Active Dynamic Weighting for Multi-Domain Adaptation

Information Science Artificial Intelligence
multi-source domain adaptation transfer learning active learning distribution alignment dynamic boundary loss

Background Introduction

Multi-source Unsupervised Domain Adaptation (MUDA) aims to transfer knowledge from multiple labeled source domains to an unlabeled target domain. However, existing methods often merely seek to blend distributions between different domains or combine multiple single-source models in the decision process through weighted fusion, without deeply investigating the differences between global and local feature distributions across different source and target domains. To address this issue, this study proposes an innovative Active Dynamic Weighting (ADW) method for multi-source domain adaptation.

Paper Source

The research work presented in this paper was carried out by a team from Xi’an University of Technology, consisting of Liu Long, Zhou Bo, Zhao Zhipeng, and Liu Zening, and was officially published online in the journal Neural Networks on May 20, 2024, under the article number 177(2024)106398.

Research Process

Overview of the Process

The research mainly includes the following parts:

  1. Design of Multi-source Dynamic Adjustment Mechanism: ADW dynamically adjusts the degree of feature alignment between source and target domains during training.
  2. Design of Dynamic Boundary Loss: To ensure the distinguishability of cross-domain categories, ADW designs a dynamic boundary loss to guide the model to focus on hard-to-classify samples near the decision boundary.
  3. Active Learning Strategy: For the first time, active learning is applied to multi-source unsupervised domain adaptation. An efficient importance sampling strategy is proposed to select hard samples from the target domain for annotation with minimal labeling budget and integrate them into the training process, thereby further optimizing domain alignment at the category level.

Detailed Explanation of Each Part

1. Multi-source Dynamic Adjustment Mechanism (Dynamic Domain Discrepancy Adjustment, DDDA):

  • Using a shared feature extractor (e.g., ResNet-50) for data mapping.
  • Introducing a Non-linear Projection Head after the shared feature extractor.
  • Calculating the Maximum Mean Discrepancy (MMD) between source and target domains.
  • Dynamically adjusting the weights of source domain samples based on the distribution distance between each source and target domain during training.

2. Dynamic Boundary Loss (Dynamic Boundary Loss):

  • Enhancing the model’s classification ability by focusing on hard-to-classify samples near the decision boundary in each source domain using the dynamic boundary loss method.
  • This loss function extends the differences of similar samples between classes, obtaining a clearer decision boundary.

3. Active Margin Sample Selection (Active Margin Sample Selection, AMSS):

  • Ranking the importance of target domain data based on a query function designed using dynamic boundary loss.
  • Selecting top-ranked samples for annotation and adding these annotated samples to the labeled sample set of the target domain.
  • Updating the model using the new sample set, optimizing the model training effect, and enhancing the classification ability of the target domain.

Main Results

Experiments were tested on multiple benchmark datasets, including Office-31, Office-Caltech 10, Office-Home, and DomainNet datasets. The experimental results show that ADW exhibits superiority in most tasks.

Results of Each Component

AB testing experiments show:

  • Each module of the dynamic domain discrepancy adjustment mechanism (DDDA) significantly improves model performance, and the modules are complementary.
  • The dynamic boundary loss function ((\mathcal{L}_{dis})) further improves the accuracy of classification tasks by focusing on samples near the boundary.
  • The active margin sample selection strategy (AMSS) significantly optimizes model training efficiency by selecting and annotating hard samples.

Comparison with Existing Methods

Overall experimental results indicate that ADW shows significant advantages over existing methods on multiple datasets (including Office-31, Office-Home, and DomainNet). Specifically:

  • On the Office-31 dataset, the ADW method performs best in single target domain tasks, especially in tasks where the target domain is Amazon, with a 1.9% improvement over the existing optimal method.
  • On the more complex DomainNet dataset, ADW achieves the current optimal performance in most task groups.
  • On the Office-Home dataset, ADW achieves the highest recognition accuracy in multiple transfer tasks, with an overall average accuracy superior to other methods.

Ablation Studies of Components

Ablation experiments further verify the contribution of each module to the overall model performance. For example:

  • The model performance improves by 1.1% when using a dynamic adjustment factor.
  • After introducing the active margin sample selection strategy, the model’s average accuracy increases by 2.8%.

Research Significance and Value

  • Scientific Value: This study innovatively introduces the active learning method to the problem of multi-source unsupervised domain adaptation, proposing multi-source dynamic weight adjustment and active margin sample selection strategies. Dynamic adjustment and active learning achieve more efficient cross-domain transfer.
  • Application Value: The effectiveness and robustness of the method have been validated on multiple real datasets. It provides new solutions for machine learning tasks in multi-source data environments, such as image classification and object detection.

Research Highlights

  • Important Findings: ADW addresses the issue of local and overall distribution differences in multi-source domain adaptation, proposing effective dynamic weight adjustment and active sample selection methods, significantly enhancing the model’s cross-domain adaptation capability.
  • Method Novelty: For the first time, dynamic adjustment and active learning are effectively combined and applied to the problem of multi-source domain adaptation, optimizing model training effects while ensuring minimal labeling cost.
  • Data Validation: Extensive experiments on multiple standard datasets demonstrate the superiority and applicability of the method.

Conclusion

The proposed Active Dynamic Weighting method (ADW) offers an innovative solution to the problem of multi-source unsupervised domain adaptation. By dynamically adjusting the feature alignment weights between source and target domains and actively selecting and annotating important target domain samples, efficient cross-domain transfer is achieved. Experimental results validate the method’s effectiveness and robustness, providing new ideas and directions for future research on multi-modal domain adaptation.