Probabilistic Tangent Subspace: A Unified View
Jianguo Lee - Department of Automation, Tsinghua University Jingdong Wang - Department of Automation, Tsinghua University Changshui Zhang - Department of Automation, Tsinghua University Zhaoqi Bian - Department of Automation, Tsinghua University
Tangent Distance (TD) is one classical method for invariant patternclassification. However, conventional TD need pre-obtain tangent vectors, which is very difficultexcept for image objects. This paper extends TD to more general pattern classification tasks.The basic assumption is that tangent vectors can be approximately represented by thepattern variations. We propose three probabilistic subspace models to encode thevariations: the linear subspace, nonlinear subspace, and manifold subspace models. These threemodels are addressed in a unified view, namely Probabilistic Tangent Subspace (PTS).Experiments show that PTS can achieve promising classification performance in non-image datasets.