PaperlayerView exact revision on arXiv

Source-linked AI summary

Maximum Classifier Discrepancy for Unsupervised Domain Adaptation

Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, Tatsuya Harada

arXiv:1712.02560v4cs.CV

TL;DR

Unsupervised domain adaptation must transfer from labeled source data to unlabeled target data despite domain shifts, while conventional domain matching can ignore task-specific boundaries and create ambiguous features. The paper uses task-specific classifiers to maximize target-output discrepancy and a feature generator to minimize it, and reports outperforming other methods across image-classification and semantic-segmentation datasets.

  • Problem

    Unsupervised domain adaptation lacks target labels, while domain-classifier matching ignores task-specific decision boundaries and can generate ambiguous features near class boundaries.

  • Method

    The method uses task-specific classifiers to detect target features far from source support and trains a feature generator to move them near that support by minimizing output discrepancy.

  • Results

    The method outperformed state-of-the-art methods in almost all image-classification and semantic-segmentation experiments.

  • Takeaways & Limitations

    Using task-specific classifier feedback enables target features to avoid class boundaries while aligning source and target distributions.

  • Takeaways & Limitations

    The paper’s theoretical insight includes an assumption involving h and h.

Abstract

from arXiv · show

In this work, we present a method for unsupervised domain adaptation. Many adversarial learning methods train domain classifier networks to distinguish the features as either a source or target and train a feature generator network to mimic the discriminator. Two problems exist with these methods. First, the domain classifier only tries to distinguish the features as a source or target and thus does not consider task-specific decision boundaries between classes. Therefore, a trained generator can generate ambiguous features near class boundaries. Second, these methods aim to completely match the feature distributions between different domains, which is difficult because of each domain's characteristics. To solve these problems, we introduce a new approach that attempts to align distributions of source and target by utilizing the task-specific decision boundaries. We propose to maximize the discrepancy between two classifiers' outputs to detect target samples that are far from the support of the source. A feature generator learns to generate target features near the support to minimize the discrepancy. Our method outperforms other methods on several datasets of image classification and semantic segmentation. The codes are available at \url{https://github.com/mil-tokyo/MCD_DA}

1. Introduction

Unsupervised domain adaptation addresses transfer from labeled source data to unlabeled target data, but domain shifts and domain-classifier matching can produce ambiguous target features. The proposed adversarial method uses task-specific classifiers and decision boundaries to align target features, with evaluations spanning multiple tasks.

  • Unsupervised domain adaptation transfers knowledge from a label-rich source domain to a label-scarce target domain without target labels.
  • Domain-classifier adaptation matches source and target feature distributions without considering task-specific categories or decision boundaries.
  • Ignoring decision boundaries can produce ambiguous target features near class boundaries rather than discriminative features.
  • The proposed method aligns target distributions using task-specific decision boundaries instead of domain labels alone.
  • Task-specific classifiers detect target features far from source support, while a feature generator learns to move them near that support.
  • The method is evaluated on digit classification, object classification, and semantic segmentation, with toy-problem analysis also reported.

2. Related Work

Prior work includes feature-learning, pseudo-labeling, image-translation, and domain-classifier approaches for domain adaptation. The paper distinguishes its two-classifier training from consensus regularization because its classifiers can be constructed from one source domain.

  • Earlier domain-adaptation strategies used autoencoders, clustering with pseudo-labels, cross-domain image translation, and middle-feature distribution matching.
  • A representative approach trains a domain classifier to predict source or target while a shared feature extractor learns features that deceive it.
  • The paper contrasts its disagreement-minimization step with consensus regularization, which maximizes agreement among classifiers in multi-source or multi-view settings.
  • Unlike consensus regularization, the proposed method can construct different classifiers using only one source domain.

3. Method

The method aligns source and target features using two task-specific classifiers whose disagreement identifies target samples outside the source support. Adversarial training maximizes this discrepancy for the classifiers and minimizes it for the generator while preserving source discriminability, with a connection to H∆H-distance theory.

  • Overall idea: The method uses a feature generator G and two task-specific classifiers F1 and F2 that output K-class probability vectors.Both classifiers receive features generated from labeled source or unlabeled target images.
  • Overall idea: Target samples far from the source support are detected through disagreement between the two distinct classifiers.Such samples are likely to lie near class boundaries and lack clearly discriminative features.
  • Adversarial training: The classifiers maximize target discrepancy while the generator minimizes it, encouraging target features to move inside the source support.Without classifier maximization, the classifiers may become too similar to detect unsupported target samples.
  • Discrepancy loss: The discrepancy loss is the average absolute difference between the classifiers’ class probabilities, using L1 distance rather than L2 distance.For class k, p1k and p2k denote the probability outputs of the two classifiers.
  • Training procedure: Training first minimizes source classification loss, then alternates classifier discrepancy maximization with generator discrepancy minimization.The source loss is crucial for learning task-specific discriminative features and maintaining correct source classification.
  • Theoretical insight: The minimax training objective is closely related to H∆H-distance, which measures disagreement between classifiers and appears in a bound on target-domain error.The relationship holds alongside source classification loss and the shared error of the ideal joint hypothesis.

4. Experiments on Classification

The experiments evaluate the method across toy, digits, traffic-sign, and VisDA classification settings, finding improved target performance and more discriminative adapted features.

  • Experimental settings: The classification experiments cover digits, traffic signs, and object adaptation, including rotated inter-twinning moons, MNIST, SVHN, USPS, SYN SIGNS, GTSRB, and VisDA.The evaluation uses multiple cross-domain settings and compares the proposed method with distribution-matching approaches.
  • Digits and traffic signs: The proposed method outperformed distribution matching-based methods in all reported digits and traffic-sign settings, although ATDA performed better in some situations.ATDA used a few labeled target samples to select hyper-parameters, unlike the unsupervised protocol used here.
  • Digits and traffic signs: As discrepancy loss diminished, accuracy improved, supporting discrepancy minimization as a route to accurate adaptation.Figure 5 reports this relationship during training.
  • Feature visualization: Adapted target features aligned with source classes and separated clearly, although the source and target distributions were not completely matched.The visualization contrasts non-adapted and adapted feature organization.
  • VisDA classification: On VisDA, the method achieved much better accuracy than other distribution-matching methods and exceeded the source-only model in every class.MMD and DANN performed worse than the source-only model for some classes, including car and plant.
  • VisDA classification: On VisDA, performance improved as the generator-update count n increased, which the authors associate with the large synthetic-to-real domain difference.The authors suggest more generator updates were needed to align the substantially different distributions.

5. Experiments on Semantic Segmentation

The method is evaluated for semantic segmentation when adapting synthetic GTA5 or Synthia images to real Cityscapes images. It improves performance despite the large synthetic-to-real domain gap and outperforms DANN in the reported comparisons.

  • Experimental settings: The segmentation experiments adapt synthetic GTA5 or Synthia data to real Cityscapes data, using Cityscapes training images for training and its validation set for testing.The setup addresses adaptation across synthetic and real image domains.
  • Experimental settings: The method was evaluated with VGG-16-based FCN-8s and DRN-D-105 architectures.Model architecture and additional hyper-parameters are described in the supplementary material.
  • Results: The quantitative and qualitative results show improved performance despite the large domain difference between synthetic and real images.The reported comparison considers the mIoU of the source-only model.
  • Results: The proposed method performs clearly better than DANN in the reported semantic-segmentation comparison.Tables 3 and 4 provide adaptation results for GTA5-to-Cityscapes and Synthia-to-Cityscapes.

6. Conclusion

The paper proposes task-specific classifiers for unsupervised domain adaptation, using them to align target features near the source support while avoiding class boundaries. Across image classification and semantic segmentation datasets, the method outperformed state-of-the-art methods in almost all experiments.

  • Conclusion: Task-specific classifiers detect target samples far from the source support, while a feature generator learns to move target features near that support.The generator uses classifier feedback during adversarial alignment.
  • Conclusion: Classifier feedback helps the generator avoid generating target features near class boundaries.
  • Evaluation: The method was extensively evaluated on image classification and semantic segmentation datasets.
  • Evaluation: In almost all experiments, the method outperformed state-of-the-art methods.

Toy Dataset Experiment

The toy-dataset experiment is documented with a specified data-generation setup and model-training configuration. The reported configuration includes a fixed rotation, optimizer settings, network architecture, and training budget.

  • Toy Dataset Experiment: The toy-dataset experiment details are provided in the supplementary section.
  • Setup: Target samples were generated with a rotation angle of 30.
  • Optimization: The model was optimized with Adam at learning rate 2.0×10^-4 using batch size 200.
  • Architecture: The feature generator used a three-layer fully connected network with 15 hidden neurons and ReLU activation.

Experiment on Digit Dataset

The digit-domain experiments use specified training and evaluation protocols across four adaptation scenarios. The protocols differ in training duration, sample availability, and dataset-specific sample selection.

  • Training Protocol: Accuracy was reported after 20,000 iterations, except for MNIST↔USPS, which stopped after 200 epochs to prevent over-fitting.The MNIST↔USPS schedule used 13 iterations per epoch.
  • SVHN→MNIST: SVHN→MNIST used the standard training set for training and the testing set for testing in both source and target domains.
  • SYN DIGITS→SVHN: SYN DIGITS→SVHN used 479400 source samples and 73257 target samples for training, with 26032 samples for testing.
  • SYN SIGNS→GTSRB: SYN SIGNS→GTSRB selected 31367 target samples for training and evaluated accuracy on the remaining samples.
  • MNIST↔USPS: MNIST↔USPS followed protocols from ADDA and PixelDA, using either selected training subsets or all training samples.The ADDA protocol used 2,000 MNIST and 1,800 USPS training samples.

Experiment on VisDA Classification Dataset

The VisDA experiment compares methods under described architectural and loss settings. The setup includes class-balance regularization, matched classification-network architectures, MMD with RBF kernels, and a DANN baseline with fixed learning rate.

  • Class Balance Loss: Class-balance loss was added to improve accuracy by assigning target samples to classes equally.The loss was also incorporated into comparable methods.
  • Class Balance Loss: The class-balance loss was weighted by λ = 0.01 and added during Steps 2 and 3 of the method.The same loss was also introduced in MMD and DANN updates.
  • Architecture: All methods used the same classification-network architecture, with 1000 neurons in fully connected layers.
  • MMD: MMD was computed at the feature generator's last layer using RBF kernels.
  • DANN: The DANN baseline used two-layer domain-classification networks with 100 hidden neurons and a fixed learning rate of 1.0 × 10^-3.

Experiments on Semantic Segmentation

The semantic-segmentation experiments evaluate adaptation across vehicle-egocentric datasets, using specified network architectures, training settings, and segmentation metrics.

  • Datasets: The experiments evaluate adaptation from synthetic GTA5 and Synthia datasets to the real-world Cityscapes dataset.GTA5 contains 24,996 images with annotations compatible with Cityscapes classes.
  • Training Details: Training uses a weight decay ratio of 2 × 10^-5 and no data augmentation, while ignoring pixels annotated as void.No predicted void label exists during testing under this setup.
  • Network Architecture: The method is applied to FCN-8s with VGG-16 and to DRN-D-105-based dilated residual networks.For FCN-8s, convolutional layers serve as the generator and fully connected layers as classifiers; for DRN-D-105, lower layers serve as the generator.
  • Evaluation Metrics: Segmentation performance is measured with intersection-over-union and pixel accuracy.IoU uses true-positive, false-positive, and false-negative pixel counts over the whole test set.
  • Evaluation Metrics: The evaluation uses code released with the VisDA challenge.The cited evaluation implementation is provided through the VisionLearningGroup taskcv-2017public repository.

Additional Results

Additional experiments show that gradient reversal enables one-step adversarial training with performance comparable to the original procedure, while generator-update effects vary across datasets.

  • Training via Gradient Reversal Layer: Gradient reversal enables one-step adversarial training, removes the generator-update hyper-parameter, and makes training faster.It flips gradient signs during back-propagation so classifier discrepancy maximization and generator discrepancy minimization occur simultaneously.
  • Results: The GRL-based model achieves the same level of semantic-segmentation performance as the proposed multi-step training procedure.This comparison is reported for the semantic-segmentation experiments.
  • Sensitivity to Hyper-Parameter: 96.0% and 96.2% average accuracy were obtained for SVHN-to-MNIST experiments with n = 5 and n = 6, respectively.Accuracy appears to increase with n but eventually saturates, while higher accuracy can require more training time.
  • Additional Results: The semantic-segmentation adaptation results are presented for GTA5-to-Cityscapes and Synthia-to-Cityscapes settings.Figure 7 qualitatively compares input, ground truth, source-only, DANN, and proposed-method results for GTA5-to-Cityscapes adaptation.
Loading 1712.02560v4…