Improving Unsupervised Domain Adaptation with Auxiliary Classifier GANs
Bayarchimeg Kalina, Jaeheung Lee · 2023
In this paper, we develop a generative-adversarial-based unsupervised domain adaptation framework to train the classification network and produce source-like images from learned embeddings of both source and target domains simultaneously.1 It is obtained by establishing an interconnection between learned features and an Auxiliary Classifier GAN (ACGAN). To produce source-like images, a deep learning model reconstructs images similar to those in the source dataset. We validated the effectiveness and generality of our method by performing experiments on four adaptation scenarios: USPS to MNIST, MNIST to USPS, SVHN to MNIST, and MNIST to MNISTM. The proposed method works well across four adaptation scenarios.