to better bound expected target error than several existing domain adaptation methods under finite effective hypothesis-space assumptions.The analysis follows the formalism of established domain adaptation generalization results.
  • Assumptions and target error: Domain adaptation assumes an ideal joint hypothesis with small combined source and target error, while adaptation seeks to reduce source-target discrepancy.The expected target-error bound combines source error, ideal-joint-hypothesis error, and distribution discrepancy.
  • Distance bounds: Theorem 1 upper-bounds source-target distribution discrepancy by FH∗-, FH-, F-, and D-distances.The FH∗-distance provides the best bound but is not computable when target instances are unlabeled.
  • Distance bounds: The FH-distance is more informative than F- and D-distance bounds and can be implemented by conditioning a loss-difference function on the task classifier.Separate domain classifiers target D-distance, conditional domain classifiers target F-distance, while integrated classifiers target FH-distance.
  • Interpretation of DADA: DADA can be formalized as category-regularized domain-adversarial training that learns explicit intra-domain discrimination and inter-domain indistinguishability.The analysis states that this may enable a better ideal joint hypothesis and better approximation of target error by source error.

C.1 Datasets

The evaluation uses Digits, Office-31, Office-Home, and ImageNet-Caltech benchmarks across closed, partial, and open set domain adaptation settings. The datasets include configurations with shared, outlier, and unknown categories.

  • Digits experiments evaluate SVHN→MNIST, MNIST→USPS, and USPS→MNIST using the same architecture and settings as ADR.
  • For partial adaptation, Office-31 and Caltech-256 share 10 target categories, while the source domain contains all 31 Office-31 categories.
  • For open set adaptation, 10 categories are known, source categories 11–20 are unknown, target categories 21–31 are unknown, and evaluation uses 11 classes.
  • Figure 4 examines how λ affects the rate of source instances failing the adversarial-training condition during the first 100 iterations for A→D and D→A.

Effect of the λ

The λ analysis studies whether source instances satisfy p_s^y > 0.5 during early DADA adversarial training. Without λ, the failure rate rises rapidly.

  • The experiment tracks source instances satisfying p_s^y > 0.5 during the early stage of DADA adversarial training.
  • Without λ, the rate of source instances failing the condition rises rapidly during early adversarial training.

Alternative Choice of Adversarial Loss for Target Instances

The alternative target-instance adversarial loss produces confused domain predictions but does not support strong unknown-category prediction. DADA’s loss instead suppresses the domain output, improving the chance of correct target classification and yielding better experiments and feature alignment.

  • The alternative symmetric adversarial loss yields p_K+1(x_t) = 0.5 when maximized over the feature extractor.
  • Because the alternative leaves the true unknown-category prediction without a chance to approach 1, it is characterized as sub-optimal.
  • DADA’s target loss yields p_K+1(x_t) = 0 when maximized over the feature extractor, giving the true category prediction a better chance to approach 1.
  • The mutually inhibitory relation between category and domain predictions makes correct target classification more likely under DADA’s loss.
  • Office-31 experiments compare DADA’s loss with the alternative DADA-DC loss using ResNet-50, with results reported in Table 6 and convergence shown in Figure 5.
  • t-SNE visualizations show DADA changing scattered target samples into category-wise clusters aligned with corresponding source samples, while source categories remain not perfectly separated.
  • Digits experiments report target-test accuracies as mean±std over five random trials to assess robustness across MNIST, SVHN, and USPS.

Office-31

On Office-31 and related benchmarks, DADA-P addresses partial adaptation by reducing negative transfer from source-only categories. Its advantage persists across category counts, backbone choices, domain discrepancy, and dataset scale.

  • DADA-P outperforms comparative methods by a large margin on Office-31 with ResNet-50, including settings with small source and target domains.
  • Whole-domain alignment methods can falsely align outlier source categories with the target domain, explaining their poor partial-adaptation performance.
  • As target categories decrease from 30 to 10, DADA-P’s relative improvement over DANN becomes larger, indicating reduced influence of negative transfer.
  • DADA-P outperforms comparative methods on Office-31 with AlexNet, Office-Home with ResNet-50, and ImageNet-Caltech with ResNet-50.
  • With 10 target categories and source categories increasing from 12 to 31, DADA significantly outperforms DANN in all settings and becomes relatively stronger as outliers increase.
  • The Office-Home and ImageNet-Caltech results extend the evaluation to larger category sets, greater domain discrepancy, and large-scale source and target domains.

C.4 Open Set Domain Adaptation.

DADA-O outperforms existing methods on both OS* and OS in Office-31 open set adaptation, supporting alignment of known instances and recognition of unknown target instances. The q study shows a trade-off: increasing q improves unknown-instance recognition while reducing OS and OS* accuracy.

  • DADA-O outperforms all comparative methods on both OS* and OS in Office-31 open set adaptation.OS* averages known-category performance, while OS also includes the unknown category.
  • DAN and DANN perform much worse than No Adaptation because marginal alignment can falsely align known source instances with unknown target instances.
  • ATI-λ and AODA effectively reduce false alignment through mechanisms that recognize unknown instances.
  • As q increases, OS and OS* accuracies decrease while Unknown accuracy increases, making target instances more likely to receive the unknown category.
  • When q = 0, aligning the whole source and target domains misclassifies all unknown target instances as known categories; DADA uses q = 0.1 as a trade-off.

D Investigation for Our Used Training Scheme

DADA is trained by alternating supervised source classification with adversarial training on labeled source and unlabeled target data. On MNIST↔USPS, this scheme keeps the required condition largely satisfied, preserves low source error, and produces target-error improvements during adversarial training.

  • Training scheme: The training scheme alternates classification training on labeled source data with adversarial DADA training on labeled source and unlabeled target data.It is designed to keep the condition p_sy_s > 0.5 satisfied during training.
  • Training dynamics: Classification training lowers the rate of source instances failing the condition, whereas adversarial training raises it as some instances cease satisfying the condition.
  • Training dynamics: The failure rate declines to a very low value oscillatory, demonstrating that the training scheme keeps the condition satisfied.
  • Training dynamics: The training error of source data remains low during adversarial-training epochs, indicating an effect similar to classification training.
  • Adaptation effects: All target-test-error valleys occur during adversarial training, and lower condition-failure rates correspond to greater performance improvement.
  • Adaptation effects: DADA performs very close to 100% on MNIST→USPS and USPS→MNIST, supporting joint feature-and-category distribution alignment.
Loading 1911.12036v2…