PaperlayerView exact revision on arXiv

Source-linked AI summary

MetaAlign: Coordinating Domain Alignment and Classification for Unsupervised Domain Adaptation

Guoqiang Wei, Cuiling Lan, Wenjun Zeng, Zhibo Chen

arXiv:2103.13575v1cs.CV

TL;DR

Alignment-based UDA can leave domain alignment and classification objectives with inconsistent optimization directions, limiting coordination between them. MetaAlign treats the objectives as meta-train and meta-test tasks; experiments report state-of-the-art performance for image-classification evaluations and validate its effectiveness on object-detection UDA.

  • Problem

    Alignment-based UDA methods may optimize domain alignment inconsistently with classification, potentially harming discriminative features and classification performance.

    Source: §1 · p. 1, §3.2 · p. 5

  • Method

    MetaAlign treats domain alignment and classification as meta-train and meta-test tasks to coordinate their optimization.

    Source: §1 · p. 1

  • Results

    Experiments report state-of-the-art performance for image-classification evaluations and validate MetaAlign on object-detection UDA.

    Source: §1 · p. 2

  • Takeaways & Limitations

    MetaAlign is applicable to alignment-based UDA for object classification and detection while preserving feature discrimination for recognition.

    Source: §1 · p. 2

Abstract

from arXiv · show

For unsupervised domain adaptation (UDA), to alleviate the effect of domain shift, many approaches align the source and target domains in the feature space by adversarial learning or by explicitly aligning their statistics. However, the optimization objective of such domain alignment is generally not coordinated with that of the object classification task itself such that their descent directions for optimization may be inconsistent. This will reduce the effectiveness of domain alignment in improving the performance of UDA. In this paper, we aim to study and alleviate the optimization inconsistency problem between the domain alignment and classification tasks. We address this by proposing an effective meta-optimization based strategy dubbed MetaAlign, where we treat the domain alignment objective and the classification objective as the meta-train and meta-test tasks in a meta-learning scheme. MetaAlign encourages both tasks to be optimized in a coordinated way, which maximizes the inner product of the gradients of the two tasks during training. Experimental results demonstrate the effectiveness of our proposed method on top of various alignment-based baseline approaches, for tasks of object classification and object detection. MetaAlign helps achieve the state-of-the-art performance.

1. Introduction

Domain shift makes source-trained models perform poorly on new domains, while alignment-based UDA can optimize domain alignment inconsistently with classification. MetaAlign uses a meta-learning scheme to coordinate these objectives, and experiments report state-of-the-art performance for image classification alongside validation on object-detection UDA.

1. Introduction

  • Alignment-based UDA methods may optimize domain alignment inconsistently with classification, potentially harming discriminative features and classification performance.Existing approaches align distributions through explicit discrepancy minimization or adversarially learned representations.Source: §1 · p. 1
  • MetaAlign treats domain alignment and classification as meta-train and meta-test tasks to encourage coordinated optimization.The figure contrasts this strategy with directly combining the two objectives, whose descent directions may be inconsistent.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • MetaAlign’s coordination maximizes the inner product of the two tasks’ gradients during training.The analysis identifies this gradient objective as the mechanism for coordinating optimization.Source: §1 · p. 2
  • The strategy is designed for alignment-based UDA in object classification and detection while preserving feature discrimination for recognition.The paper evaluates it on image classification, including UDA and domain generalization, and object-detection UDA.Source: §1 · p. 2
  • MetaAlign achieves state-of-the-art performance in image-classification evaluations and is validated on object-detection UDA.The image-classification evaluations include UDA and domain generalization.Source: §1 · p. 2

2. Related Work

Related UDA work commonly learns domain-invariant representations through explicit distribution-discrepancy measures or adversarial training. MetaAlign targets optimization inconsistency in alignment-based methods and is also positioned for object-detection UDA and the domain-invariant-feature category of domain generalization.

Unsupervised Domain Adaptation

  • UDA approaches to domain-invariant representations mainly reduce explicit distribution discrepancy or use adversarial training.Examples include MMD and higher-order statistics for explicit alignment, and domain discriminators for adversarial alignment.Source: §2 · p. 2
  • Existing alignment approaches optimize alignment and classification directly, while MetaAlign uses meta-optimization to mitigate their inconsistency.The paper describes the inconsistency as a potential challenge to better UDA.Source: §2 · pp. 2–3

UDA for Object Detection

  • Object-detection UDA includes adversarial image- and instance-level alignment, global/local feature alignment, and region-based approaches.The cited examples include DA-Faster, SW-DA, and K-means-clustered region alignment.Source: §2 · p. 3

Domain Generalization

  • Domain generalization differs from UDA because the target domain is inaccessible during training, and its methods include domain-invariant features, augmentation, meta-learning, and auxiliary tasks.MetaAlign is described as applicable to the domain-invariant-feature category.Source: §2 · p. 3

Neural Network Meta Learning

  • The paper positions MetaAlign as a meta-optimization strategy that treats alignment and classification objectives as meta-train and meta-test tasks.Its stated aim is to mitigate the underlying optimization inconsistency in UDA.Source: §2 · p. 3

3. Proposed MetaAlign for UDA

UDA transfers knowledge from labeled source data to unlabeled target data, where domain alignment may conflict with classification optimization. MetaAlign uses meta-optimization to coordinate these tasks.

3. Proposed MetaAlign for UDA

  • Existing UDA alignment methods may not work harmoniously with classification, potentially damaging feature discrimination; MetaAlign is proposed to encourage consistency between the tasks.Source: §3 · p. 3

3.1. Recap of Alignment Based UDAs

  • Alignment-based UDA methods use adversarial training to fool a domain discriminator or explicit metrics such as MMD to reduce feature-distribution discrepancy.The section describes DANN and MMD as representative frameworks.Source: §3.1 · pp. 3–4
  • For adversarial adaptation, a feature generator and object classifier are optimized with classification and domain-alignment objectives, while a discriminator distinguishes source from target features.The domain alignment loss is defined as the negative domain-classification loss; the baseline objective omits the balancing hyperparameter for simplicity.Source: §3.1 · p. 4

3.2. Meta-learning to Align Lcls and Ldom

  • MetaAlign treats domain alignment and classification as meta-train/meta-test tasks on the same samples, then evaluates both objectives after a gradient step on the meta-train task.The tasks can exchange roles during training; the method’s analysis links coordination to maximizing the inner product of their gradients.Source: §3.2 · p. 5
  • The method learns a scalar β_m for each parameter group to weight optimization consistency and adds an L1 constraint to avoid a trivial solution.Groups can correspond to convolutional blocks; the training objective includes a penalty on the weights.Source: §3.2 · p. 6
  • MetaAlign can alternate which task is meta-train and which is meta-test; the described procedure illustrates domain alignment as meta-train and classification as meta-test.Source: §3.2 · p. 6

4. Experiments

The experiments evaluate MetaAlign on UDA classification, object detection, and domain generalization. They cover multiple alignment-based baselines and report improved results across these settings.

4. Experiments

  • The experiments cover UDA classification, UDA object detection, and domain-generalization classification.The paper directs readers to supplementary material for further experimental details.Source: §4 · p. 6, §4.2 · p. 8

4.1.1 Datasets and Settings

  • UDA classification is evaluated on Office31 and Office-Home, including all 12 one-source-to-one-target Office-Home cases; reported results average multiple runs.Office31 has three domains and 31 categories; Office-Home has four domains and 65 categories.Source: §4.1.1 · p. 6

4.1.2 Ablation Study

  • On Office-Home, MetaAlign improves average accuracy for MMD, DANN, CDAN, DANNPE, and GVB by 1.0%, 3.9%, 2.0%, 1.8%, and 0.9%, respectively.The improvements are reported regardless of differences in the domain-alignment loss design.Source: §4.1.2 · p. 7
  • On Office-Home with DANNPE, learned β yields a 1.8% accuracy improvement versus 1.4% without β; alternative meta-train/task-role settings differ by less than 0.3% accuracy.The passage attributes the role-setting similarity to those settings sharing the same optimization objective as Eq. (7).Source: §4.1.2 · p. 7

4.1.3 Comparisons with State-of-the-Arts

  • GVB+MetaAlign outperforms GVB and achieves the best performance on both Office-Home and Office31.Table 2 reports Office31 classification accuracy with a ResNet-50 backbone.Source: §4.1.3 · p. 7, Table 2 · p. 7

4.1.4 Feature Visualization

  • Grad-CAMs show MetaAlign focusing domain alignment on foreground-related regions, while t-SNE shows more compact clusters and fewer boundary-scattered points than DANNPE.The t-SNE comparison is for Pr → Cl and includes Source-Only, DANNPE, and DANNPE+MetaAlign.Source: §4.1.4 · pp. 7–8, Figure 4 · p. 7

4.2. Experiments on UDA for Object Detection

  • MetaAlign improves Pascal VOC-to-Watercolor2k mAP by 2.3 (4.6%) over W-DA and 1.9 (3.5%) over SW-DA; SW-DA with MetaAlign achieves state-of-the-art performance.Bird results are unstable because bird bounding boxes comprise about 2.6% of all boxes.Source: §4.2 · p. 8, Table 5 · p. 8

4.3. Experiments on Domain Generalization

  • On PACS, DANNPE improves over AGG by about 1.6%, and MetaAlign adds 1.2%; DANNPE+MetaAlign outperforms prior methods in average accuracy, especially on Sketch.Sketch has the largest domain gap among the domains, and the passage says it therefore requires more powerful domain-invariant features.Source: §4.3 · p. 8, Table 4 · p. 8

5. Conclusion

The paper identifies optimization inconsistency between domain alignment and classification in alignment-based UDA, and proposes MetaAlign to coordinate the tasks. Experiments validate its applicability to classification and detection.

5. Conclusion

  • MetaAlign is applicable to various alignment-based UDA methods for classification and detection.The paper identifies inconsistency between domain alignment and classification as the problem it addresses.Source: §5 · p. 8
  • MetaAlign treats one task as meta-train and the other as meta-test, and its objective analysis indicates coordinated optimization of the two tasks.Source: §5 · p. 8

1. More Details of Baselines

The baseline descriptions distinguish DANNPE’s probability-based discriminator and entropy weighting from DANN, and describe MMD as a feature-alignment constraint. They also specify the kernel used for MMD and introduce a MetaAlign objective for MMD-based UDA.

DANNPE

  • DANNPE feeds predicted class probabilities to its domain discriminator and entropy-weights samples, and the paper reports that it significantly outperforms DANN.The weighting prioritizes easy-to-transfer samples; the pipeline is shown in Figure 5.Source: §1 · p. 12

MMD

  • MMD directly constrains the output of G to encourage feature alignment between source- and target-domain data.Source: §1 · p. 12
  • The MMD description uses the characteristic RBF kernel, with f_i=G(x_i) and K denoting a kernel function.The kernel bandwidth is σ.Source: §1 · p. 12
  • The paper gives a MetaAlign optimization objective for MMD-based UDA, analogous to its Eq. (8).Source: §1 · p. 12

2. Experiments

Experiments include UDA classification and detection, plus domain generalization on PACS. In Watercolor2k visualizations, SW-DA+MetaAlign has fewer false and missing detections than SW-DA.

2. Experiments

  • The experiments section presents additional implementation, dataset, setting, competitor, and result details.Source: §2 · p. 12

2.1. UDA for Classification

  • The experiments provide implementation details and a DANNPE pipeline overview, and show classification-task Grad-CAMs focused on foreground objects.The classification setup uses an ImageNet-pretrained ResNet-50 feature extractor and divides its convolutional layers into four groups.Source: §2.1 · p. 12, Figure 5 · p. 12, Figure 6 · p. 13

2.2. UDA for Object Detection

  • The detection experiments use Pascal VOC as source and Watercolor2k as target, with SW-DA as the MetaAlign baseline among the listed competitors.Watercolor2k has 2K images, including 1K for training and 1K for testing; images are resized to a shorter side of 600 pixels.Source: §2.2 · pp. 12–13
  • SW-DA+MetaAlign yields more accurate Watercolor2k detections, with false and missing detections largely reduced relative to SW-DA.Figure 7 compares SW-DA in its first two rows with SW-DA+MetaAlign in its last two rows.Source: §2.2 · pp. 13–14, Figure 7

2.3. Domain Generalization

  • The domain-generalization experiments use PACS under leave-one-out evaluation, training on three domains and testing on the remaining domain.The DANNPE domain discriminator distinguishes the three source domains.Source: §2.3 · p. 14

Related papers

Loading 2103.13575v1…