Matrix-pattern-oriented Ho–Kashyap classifier with regularization learning

Abstract

Existing classifier designs generally base on vector pattern, hence, when a non-vector pattern such as a face image as the input to the classifier, it has to be first concatenated to a vector. In this paper, we, instead, explore using a set of given matrix patterns to design a classifier. For this, first we represent a pattern in matrix form and recast existing vector-based classifiers to their corresponding matrixized versions and then optimize their parameters. Concretely, considering its similar principle to the support vector machines of maximizing the separation margin and superior generalization performance, the modified HK algorithm (MHKS) is chosen and then a matrix-based MHKS classifier (MatMHKS) is developed. Experimental results on ORL, Letters and UCI data sets show that MatMHKS is more powerful in generalization than MHKS. This paper focuses on: (1) purely exploring the classification performance discrepancy between matrix- and vector-pattern representations; more importantly, (2) developing a new classifier design directly for matrix pattern.