Manifold Guided Label Transfer for Deep Domain Adaptation

We propose a novel domain adaptation method for deep learning that combines adaptive batch normalization to produce a common feature-space between domains and label transfer with subspace alignment on deep features. The first step of our method automatically conditions the features from the source/target domain to have similar statistical distributions by normalizing the activations in each layer of our network using adaptive batch normalization. We then examine the clustering properties of the normalized features on a manifold to determine if the target features are well suited for the second of our algorithm, label-transfer. The second step of our method performs subspace alignment and k-means clustering on the feature manifold to transfer labels from the closest source cluster to each target cluster. The proposed manifold guided label transfer methods produce state of the art results for deep adaptation on several standard digit recognition datasets.