正規直交化法による多層ニューラルネットワークの学習アルゴリズム

抄録

Recently, the author proposed a new learning algorithm derived algebraically which is composed of the error back propagation to give a fictitious teacher to each hidden layer and the update rule of each connection. The details of the algorithm with some comments are demonstrated in this paper using the Exclusive-OR problem as an example. Some update rules of each connection are also proposed here. Among them, the update rule using an orthonormalization method with some scheme is most effective. Simulation results show thst 5 steps convergence can be obtained in many cases of different initial values. The exceptional cases which do not show the convergence are also made convergent by some improvement in the nonlinear characteristic of the unit.