Unsupervised Domain Adaptation with Multi-kernel MMD

In this work, we propose a method to solve the problem of unsupervised domain adaptation. Most of existing works are based on adversarial learning method, which obtains the features of inputs through a feature extraction network, and distinguishes the features by a domain classifier that can generate domain-invariant features. The domain-invariant features is used as the input of a classifier (fully connected network) to get the final category of the image. However, in fact, each layer of the fully connected network can be regarded as the domain invariant feature. Consequently, we employ the last layer as domain-invariant features which can also be considered as a probability distribution. Finally, the domain discrepancies between the distribution in source domain and target domain is measured by multi-kernel MMD, which is what we need to minimize. Experimental evidences show that the proposed method achieves satisfactory results on standard domain adapatation benchmarks.