Linear discriminant analysis (LDA) is a basic tool of pattern recognition,
and it is used in extensive fields, e.g. face identification. However, LDA
is poor at adaptability since it is a batch type algorithm. To overcome thi
s, new algorithms of online LDA are proposed in the present paper. In face
identification task, it is experimentally shown that the new algorithms are
about two times faster than the previously proposed algorithm in terms of
the number of required examples, while the previous algorithm attains bette
r final performance than the new algorithms after sufficient steps of learn
ing. The meaning of new algorithms are also discussed theoretically, and th
ey are suggested to be corresponding to combination of PCA and Mahalanobis
distance.