Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
PROGRAM, INFORMATION PROCESSING DEVICE, INFORMATION PROCESSING METHOD, AND TRAINED DNN MODEL
Document Type and Number:
WIPO Patent Application WO/2024/042714
Kind Code:
A1
Abstract:
The present invention: calculates a first importance level with respect to a DNN model of weights for a case in which in-distribution data is used as the input for the DNN model, by using a loss function for the in-distribution data and the weight thereof, with respect to one of a plurality of weights indicating the strength of the connection between a pair of neurons constituting the DNN model; uses out-of-distribution data and the loss function to calculate a second importance level with respect to the DNN model of weights for a case in which the out-of-distribution data is used as the input for the DNN model; and deletes a weight from the DNN model if a new importance level calculated by combining the first importance level and the second importance level indicates a relationship less than a threshold value, thereby compressing the DNN model by deleting a weight that is unnecessary for both the in-distribution data and the out-of-distribution data in the DNN model.

Inventors:
KODA SATORU (JP)
Application Number:
PCT/JP2022/032243
Publication Date:
February 29, 2024
Filing Date:
August 26, 2022
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
FUJITSU LTD (JP)
International Classes:
G06N3/082
Foreign References:
US20180260695A12018-09-13
Other References:
HENDRYCKS DAN, LIU XIAOYUAN, WALLACE ERIC, DZIEDZIC ADAM, KRISHNAN RISHABH, SONG DAWN: "Pretrained Transformers Improve Out-of-Distribution Robustness", PROCEEDINGS OF THE 58TH ANNUAL MEETING OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, STROUDSBURG, PA, USA, 1 January 2020 (2020-01-01) - 10 July 2020 (2020-07-10), Stroudsburg, PA, USA, pages 2744 - 2751, XP093142629, DOI: 10.18653/v1/2020.acl-main.244
KODA SATORU, IKUYA MORIKAWA : "Out-of-distribution Vulnerability and Detection", SCIS 2022 2022 SYMPOSIUM ON CRYPTOGRAPHY AND INFORMATION SECURITY OSAKA, JAPAN & ONLINE, 18 August 2022 (2022-08-18) - 21 January 2022 (2022-01-21), XP093142630
Attorney, Agent or Firm:
SANADA, Tamotsu et al. (JP)
Download PDF: