Title:
DEVICE AND METHOD FOR LEARNING NEURAL NETWORK
Document Type and Number:
Japanese Patent JP3042927
Kind Code:
B2
Abstract:
PURPOSE: To execute learning of a multi-layer neural network more efficiently, and also, with higher accuracy.
CONSTITUTION: In an operation arithmetic part 2, with respect to a function &phiv (v1, x) for approximating teacher data f(x), a linear operation element O for satisfying a prescribed condition is found out, and the linear operation element O is operated on the teach data f(x). In a linear prediction replacement arithmetic part 3, based on a result of operation O.f(x) from the operation arithmetic part 2, learning of a neural network is replaced with a problem of a linear prediction and a solution of the problem of the linear prediction is derived. In a first learning result calculating part 4, based on the solution from the linear prediction replacement arithmetic part 3, a parameter v1 of the function &phiv (v1, x) is derived as a first learning result. In a second learning result calculating part 5, based on the teacher data f(x) approximated by the function &phiv (v1, x) having a first learning result as the parameter v1, a parameter c1 is derived as a second learning result.
Inventors:
Sumio Watanabe
Application Number:
JP4752592A
Publication Date:
May 22, 2000
Filing Date:
February 03, 1992
Export Citation:
Assignee:
株式会社リコー
International Classes:
G06F15/18; G06G7/60; G06N3/08; G06N99/00; (IPC1-7): G06F15/18; G06G7/60
Other References:
電子情報通信学会技術研究報告ＶＯＬ．９１，ＮＯ．３０２ ｐ１１１−１１８ １９９１
電子情報通信学会技術研究報告ＶＯＬ．９３，ＮＯ．３４１ ｐ７１−７８ １９９３
電子情報通信学会技術研究報告ＶＯＬ．９３，ＮＯ．３４１ ｐ７１−７８ １９９３
Attorney, Agent or Firm:
Masaharu Uemoto