Subspace dimension selection and averaged learning subspace method handwritten digit classification

Abstract

We present recent improvements in using subspace classifiers in recognition of handwritten digits. Both non-trainable CLAFIC and trainable ALSM methods are used with four models for initial selection of subspace dimensions and their further error-driven refinement. The results indicate that these additions to the subspace classification scheme noticeably reduce the classification error.