Parameterization of a machine learning system for a control system

1. A method for computer-assisted parameterization of a machine learning system configured to ascertain as an output signal a class from a plurality of classes based on an input signal, the method comprising the following steps:

training the machine learning system once with correctly labeled training data, and once with not correctly labeled training data; and

selecting hyperparameters of the machine learning system in such a way that the trained machine learning system reproduces actual classifications of the correctly labeled training data better than actual classifications of the not correctly labeled training data,

wherein the hyperparameters are selected in such a way that for the trained machine learning system, a statistical frequency distribution of margins that result when the machine learning system is supplied with the not correctly labeled training data has two maxima as a function of the margins.

2. The method as recited in claim 1 , wherein the machine learning system is a neural network.

3. The method as recited in claim 1 , wherein the actual classifications of the not correctly labeled training data are a permutation of the actual classifications of the correctly labeled training data.

4. The method as recited in claim 3 , wherein the not correctly labeled training data are ascertained from the correctly labeled training data by a random permutation of the actual classifications.

5. The method as recited in claim 1 , wherein the hyperparameters are selected in such a way that a portion of the actual classifications of the correctly labeled training data that are correctly reproduced by the trained machine learning system is greater than a portion of the actual classifications of the not correctly labeled training data that are correctly reproduced by the trained machine learning system.

6. The method as recited in claim 5 , wherein the hyperparameters are selected in such a way that the portion of the actual classifications of the correctly labeled training data that are correctly reproduced by the trained machine learning system is greater than a first predefinable threshold value, and/or the portion of the actual classifications of the not correctly labeled training data that are correctly reproduced by the trained machine learning system is less than a second predefinable threshold value.

7. The method as recited in claim 1 , wherein the hyperparameters are selected in such a way that for the trained machine learning system, a statistical frequency distribution of margins that result when the machine learning system is supplied with unlabeled training data has only one maximum as a function of the margins, the maximum being assumed for a value of the margins that is greater than zero.

8. A training system for computer-assisted parameterization of a machine learning system configured to ascertain as an output signal a class from a plurality of classes based on an input signal, the training system configured to:

train the machine learning system once with correctly labeled training data, and once with not correctly labeled training data; and

select hyperparameters of the machine learning system in such a way that the trained machine learning system reproduces actual classifications of the correctly labeled training data better than actual classifications of the not correctly labeled training data,

wherein the hyperparameters are selected in such a way that for the trained machine learning system, a statistical frequency distribution of margins that result when the machine learning system is supplied with the not correctly labeled training data has two maxima as a function of the margins.

9. A method for operating a control system including a machine learning system that has been trained to ascertain as an output signal a class from a plurality of classes based on an input signal, by training the machine learning system once with correctly labeled training data, and once with not correctly labeled training data, and selecting hyperparameters of the machine learning system in such a way that the trained machine learning system reproduces actual classifications of the correctly labeled training data better than actual classifications of the not correctly labeled training data, wherein the hyperparameters are selected in such a way that for the trained machine learning system, a statistical frequency distribution of margins that result when the machine learning system is supplied with the not correctly labeled training data has two maxima as a function of the margins, the method comprising:

ascertaining a control signal as a function of an output signal that is ascertained using the machine learning system.

10. The method as recited in claim 9 , wherein an at least semi-autonomous robot and/or a manufacturing system and/or a personal assistant and/or an access system and/or a monitoring system or a medical imaging system, is controlled as a function of the ascertained control signal.

11. A control system, comprising:

one or multiple processors; and

at least one machine-readable memory medium on which instructions are stored, which when executed on the processors prompt the control system to ascertain a control signal as a function of an output signal that is ascertained using an machine learning system, the machine learning system having been trained to ascertain as an output signal a class from a plurality of classes based on an input signal, by training the machine learning system once with correctly labeled training data, and once with not correctly labeled training data, and selecting hyperparameters of the machine learning system in such a way that the trained machine learning system reproduces actual classifications of the correctly labeled training data better than actual classifications of the not correctly labeled training data, wherein the hyperparameters are selected in such a way that for the trained machine learning system, a statistical frequency distribution of margins that result when the machine learning system is supplied with the not correctly labeled training data has two maxima as a function of the margins.

12. A method for operating a control system, the control system including a machine learning system configured to ascertain as an output signal a class from a plurality of classes based on an input signal, the method comprising the following steps:

training the machine learning system once with correctly labeled training data, and once with not correctly labeled training data;

selecting hyperparameters of the machine learning system in such a way that the trained machine learning system reproduces actual classifications of the correctly labeled training data better than actual classifications of the not correctly labeled training data, wherein the hyperparameters are selected in such a way that for the trained machine learning system, a statistical frequency distribution of margins that result when the machine learning system is supplied with the not correctly labeled training data has two maxima as a function of the margins; and

ascertaining a control signal as a function of an output signal that is ascertained using the machine learning system.

13. A machine learning system whose hyperparameters are selected in such a way that the machine learning system is trained to reproduce actual classifications of correctly labeled training data better than actual classifications of not correctly labeled training data, wherein the hyperparameters are selected in such a way that for the trained machine learning system, a statistical frequency distribution of margins that result when the machine learning system is supplied with the not correctly labeled training data has two maxima as a function of the margins.

14. A non-transitory machine-readable memory medium on which is stored a computer program for computer-assisted parameterization of a machine learning system configured to ascertain as an output signal a class from a plurality of classes based on an input signal, the computer program, when executed by one or more processors, causing the one or more processors to perform:

training the machine learning system once with correctly labeled training data, and once with not correctly labeled training data; and

selecting hyperparameters of the machine learning system in such a way that the trained machine learning system reproduces actual classifications of the correctly labeled training data better than actual classifications of the not correctly labeled training data,

wherein the hyperparameters are selected in such a way that for the trained machine learning system, a statistical frequency distribution of margins that result when the machine learning system is supplied with the not correctly labeled training data has two maxima as a function of the margins.