Partial Oblique Projection Learning for Optimal Generalization

In practice, it is necessary to implement an incremental and active learning for a learning method. In terms of such implementation, this paper shows that the previously discussed S-L projection learning is inappropriate to constructing a family of projection learning, and proposes a new version called partial oblique projection (POP) learning. In POP learning, a function space is decomposed into two complementary subspaces, so that functions belonging to one of the subspaces can be completely estimated in noiseless case; while in noisy case, the dispersions are set to be the smallest. In addition, a general form of POP learning is presented and the results of a simulation are given.