Transfer Learning with Dynamic Adversarial Adaptation Network

Chaohui Yu, Jindong Wang, Yiqiang Chen, Meiyu Huang

Introduction

In deep learning, the availability of labeled data is often a bottleneck. While we might possess a vast labeled "source" dataset for one task, we frequently encounter "target" domains with only unlabeled data. Unsupervised Domain Adaptation (UDA) aims to bridge this gap. Traditional adversarial learning methods often apply a static approach to alignment. The Dynamic Adversarial Adaptation Network (DAAN) addresses this by allowing the model to dynamically prioritize different levels of distribution alignment.

The Challenge of Domain Alignment

Effective domain adaptation requires aligning distributions across different levels. Specifically, models must consider:

  1. Marginal Distribution: The global alignment of data features, irrespective of class labels.
  2. Conditional Distribution: The alignment of class-specific features to ensure categories like "chairs" in the source match "chairs" in the target.

Existing methods often focus on one at the expense of the other. For instance, DANN emphasizes global alignment, while MADA targets subdomains. However, the relative importance of marginal versus conditional alignment varies by task.

Motivation for dynamic adaptation: choosing between marginal and conditional alignment.

Dynamic Adversarial Adaptation Network (DAAN)

DAAN introduces a framework to quantitatively evaluate and weight these alignments on the fly. The architecture consists of four primary components:

  • Feature Extractor ($G_f$): A deep backbone (e.g., ResNet-50) for high-level feature extraction.
  • Label Classifier ($G_y$): Trained on labeled source data for class prediction.
  • Global Domain Discriminator ($G_d$): An adversarial module for global (marginal) distribution alignment.
  • Local Subdomain Discriminators ($G_d^c$): A set of $C$ discriminators (one per class) for category-specific (conditional) alignment.
The DAAN architecture featuring global and local discriminators with a dynamic adversarial factor.

The Dynamic Adversarial Factor ($\omega$)

The core innovation is the Dynamic Adversarial Factor ($\omega$). Instead of a fixed weight, DAAN calculates $\omega$ dynamically using the $\mathcal{A}$-distance, which measures distributional discrepancy. If global distributions differ significantly, $\omega$ prioritizes marginal alignment. As they converge, the focus shifts to local subdomains for fine-grained class alignment. This is implemented via a Gradient Reversal Layer (GRL), enabling end-to-end training.

Experimental Evaluation

The authors benchmarked DAAN on two major datasets:

  • ImageCLEF-DA: 12 categories across Caltech-256, ImageNet, and Pascal VOC.
  • Office-Home: A challenging set of 15,588 images across 65 categories and 4 domains (Artistic, Clip Art, Product, Real-World).

Key Findings

  • Distribution "Attention": The factor $\omega$ acts as an attention mechanism for distributions, focusing on the most critical discrepancies.
  • Efficiency: Unlike previous kernel-based methods (e.g., MEDA), DAAN is purely deep-learning-based and scales linearly with data size.
  • Performance: DAAN achieved state-of-the-art results, reaching 61.8% accuracy on Office-Home (surpassing MEDA's 60.2%) and 86.8% on ImageCLEF-DA.
  • Ablation & Visualization: Studies confirmed that dynamic weighting outperforms any fixed $\omega$. t-SNE visualizations showed DAAN produces tighter, more distinct class clusters in the target domain.

Conclusion and Future Work

DAAN provides a robust, efficient method for unsupervised domain adaptation by balancing global and local feature alignment. While it relies on pseudo-labels for target data—which can be noisy early in training—it sets a new standard for adversarial transfer learning. Future research may extend this dynamic alignment to more complex visual tasks like object detection and image segmentation.

Summary updated

BibTeX
@inproceedings{yu2019transfer,
  author = {Chaohui Yu and Jindong Wang and Yiqiang Chen and Meiyu Huang},
  title = {{Transfer Learning with Dynamic Adversarial Adaptation Network}},
  booktitle = {ICDM},
  year = {2019},
  url = {https://openreview.net/forum?id=yhtVKoeyBZ},
}
Figure