US 12,236,672 B2
Processing of learning data sets including noisy labels for classifiers
William Harris Beluch, Stuttgart (DE); Jan Mathias Koehler, Stuttgart (DE); and Maximilian Autenrieth, Eislingen/Fils (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Apr. 16, 2021, as Appl. No. 17/233,410.
Claims priority of application No. 102020205542.4 (DE), filed on Apr. 29, 2020.
Prior Publication US 2021/0342650 A1, Nov. 4, 2021
Int. Cl. G06N 3/08 (2023.01); G06F 18/21 (2023.01); G06F 18/214 (2023.01); G06F 18/241 (2023.01); G06V 10/774 (2022.01); G06V 10/82 (2022.01)
CPC G06V 10/774 (2022.01) [G06F 18/2148 (2023.01); G06F 18/2193 (2023.01); G06F 18/241 (2023.01); G06N 3/08 (2013.01); G06V 10/82 (2022.01)] 14 Claims
OG exemplary drawing
 
1. A method for processing of learning data sets for a classifier, the learning data sets each encompassing measured data, which were obtained by a physical measuring process, or by a partial or complete simulation of a physical measuring process, or by a partial or complete simulation of a technical system observable using a physical measuring process, as learning input variable values, the learning data sets additionally each encompassing learning output variable values to which the classifier is to map the learning input variable values, the method comprising the following steps:
processing the learning input variable values of at least one of the learning data sets multiple times in a non-congruent manner by the classifier trained up to an epoch E2 so that the learning input variable values are mapped to different output variable values;
ascertaining a measure for an uncertainty of the output variable values from deviations of the output variable values;
in response to the uncertainty meeting a predefined criterion, ascertaining at least one updated learning output variable value for a learning data set of the at least one of the learning data sets from one or multiple further output variable value(s) to which the classifier maps the learning input variable values after a reset to an earlier training level with epoch E1<E2 such that an accuracy of the ascertained at least one updated learning output variable value is greater than the accuracy ascertained based on the training level with epoch E2;
wherein the learning input variable values are mapped by multiple modifications of the classifier, wherein the multiple modifications differ from one another to such an extent that they are not congruently merged into one another with advancing training of the classifier, to different output variable values, and
wherein a separate uncertainty is ascertained for each output variable value ascertained by each of the modifications, only an established number of the modifications which generated the output variable values having lowest uncertainties contributing to the ascertainment of the at least one updated learning output variable value.