Maximal Domain Independent Representations Improve Transfer Learning

Domain adaptation (DA) adapts a training dataset from a source domain for use in a learning task in a target domain in combination with data available at the target. One popular approach for DA is to create a domain-independent representation (DIRep) learned by a generator from all input samples and then train a classifier on top of it using all labeled samples. A domain discriminator is added to train the generator adversarially to exclude domain specific features from the DIRep. However, this approach tends to generate insufficient information for accurate classification learning. In this paper, we present a novel approach that integrates the adversarial model with a variational autoencoder. In addition to the DIRep, we introduce a domain-dependent representation (DDRep) such that information from both DIRep and DDRep is sufficient to reconstruct samples from both domains. We further penalize the size of the DDRep to drive as much information as possible to the DIRep, which maximizes the accuracy of the classifier in labeling samples in both domains. We empirically evaluate our model using synthetic datasets and demonstrate that spurious class-related features introduced in the source domain are successfully absorbed by the DDRep. This leaves a rich and clean DIRep for accurate transfer learning in the target domain. We further demonstrate its superior performance against other algorithms for a number of common image datasets. We also show we can take advantage of pretrained models.