METHOD FOR TRAINING A NEURAL NETWORK, CLASSIFICATION METHOD AND CORRESPONDING COMPUTER PROGRAM

The method improves neural network training by using distance-transformed matrices to modulate performance indicators, addressing annotation errors and enhancing accuracy in classification tasks.

FR3170670A1Pending Publication Date: 2026-06-26THALES SA
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
FR · FR
Patent Type
Applications
Current Assignee / Owner
THALES SA
Filing Date
2024-12-23
Publication Date
2026-06-26

Smart Images

  • Figure 00000000_0000_ABST
    Figure 00000000_0000_ABST
Patent Text Reader

Abstract

METHOD FOR TRAINING A NEURAL NETWORK, CLASSIFICATION METHOD AND CORRESPONDING COMPUTER PROGRAM S A method for training (10) a neural network (15) includes the initialization (110) of a neural network (15_in); the provision (120) of matrix pairs (51) comprising an input matrix (51A) and a target label matrix (51B) each term of which is a class number (num_i) of a term of the input matrix for the training (130) of training sets (65) and intermediate (66) and final (67) validation sets; and the training of the neural network with said sets.For a selected class (Ci,sel), for each target label matrix and a respective estimated label matrix (51C_Ent, 51C_VI, 51C_VF) from the current neural network in response to the respective input matrix, a distance transformed matrix is ​​determined to apply a modulation function (f_mod) in order to estimate (144, 154) a performance indicator of the current neural network. Figure for the abstract: 3.
Need to check novelty before this filing date? Find Prior Art