Title:
SELF-SUPERVISED LEARNING METHOD AND DEVICE FOR LEARNING BINARY NEURAL NET BY USING VARIABLE FLOATING POINT NEURAL NETWORK AS MOVING TARGET, AND TESTING METHOD AND DEVICE USING SAME
Document Type and Number:
WIPO Patent Application WO/2024/085337
Kind Code:
A1
Abstract:
Disclosed is a self-supervised learning (SSL) method for learning a binary neural network by using a variable floating point (FP) neural network as a moving target, comprising: (a) a step in which a learning device, when a learning vector is obtained, (i) causes the binary neural network including binary parameters in binary format to apply a binary operation to the learning vector, and (ii) causes the variable FP neural net, which includes FP parameters in an FP format learned according to the SSL methodology and of which the FP parameters are adjusted in the process of learning the binary parameters, to apply an FP operation to the learning vector; and (b) a step in which the learning device learns the FP parameters while learning the binary parameters by using a coupling loss between the binary neural net and the variable FP neural network generated with reference to (i) the learning binary output obtained from the binary neural network and (ii) the learning FP output obtained from the variable FP neural net.
Inventors:
CHOI JONG HYUN (KR)
KIM DA HYUN (KR)
KIM DA HYUN (KR)
Application Number:
PCT/KR2023/007657
Publication Date:
April 25, 2024
Filing Date:
June 02, 2023
Export Citation:
Assignee:
UIF UNIV INDUSTRY FOUNDATION YONSEI UNIV (KR)
International Classes:
G06N3/08; G06N3/04
Attorney, Agent or Firm:
SU INTELLECTUAL PROPERTY (KR)
Download PDF:
Previous Patent: METHOD FOR PREPARING SUPERABSORBENT POLYMER
Next Patent: PREVENTING SENSOR CONTAMINATION
Next Patent: PREVENTING SENSOR CONTAMINATION