力学系を用いた階層型NNの学習の高速化に関する一検討

抄録

This paper proposes a new method, which uses dynamics of dynamical system, for improving learning efficiency in multilayered neural networks. The shape of the error function is analyzed using Monte-Carlo simulation, and the learning efficiency is improved on the basis of the shape analysis and the adaptive learning rate.
The proposed method uses the adaptive learning rate, which is determined from the gradient of the error function. The proposed method is applied to XOR and 5-parity check problems and the effectiveness and feasibility of the proposed method are verified.