Source-Free Unsupervised Domain Adaptation

This chapter discusses SFDA, where models trained on labeled source data need to adapt to unlabeled target data without accessing the original source data. Two perspectives are then explored for SFDA: data perturbation and model perturbation. For data perturbation, the idea is to extend the classification boundary of a source-trained model using adversarial data examples. This is inspired by the connection between target samples and adversarial examples, which both violate the i.i.d. assumption. A framework is proposed to generate smooth and diverse adversarial examples to explore unseen distributions and improve generalization. For model perturbation, the weights of a source model are slightly perturbed to introduce uncertainties that allow adapting to target distributions. This converts a deterministic network into a Bayesian neural network with fixed means and learnable variances. Parameter sharing further reduces the number of learnable parameters for efficient adaptation. Model perturbation avoids distorting weights like fine-tuning and is more flexible than only updating batch normalization statistics. Experiments demonstrate the effectiveness of both data and model perturbation for SFDA.