Method for the safe operation of a machine learning device and a suitably configured apparatus

Abstract

The present invention relates to a method (1) for operating a machine learning device. In this method, the device is used for a predetermined task and learns independently. A non-self-learning monitoring device, separate from the machine learning device, monitors whether the machine learning device begins to behave undesirably due to its independent learning. If this occurs, the machine learning device is automatically switched off. The invention also relates to a corresponding device configured to carry out such a method (1).

Description

[0001] The present invention relates to a method for operating a machine learning device and a correspondingly configured device.

[0002] Machine learning offers great potential for a wide variety of applications. However, it can also pose risks, as it is typically impossible to precisely specify or verify which often highly abstract features or details a machine learning system actually learns from the input data. As a result, unexpected behavior from such a system can occur, which, depending on the application, can lead to reduced user experience or even pose a security risk. Therefore, there is a need for improvements to address this issue.

[0003] For example, DE 10 2018 206 712 B4 describes an operating method for an autonomously operable device. In this method, sensor data relating to the current environmental state of the device are fed to a control algorithm that is implemented as a self-learning algorithm and is fully trained. Based on this data, the control algorithm makes a control decision. A monitoring algorithm independent of the control algorithm determines the quality of the control decision. Depending on the determined quality, the device is operated according to the control decision, or the control decision is rejected and the device is placed in a safe operating state.

[0004] The object of the present invention is to enable a particularly helpful, flexible and at the same time secure use of machine learning.

[0005] This object is achieved by the subject matter of the independent claims. Further possible embodiments of the invention are disclosed in the subclaims, the description, and the figures. Features, advantages, and possible embodiments presented in the description for one of the subject matter of the independent claims are to be regarded at least analogously as features, advantages, and possible embodiments of the respective subject matter of the other independent claims, as well as any possible combination of the subject matter of the independent claims, optionally in conjunction with one or more of the subclaims.

[0006] The method according to the invention serves to operate, utilize, or use a machine learning device trained or pre-trained for a specific, predetermined task or application. Such a device can, for example, be, comprise, or implement an artificial neural network. However, the device can also be, comprise, or implement another type of machine learning model or models, such as a machine learning model with linear or logistic regression, an adaptable decision tree, an SVM (support vector machine), a random forest, a dimensionality reduction algorithm, a gradient boosting algorithm, and / or the like.

[0007] In one method step of the method according to the invention, the machine learning device is used, i.e., applied or operated, for the specified task. Inputs, i.e., input data, can be continuously or repeatedly fed to the device. Outputs generated by the device from this data can then be used or further processed depending on the task or application, for example, to control another device or as a basis for a decision or the like. For example, the device can be used in a motor vehicle, wherein the input data can then be or include, for example, sensor data and / or status data, which, for example, indicate a current operating state of the motor vehicle.Based on the outputs, for example, a longitudinal and / or lateral guidance of the motor vehicle and / or an operation or control of one or more other devices or systems of the motor vehicle can be carried out and / or, for example, an output to an occupant of the motor vehicle can be determined or carried out and / or the like.

[0008] According to the method according to the invention, the machine learning device continues to learn independently while it is being used for the given task. In other words, the device can actively develop, adapt, or attempt to optimize itself during operation or use by an end user. The machine learning device is therefore not provided by a manufacturer, for example, in an immutable final state, but with the option and functions for independent automatic adaptations. How the device changes during this independent automatic adaptation or learning can then depend on how it or each individual instance of the device is specifically used, for example, on which inputs are fed to it.

[0009] Furthermore, according to the method according to the invention, a non-self-learning safety or monitoring device different from the machine learning device is used to monitor whether the machine learning device, due to its independent learning, begins to behave in an undesirable manner, i.e., incorrectly or contrary to at least one predetermined goal or principle, or the like. In other words, it can be monitored or regularly checked, for example, whether the machine learning device is developing in an undesirable direction or in an undesirable manner, i.e., degenerating in this sense. The monitoring device can, for example, comprise a fixed, i.e., unchangeable, corresponding monitoring or evaluation algorithm or the like.Likewise, the monitoring device can be configured, for example, for a fixed statistical evaluation of the outputs of the machine learning device and / or the resulting or based measures, events, consequences, or the like. Likewise, the monitoring device can, for example, use a predefined reference or threshold value to recognize or detect undesirable behavior of the machine learning device, i.e., perform appropriate comparisons and / or evaluate one or more predefined criteria, and / or the like.

[0010] If the monitoring device recognizes or detects undesirable behavior of the machine learning device, i.e., a corresponding malfunction or a corresponding undesirable development of the machine learning device, then, in a further method step of the method according to the invention, the machine learning device, i.e., at least a current version or instance of a machine learning model used therein, is automatically deactivated—in whole or in part—for example, by the monitoring device or by a deactivation device coupled thereto, for example, triggered by the monitoring device or its output, or the like. Deactivating the machine learning device can, for example, mean that the machine learning device is deactivated completely or only partially.In the latter case, the device can be deactivated or switched off to such an extent that it no longer performs the undesired behavior. For this purpose, for example, a function or module or model of the device with which the supplied input data is processed in such a way that the undesired behavior results can be deactivated or switched off. At the same time, however, a basic functionality of the device can be retained, which, for example, makes it possible to control, configure, or reset the machine learning device and / or read data from the device and / or the like. The machine learning device within the meaning of the present invention can comprise software and / or hardware components.

[0011] The autonomous learning of the machine learning device provided according to the invention can enable or achieve this adaptation particularly precisely to a specific application or to a usage behavior or usage profile that changes over time. This means that the device can then be correspondingly flexible and therefore particularly helpful or useful for the respective user. This can therefore enable a particularly high degree of convenience. The automatic monitoring of behavior and, if necessary, the automatic shutdown of the machine learning device in the event of undesirable behavior, which is also provided, can at the same time ensure that negative, i.e. undesirable, consequences of the use or application of the device do not occur, or at least do not occur continuously. This can create a particularly high degree of security.

[0012] In one possible embodiment of the present invention, to detect undesirable behavior of the machine learning device, at least one predefined characteristic or key figure is automatically recorded or determined, i.e., in particular continuously or regularly repeated, and compared with a respective predefined comparison value. Depending on which characteristic, i.e., for example, which KPI (Key Performance Indicator) is used, undesirable behavior can be detected, for example, if the predefined comparison value is exceeded or undershot. The comparison value can therefore be a predefined reference value or threshold for a desired or acceptable behavior of the machine learning device.

[0013] For example, when using the machine learning device in a motor vehicle or to control a motor vehicle or a vehicle function, the parameter can be or relate to, for example, maintaining safety distances and / or a relative dimming time for oncoming or preceding vehicles and / or fuel or energy consumption of the motor vehicle and / or route selection and / or a frequency of corrective interventions by a driver during at least partially automated operation of the motor vehicle and / or a tenor of statements or behaviors of the driver or a vehicle occupant regarding automated measures carried out by the device or based on its outputs, in particular in connection with a corresponding frequency of the statements or behaviors, and / or the like.Undesirable behaviors can include, for example, choosing unsuitable or suboptimal routes, avoiding certain types of roads without reason, or accepting or provoking dangerous situations, for example, by automatically dimming the headlights too late or similar. Different comparison values ​​can be specified for different parameters or different categories of parameters. Likewise, undesirable behaviors can be classified into different predefined categories, with different comparison values ​​or an additional threshold value, above which the device is deactivated, being specified for these different categories of behavior.Such categories of undesirable behavior can include, for example, disruptive behavior, uneconomical behavior, safety-endangering behavior, life-threatening behavior, and / or other similar behaviors. For undesirable behaviors in the latter categories, stricter comparison or threshold values ​​can then be specified or used than for undesirable behaviors in the former categories.

[0014] The embodiment of the present invention proposed here can enable simple, reliable and objective detection of undesirable behavior despite the possible high complexity or the possible high degree of abstraction within the machine learning facility.

[0015] In a further possible embodiment of the present invention, backup copies of the machine learning device are regularly and automatically created. After the device is automatically switched off due to undesired behavior, the device is then automatically reset to the state given or saved in one of the previously created backup copies, in which the undesired behavior had not yet occurred, i.e. to a correspondingly earlier, secure version. In other words, it is possible to force the device to forget any learning progress that led to the undesired behavior. After this reset to the earlier, secure state, the machine learning device can then be reactivated, in particular automatically, and used for its task. The data saved automatically orSelf-learned behaviors depend on the inputs supplied to the device. In typical applications, however, it is very unlikely that exactly the same inputs that ultimately led to the undesired behavior will repeat after the device has been reset. Thus, the embodiment of the present invention proposed here enables continued use of the device and thus a correspondingly high level of user comfort, in particular without the need for a visit to a workshop, manual maintenance, or replacement of the device by specialist personnel, for example. By regularly creating backup copies of the device, any desired learning progress made by the device up to that point can be retained or further utilized.This allows a particularly advantageous combination of high security and high user comfort to be achieved, since the desired learning progress does not have to be completely relearned or rebuilt after the reset and, in addition, an excessively sudden change in the behavior of the device can be avoided if necessary.

[0016] In a possible further development of the present invention, the machine learning device is automatically reactivated after being reset to perform the specified task, but without the ability or function for independent learning. In other words, further independent learning of the device based on the status of the backup copy used for the reset is prevented. This allows the device to continue to perform its task and thus offer a correspondingly high level of user comfort. At the same time, however, a recurrence of the undesired behavior can be particularly reliably prevented. Depending on the application, for example, automatic adaptation of the device to the respective application can be achieved, but undesired overadaptation, which can be shown or manifested in the undesired behavior, can be avoided or limited.

[0017] In a further possible embodiment of the present invention, if undesirable behavior of the machine learning device is detected, its current status or state at which the undesirable behavior occurred, i.e., a corresponding current version of the device that exhibits or exhibited the undesirable behavior, is automatically saved for further investigation. In particular, this status or version of the device can be sent to a server device provided for this purpose. Depending on the application, an at least partially or partially wireless data connection, such as a cellular connection or a WLAN connection or the like, or a wired data connection, such as an internet connection or the like, can be used for this purpose.Here, for example, the corresponding program code of the machine learning device can be quarantined, in a similar way to what is possible, for example, with conventional antivirus programs with malicious code or infected files. The proposed storage or retention of at least one instance of the machine learning device exhibiting or showing the undesired behavior, i.e., at least its machine learning model, can be carried out, for example, before the reset to an earlier state mentioned elsewhere. Based on the problematic instance of the machine learning device or machine learning model stored in this way, it can then be examined at a later point in time, for example, how the undesired behavior came about in detail and / or how it can be prevented or corrected and / or the like.This can be done, in particular, without further impairing the actual machine learning device or a user or operator of the original device, i.e., in parallel with its operation. The embodiment of the present invention proposed here thus enables a correspondingly high level of user comfort and simultaneously opens up the possibility for improved future security. For example, it might be possible to adapt the machine learning device and / or other or future machine learning devices, for example by means of a software patch or the like, based on the knowledge gained from the stored problematic instance, in order to prevent them from manifesting the undesired behavior even with continued independent learning.

[0018] In a further possible embodiment of the present invention, the machine learning device has an identification feature. Based on this feature, upon detection of undesirable behavior of the device, other instances of this device are automatically searched for or located and also deactivated or, as described elsewhere, reset to an earlier state and / or prevented from further independent learning in order to prevent their corresponding development toward the observed undesirable behavior. Such other instances of the device can, for example, be present or used in other motor vehicles or systems than the machine learning device considered here.The identification feature can, for example, specify at least a basic model or a basic version of the device, based on which the device developed the undesired behavior through its independent learning. The embodiment of the present invention proposed here can thus even proactively or preventively prevent the other instances from also developing or learning the undesired behavior. This enables or implements a widespread or distributed use of corresponding machine learning devices with a correspondingly high degree of security.To locate the other instances, for example, the identification feature and, if applicable, a corresponding instruction to deactivate or stop the autonomous learning or the like can be disseminated via a broadcast system and / or, when using the devices in motor vehicles, via car-to-car data connections and / or centrally controlled by a server device or the like mentioned elsewhere. Accordingly, devices equipped with such machine learning devices, such as motor vehicles or electronic devices or systems or the like, can be equipped with a corresponding receiving device for such identification features or instructions.

[0019] To achieve even greater security, if undesired behavior of a device is detected, it can also be provided that other machine learning devices whose identifying feature cannot be determined or read, and for which it cannot be verified whether they correspond to the device that exhibited the undesired behavior, are then switched off across the board or prevented from further independent learning. This can be provided at least for machine learning devices that are used in the same type of device. This allows the spread of the undesired behavior, its occurrence, or its effects to be particularly reliably prevented or limited.This can be used, for example, to respond to cases in which the identifiability of self-learning machine learning devices is limited - be it by a user, a third party, or the device itself.

[0020] In a possible development of the present invention, the identification feature cannot be changed by the machine learning device, for example, its process, module, or program code for independent learning and thus for adapting or modifying the device. For this purpose, the identification feature can be cryptographically secured, in particular. The latter can mean, for example, that the identification feature or access to it is encrypted, protected by a password or a cartographic key, or requires authentication, or the like. This can also contribute to security, as other instances of such a machine learning device that exhibit or have exhibited the undesirable behavior can thus be found or identified particularly reliably.For example, it can be avoided or at least made more difficult to change the identification feature in an unauthorized or unintended manner in order to avoid or circumvent a planned shutdown of the device or a planned stopping of autonomous learning. Blockchain technology can be used for this purpose if necessary. For example, all circulating instances of the machine learning device and their identification feature could be stored in a blockchain or linked to a blockchain, so that a change to the identification feature of a device would have to be approved, for example, by a blockchain operator or a plurality of corresponding nodes and / or the current identification feature of each device or instance is documented in a tamper-proof manner at all times, especially even outside the respective device or instance.

[0021] In a possible further development of the present invention, the identification feature specifies both an initial base version of the machine learning device and its respective current state. The initial base version can be the version or state before the start of independent learning during operation of the device, for example by an end customer or end user or the like. This can therefore be, for example, a delivery state of the device in which it was manufactured or provided by a manufacturer. The respective current state, on the other hand, can be or indicate, for example, a current generation or version or current weights or even a development or change path from the initial base version of the device and / or the like.The current state can therefore - in contrast to the initial base version - change over time as the device continues to learn independently. Specifying the current state can be useful, for example, if modified versions of the machine learning device or corresponding changes are distributed via a network as a result of the independent learning, for example to other devices in which corresponding machine learning devices are used or to correspondingly distributed instances of the device or the like. This can therefore be used to achieve or ensure, for example, that only those other devices or instances that correspond to or are similar to the device that exhibited the undesired behavior are automatically switched off or reset, or prevented from further learning.This can, for example, affect devices or instances that have the same base version or the same or a sufficiently similar current state according to a predefined criterion or threshold. This can ensure continued security and protection against the undesired behavior while simultaneously avoiding unnecessary restrictions on the usability of users or users of other devices or instances.

[0022] In a further possible embodiment of the present invention, the machine learning device is used in a motor vehicle. Here, the device can therefore be used in particular for controlling or as part of the control of at least one vehicle function of a motor vehicle. For this purpose, the device can be designed, for example, as a vehicle control unit or as a module or subsystem of a vehicle control unit or of a vehicle network or the like. This can represent a particularly useful application of the present invention, since in the case of series or mass use of this type, there can be a large number of machine learning devices active in public spaces that could potentially lead to or contribute to safety-critical situations. At the same time, individual instances orDepending on the habits of the drivers of the motor vehicles or depending on the use or area of ​​application of the motor vehicles, facilities receive significantly different inputs and thus learn correspondingly different behaviors.

[0023] The present invention also relates to a device having a machine learning device configured for autonomous learning. Furthermore, the device according to the invention comprises a safety or monitoring device for detecting undesired behavior and, if necessary, for deactivating the machine learning device. The device according to the invention is configured for executing the method according to the invention, in particular automatically. For this purpose, the device can comprise, for example, a processing device, such as a microchip, microcontroller, microprocessor, or the like, and a computer-readable data memory coupled thereto.A corresponding operating or computer program can be stored in this data memory, which can implement the method steps, measures, or sequences mentioned in connection with the method according to the invention, or corresponding control instructions, as well as, if appropriate, the program code or software components of the mentioned devices. This operating or computer program can then be executable by the process device in order to carry out the corresponding method or to effect its execution. The device according to the invention can be, for example, a control unit, a motor vehicle, a system, an electronic device, or the like.

[0024] Further features of the invention can be derived from the following description of the figures and from the drawings. The features and combinations of features mentioned above in the description, as well as the features and combinations of features shown below in the description of the figures and / or in the figures alone, can be used not only in the respective combinations specified, but also in other combinations or on their own, without departing from the scope of the invention.

[0025] The drawing shows in the only figure an exemplary schematic flow chart for illustrating a method for using a machine learning device.

[0026] Machine learning systems, i.e., AI systems (class: artificial intelligence), can learn to perform or fulfill a specific, predefined task. In this case, impulses or specific characteristics of input data that lead to further development, i.e., learning, as well as the knowledge ultimately stored within the respective system, can be viewed from the outside or at least largely unknown to outsiders. There is therefore a latent danger that such a self-learning system could become autonomous and develop in an undesirable direction.When such a self-learning device is used or applied in a motor vehicle, this could, for example, mean or result in the device beginning to provoke accidents or dangerous situations, or generating outputs that, when used by other devices or systems in the motor vehicle, could lead to correspondingly critical situations. The following describes a method that can address this problem.

[0027] As a starting point, a machine learning device is initially provided that is trained or pre-trained for a specific, predetermined task and is set up for continuous, independent learning. This device also has an unfalsifiable identification feature. Such an identification feature can, for example, be implemented analogously to or using blockchain technology and / or cryptography technology. Such machine learning devices can be used as subsystems - implemented wholly or partially in software and / or hardware - for example in or by control units or a vehicle network of a motor vehicle. This means that such devices can be used, at least initially in the same form, in many, for example thousands or even millions of motor vehicles. The respective identification feature can then ensure that the devices orA corresponding AI module is always and everywhere uniquely identifiable. This can include both a respective base version or baseline as well as a current generation or development status. The identification feature can be secured using cryptographic methods, for example, so that forgery can be ruled out or at least made more difficult.

[0028] Fig.1 shows an exemplary schematic flowchart 1 for a corresponding method for using or operating corresponding machine learning devices. The method can be started in a method step S1. Here, for example, a device equipped with a corresponding machine learning device, for example, the respective motor vehicle, and the respective device can be put into operation. Furthermore, a security or monitoring device of the respective device assigned to the machine learning device can also be put into operation. This device can, for example, read the identification feature of the machine learning device in a method step S2.

[0029] During continuous operation of the device or the machine learning device, a method step S3 can then be used to monitor whether the device begins to behave in an undesirable manner due to its independent learning. In other words, it can be checked or monitored here whether the machine learning device begins to degenerate. For this purpose, the behavior of the machine learning device can be subject to constant or ongoing automatic observation, for example, by the monitoring device. For example, the monitoring device can continuously or regularly automatically collect one or more predefined key figures and compare them with predefined comparison or threshold values.Likewise, the behavior of the machine learning device can also be observed or monitored directly or indirectly by a respective user, for example a driver or vehicle occupant of the respective motor vehicle. The monitoring device can then detect undesirable behavior of the machine learning device, for example based on a reaction or behavior of the driver or vehicle occupant. Likewise, the respective user, in this case for example the driver or vehicle occupant, can make a corresponding user input to indicate the undesirable behavior. The monitoring device can then be configured to record or evaluate corresponding user inputs and, if necessary, to detect the undesirable behavior of the machine learning device based on this.

[0030] In one possible use case, the machine learning system could, for example, be configured to monitor driver statements and, based on these statements, execute or adjust route planning for the vehicle. If the driver then says, for example, that they could always drink coffee, the machine learning system could then plan intermediate stops for the vehicle at appropriate coffee outlets or, if necessary, even automatically order coffee. This behavior may initially correspond to the driver's request and be approved or reinforced by the driver accordingly. If this is repeated several times, a corresponding learning experience or learning progress can become established or fixed in the machine learning system.However, this can then lead the machine learning engine to attempt to further optimize or maximize its behavior with regard to this aspect. This can mean, for example, that the machine learning engine causes the vehicle to stop at every single coffee outlet or order coffee at every possible opportunity, which then, overall or in sum, no longer corresponds to the driver's wishes. However, a contrary or negative statement from the driver in this regard may not lead to a correction of the machine learning engine's behavior, because the learned behavior, i.e., a corresponding imprint, is already too firmly established or imprinted within it.The corresponding undesirable behavior can then be automatically detected, for example, based on corresponding negative or angry comments from the driver and / or on costs for automatic orders of coffee or the like that exceed a predetermined comparison value.

[0031] As long as no such undesirable behavior is detected, the monitoring of the behavior can continue, which is indicated here by a corresponding loop-shaped path.

[0032] If, however, it is detected in method step S3 that the machine learning device is beginning to display undesired behavior, i.e. to degenerate, the method can be continued in method step S4. In method step S4, one or more predetermined measures can be carried out in order to put the machine learning device into a safe state. This can in particular be a state in which a continued or repeated occurrence of the undesired behavior no longer occurs, i.e. is avoided. For example, the machine learning device can be switched off or shut down in whole or in part. The current status or version of the machine learning device can then be saved separately, for example in quarantine or a secured container or the like, as a backup copy. In addition, the device orTheir learning progress can be reset to the last known state in which the undesired behavior did not occur. Likewise, further independent learning of the machine learning device can be deactivated, at least until further notice. If necessary, the machine learning device can then be reactivated in the reset state, in particular with the independent learning function deactivated.

[0033] In addition, an error message or report can be automatically output in a method step S5. This can be used, for example, to inform the user and / or a manufacturer and / or a regulatory authority and / or the like about the undesirable behavior of the machine learning device that has occurred. Likewise, if the machine learning device failed to enter the safe state in method step S4, a corresponding error message can be automatically output in method step S5.

[0034] Furthermore, in a method step S6, for example based on the read-out identification feature, other instances of the machine learning device can be located and also deactivated or reset to a previous, secure state and / or prevented from further independent learning. If necessary, machine learning devices or their changes or learning progress can automatically spread via corresponding data networks, which can fundamentally be a desired behavior in the sense of swarm intelligence. However, this can also be prevented or deactivated here upon prior detection of the undesired behavior. For example, the identification feature of the affected devices or components in the entire vehicle fleet can be provided with a corresponding blocking or deactivation flag or command or the like.

[0035] If a device or instance is successfully put into a safe state, a corresponding status message can be output in a respective method step S7, for example. These status messages can be managed or collected, for example, in the network of devices or apparatuses and / or by a central server device or the like. Based on this, it can then be checked, for example, whether all of the corresponding devices or instances have been successfully put into a safe state. For example, the detection and putting of the corresponding devices or instances into a safe state can be carried out in accordance with method step S6, i.e. effectively ora corresponding search or security call should be continued or remain active until, and in particular only until, it is recognized from the status messages that all facilities or instances have been successfully put into a secure state.

[0036] In a method step S8, or for example already in parallel to the previous method steps, the backup copy of the machine learning device can then be examined or made available for examination in the state in which it exhibited the undesired behavior. For this purpose, this backup copy can be sent, for example, to a corresponding server device or a laboratory or the like. Based on this, the causes of the undesired behavior and / or possible countermeasures or corrections can then be determined. For this purpose, for example, an appropriately secured device, for example a motor vehicle without a communications module or without a connection to a data network or the like, can be equipped and operated with the backup copy, i.e. the machine learning device that exhibits or exhibits the undesired behavior.Depending on the type of undesirable behavior, the device can be operated in a secure environment, such as a dedicated test site or the like, to avoid dangerous situations as much as possible. Likewise, this device can be equipped with special measurement technology or the like to enable the behavior of the machine learning device to be observed or investigated in more detail.

[0037] If necessary, a countermeasure, such as a software fragment, a learning pattern, a set of input data, or the like, can also be developed here, which can be used to adapt other machine learning instances or devices to prevent them from continuing the undesired behavior or from learning or developing it in the future. A corresponding program, software patch, or the like can then be distributed to the other devices or instances, for example, also via a corresponding network or as part of a respective maintenance measure or the like.

[0038] Overall, the examples described show how measures for detecting, shutting down, and storing machine learning facilities, especially distributed AI systems, can be used to control them. List of reference symbols 1 Schedule S1 - S9 process steps QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2018 206 712 B4

[0003]

Claims

[1] Method for operating a machine learning facility in which automatically - the machine learning facility is used for a given task, - the machine learning facility continues to learn independently, - by means of a non-self-learning monitoring device different from the machine learning device, it is monitored whether the machine learning device starts to behave in an undesirable manner as a result of its independent learning, and if this is the case, - the machine learning facility is switched off. [2] Method according to claim 1, characterized by that in order to detect the undesirable behavior of the machine learning device, at least one predetermined characteristic value is automatically determined and compared with a respective predetermined comparison value. [3] Method according to one of the preceding claims, characterized by that backup copies of the machine learning facility are automatically created on a regular basis and that, after the machine learning facility is switched off due to undesired behavior, it is reset to the state saved in one of the backup copies at which the undesired behavior had not yet occurred. [4] Method according to claim 3, characterized by that after resetting the machine learning setup, it will automatically be reactivated to perform the given task, but without the self-learning function. [5] Method according to one of the preceding claims, characterized bythat if undesirable behavior of the machine learning facility is detected, its current state at which the undesirable behavior occurred is automatically saved for further investigation, in particular sent to a server facility set up for this purpose. [6] Method according to one of the preceding claims, characterized by that the machine learning facility has an identification feature and, based on this, upon detection of undesirable behavior, other instances of this machine learning facility are automatically searched for and also switched off and / or prevented from further independent learning. [7] Method according to claim 6, characterized by that the identification feature cannot be changed by the machine learning facility, in particular that it is cryptographically secured. [8] Method according to claim 6 or 7, characterized bythat the identifier specifies both an initial base version of the machine learning facility and its current state. [9] Method according to one of the preceding claims, characterized by that the machine learning facility is used in a motor vehicle. [10] Device comprising a machine learning device configured for autonomous learning and a monitoring device for detecting undesired behavior and, if necessary, for switching off the machine learning device, wherein the device is configured to carry out the method according to one of the preceding claims.

Images