US 12,230,010 B2
Training of classifiers and/or regressors on uncertain training data
Kanil Patel, Stuttgart (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Mar. 25, 2022, as Appl. No. 17/656,453.
Claims priority of application No. 21171428 (EP), filed on Apr. 30, 2021.
Prior Publication US 2022/0366672 A1, Nov. 17, 2022
Int. Cl. G06V 10/764 (2022.01); G06N 20/00 (2019.01)
CPC G06V 10/764 (2022.01) [G06N 20/00 (2019.01)] 13 Claims
OG exemplary drawing
 
1. A method for training a classifier or a regressor that is configured to map an input sample of measurement data to: (i) a set of classification scores which are each a score with respect to a different respective class of available classes, or (ii) a set of regression scores which are each a score with respect to a different regression value of available regression values, the method comprising the following steps:
providing a set of training samples, each of the training samples being labelled with: (i) ground truth classification scores which are each a score with respect to a different respective class of the available classes, or (ii) ground truth regression scores which are each a score with respect to a different regression value of the available regression values;
for each training sample from at least a subset of the training samples:
determining a confidence score that quantifies: (i) an uncertainty of the training sample, or (ii) an ease or difficulty of classifying the training sample,
reducing, by an amount that is dependent on the confidence score: (i) a largest ground truth classification score of the ground truth classification scores with which the training sample is labeled, or (ii) a largest ground truth regression score of the ground truth regression scores with which the training sample is labeled, and
distributing the amount to: (i) others of the ground the truth classification scores of the ground truth classification scores with which the training sample is labeled, or (ii) others of the regression scores with which the training sample is labeled,
so as to obtain updated ground truth classification scores for the training sample, or updated ground truth regression scores for the training sample;
mapping, by the classifier or the regressor, each of the training samples to classification scores or regression scores;
rating, using a predetermined loss function, a deviation of the classification scores to which each training sample is mapped or the regression scores to which each training same is mapped, from the updated ground truth classification scores of the training sample or the updated ground truth regression scores of the training sample; and
optimizing parameters that characterize a behavior of the classifier of the regressor with an objective that, when further training samples are supplied to the classifier or the regressor, the rating by the loss function is likely to improve.