Unsupervised Domain Adaptation via Regularized Conditional Alignment

Safa Cicek, Stefano Soatto

Abstract

We propose a method for unsupervised domain adaptation that trains a shared embedding to align the joint distributions of inputs (domain) and outputs (classes), making any classifier agnostic to the domain. Joint alignment ensures that not only the marginal distributions of the domains are aligned, but the labels as well. We propose a novel objective function that encourages the class-conditional distributions to have disjoint support in feature space. We further exploit adversarial regularization to improve the performance of the classifier on the domain for which no annotated data is available.

Introduction

In modern machine learning, models often encounter a "domain shift" when the distribution of the training (source) data differs from the deployment (target) data. Unsupervised Domain Adaptation (UDA) aims to mitigate this by leveraging labeled source data and unlabeled target data. Traditional approaches, such as Domain-Adversarial Neural Networks (DANN), typically focus on aligning marginal distributions—making source and target features indistinguishable. However, this often ignores the underlying semantics, potentially leading to incorrect class alignment across domains. This paper introduces a novel approach that aligns the joint distribution of inputs and outputs, ensuring class-consistent and discriminative feature mapping.

Proposed Methodology

The proposed framework utilizes a three-pronged architecture: an Encoder ($g$), a Class Predictor ($h_c$), and a Joint Predictor ($h_j$).

The proposed architecture featuring an Encoder, Class Predictor, and Joint Predictor for conditional alignment.

Joint Distribution Alignment

Unlike standard binary discriminators, the Joint Predictor performs a $2K$-way classification task. For $K$ classes, it outputs probabilities for $2K$ categories (e.g., "Source Dog", "Target Dog", "Source Cat", etc.). The Encoder is trained to "fool" this predictor by mapping corresponding classes from different domains to the same region in feature space, effectively minimizing the distance between class-conditional distributions.

Regularization and Semi-Supervised Learning

Since target labels are unavailable, the model employs its own Class Predictor to generate pseudo-labels. To ensure these predictions are robust and the decision boundaries are well-defined, the authors incorporate Virtual Adversarial Training (VAT) and Entropy Minimization. These semi-supervised learning (SSL) techniques push decision boundaries into low-density regions of the data space, preventing the model from making ambiguous predictions.

Visualization of adversarial feature matching with and without input smoothing.

Experimental Evaluation

The method was rigorously tested on standard UDA benchmarks, including digit datasets (MNIST, SVHN, SYN-DIGITS) and object datasets (CIFAR-10, STL-10).

The results demonstrate significant performance gains:

  • MNIST → SVHN: The model achieved an accuracy of 89.19%, a massive improvement over standard DANN (60.6%) and other advanced methods like VADA (73.3%).
  • STL → CIFAR: The approach reached 77.76% accuracy, outperforming the previous state-of-the-art.

Key Insights

  1. Semantic Consistency: Aligning joint distributions ensures that "meaning" is preserved across domains, preventing the semantic scrambling common in marginal alignment methods.
  2. Disjoint Support: The proposed objective function mathematically encourages class-conditional distributions to have disjoint support, which prevents "class collapse" and maintains discriminative power.
  3. Synergistic Regularization: Adversarial domain adaptation and SSL regularization are complementary. While the former handles global domain shifts, the latter refines local decision boundaries and handles noise.

Conclusion

This work presents a robust framework for unsupervised domain adaptation through regularized conditional alignment. By focusing on joint distributions and leveraging adversarial regularization alongside semi-supervised learning, the proposed method achieves state-of-the-art results across various challenging domain shifts. Future work may explore combining this approach with ensemble techniques or advanced data augmentation to further enhance performance on high-resolution datasets.

Summary updated

Abstract

We propose a method for unsupervised domain adaptation that trains a shared embedding to align the joint distributions of inputs (domain) and outputs (classes), making any classifier agnostic to the domain. Joint alignment ensures that not only the marginal distributions of the domains are aligned, but the labels as well. We propose a novel objective function that encourages the class-conditional distributions to have disjoint support in feature space. We further exploit adversarial regularization to improve the performance of the classifier on the domain for which no annotated data is available.
BibTeX
@inproceedings{cicek2019unsupervised,
  author = {Safa Cicek and Stefano Soatto},
  title = {{Unsupervised Domain Adaptation via Regularized Conditional Alignment}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2019},
  url = {https://openreview.net/forum?id=vAJba12o4S},
}
Figure