Kernelized Transfer Joint Matching for Unsupervised Domain Adaptation

Abstract

Transfer learning is an emerging technique through which the machine can learn a new task from the previous experience of another related task to solve the problem of insufficient labelled data. Existing work has explored two transfer learning strategies: feature matching and instance reweighting independently or jointly. However, when we deal with nonlinear training and test data, these strategies may fail. Therefore, to handle this situation, in this paper, we propose a novel kernelized transfer joint matching (KTJM) approach. More specifically, KTJM aims to minimize the distribution gap between the source and the target domains by jointly matching the features and reweighting the instances across domains with Laplacian regularization in a kernelized unified framework. Basically, KTJM constructs a new invariant feature space for both irrelevant instances and distribution gap in an RKHS using d-dimensional embeddings extracted by the kernelized PCA. Extensive experimental results have verified that the proposed kernelized framework outperforms several state-of-the-art transfer learning or domain adaptation methods on the Office-Caltech10 and PIE face data sets. Particularly, KTJM achieved an average accuracy of 90.2% and 79.342% for all classification tasks of Office-Caltech10 data set using Decaf features and PIE face data set, respectively.