Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A. Efros, Trevor Darrell, ICML-2018
This paper proposes a novel discriminatively-trained Cycle-Consistent Adversarial Domain Adaptation model. Leveraging the cycle-consistency, the model does not require aligned pairs and the author claims state-of-the-art results across multiple domain adaptation tasks.
The main contribution of the paper is its novel techique for domain adaptation using cycle-consistency losses, taking inspiration from the CycleGAN paper. But while CycleGAN produced task-agnostic domain transfer, this model has been trained for various particular tasks.
Provided is source data Xs, source labels Ys, target data Xt, but no target labels. The goal is to learn a model f to correctly predict label for target data Xt.
What are the cases to be considered in the cases of basict things that we are