Multi-Step Online Unsupervised Domain Adaptation

Abstract

In this paper, we address the Online Unsupervised Domain Adaptation (OUDA)\nproblem, where the target data are unlabelled and arriving sequentially. The\ntraditional methods on the OUDA problem mainly focus on transforming each\narriving target data to the source domain, and they do not sufficiently\nconsider the temporal coherency and accumulative statistics among the arriving\ntarget data. We propose a multi-step framework for the OUDA problem, which\ninstitutes a novel method to compute the mean-target subspace inspired by the\ngeometrical interpretation on the Euclidean space. This mean-target subspace\ncontains accumulative temporal information among the arrived target data.\nMoreover, the transformation matrix computed from the mean-target subspace is\napplied to the next target data as a preprocessing step, aligning the target\ndata closer to the source domain. Experiments on four datasets demonstrated the\ncontribution of each step in our proposed multi-step OUDA framework and its\nperformance over previous approaches.\n