Title:
MACHINE LEARNING METHOD AND MACHINE LEARNING DEVICE
Document Type and Number:
Japanese Patent JP2023117922
Kind Code:
A
Abstract:
To provide a machine learning method and a machine learning device in a neural network including a physical system that do not require a large operation cost and accurate information of the physical system for a training.SOLUTION: The present invention relates to a machine learning method by a neural network operation using an operational element of a physical system. The machine learning method employs a training method which uses an operation by DFA, and differentiation information f'(x) of a nonlinear function f(x) is replaced by an arbitrary nonlinear mapping g(x) in an update rule of an operation by a DFA operation, and a value of a linear combination is updated by using the arbitrary nonlinear mapping g(x) and a linear random mapping B(l)e(L) of an error signal of a final layer. There is also provided a machine learning device for equipping the machine learning method.SELECTED DRAWING: Figure 2
Inventors:
NAKAJIMA MITSUMASA
HASHIMOTO TOSHIKAZU
INOUE KATSUMA
NAKAJIMA KOHEI
KUNIYOSHI YASUO
HASHIMOTO TOSHIKAZU
INOUE KATSUMA
NAKAJIMA KOHEI
KUNIYOSHI YASUO
Application Number:
JP2022020744A
Publication Date:
August 24, 2023
Filing Date:
February 14, 2022
Export Citation:
Assignee:
NIPPON TELEGRAPH & TELEPHONE
UNIV TOKYO
UNIV TOKYO
International Classes:
G06N3/08; G06G7/60; G06N3/06
Attorney, Agent or Firm:
Patent Attorney Tani / Abe Patent Office