We present a novel and unified deep learning framework which is capable of learning domain-invariant representation from data across multiple domains. Realized by adversarial training with additional ability to exploit domain-specific information, the proposed network is able to perform continuous cross-domain image translation and manipulation, and produces desirable output images accordingly. In addition, the resulting feature representation exhibits superior performance of unsupervised domain adaptation, which also verifies the effectiveness of the proposed model in learning disentangled features for describing cross-domain data.

Paper summary: The authors propose a VAE-GAN architecture for learning multi-domain image representations where the domain information is disentangled from the domain invariant information, while simultaneously generating images of all considered domains. A domain classifier D_v pushes the VAE encoder E to learn a domain-invariant embedding z by adversarial training. This embedding is then concatenated with a one-hot encoding v of the domain before being fed to the VAE decoder G, which is then able to reconstruct domain-appropriate images. The VAE is further paired with an adversarial real / fake classifier, which improves image sharpness, and a domain classifier, which promotes mutual information between the generated images and the domain encoding v. Experiments are performed on human face generation and semi-supervised digits classification. Comments: The proposed approach is pretty straightforward when compared to recent works on image generation / feature disentanglement (which is good), and the "unified" nature of the proposed architecture must be appreciated.

We present a novel and unified deep learning framework which is capable of learning domain-invariant representation from data across multiple domains. Realized by adversarial training with additional ability to exploit domain-specific information, the proposed network is able to perform continuous cross-domain image translation and manipulation, and produces desirable output images accordingly. In addition, the resulting feature representation exhibits superior performance of unsupervised domain adaptation, which also verifies the effectiveness of the proposed model in learning disentangled features for describing cross-domain data. Papers published at the Neural Information Processing Systems Conference.