Model Adaptation: Unsupervised Domain Adaptation without Source Data

TL;DR

This work tackles unsupervised model adaptation without access to source data by introducing 3C-GAN, a collaborative framework where a generator conditioned on class labels ($x_g = G(y,z)$) and a fixed pre-trained classifier $C$ co-evolve with a discriminator to produce target-style samples guiding adaptation. The model augments training with a weight regularization term $\ obreakspace\ell_{wReg}$ to keep the adapted predictor close to the source, and a clustering-based regularization $\ell_{cluReg}$ (with a VAT-style perturbation) to enforce local smoothness and improve decision boundaries. The approach is validated on multiple benchmarks (digits, traffic signs, Office-31, VisDA17) and consistently outperforms or matches state-of-the-art methods while not requiring source data during adaptation. This demonstrates practical, data-privacy-friendly domain adaptation with strong generalization across diverse visual domains.

Abstract

In this paper, we investigate a challenging unsupervised domain adaptation setting -- unsupervised model adaptation. We aim to explore how to rely only on unlabeled target data to improve performance of an existing source prediction model on the target domain, since labeled source data may not be available in some real-world scenarios due to data privacy issues. For this purpose, we propose a new framework, which is referred to as collaborative class conditional generative adversarial net to bypass the dependence on the source data. Specifically, the prediction model is to be improved through generated target-style data, which provides more accurate guidance for the generator. As a result, the generator and the prediction model can collaborate with each other without source data. Furthermore, due to the lack of supervision from source data, we propose a weight constraint that encourages similarity to the source model. A clustering-based regularization is also introduced to produce more discriminative features in the target domain. Compared to conventional domain adaptation methods, our model achieves superior performance on multiple adaptation tasks with only unlabeled target data, which verifies its effectiveness in this challenging setting.

Figures (6)

• Figure 1: Comparison between conventional data-based adaptation (left) and our model adaptation (right). Conventional unsupervised domain adaptation methods require labeled source data during adaptation, while our proposed model adaptation method only relies on unlabeled target data.
• Figure 2: An overview of the proposed architecture. During target generation (top), we aim to learn a class conditional generator $G$ for producing target-style training samples $x_g = G(y,z)$ via the discriminator $D$ and the prediction model $C$ (which is fixed as denoted by dashline). The generated images and proposed regularizations are used for model adaptation (bottom). These two procedures are repeated, with $G$ and $C$ collaborating with each other. (See text for details)
• Figure 3: Class conditional generation in (a) MNIST$\rightarrow$MNIST-M and (b) SVHN$\rightarrow$MNIST. The top row indicates the samples generated with pre-trained source model, and the bottom row refers to the samples generated during the last adaptation stage.
• Figure 4: Class-conditional generation in (a) Syn.Digits$\rightarrow$SVHN and (b) Syn.Sign$\rightarrow$GTSRB (shows the first 19 out of 43 classes). Each column has the same class $y$ and the rows share the same noise vector $z$.
• Figure 5: The t-SNE projection of the last hidden layer of target features (a) before adaptation and (b) after adaptation in the task of Syn.Sign$\rightarrow$GTSRB. Different colors represent different classes.