The author believes that there are public characteristics and private characteristics between domains
If private features are also migrated, it will cause negative migration.
Therefore, the author proposed Domain Separation Networks
2. Deep Reconstruction-Classification Networks for Unsupervised Domain Adaptation (DRCN) deep reconstruction classification network
DRCN performs supervised classification on labeled source data
Learn the feature expression of unlabeled data in the target domain and convert the images in the source domain into images similar in appearance to the target data set
Encoding parameters are shared between the two tasks, while decoding parameters are separate
Multiple adversarial networks (each network corresponds to a class, k in total), each class in the source domain is assigned a domain discriminator
Which discriminator does the sample in the target domain correspond to? Probability of the kth class x Characteristics of the sample -> Domain discriminator of the kth class
Each discriminator is more specific to the class it focuses on, thereby performing fine-grained alignment of different data distributions.
4. Generate To Adapt: Aligning Domains using Generative Adversarial Networks Generate Adapt: Use Generative Adversarial Networks to align domains
propose an adversarial image generation method that directly learns joint feature space (shared feature) embeddings using labeled data from the source domain and unlabeled data from the target
G generates data that is similar but not identical to the source domain, and facilitates the learning of the F feature extractor by confronting D
D Discrimination - Distinguish between real data and generated data
5. Learning Semantic Representations for Unsupervised Domain Adaptation (moving semantic transfer network-MSTN)
How to use pseudo-labels for domain adaptation?
Align the centers of the same category in the source domain (with labels) and the target domain (pseudo-labels, the network predicts a label) to learn semantic information.
The semantic transfer loss is calculated in each iteration of the model.
