PaperlayerView exact revision on arXiv

Source-linked AI summary

Active Adversarial Domain Adaptation

Jong-Chyi Su, Yi-Hsuan Tsai, Kihyuk Sohn, Buyu Liu, Subhransu Maji, Manmohan Chandraker

arXiv:1904.07848v2cs.CVcs.LG

TL;DR

Source-trained models can perform poorly under domain shift, and unsupervised adaptation often lacks the performance of supervised methods when target labels are scarce. AADA unifies adversarial domain adaptation with active learning by using discriminator-derived importance weights that favor diverse and uncertain target samples. Across recognition and detection tasks, it improves over baseline approaches, including a 6% gain over fine-tuning with 50 labeled target images.

  • Problem

    Domain shift can hurt source-trained predictors, while unsupervised domain adaptation often trails supervised methods and target annotation is laborious and inefficient to collect naively.

  • Method

    AADA jointly uses domain adversarial learning for adaptation and an importance-weight selection criterion combining target-source diversity with classification uncertainty.

  • Results

    AADA improves over baseline adaptation, transfer, and sampling methods across recognition and detection tasks, including a 6% advantage over fine-tuning with 50 labeled target images.

  • Takeaways & Limitations

    The framework unifies domain adaptation and active learning, retaining improvements across object recognition and detection while reducing target-domain labeling requirements.

  • Takeaways & Limitations

    Adversarial adaptation becomes counter-productive after many target labels in the large-gap VisDA setting, where fine-tuning adds no further improvement.

Abstract

from arXiv · show

We propose an active learning approach for transferring representations across domains. Our approach, active adversarial domain adaptation (AADA), explores a duality between two related problems: adversarial domain alignment and importance sampling for adapting models across domains. The former uses a domain discriminative model to align domains, while the latter utilizes it to weigh samples to account for distribution shifts. Specifically, our importance weight promotes samples with large uncertainty in classification and diversity from labeled examples, thus serves as a sample selection scheme for active learning. We show that these two views can be unified in one framework for domain adaptation and transfer learning when the source domain has many labeled examples while the target domain does not. AADA provides significant improvements over fine-tuning based approaches and other sampling methods when the two domains are closely related. Results on challenging domain adaptation tasks, e.g., object detection, demonstrate that the advantage over baseline approaches is retained even after hundreds of examples being actively annotated.

1. Introduction

Domain shift can degrade source-trained predictors, while unsupervised adaptation often trails supervised performance. AADA addresses limited target labels by combining adversarial adaptation with active sample selection based on diversity and uncertainty.

  • Covariate shift occurs when source and target input distributions differ, potentially causing source-trained predictors to perform poorly on the target domain.
  • Unsupervised domain adaptation often underperforms supervised methods, motivating effective selection and adaptation of a limited number of labeled target examples.
  • AADA combines domain adversarial learning with active learning to adapt representations and select target samples for annotation.
  • Its selection criterion combines diversity from importance weights estimated by a domain discriminator with uncertainty from classification entropy.
  • AADA outperforms fine-tuning by 6% when only 50 target-domain images are labeled and reports improvements across classification and detection tasks.

2. Related Work

Domain adaptation reduces source–target discrepancies, while active learning seeks informative labels under a limited budget. AADA unifies these goals by using adversarial training and discriminator-based sampling for target-domain annotation.

  • Domain adaptation: Domain adaptation seeks source–target invariance by reducing domain discrepancy in deep representations while preserving source-label predictiveness.
  • Active learning: This work considers selecting a few target examples to label without assuming a prior target-label distribution.
  • Active learning: Active learning methods quantify informativeness to maximize performance under a limited annotation budget using cues such as uncertainty, diversity, and representativeness.
  • Prior work: Earlier work combined importance weighting with MMD-based selection, but its integration with deep models and domain adversarial training was unclear.
  • AADA: Unlike ALDA, AADA uses one discriminator both to train the recognition model adversarially and to support sampling through diversity and entropy-based uncertainty.

3. Proposed Algorithm

AADA combines domain-adversarial adaptation with active sample selection, using discriminator-based diversity and classifier uncertainty to iteratively choose target examples for labeling and retraining.

  • Domain adaptation: AADA uses a domain adversarial neural network with a feature extractor, class predictor, and domain discriminator to align source and target representations.The discriminator separates source and target domains while the feature extractor is trained to fool it, forming an adversarial alignment objective.
  • Active learning process: AADA alternates model training, scoring, oracle labeling, and retraining across labeled source, labeled target, and remaining unlabeled target data.Each round computes s(x), selects a budgeted set of target images, obtains labels, updates the labeled target set, and retrains the model.
  • Sample selection: The sample-selection objective combines importance-based diversity with prediction entropy to identify informative unlabeled target examples.Importance reflects target-versus-source density, while entropy provides an uncertainty cue related to classification risk.
  • Collaborative roles: Adversarial training makes the discriminator useful for selection because independently trained discriminators may score most target samples similarly or focus on properties unrelated to the classifier.The paper states that joint learning is necessary for selected samples to benefit classifier updates.
  • Semi-supervised adaptation: With labeled data from both domains, AADA trains on their union and uses uniform sampling to obtain the tightest stated generalization bound under the given zero-error assumption.The bound uses source and target error weights and labeled-example proportions, with tightness when αs = βs and αt = βt.

4. Experiments on Digit Classification

Experiments on SVHN→MNIST compare target-sample selection and training schemes for active domain adaptation. AADA combines importance-weight sampling with adversarial training, performing best when few labeled targets are available, while fine-tuning becomes preferable as labeled targets increase.

  • Experimental setting and methods: AADA evaluates adversarial, joint, fine-tuning, and target-only training alongside importance-weight, clustering, core-set, diversity, BvSB, and random sampling strategies.The experiments simulate oracle annotations on SVHN→MNIST, with SVHN as labeled source and MNIST as initially unlabeled target.
  • Comparison of sampling methods: 95% accuracy is reached with 160 labeled samples using importance-weight sampling and adversarial training, while random selection requires twice as many annotations for similar performance.Importance-weight sampling also remains more stable across rounds than the other sampling baselines.
  • Comparison of different cues: Combining diversity and uncertainty cues produces the best sample-selection performance, exceeding either cue alone.The diversity cue uses discriminator predictions, whereas uncertainty uses classifier-prediction entropy.
  • Comparison of training schemes: Fine-tuning is best with random sampling, but adversarial training becomes strongest when importance-weight sampling addresses sampling imbalance.The ablation identifies mode collapse as a problem for adversarial training with random sampling.
  • Comparison of training schemes: With importance-weight sampling, adversarial training is better below 250 labeled targets, whereas fine-tuning performs better above that threshold.This indicates that the preferred training scheme changes as more target labels are acquired.
  • Comparison with ALDA: AADA outperforms the joint-training importance-weight baseline by 7.6% on average over the first 25 rounds.Without adversarial loss, importance weights can lack diversity because they become overly confident.

5. More Experimental Results

Experiments on object recognition and detection compare sampling methods, training schemes, and domain shifts. AADA performs competitively or best across the reported Office, KITTI-to-Cityscapes, and VisDA-18 settings, while its benefits depend on the domain relationship and available labels.

  • Object Recognition: AADA outperforms random selection by around 3% from rounds 10 to 20 on Office D →A object recognition.AADA reaches 85% accuracy with 800 labeled targets, while random selection requires 200 more labeled targets for similar performance.
  • Object Recognition: 81.3% and 80.7% mean accuracy are achieved by BvSB and AADA, respectively, over 20 rounds on Office D →A.
  • Object Detection: 60.4% accuracy is achieved with 100 labeled Cityscapes targets selected by AADA, while other baselines require about twice as many annotations for similar performance.The task adapts KITTI to Cityscapes using adversarial training and importance-weight sampling.
  • Object Detection: AADA outperforms all other baselines on KITTI →Cityscapes object detection, including alternative adversarial-training sampling methods and random sampling with different training schemes.
  • VisDA-18 Challenge: 79.8% and 80.1% mean accuracies are obtained by importance-weight sampling and BvSB, respectively, on VisDA-18 after each round.Importance-weight sampling performs equally well as BvSB and k-center, and outperforms k-means and random baselines under fine-tuning.
  • VisDA-18 Challenge: Adversarial training stops introducing further improvement after a few labeled targets on VisDA-18, where the synthetic-to-real domain gap is large.The authors attribute this setting-specific behavior to the relative benefit of adding annotated targets and to the target domain being closer to ImageNet pre-training than the synthetic source.

6. Conclusion

AADA unifies domain adaptation and active learning through adversarial training. The framework uses domain adversarial learning for adaptation with few labeled targets and discriminator-derived importance weights for active target selection, improving reported object recognition and detection results while reducing labeling cost.

  • AADA is a unified framework for domain adaptation and active learning via adversarial training.
  • With few labeled targets, domain adversarial learning improves classification while the discriminator supplies importance weights for selecting target samples.
  • AADA improves over other baselines across different training and sampling schemes on object recognition and detection tasks.
  • The framework is reported to reduce target-domain labeling cost on object classification and detection tasks.

A.1. Comparison to ALDA [52]

The supplementary comparison evaluates original online ALDA against AADA on SVHN →MNIST digit classification. O-ALDA selects samples online and updates its classifier after each selection.

  • O-ALDA selects one sample at a time and updates the classifier online after each selection.The supplementary procedure separately trains a perceptron classifier, source classifier, and domain separator from features extracted by the domain adversarial model.
  • The comparison concerns AADA and O-ALDA on SVHN →MNIST digit classification.

A.2. More Object Detection Results

Additional object-detection experiments extend evaluation to nine selection rounds and up to 1,000 labeled targets. AADA remains ahead of alternative sampling and training baselines across this larger annotation budget.

  • AADA improves over adversarial-training sampling strategies and random sampling with different training schemes when up to 1,000 labeled targets are available.
  • The extended evaluation uses nine rounds with batch sizes of 10, 10, 10, 20, 50, 100, 100, 200, and 500.The x-axis is plotted on a logarithmic scale.

A.3. Comparison of Training Schemes on Office

Across the Office, object-detection, and VisDA comparisons, AADA's advantage depends on the training scheme and domain relationship. Importance-weight sampling consistently improves selection, while adversarial training is strongest in the Office setting.

  • Object detection: AADA outperforms all other baselines for KITTI→Cityscapes object detection with up to 1000 labeled targets.The reported comparison includes the initial round with no labeled target data.
  • Office dataset: AADA outperforms other training-scheme and sampling-strategy combinations on the Office D→A task.Importance-weight sampling improves both adversarial training and fine-tuning over random selection.
  • VisDA dataset: On VisDA, adversarial training does not improve as more labeled targets are added, whereas fine-tuning benefits from additional labels.The target domain is closer to ImageNet pre-training images, and importance-weight sampling beats random sampling with fine-tuning.

A.4. Comparison of Training Schemes on VisDA

VisDA favors fine-tuning because its target domain is closer to the ImageNet images used for pre-training. Adversarial training helps without labeled targets but declines after labeled targets are added.

  • Training strategy: VisDA's target domain is closer to ImageNet pre-training images, making fine-tuning more suitable than adversarial training.The training strategy is chosen according to this dataset property.
  • Adversarial training: Adversarial training raises target accuracy from 57.0% to 62.5% without labeled target data, compared with source-only training.This is the unsupervised domain-adaptation setting with no labeled target available.
  • Effect of labeled targets: With additional labeled targets, adversarial-training accuracy decreases regardless of sampling strategy, while fine-tuning accuracy increases.Fine-tuning with the proposed importance weight still performs better than random sampling.

A.5. Effect of Source Data Number

Importance-weight sampling remains beneficial when labeled source data is limited. The reported effect is strongest with smaller source-data subsets and is also evaluated in SVHN→MNIST experiments.

  • Limited source data: Importance-weight sampling improves over random sampling for source-data subsets of 5%, 20%, and 50%.The comparison selects 50 labeled targets per round over 10 rounds.
  • Limited source data: The improvement over random sampling is especially large when the labeled source subset is smaller.This pattern is reported for SVHN→MNIST with a subset of labeled source data.
  • SVHN→MNIST: In SVHN→MNIST, the CDAN-based experiment selects 10 target samples per round and reaches performance saturation after 50 rounds.The implementation reports 87.8% accuracy when no labeled target is available.
Loading 1904.07848v2…