Abstract:
Source-free unsupervised domain adaptation aims to learn a model that generalizes well on a target domain given the pre-trained source model and unlabeled target data. Traditional unsupervised domain adaptation methods are mostly not applicable to this setting since no source data are available. To tackle this problem, we propose to generate labeled surrogate source training data from the source model by fixing the model and optimizing the inputs. To avoid naive local fittings to individual instances and in light of the model optimization process, we further enforce model gradient based global fitting constraints on the whole dataset generation and solve the formulated optimization problem using an ADMM algorithm. The generated labeled source training data can then be used to deploy existing unsupervised domain adaptation methods. Furthermore, we propose to incorporate the unlabeled target data into the domain adaptation process to improve generalization in the target domain with a mutual information loss. Experiments show that our proposed method can achieve the state-of-the-art results on benchmark datasets.
