PaperlayerView exact revision on arXiv

Source-linked AI summary

Adversarial Discriminative Domain Adaptation

Eric Tzeng, Judy Hoffman, Kate Saenko, Trevor Darrell

arXiv:1702.05464v1cs.CV

TL;DR

Unsupervised domain adaptation must classify unlabeled target data despite domain shift, while prior approaches had limitations in discriminative modeling, weight sharing, or adversarial losses. ADDA unifies these ideas through asymmetric, untied discriminative mappings and achieves strong results across benchmark and challenging cross-modality tasks.

  • Problem

    Unsupervised adaptation must learn target representations and classifiers from unlabeled target images, while the best way to parameterize source and target mappings remains open.

  • Method

    ADDA unifies adversarial adaptation choices and learns a supervised source representation plus an independently initialized target mapping aligned through a domain-adversarial loss.

  • Results

    ADDA generalizes well across four domain shifts, achieving strong results on benchmark datasets and a challenging cross-modality adaptation task.

  • Takeaways & Limitations

    ADDA provides a cohesive framework for comparing adversarial adaptation methods and supports effective unsupervised adaptation across varied tasks.

  • Takeaways & Limitations

    Adversarial training can suffer vanishing gradients when the discriminator converges quickly early in training.

Abstract

from arXiv · show

Adversarial learning methods are a promising approach to training robust deep networks, and can generate complex samples across diverse domains. They also can improve recognition despite the presence of domain shift or dataset bias: several adversarial approaches to unsupervised domain adaptation have recently been introduced, which reduce the difference between the training and test domain distributions and thus improve generalization performance. Prior generative approaches show compelling visualizations, but are not optimal on discriminative tasks and can be limited to smaller shifts. Prior discriminative approaches could handle larger domain shifts, but imposed tied weights on the model and did not exploit a GAN-based loss. We first outline a novel generalized framework for adversarial adaptation, which subsumes recent state-of-the-art approaches as special cases, and we use this generalized view to better relate the prior approaches. We propose a previously unexplored instance of our general framework which combines discriminative modeling, untied weight sharing, and a GAN loss, which we call Adversarial Discriminative Domain Adaptation (ADDA). We show that ADDA is more effective yet considerably simpler than competing domain-adversarial methods, and demonstrate the promise of our approach by exceeding state-of-the-art unsupervised adaptation results on standard cross-domain digit classification tasks and a new more difficult cross-modality object classification task.

1. Introduction

Domain shift limits the transfer of deep representations to novel datasets and tasks, while obtaining enough labeled data for fine-tuning is often prohibitively difficult. The paper introduces a unified adversarial adaptation framework and proposes ADDA, which uses discriminative representations and asymmetric mappings to address domain differences.

  • Motivation: Deep convolutional representations trained on one large dataset often fail to generalize to novel datasets and tasks because of dataset bias or domain shift.This motivates domain adaptation methods that reduce differences between source and target domains.
  • ADDA: ADDA learns a discriminative mapping of target images into the source feature space by fooling a domain discriminator that distinguishes encoded target images from source examples.This mechanism is illustrated in Figure 1 through a target encoder and domain discriminator.
  • Motivation: Fine-tuning deep networks for new tasks is often prohibitively difficult and expensive because it requires enough labeled data for many parameters.The introduction identifies limited labeled data as a practical barrier to adapting pretrained models.
  • Unified framework: The proposed unified framework organizes adversarial domain adaptation around choices including weight sharing, base models, and adversarial losses, while subsuming previous work.It is intended to expose similarities among existing approaches and support new instantiations.
  • ADDA: ADDA is a previously unexplored unsupervised adversarial adaptation method that prioritizes discriminative representation learning and asymmetric mappings over generative input modeling.The authors argue that generative modeling of input image distributions is unnecessary for the ultimate discriminative task, while asymmetric mappings can better model low-level feature differences.

2. Related work

Related work on unsupervised domain adaptation primarily learns representations that reduce differences between source and unlabeled target feature distributions. Approaches include MMD-based alignment, adversarial domain confusion, and GAN-based transfer methods such as CoGAN.

  • Domain adaptation: Unsupervised domain adaptation commonly guides feature learning by minimizing differences between source and target feature distributions when target labels are sparse or absent.This work focuses on unlabeled target domains.
  • MMD-based methods: MMD-based methods align domains using distributional statistics while retaining source-domain classification objectives.DDC combines MMD with source classification, whereas DAN applies MMD in a reproducing kernel Hilbert space to match higher-order statistics.
  • Adversarial adaptation: Adversarial adaptation methods learn source-discriminative representations that make source and target domains difficult to distinguish.One approach adds a binary domain classifier and uses domain confusion loss to push predictions toward a uniform distribution.
  • Generative adversarial methods: GAN-based related work pits a generative model against a discriminator that distinguishes generated samples from training data using binary-label prediction loss.Conditional GANs additionally provide both networks with information such as class labels.
  • CoGAN: CoGAN performs domain transfer by training separate GANs for source and target images while tying their high-level parameters to obtain a domain-invariant feature space.It also uses the same noise input to generate corresponding image pairs from the two distributions.

3. Generalized adversarial adaptation

The framework casts unsupervised adversarial adaptation as learning source and target mappings whose representations are aligned so a source classifier can recognize target data. It unifies methods through mapping parameterization, source–target constraints, and adversarial mapping losses, motivating choices such as untied mappings and GAN losses.

  • General framework: Unsupervised adaptation learns a source mapping and classifier, then adapts them for unlabeled target images and target classification.The target domain provides images but no label observations.
  • General framework: Adversarial adaptation minimizes the distance between source and target mapping distributions, allowing the source classifier to be reused directly on target representations.Under this alignment, the classifier is shared: C = Cs = Ct.
  • General framework: The unified formulation alternates a domain discriminator loss with a constrained adversarial objective that optimizes the source and target mappings.The discriminator predicts whether representations originate from the source or target domain, while mapping constraints and losses vary across methods.
  • Mapping parameterization: Mapping constraints range from exact layerwise equality through weight sharing to partially or fully untied source and target mappings.Fully shared mappings reduce parameters and preserve source discriminativeness, whereas untying allows domain-specific parameters but can change optimization behavior.
  • Adversarial losses: The GAN loss preserves minimax fixed points while providing stronger gradients than direct minimax optimization, whose discriminator can converge quickly and cause vanishing gradients.When both distributions change, however, the GAN objective can produce oscillation, motivating alternative objectives such as domain confusion.

4. Adversarial discriminative domain adaptation

ADDA is a discriminative, adversarial domain-adaptation method defined by unshared source and target mappings and a standard GAN loss. It trains a labeled source model first, then adversarially adapts the target encoder while using the source classifier.

  • Method design: ADDA combines a discriminative base model, unshared source and target weights, and the standard GAN loss.The generalized framework reduces method design to choosing the base model, weight sharing, and adversarial objective.
  • Method design: Unshared mappings enable domain-specific feature extraction, so the target model is initialized from the pretrained source model to avoid degenerate solutions.The target domain has no labels, making initialization and training procedures important when weights are untied.
  • Adversarial adaptation: ADDA learns an asymmetric mapping by modifying the target model to match the fixed source distribution, motivating the inverted-label GAN loss.This setup parallels adversarial learning in which a generated space is updated toward an indistinguishable fixed real space.
  • Training procedure: Training proceeds in stages: optimize the source classifier and encoder with labeled source data, then optimize discriminator and target-mapping adversarial losses while fixing the source encoder.Because the source encoder remains fixed during target learning, the source classification objective need not be revisited.
  • Training procedure: During testing, target images are encoded by the adapted target encoder and classified by the source classifier in a shared feature space.The procedure first pre-trains the source encoder, then makes encoded source and target examples indistinguishable to a domain discriminator.
  • Motivation and scope: The unified framework motivates ADDA and supports informed design decisions for unsupervised adaptation across digits and modalities.The paper next evaluates the method on unsupervised adaptation benchmark tasks spanning digit and cross-modality settings.

5. Experiments

Experiments evaluate ADDA on unsupervised adaptation across digit domain shifts and a challenging RGB-to-depth object-classification task. ADDA generally matches or exceeds competing methods and substantially improves recognition without labeled target data.

  • Experimental scope: ADDA is evaluated on four domain shifts spanning MNIST, USPS, SVHN, and NYUD RGB-to-depth adaptation.The experiments cover digit datasets of varying difficulty and a cross-modality 19-way object-classification task.
  • Digit adaptation: In 3 of 4 digit setups, ADDA outperforms every competing method, while in the remaining setup it trails only one competitor.The comparisons use state-of-the-art unsupervised domain-adversarial learning methods.
  • Cross-modality adaptation: 50%+ relative improvement over the nonadaptive baseline is achieved on NYUD RGB-to-depth adaptation without labeled depth data.ADDA learns a useful depth representation despite the large modality shift.
  • Digit adaptation: On easier MNIST and USPS shifts, ADDA achieves performance comparable to CoGANs despite being considerably simpler.The result suggests image-generation machinery is not required for effective adaptation on these shifts.
  • Cross-modality adaptation: Average per-category accuracy rises from 13.9% to 21.1%, while counter accuracy increases from 2.9% to 44.7% after adaptation.Three classes remain unrecovered because they have no correctly labeled target images before adaptation.
  • Cross-modality adaptation: ADDA predicts a wider variety of depth-image classes, sacrificing pillow accuracy while substantially improving many other classes through reasonable confusions.The source-only model overpredicts pillow, whereas ADDA learns a representation useful for recognition.

6. Conclusion

The paper presents a unified adversarial framework for unsupervised domain adaptation and introduces ADDA by combining strategies from prior approaches. ADDA generalizes across four domain shifts, achieves strong benchmark and cross-modality results, and learns target-like representations.

  • 6. Conclusion: The unified framework clarifies similarities and differences among recently proposed adversarial adaptation methods.It provides a cohesive view of each approach’s benefits and key ideas.
  • 6. Conclusion: ADDA combines strategies identified through the framework into a new unsupervised adaptation method.The conclusion presents ADDA as the resulting adaptation approach.
  • 6. Conclusion: Across four domain shifts, ADDA generalizes across varied tasks with strong benchmark and challenging cross-modality adaptation results.The evaluation covers benchmark adaptation datasets and a challenging cross-modality adaptation task.
  • 6. Conclusion: ADDA representations more closely resemble supervised target-domain features than unadapted features.This additional analysis provides further evidence of ADDA’s effectiveness.
Loading 1702.05464v1…