Learning using statistical invariants with privileged information

Xueqin Yan, Chun-Na Li, Yuan-Hai Shao, Yanhui Meng

Abstract: Learning using privileged information has been widely applied across various fields, with most existing research based on the classical SVM+ model. Inspired by recent advances in learning using statistical invariants, this paper introduces a new paradigm for representing privileged information as statistical invariants, enabling its integration into the model in the mode of weak convergence. This method is denoted as learning using statistical invariants with privileged information (LUSIPI). In LUSIPI, privileged information establishes a connection between the decision functions in the input space and privileged space through statistical invariants, which helps to select a suitable set of admissible functions. Additionally, this method enables the simultaneous utilization of multiple types of privileged information. Experimental results on the UCI, MNIST, and 20 newsgroups datasets demonstrate that the method improves the classification performance and exhibits enhanced stability with variations in privileged information compared to SVM+.