誤差逆電搬学習アルゴリズムの初期値設定法
Year: 1993
Type: article
Abstract: This paper proposes a new initialization method of the error back propagation learning algorithm for multilayer neural networks the initial biases of the output units are set in terms of target averages easily obtained a priorifrom a set of examples, and the initial weights of the output units are set zero The learning by the proposed method starts from an initially disconnected network, and its process coincides with a neurophysiological hypothesis on the formation of the synapse connections in the biological nervous systems. neuroblasts outgrowtheir axons to form new synapse conections with their pioneer neurons. From simulation results for the symmetry and the parity problems, it was confirmed that the proposed initialization is useful to improve the learning convergency
Institution: Kindai University
Language: Japanese
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222853913
