Learning Transferable Features with Deep Adaptation Networks

This paper addresses domain adaptation, where a model trained on one domain must generalize to another. The authors note that although deep neural networks learn transferable features, those features move from general to specific in higher layers, so transferability drops sharply as domain discrepancy increases. To counter this, they propose the Deep Adaptation Network (DAN), which extends deep convolutional neural networks to the domain adaptation setting. In DAN, the hidden representations of all task-specific layers are embedded in a reproducing kernel Hilbert space, where the mean embeddings of the different domain distributions can be explicitly matched, and an optimal multikernel selection method is used to further reduce domain discrepancy.

DAN can learn transferable features with statistical guarantees and scales linearly thanks to an unbiased estimate of the kernel embedding. Extensive experiments show that the architecture yields state-of-the-art image classification error rates on standard domain adaptation benchmarks. By formally reducing dataset bias in task-specific layers, the work provided a principled kernel-based approach to making deep features transferable across domains.