In this paper, we propose a new learning algorithm for the Subspace Pa
ttern Recognition Method (SPRM) called the Hebbian Learning Subspace M
ethod (HLSM). It uses the notion of a weighted squared orthogonal proj
ection distance which gives different weightages to different basis ve
ctors in the computation of the orthogonal projection distance. The pr
inciple applied during learning is the same as that used in the earlie
r Learning Subspace Method (LSM): the projection on the wrong subspace
is always decreased and the one on the correct subspace is always inc
reased. We also propose a neural implementation for the HLSM. Experime
nts have been conducted on an extensive numeric set of handprinted cha
racters involving 16659 samples using the SPRM, the HLSM and the Avera
ged LSM. Excellent results have been obtained using all the subspace m
ethods thus demonstrating the suitability of subspace methods for this
application. Copyright (C) 1996 Pattern Recognition Society.