For handling cross-domain distribution mismatch, a specially designed subspace and reconstruction transfer functions bridging multiple domains for heterogeneous knowledge sharing are wanted. In this paper, we propose a novel reconstruction-based transfer learning method called Latent Subspace Transfer Network (LSTN). We embed features/pixels of source and target into reproducing kernel Hilbert space (RKHS), in which the high dimensional features are mapped to nonlinear latent subspace by feeding them into MLP network. This approach is very simple but effective by combining both advantages of subspace learning and neural network. The adaptation behaviors can be achieved in the method of joint learning a set of hierarchical nonlinear subspace representation and optimal reconstruction matrix simultaneously. Notably, as the latent subspace model is a MLP Network, the layers in it can be optimized directly to avoid a pre-trained model which needs large-scale data. Experiments demonstrate that our approach outperforms existing non-deep adaptation methods and exhibits classification performance comparable with that of modern deep adaptation methods.
展开▼
