Method and apparatus for operating a control system

Abstract

A method for parameterizing a machine learning system (60), in particular a neural network, the machine learning system being set up to determine the assigned class (y) in a plurality of classes respectively based on input data (x), wherein the machine learning system (60) is trained once with correctly labeled training data and once with incorrectly labeled training data, wherein the hyperparameters (θ H ) are selected such that the corresponding trained machine learning system (60) can reproduce the actual classification (y T ) of the correctly labeled training data better than the actual classification (y T ) of the incorrectly labeled training data.

Description

Technical Field

[0001] The invention relates to a method for parameterizing a machine learning system, a training system, a control system, a method for operating a control system, a computer program and a machine-readable storage medium. Background Art

[0002] A method for training a machine learning system is known from DE 20 2017 102 238 U, wherein parameters of the machine learning system are trained using a training set having observations and associated desired first output values ​​for the machine learning system. Summary of the invention

[0003] Advantages of the present invention

[0004] In the following, X represents the input space and Y represents the output space (target space). Given a training data record S = ((x 1 , y 1 ), ..., (x m , y m )). In this case, the data is taken according to a fixed and unknown distribution P about (X,Y). We also give a loss function For a function f: X → Y, by R[f]: = to define expected loss. Empirical risk is defined by This parameter is also referred to as the characteristic parameter below. .

[0005] The task of the machine learning method is to learn a function f: X→ Y from a class of functions F that minimizes the expected loss R[f]. Since this is usually not possible, the task is to learn a function f from a class of functions F that minimizes the expected risk based on the training data records S. S : Replaced by the task of X → Y.

[0006] If the function f S is obtained in this way, then the central question is f S For new data points How general is the degree of To characterize.

[0007] Machine learning systems, such as neural networks, can be described using a number of parameters. These parameters can be divided into: architectural parameters, such as depth, number and form of filters, choice of nonlinearity and connections; and optimization parameters, such as step size, batch size, and number of iterations. In other words, architectural parameters characterize the function f, while optimization parameters characterize the optimization method used during training.

[0008] The selection of these parameters can be based, for example, on empirical knowledge or the performance capabilities of the neural network on validation data records. Since neural networks are usually parameterized by many more parameters than are given as data points, there is a risk that the network learns the training data with great effort. A network trained in this way may not be well suited for use in safety-critical functions, such as functions for automated driving, since the output for new data points may not be well understood.

[0009] The method with the features of independent claim 1 prevents the machine learning system from having to learn the training data in a complex manner during training, i.e., architecture parameters and / or optimization parameters can be automatically determined by parameter search. The architecture parameters and / or optimization parameters determined in this way result in that the machine learning system can be more generally applicable. Summary of the invention

[0011] In a first aspect, the invention relates to a method having the features of independent claim 1. Further aspects of the invention are the subject matter of the independent, parallel claims. Advantageous developments are the subject matter of the dependent claims.

[0012] Consider machine learning systems, especially artificial neural networks. In order to solve classification tasks, the input space The input signal x is assigned to a class y from a number k of further classes. This assignment is done, for example, by means of the function to carry out, among which is a Euclidean space. The components of f(x) each correspond to one of these classes and in this case each characterize the probability that the class y to which it belongs is the correct classification of the input signal x. In order to assign the input signal x to a specific class, the argmax function can be used. The argmax function outputs the coordinates of the maximum value of f, that is, .

[0013] The training data includes multiple training points (x T , y T ), these training points are exemplary input data x T and the desired category y T If the classification N(x T ) is correct, then f(x T) is the maximum value among the values. That is, Applicable to all .

[0014] Margin m If the margin is positive, the classification is correct; if the margin is negative, the classification is wrong.

[0015] The present invention is based on the observation that the histogram of the margins of the training data follows different distributions, depending on whether the machine learning system has correctly learned the input data x T With classification y T The basic structure of the relationship between them. Fig.11 Such a distribution is shown as an example. The margin m is plotted on the abscissa and the frequency h is plotted on the ordinate.

[0016] If we use the training data (x T , y T ) to train the machine learning system, then when the input data x T and classification y T The distribution of Fig.11 A or Fig.11 The distribution shown in B, and when the input data x T and classification y T The distribution of can be correctly learned when Fig.11 The distribution shown in C. Which of these distributions is obtained depends on the hyperparameter (hereinafter θ H to indicate).

[0017] Hyperparameter θ H can be characterized in that these hyperparameters remain constant while training the machine learning system. H For example, the architectural parameters θ may be included A and / or optimize the parameter θ O .

[0018] Architecture parameters θ A Characterizes the structure of the machine learning system. If the machine learning system is a neural network, then the architecture parameter θ A For example, it includes the depth of the neural network (that is to say the number of layers) and / or the number of filters and / or parameters characterizing the filter form and / or parameters characterizing the nonlinearity of the neural network and / or parameters characterizing which layers of the neural network are connected to which other layers of the neural network.

[0019] Optimization parameter θ Oare parameters that characterize the optimization algorithm for adapting the parameters θ, which are adapted during training. For example, these parameters may include the numerical step size and / or the size of the data batches (English: “batches”) and / or the minimum or maximum number of iterations.

[0020] For the hyperparameter θ H For different values ​​of , the learning success rate of the training of the machine learning system is different. In order to design the generalization of the machine learning system as best as possible, it has been recognized that: the hyperparameter θ H Select so that: Correctly labeled data records can be learned well, that is, Fig.11 The distribution of the margin m is similar to that shown in C, but for the classes whose labels are incorrectly and especially randomly marked Relative to the correctly labeled data record X C Data records that have changed, especially those that have been replaced in sequence For example, get Fig.11 A or 11B are similar to the distribution of the margin m. Preferably, for example, including regularization parameters, especially l 1 -Hyperparameter θ of the regularization parameter λ H is chosen to be so large that Fig.11 The distribution of the margin m shown in B has a second maximum as a function of the margin (that is, the regularization parameter is chosen so large that the distribution has the second maximum, but does not have the second maximum if the regularization parameter is chosen smaller). BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Subsequently, the embodiments of the present invention are further described with reference to the accompanying drawings. In the drawings:

[0022] Figure 1 Schematically shows the construction of an embodiment of the present invention;

[0023] Figure 2 Schematically illustrates an embodiment for controlling an at least partially autonomous robot;

[0024] Figure 3 An embodiment for controlling a production system is schematically shown;

[0025] Figure 4 An embodiment for controlling a personal assistant is schematically shown;

[0026] Figure 5 An embodiment of a system for controlling access is schematically shown;

[0027] Figure 6An embodiment for controlling a monitoring system is schematically shown;

[0028] Figure 7 An embodiment for controlling a medical imaging system is schematically shown;

[0029] Figure 8 The training system is schematically shown;

[0030] Fig. 9 A possible flow of a method for determining optimal hyperparameters is shown in a flowchart;

[0031] Fig.10 A possible process for running a machine learning system is shown in a flowchart;

[0032] Fig.11 The statistical distribution of the margins is shown exemplarily. DETAILED DESCRIPTION

[0033] Figure 1 An actuator 10 is shown interacting with a control system 40 in its surroundings 20. The actuator 10 and the surroundings 20 are also collectively referred to as an actuator system. The state of the actuator system is detected at preferably uniform time intervals by means of a sensor 30, which can also be provided by a plurality of sensors. The sensor signal S of the sensor 30 or each sensor signal S in the case of a plurality of sensors is transmitted to the control system 40. The control system 40 thus receives a sequence of sensor signals S. The control system 40 determines therefrom a control signal A, which is transmitted to the actuator 10.

[0034] The control system 40 receives a sequence of sensor signals S of the sensor 30 in an optional receiving unit 50, which converts the sequence of sensor signals S into a sequence of input signals x (alternatively, the sensor signals S can also be used directly as input signals x). The input signal x can be, for example, a segment of the sensor signal S or a further processing of the sensor signal S. The input signal x can, for example, include image data or an image, or include individual frames of a video recording. In other words, the input signal x is determined from the sensor signal S. The input signal x is fed to a machine learning system 60, which is, for example, a neural network.

[0035] The machine learning system 60 is preferably parameterized by parameters θ, which are stored in a parameter memory P and provided by the latter.

[0036] The machine learning system 60 determines the output signal y from the input signal x. The output signal y is supplied to an optional transformation unit 80 which determines therefrom an actuation signal A which is supplied to the actuator 10 in order to actuate the actuator 10 accordingly.

[0037] The actuator 10 receives the control signal A, is controlled accordingly and performs a corresponding action. In this case, the actuator 10 may include a control logic (not necessarily structurally integrated) which determines a second control signal from the control signal A, which is then used to control the actuator 10 .

[0038] In other embodiments, the control system 40 includes the sensor 30. In still other embodiments, the control system 40 also includes the actuator 10, alternatively or additionally.

[0039] In other preferred embodiments, the control system 40 includes one or more processors 45 and at least one machine-readable storage medium 46 on which commands are stored, which, when executed on the processor 45, cause the control system 40 to implement the method according to the present invention.

[0040] In an alternative embodiment, a display unit 10 a is provided as an alternative to the actuator 10 or in addition to the actuator 10 .

[0041] Figure 2 An exemplary embodiment is shown in which a control system 40 is used to control an at least partially autonomous robot, here an at least partially autonomous motor vehicle 100 .

[0042] Sensor 30 may be, for example, one or more video sensors and / or one or more radar sensors and / or one or more ultrasound sensors and / or one or more LiDAR (laser radar) sensors and / or one or more position sensors (e.g. GPS), which are preferably arranged in motor vehicle 100. Alternatively or in addition, sensor 30 may also include an information system that determines information about the state of the actuator system, such as a weather information system that determines the current or future state of the weather in surroundings 20.

[0043] The machine learning system 60 may detect objects, for example, in the environment of the at least partially autonomous robot based on the input data x. The output signal y may be information characterizing the location of the object in the environment of the at least partially autonomous robot. The output signal A may then be determined based on and / or corresponding to the information.

[0044] The actuator 10 preferably arranged in the motor vehicle 100 may be, for example, a brake device, a drive device or a steering device of the motor vehicle 100. Subsequently, the control signal A may be determined such that the actuator or actuators 10 are controlled such that, in particular when certain classes of objects are involved, such as pedestrians, the motor vehicle 100, for example, avoids a collision with an object identified by the machine learning system 60. In other words, the control signal A may be determined according to and / or corresponding to the determined class.

[0045] Alternatively, the at least partially autonomous robot may also be another mobile robot (not shown), for example, a robot that moves by flying, floating, diving or walking. The mobile robot may also be, for example, an at least partially autonomous lawn mower or an at least partially autonomous cleaning robot. In these cases, the control signal A may also be determined so that the drive device and / or the steering device of the mobile robot are controlled so that the at least partially autonomous robot, for example, avoids a collision with an object identified by the machine learning system 60.

[0046] In another alternative, the at least partially autonomous robot can also be a garden robot (not shown) which uses an imaging sensor 30 and a machine learning system 60 to determine the type or state of plants in the surroundings 20. The actuator 10 can then be, for example, a feeder of chemicals. The control signal A can be determined as a function of the determined type of plant or the determined state of the plant so that an amount of chemical corresponding to the determined type or the determined state is applied.

[0047] In other alternatives, the at least partially autonomous robot can also be a household appliance (not depicted), in particular a washing machine, a stove, an oven, a microwave oven or a dishwasher. The state of an object processed by a household appliance can be detected by a sensor 30, such as an optical sensor, for example, the state of the laundry in the washing machine can be detected in the case of a washing machine. Then, the type or state of the object can be determined using a machine learning system 60 and the type or state of the object can be characterized by an output signal y. Then, the control signal A can be determined to make the household appliance control according to the determined type or determined state of the object. For example, in the case of a washing machine, the washing machine can be controlled according to what material the laundry is composed of. Then, the control signal A can be selected according to which material the laundry is determined.

[0048] Figure 3 The following exemplary embodiment is shown in which a control system 40 is used to control a production machine 11 of a production system 200 by controlling an actuator 10 which controls the production machine 11. The production machine 11 can be, for example, a machine for punching, sawing, drilling and / or cutting.

[0049] The sensor 30 can then be, for example, an optical sensor which detects, for example, properties of the finished product 12. It is possible that the actuator 10 controlling the production machine 11 is actuated as a function of the determined properties of the finished product 12 so that the production machine 11 accordingly carries out the subsequent processing steps of the finished product 12. It is also possible that the sensor 30 determines the properties of the finished product 12 processed by the production machine 11 and, as a function of this, adapts the actuation of the production machine 11 for the next finished product.

[0050] Figure 4 The following exemplary embodiment is shown in which the control system 40 is used to control the personal assistant 250. Preferably, the sensor 30 is a sound sensor which receives a speech signal of the user 249. Alternatively or additionally, the sensor 30 can also be set up to receive an optical signal, for example a video image of the gesture of the user 249.

[0051] Based on the signal of the sensor 30, the control system 40 determines the control signal A of the personal assistant 250, for example, by means of a machine learning system performing gesture recognition. The determined control signal A is then transmitted to the personal assistant 250 and the personal assistant is controlled accordingly. The determined control signal A can in particular be selected so that the control signal corresponds to the desired control envisioned by the user 249. The envisioned desired control can be determined based on the gesture recognized by the machine learning system 60. The control system 40 can then select the control signal A for transmission to the personal assistant 250 based on the envisioned desired control and / or select the control signal A for transmission to the personal assistant 250 corresponding to the envisioned desired control.

[0052] The corresponding control may include, for example, personal assistant 250 calling up information from a database and reproducing this information in a form that is acceptable to user 249 .

[0053] Instead of personal assistant 250 , a domestic appliance (not shown), in particular a washing machine, a stove, an oven, a microwave or a dishwasher, can also be provided in order to be controlled accordingly.

[0054] Figure 5The following embodiment is shown, in which a control system 40 is used to control an access system 300. The access system 300 may, for example, include a physical access control device, such as a door 401. The sensor 30 may, for example, be an optical sensor (e.g. for detecting images or video data), which is configured to detect faces. With the aid of a machine learning system 60, the detected image may be interpreted. For example, the identity of a person may be determined. The actuator 10 may be a lock, which, depending on a control signal A, opens the access control device or does not open the access control device, for example, opens the door 401 or does not open the door 401. For this purpose, the control signal A may be selected depending on the interpretation of the machine learning system 60, for example, depending on the determined identity of the person. Instead of a physical access control device, a logical access control device may also be provided.

[0055] Figure 6 The following embodiment is shown, in which the control system 40 is used to control the monitoring system 400. Figure 5 The embodiment shown in FIG. 1 differs from this embodiment in that a display unit 10 a is provided instead of the actuator 10, which is controlled by the control system 40. For example, the machine learning system 60 can determine whether an object recorded by the optical sensor is suspicious, and the control signal A can then be selected so that the object is highlighted in color by the display unit 10 a.

[0056] Figure 7 The following embodiment is shown, in which the control system 40 is used to control a medical imaging system 500, such as an MRT apparatus, an X-ray apparatus, or an ultrasound apparatus. The sensor 30 may be provided, for example, by an imaging sensor, and the display unit 10a is controlled by the control system 40. For example, the machine learning system 60 may determine whether an area recorded by the imaging sensor is abnormal, and the control signal A may then be selected so that the area is highlighted by the display unit 10a with a color.

[0057] Figure 8 An embodiment of a training system 140 for training a machine learning system 60 is schematically shown. A training data unit 150 determines suitable input signals x, which are fed to the machine learning system 60. For example, the training data unit 150 accesses a computer-implemented database in which training data records are stored and, for example, randomly selects an input signal x from the training data records. Optionally, the training data unit 150 also determines a desired or “actual” output signal y assigned to the input signal x. T , these output signals are supplied to an evaluation unit 180 .

[0058] The artificial neural network x is designed to determine associated output signals y as a function of the input signals x supplied to it. These output signals y are supplied to an evaluation unit 180 .

[0059] The training system 140 includes a second parameter memory Q in which hyperparameters θ are stored. H .

[0060] The modification unit 160 uses, for example, Fig. 9 The method illustrated in FIG. 1 determines new parameters θ′ and supplies these new parameters to the parameter memory P, where they replace the parameters θ. The modification unit 160 uses, for example, Fig. 9 The method illustrated in the figure determines the new hyperparameter θ' H And these new hyperparameters are transmitted to the second parameter memory Q.

[0061] The evaluation unit 180 can, for example, use a method that depends on the output signal y and the desired output signal y. T The cost function (loss function) is used to determine the feature parameters , which characterizes the performance capability of the machine learning system 60. The parameter θ can be calculated based on the characteristic parameter to be optimized.

[0062] In other preferred embodiments, the training system 140 includes one or more processors 145 and at least one machine-readable storage medium 146 on which commands are stored that, when executed on the processor 145, cause the control system 140 to implement the method according to the present invention.

[0063] Fig. 9 The flowchart shows a method for determining the optimal hyperparameter θ according to an embodiment. H The process of the method.

[0064] First (1000), the hyperparameter θ H is initialized, for example, randomly or to a fixedly predeterminable value. Then, the training data unit 150 provides the correctly labeled training data X C = (x T ,y T ) records. The parameters θ are set to pre-given initial values, for example, these parameters can be set to randomly selected values.

[0065] Next (1100), a real random number generator or a pseudo random number generator is used to generate a random number for the input signal x T and the corresponding output signal yT , that is, a random permutation of the distribution of categories. According to this random permutation, by T Permutation to determine randomization classification , and produce incorrectly labeled data records .

[0066] Alternatively, in step ( 1100 ), the randomized classification can also be generated by means of a real random number generator or a pseudo-random number generator by randomly selecting from a set of possible classes. , and this produces incorrectly labeled data records .

[0067] Next (1200), record X based on the correctly labeled training data C To determine the input signal x=x T , these input signals are fed to a machine learning system 60 and output signal y is determined therefrom. To this end, for each input signal x, a function k-dimensional parameter f(x) is determined and output signal y is determined as the component of f(x) having the maximum value.

[0068] Next ( 1300 ), the output signal y and the actual, ie desired, output signal y assigned to the input signal x are provided in the evaluation unit 180 . T .

[0069] Next (1400), according to the determined output signal y and the desired output signal y T To determine the characteristic parameters Next, new parameters θ' are determined by means of an optimization method, such as the gradient descent method, which have an influence on the characteristic parameters The optimization is performed by executing steps (1200) and (1300) with new parameters θ', possibly multiple times, until the best new parameters θ' are determined. These best new parameters are then stored in a first parameter memory P.

[0070] Finally (1500), the margin m and an indicator characterizing the statistical distribution of the margin m, for example the proportion of those margins m that are positive, are determined with the aid of the determined k-dimensional parameter f(x) obtained for the optimal new parameter θ'. The parameters θ are reset to predeterminable initial values, for example, these parameters can be set to randomly selected values.

[0071] Now, steps (1200) to (1500) are correspondingly performed for the incorrectly labeled data record X r to be repeated.

[0072] To do this, first (1600) record X based on the incorrectly labeled training datar To determine the input signal x=x T , these input signals are fed to a machine learning system 60 and output signal y is determined therefrom. To this end, for each input signal x, a function k-dimensional parameter f(x) is determined and output signal y is determined as the component of f(x) having the maximum value.

[0073] Next (1700), the output signal y and the actual, ie desired, output signal associated with the input signal x are provided in the evaluation unit 180. .

[0074] Next (1800), according to the determined output signal y and the desired output signal To determine the characteristic parameters Next, new parameters θ' are determined by means of an optimization method, such as the gradient descent method, which have an influence on the characteristic parameters The optimization is performed by carrying out steps ( 1600 ) and ( 1700 ) with new parameters θ′, optionally multiple times, in each case, until the optimum new parameters θ′ are determined.

[0075] Finally (1900), with the aid of the determined training data X that has been incorrectly labeled r The k-dimensional variable f(x) obtained by taking the optimal new parameter θ′ as an example, determines the second margin m′ and an indicator characterizing the statistical distribution of the second margin m′, for example, a second proportion of the second margin m′ that is positive.

[0076] Now (2000) it is checked whether the determined share is greater than a first threshold and whether the determined second share is less than a second threshold. If this is not the case, the hyperparameter θ H is changed, for example randomly, to a predeterminable discrete grid, and branches back to step (1000) and the method uses the changed hyperparameter θ H is re-executed (in this case at step (1000) with the hyperparameter θ H is not reinitialized). If the hyperparameter θ H All possible values ​​of have been checked, the method terminates with an error report.

[0077] If the determined share is greater than the first threshold and the determined second share is less than the second threshold, the hyperparameter θ is set to H The current value of is stored in the second parameter memory Q. The method ends here.

[0078] Instead of adjusting the hyperparameter θ H Gradually changing and evaluating the hyperparameter θ HThe share and the second share are determined respectively by all possible values ​​of on a pre-given discrete network; and from the hyperparameter θ H Of all possible values ​​of , those values ​​are selected for which the condition that the determined share is greater than a first threshold value and the determined second share is less than a second threshold value is fulfilled best, for example Pareto-optimally.

[0079] Fig.10 The flow of the method for operating the machine learning system 60 is shown. The machine learning system 60 receives (2000) input data x and determines associated output data y by means of a function f parameterized by the parameter θ. A control signal A can be determined from the output data y (2100). The method ends here.

[0080] It should be understood that these methods can be implemented not only in software as described, but also in hardware, or in a hybrid form of software and hardware.

Claims

1. A method for computer-aided parameterization of a machine learning system (60) for a control system (40), the machine learning system being configured to determine, as an output signal y, a respective class from a plurality of classes from an input signal x, wherein the input signal is an image recorded by at least one sensor and to determine, from the output signal, an actuation signal A for actuating an actuator, The machine learning system (60) uses correctly labeled training data X C is trained once and uses incorrectly labeled training data X r is trained once, wherein the hyperparameter θ of the machine learning system (60) H is selected so that the corresponding trained machine learning system (60) can make the correctly labeled training data X C The actual classification y T Than the incorrectly labeled training data X r The actual classification y T To reproduce better, Where the hyperparameter θ H is selected so that for a trained machine learning system (60), when the incorrectly labeled training data X is fed to the machine learning system (60) r The statistical frequency distribution of the margin m obtained as a function of the margin m has two maxima, The margin m is obtained by To define, where f(x T ) is the input signal x T The probability of correct classification and i and j are f(x T ), wherein the training data includes a plurality of training points (x T ,y T ), the training point is the input data x T and the category y T of pairs, and if the classification N(x T )=argmax f(x T ) is correct, then f(x T ) is f(x T )'s value.

2. The method of claim 1, wherein the machine learning system (60) is a neural network.

3. The method according to claim 1, wherein the incorrectly labeled training data X r The actual classification y T is the correctly labeled training data X C The actual classification y T replacement.

4. The method according to claim 3, wherein the incorrectly labeled training data X r By using the correctly labeled training data X C The actual classification y in T is determined by random permutation.

5. The method according to any one of the preceding claims, wherein the hyperparameter θ H Correct training data X selected so that it is correctly reproduced by the trained machine learning system (60) C The actual classification y T The fraction of incorrectly labeled training data X that is correctly reproduced by the trained machine learning system (60) is greater than r The actual classification y T share.

6. The method according to claim 5, wherein the hyperparameter θ H The correctly labeled training data X is selected so as to be correctly reproduced by the trained machine learning system (60) C The actual classification y T The proportion of the incorrectly labeled training data X is greater than a predetermined first threshold; and / or the incorrectly labeled training data X is correctly reproduced by the trained machine learning system (60) r The actual classification y T The proportion is less than a predetermined second threshold value.

7. The method according to any one of claims 1 to 4, wherein the hyperparameter θ H is selected so that for a trained machine learning system (60), the statistical frequency distribution of the margin m obtained when unlabeled training data is supplied to the machine learning system (60) has only one maximum value as a function of the margin m, wherein the maximum value is taken when the value of the margin m is greater than zero.

8. A training system (140) configured to carry out the method according to any one of claims 1 to 7.

9. A method for operating a control system (40) comprising a machine learning system (60), wherein the machine learning system is trained by means of a method according to any one of claims 1 to 7, and wherein a control signal A is then determined based on an output signal y determined by means of the machine learning system (60).

10. A method according to claim 9, wherein at least a partially autonomous robot (100) and / or a production system (200) and / or a personal assistant (250) and / or an access system (300) and / or a monitoring system (400) or a medical imaging system (500) are controlled according to the determined control signal A.

11. A control system (40), comprising one or more processors (45) and at least one machine-readable storage medium (46), wherein commands are stored on the machine-readable storage medium, and when the commands are executed on the processor (45), the commands cause the control system (40) to implement the method according to any one of claims 9 or 10.

12. A method for operating a control system (40) according to claim 11, wherein a machine learning system (60) included in the control system (40) is first trained with the aid of a method according to any one of claims 1 to 7 and then a method according to claim 9 or 10 is implemented using the machine learning system (60) trained in this way.

13. A machine learning system (60), wherein a hyperparameter θ of the machine learning system is H is selected so that the machine learning system (60) can be trained to correctly label the training data X C The actual classification y T Than incorrectly labeled training data X r The actual classification y T Better reproduction, wherein the machine learning system (60) is trained with the aid of a method according to any one of claims 1 to 7.

14. A computer program product comprising a computer program, which is configured to implement the method according to any one of claims 1 to 7 or 9 or 10 or 12.

15. A machine-readable storage medium (46, 146) on which a computer program is stored, the computer program being designed to implement the method according to any one of claims 1 to 7 or 9 or 10 or 12.

Images