Title:
IMPROVEMENT OF COMPUTATION GRAPH
Document Type and Number:
WIPO Patent Application WO/2023/210665
Kind Code:
A1
Abstract:
According to the present invention, an information processing method is executed by one or a plurality of processors included in an information processing device, the method comprising: performing training by inputting prescribed data to a prescribed learning model using a neural network that is represented by a prescribed computation graph; changing the prescribed data and/or the prescribed computation graph, wherein the changing includes changing a function, which is associated with a prescribed node of a prescribed layer in the prescribed computation graph, between the function and an activation function to which an output value of the function is input, and acquiring a trained result of the training using the changed prescribed data and/or the changed prescribed computation graph; performing supervised learning using training data including arbitrary data and an arbitrary computation graph by which training has been performed and a trained result when the training is performed with the arbitrary data and the arbitrary computation graph; and generating, as a prediction model generated by the supervised learning, a prediction model that is input with prescribed data and outputs a specific computation graph.
Inventors:
KUBOTA NOZOMU (JP)
Application Number:
PCT/JP2023/016361
Publication Date:
November 02, 2023
Filing Date:
April 25, 2023
Export Citation:
Assignee:
KUBOTA NOZOMU (JP)
International Classes:
G06N3/0985
Foreign References:
| JP7199121B1 | 2023-01-05 | |||
| US20200074296A1 | 2020-03-05 |
Attorney, Agent or Firm:
INABA, Yoshiyuki et al. (JP)
Download PDF: