WO2022013124A1

WO2022013124A1 - Method and device for providing a database for assessing the robustness of at least one ai-based information processing system

Info

Publication number: WO2022013124A1
Application number: PCT/EP2021/069268
Authority: WO
Inventors: Fabian HÜGER; Peter Schlicht; Nico Schmidt; Florens GRESSNER; Wiebke Günther; Felix ASSION
Original assignee: Volkswagen Aktiengesellschaft; neurocat GmbH
Priority date: 2020-07-13
Filing date: 2021-07-12
Publication date: 2022-01-20
Also published as: DE102020208736A1; EP4179470A1

Abstract

The invention concerns a method for providing a database for assessing the robustness of at least one AI-based information processing system (10), wherein at least one AI-based information processing system (10), at least one dataset (11), at least one data augmentation definition (12) and at least one difference measure definition (13) are received as input parameters (9), wherein a multidimensional data structure (20) is produced, and wherein each data point (21) of the multidimensional data structure (20) comprises a difference value (22) determined by means of the at least one defined difference measure (13a), said difference value being determined by forming the at least one defined difference measure (13a) between output data that have been produced by the at least one AI-based information processing system (10) for data (11a) and for augmented data, and wherein the multidimensional data structure (20) produced is provided, as a result of which a robustness of the at least one AI-based information processing system (10) can be assessed on the basis of the difference values (22) that the multidimensional data structure (20) comprises. The invention also relates to a device (1).

Description

description

Method and device for providing a database for the robustness assessment of at least one AI-based information processing system

The invention relates to a method and a device for providing a database for evaluating the robustness of at least one AI-based information processing system.

Machine learning, for example based on neural networks, has great potential for use in modern driver assistance systems and automated vehicles. Functions based on deep neural networks process sensor data (e.g. from cameras, radar or lidar sensors) in order to derive relevant information from it. This information includes, for example, a type and a position of objects in the area surrounding the motor vehicle, a behavior of the objects, or a roadway geometry or topology.

An essential feature in the development of AI-based information processing systems (the training) lies in the purely data-driven parameter fitting without expert intervention. For example, in the case of deep neural networks, a deviation of an output (for a given parameterization) of the neural network from a ground truth is determined (the so-called loess). The loss function used here is selected in such a way that the parameters of the neural network depend on it in a differentiable manner. As part of the gradient descent method, the parameters of the neural network are adjusted in each training step depending on the derivation of the deviation (determined on several examples). These training steps are repeated very often until the loess no longer decreases.

In this procedure, the parameters of an AI-based information processing system, in particular a neural network, are determined without an expert assessment or semantically motivated modeling. This can have significant consequences for the properties of the AI-based information processing system, in particular the neural network. In particular, deep neural networks are largely non-transparent for humans and their calculations are difficult to interpret. This poses a massive limitation to systematic testing or formal verification.

Furthermore, deep neural networks in particular are susceptible to harmful interference, so-called adversarial perturbations: small manipulations of the input data that are hardly perceptible to humans or do not change a semantic content can lead to completely different output data. Such disruptive influences can be both intentional changes to the data ("neural hacking") and random image changes (sensor noise, weather influences, certain colors or contrasts).

Furthermore, it is particularly unclear to which input characteristics a neural network sensitizes. This ensures that data generated synthetically, for example by simulation, has so far hardly been used successfully for training neural networks: neural networks trained in simulation or on otherwise synthetic data have an astonishingly poor performance on real sensor data. Executing neural networks in a different data domain (training in summer, execution in winter, etc.) sometimes drastically reduces the functional quality. One of the consequences of this is that the possibility of developing and releasing neural networks in simulation (elimination of expensive labeling and complex real tests), which sounds very attractive from the cost point of view, does not seem realistic.

The second point in particular is of great importance for possible limitations of potent neural networks in the area of functional safety. In order to measure the latter, it is essential to measure an assessment of the robustness of the network design against minor changes (augmentations) of the input data. Since such changes can be manifold (sensor noise, weather influences, image manipulations, semantically meaningless content changes, e.g. the wall color of background buildings), there is no clear and accepted measure of robustness. Rather, many robustness values against disturbances (i.e. augmentations) of various types and intensities must be measured. Furthermore, the robustness of neural networks is not an absolute value, but rather dependent on current input data.

Therefore, a sustainable and reproducible measurability of the robustness of AI-based information processing systems, in particular of neural networks, is desirable. In particular, it is desirable to have a flexible and modular measurement of the influence of data disturbances (augmentations) on the prediction quality of AI-based data in the most efficient way possible in terms of computing, storage and data

Information processing systems, in particular of neural networks to provide.

The invention is therefore based on the object of providing a method and a device for providing a database for the robustness assessment of at least one AI-based information processing system.

The object is achieved according to the invention by a method having the features of patent claim 1 and a device having the features of patent claim 10 . Advantageous configurations of the invention result from the dependent claims.

In particular, a method for providing a database for the robustness assessment of at least one AI-based information processing system is made available, the AI-based information processing system having a function for automated driving of a motor vehicle and/or for driver assistance of the motor vehicle and/or for environment detection and/or or environmental perception or for another application, with at least one Kl-based information processing system, at least one data set, at least one data augmentation definition and at least one difference measure definition being received as input parameters, with a multidimensional data structure being generated based on the input parameters, with the dimensions and value ranges of Dimensions of the multidimensional data structure are defined by the received input parameters, and each data point of the multidimensional data structure has a middle ls of the at least one defined difference measure, which is determined by forming the at least one defined difference measure between output data, which is formed by the at least one KI-based information processing system for data of the at least one data set and for the same by means of the at least one defined data augmentation augmented data were generated, and wherein the generated multidimensional data structure is provided, so that a robustness of the at least one AI-based information processing system can be assessed based on the difference values comprised by the multidimensional data structure.

Furthermore, in particular a device for providing a database for the robustness assessment of at least one AI-based information processing system created, wherein the AI-based information processing system provides a function for the automated driving of a motor vehicle and/or for driver assistance of the motor vehicle and/or for an environment detection and/or environment perception or for another application, comprising a data processing device, the data processing device being set up for this purpose is to receive at least one Kl-based information processing system, at least one data set, at least one data augmentation definition and at least one differential dimension definition as input parameters, and to generate a multidimensional data structure based on the input parameters, with the dimensions and value ranges of the dimensions of the multidimensional data structure being determined by the received input parameters are defined, and wherein each data point of the multidimensional data structure has a difference determined by means of the at least one defined difference measure zwert, which is determined by forming the at least one defined difference measure between output data that was generated by inference from the AI-based information processing system for data in the at least one data set and for the same data augmented by means of the at least one defined data augmentation, and the provide generated multidimensional data structure, so that a robustness of the at least one AI-based information processing system can be assessed based on the difference values comprised by the multidimensional data structure.

The method and the device make it possible to provide a database for the robustness assessment of at least one AI-based information processing system. For this purpose, a multidimensional data structure is generated in which differential values are stored as data points. The dimensions of the multidimensional data structure, which in particular can also be referred to as a hypercube, include at least the following dimensions: AI-based information processing system, data set, data augmentation and difference measure. Therefore, at least one AI-based information processing system, at least one data set, at least one data augmentation definition and at least one difference measure definition are received for generating the multidimensional data structure. For each combination of these dimensions, i.e. for each possible characteristic within the dimensions, a difference value is calculated using the respective difference measure and stored for the associated data point. The difference value defined via the respective difference measure definition is determined from the output data that is generated by the at least one CI-based information processing system for (non-augmented) data of the data set and for augmented data. The augmented data is defined using the for the data point via the data augmentation definition data augmentation generated. In other words, for each data point within the multidimensional data structure, there is a combination of at least the dimensions Kl-based information processing system, data set, data augmentation definition and difference measure definition, i.e. for each data point there is a Kl-based information processing system, a data set, a data augmentation (or a data augmentation method ) and a difference measure is defined. Based on this, data of the data set are augmented for the data point by means of data augmentation (e.g. disturbed) and a difference value between the non-augmented data and the augmented data of the data set is determined using the difference measure by applying the AI-based information processing system to the data. The determined difference value is stored in the data point. This procedure is carried out for all data points until a difference value has been determined and stored for each of the data points within the multidimensional data structure. The difference values can later be retrieved from the data structure at any time in a targeted manner, i.e. by specifying the combination of at least the Kl-based information processing system (or a link to it), the data set, the data augmentation definition and the difference measure definition, so that a robustness of the at least one Kl -based information processing system can be assessed based on the difference values comprised by the multidimensional data structure. This makes it possible to determine robustness at any time, ie also afterwards, and to do this, for example, only consider a subset of the multidimensional data structure. As a result, the determination and evaluation of the robustness of the at least one AI-based information processing system can be improved, in particular with regard to flexibility.

One advantage of the method and the device is that the robustness of a AI-based information processing system can be measured and evaluated even without the AI-based information processing system, i.e. without a model description (structure,

parameters, activation functions, etc.), and can be done without at least one data set. This is particularly advantageous if the AI-based information processing system and/or possibly sensitive data itself is not to be published or made available. For example, the method allows a KI-based information processing system to be certified without the KI-based information processing system itself having to be part of the certification method. For this purpose, only the multidimensional data structure has to be provided. In particular, the method and the device allow meaningful robustness metrics to be flexibly generated based on the multidimensional data structure. In particular, the method and the device can provide a metric generator for assessing a robustness of the AI-based information processing system.

Provision is made for the at least one AI-based information processing system to provide a function for automated driving of a motor vehicle and/or for driver assistance of the motor vehicle and/or for detecting and/or perceiving the surroundings. In particular, it is provided that a Kl-based

Information processing system, for example a trained neural network, is loaded into the memory of at least one control unit and is executed or used there in order to evaluate and process sensor data recorded in particular by means of a sensor system and, for example, to generate and provide control signals, for example for an actuator system. In addition, the generated at least one multidimensional data structure can also be stored or have been stored in the memory of the at least one control unit. As a result, robustness can be checked and/or verified at any time.

In principle, alternatively or additionally, however, it can also be provided that the method and the device are used in other applications. This can be, for example, automated fleet control, interior surveillance, driver monitoring, production control, video surveillance, robotic applications, automated flying, automated rail vehicles or space travel applications. In these applications, too, provision is made for a AI-based information processing system to be loaded into the memory of at least one control device, with the at least one control device being used to execute the AI-based information processing system, for example for evaluating recorded sensor data and for generating and providing control signals, for example for a production process and/or for at least one actuator. In addition, the generated at least one multidimensional data structure can also be stored or stored in the memory of the control device.

An AI-based information processing system is in particular an information processing system that is based on a method of artificial intelligence (AI). For example, the AI-based information processing system can be designed as a deep neural network. In principle, however, the method described in this disclosure can also be used in other AI-based information processing systems be used, for example in rule-based information processing systems. Receiving the at least one AI-based information processing system should mean in particular that a link or a reference is received that uniquely identifies the at least one AI-based information processing system (eg a name, a version number, etc.). Alternatively or additionally, receiving the AI-based information processing system can also include receiving a structural description and/or parameters of the at least one AI-based information processing system, so that the at least one AI-based information processing system is then executed directly or used as part of the method can.

The at least one AI-based information processing system is in particular trained and/or finally parameterized. For example, the AI-based information processing system can be a trained neural network. The at least one AI-based information processing system is then executed on the data and the augmented data, in particular by means of the data processing device, and delivers results in each case. In principle, it is also possible to run the at least one AI-based information processing system on another device, for example a control device or an external data processing device, with the other device then providing the results for the data and the augmented data.

A trained neural network includes, in particular, a structural description and parameters (e.g. filter parameters, weightings, activation functions, etc.) of the neural network. When receiving, in particular the structure description and the parameters are received. The trained neural network is then executed on the data and the augmented data using the data processing device.

A data set comprises in particular data, in particular (acquired) sensor data. In particular, the data can be one-dimensional or multi-dimensional, in particular two-dimensional. For example, the data can be images from a camera or a lidar sensor. In principle, however, any sensor data can be used.

A data augmentation definition defines in particular a data augmentation or a data augmentation method. The data augmentation definition specifies how data in the dataset is to be modified. A large number of changes can be provided here. Examples include: adding noise, adding one or more adversarial disorders, changing a contrast, changing a brightness, Changing colors, changing a weather condition (e.g. adding snow or rain to a camera image captured in summer). A data augmentation or a data augmentation method is configured or defined in particular as a function of physical sensor properties (faults, etc.) and/or possible physical and/or technical faults in the sensor system and/or possible adversarial faults.

A difference measure definition defines in particular a difference measure. The difference measure specifies in particular how output data from the at least one KI-based information processing system, which was generated for (non-augmented) data in the dataset, with output data from the at least one KI-based

Information processing system that were generated for augmented data to be compared. If the AI-based information processing system outputs a vector as output data, for example, then a difference measure can include the comparison of the vectors, for example by determining a difference between the vectors. A simple example of a further difference measure is the following: If the AI-based information processing system outputs, for example, as output data how many pedestrians are present in a captured camera image, the number output for the data and the augmented data can be compared with one another (e.g 3 pedestrians versus 5 pedestrians, so the difference value is equal to 2 pedestrians).

If chronologically sequential data is stored in a data set, a difference measure can also refer to chronologically sequential, ie temporally adjacent, data. In this way, a database for robustness assessments when processing video sequences (or other temporally sequential data) can be generated and made available by an AI-based information processing system. For example, as part of the robustness assessment, it can be checked whether or not a pedestrian in a video sequence is reliably recognized as a pedestrian by the AI-based information processing system over a number of individual video frames.

Before the multidimensional data structure is provided, all artefacts are also stored in the multidimensional data structure as metadata and/or headers. These artifacts include, for example: references to software code used, references to the at least one Kl-based information processing system (e.g. to a trained neural network) and hyperparameters used for training, references to one or more data sets used (possibly incl descriptive data) and/or initial values used for random generators used (“Random Seeds”). In particular, the artifacts include neither the at least one data set used nor the AI-based information processing system. In particular, therefore, no data record and no AI-based information processing system are stored in the multidimensional data structure.

Parts of the data processing device can be designed individually or combined as a combination of hardware and software, for example as program code that is executed on a microcontroller or microprocessor.

Provision can be made for the robustness of the at least one AI-based information processing system to be evaluated on the basis of the multidimensional data structure provided. For this purpose, the difference values of the individual data points are evaluated and a robustness is determined. For example, the difference values can be compared with at least one robustness requirement. Such a robustness requirement can be, for example, a threshold value that is predetermined as a function of the respective augmentation method and that, for example, must not be exceeded by (averaged) difference values.

For example, based on the difference values stored in the multidimensional data structure and specifying at least one robustness metric, a report for assessing the robustness of the AI-based information processing system, for example a trained neural network, can be created and output. In particular, a metric generator for assessing the robustness of the AI-based information processing system, for example the trained neural network, can be provided using the method and the device.

In a further development, it can be provided that the at least one AI-based information processing system is discarded based on the evaluation, the method is repeated with changed input parameters, or the AI-based one is repeated

information processing system is certified as robust. Furthermore, it can also be provided that the output data of an AI-based information processing system certified as robust, for example a trained neural network, is likewise certified and/or marked as robust. Provision can be made here for the AI-based information processing system, for example the trained neural network, to be loaded into a memory of at least one control unit after certification. There, the at least one AI-based information processing system, for example the trained neural network, can then be used to process sensor data, for example when detecting the surroundings in a motor vehicle.

In one embodiment it is provided that the at least one AI-based information processing system is a neural network and/or comprises at least one neural network. The neural network can in particular be a deep neural network, for example a convolutional network. A neural network includes in particular a structural description and parameters (e.g. filter parameters, weightings, activation functions, etc.) of the neural network. The neural network is in particular a trained neural network.

In one embodiment it is provided that the provision includes providing an interface for the targeted retrieval of data points of the multidimensional data structure. In particular, such an interface enables individual data points or areas or sets of data points to be retrieved or queried as a function of combinations of at least one AI-based information processing system, for example a trained neural network, a data set, a data augmentation definition and a difference measure definition. In this way, robustness (in particular robustness values) can be determined flexibly and, for example, averaged over a number of data points or a number of data augmentations or data augmentation methods. The interface can also be a remote interface, for example provided by a central server on which multidimensional data structures are stored and from which data points can be retrieved via the remote interface.

Provision can also be made here for the interface to be set up so that a query can be specifically restricted by further parameters. Possible queries with specification of the corresponding parameters can be, for example, the following:

- Queries of concrete entries (differential values) of the multidimensional data structure,

- Queries of maximum, minimum or average values over sections of the multidimensional data structure, - Querying of variances, bandwidths etc. of the difference values in sections of the multidimensional data structure,

- Query weighted sums (integrals) over sections of the multidimensional data structure,

- Querying histograms about the difference values and/or the aforementioned values,

- Queries about functional curves of the difference values and/or the aforementioned values with variation of section parameters,

- Queries from heat maps or other data visualization methods about the difference values and/or the aforementioned values or curves,

- Querying of the time curves of the difference values and/or the aforementioned values in the case of a time-sequential data set.

In one embodiment it is provided that the provision includes transmitting the multidimensional data structure to a certification service provider and/or a user of the AI-based information processing system and/or loading the multidimensional data structure into a memory of at least one control unit. As a result, robustness can also be determined afterwards, i.e. after delivery and use or during the entire service life of the at least one AI-based information processing system in the field, for example by means of new or changed robustness measures, without the Difference values must be determined. Provision can be made here for an associated multidimensional data structure to be stored in the memory in order to operate the AI-based information processing system. For example, it can be provided that before operating the AI-based information processing system, the control unit checks whether an associated multidimensional data structure is stored in the memory or not, with the AI-based information processing system only being able to be operated or used if such a multidimensional data structure is stored. In addition, the robustness of the AI-based information processing system can also be checked using the control unit, in particular as a function of the difference values stored in the multidimensional data structure.

It is also possible for multidimensional data structures to be stored in a central archive, for example in a memory on a central server, in order to be able to check and/or check the robustness of AI-based information processing systems used in the field at any time. In one embodiment, it is provided that a set of sub-parameters for the at least one data augmentation definition is additionally received as an input parameter, with the multidimensional data structure being generated and/or the data being augmented taking into account the received set of sub-parameters. In this way, the data augmentation or the data augmentation method or methods can be further specified or parameterized. In the case of a brightness variation in camera images, a range of brightness values can be specified, for example, within which the camera images are to be augmented, ie their brightness should be varied. For each of the sub-parameters to be taken into account, the multidimensional data structure expands accordingly. In the case of a brightness variation with three sub-parameters (e.g. -20%, 0 and +20%), there are three corresponding values for the brightness variation in the data augmentation definition dimension. Another example is specifying various noise parameters (e.g. target value for a signal-to-noise ratio, etc.).

In one embodiment, it is provided that a set of filter criteria for individual input parameters is additionally received as an input parameter, with the multidimensional data structure being generated taking into account the received set of filter criteria. In this way, a multidimensional data structure can be created in a targeted manner that takes the filter criteria into account. In this way, specific test scenarios in particular can be mapped or prepared. For example, filter criteria can be tags on the data that allow binning and/or resolving a certain robustness on the criteria. For example, these can be context values (e.g. for the contexts city, country, corner case, holiday, big event nearby, weather,...) or data properties (e.g. pollution, motion blur, blinding... ).

One embodiment provides that a selection of statistical distribution functions for parameter distributions for the at least one data augmentation definition is additionally received as an input parameter, the multidimensional data structure being generated taking into account the selection of statistical distribution functions for parameter distributions. This allows statistical distributions to be taken into account when augmenting (eg disruption) the data in the dataset. For example, for a data augmentation that includes a brightness variation in images, a distribution function can be transferred from which statistical sampling is performed when generating the multidimensional data structure (e.g. 10 samples from the following distribution: [-30% to -10%] with p=0.3, [-10% to +10%] with p=0.5 and [+10% to +30%] with p=0.2). In one embodiment it is provided that a selection of distributions for combinations of parameters of data augmentation definitions and data of the data set is additionally received as an input parameter, the multidimensional data structure being generated taking into account the received set of distributions for the combinations of the parameters. This allows distributions for the combinations to be taken into account. A distribution can, for example, be an 'exposure' to the individual input data streams or tags, which allows aggregating "against" the distribution (ie it is summed up over (the individual data)x(exposure)). In this way, a "realistically expected robustness risk" can be calculated.

One embodiment provides that a selection for a relevance of individual data points is received as an input parameter, with the multidimensional data structure being generated taking into account the received relevance. In this way, particularly relevant data points, that is to say particularly relevant combinations of the input parameters, can be identified so that data points and difference values are generated or determined for these combinations. This is particularly advantageous if a robustness check for specific combinations of the input parameters is required by law, for example, or has proven itself in the robustness assessment.

In one embodiment, it is provided that the results obtained from the at least one AI-based information processing system are additionally stored in the multidimensional data structure for each data point. As a result, the results generated for the data and the augmented data can still be used later. In particular, when the multidimensional data structure is expanded to include further data augmentations, results for the non-augmented data do not have to be generated again, instead the results already stored in the multidimensional data structure can be accessed directly. In this way, computing power and computing time can be saved.

One embodiment provides for the multidimensional data structure to be expanded after it has been provided by inserting at least one additional dimension and/or by expanding a value range of at least one dimension, with the expanded multidimensional data structure being provided. As a result, additional or new data augmentation methods, additional or new data sets and/or additional or new difference measures can also be taken into account subsequently. In particular, they can already existing data points can be reused and do not have to be recalculated, since an extension of the multidimensional data structure is easily possible.

Further features for the configuration of the device result from the description of configurations of the method. The advantages of the device are in each case the same as in the embodiments of the method.

Furthermore, a control unit is also created, comprising a memory, a multidimensional data structure being stored in the memory, which is suitable as a database for a robustness assessment for at least one AI-based

enable information processing system. The multidimensional data structure was created in particular by means of the method described in this disclosure or by means of the device described in this disclosure.

A use of the control unit described above for evaluating and certifying at least one AI-based information processing system is then proposed. As a result, the multidimensional data structure can be provided in a manner that is technically easy to implement. It can be provided here that the control unit also provides the AI-based information processing system. If both are provided by the control unit, the robustness of the AI-based information processing system can be checked and/or verified at any time, in particular after delivery of the control unit to a certifier and/or an end customer.

Furthermore, a computer program is also provided, comprising instructions which, when the computer program is executed by a computer, cause it to carry out the method steps of the method according to any of the described embodiments.

In addition, a data carrier signal is also created that transmits such a computer program.

The invention is explained in more detail below on the basis of preferred exemplary embodiments with reference to the figures. Here show: Fig. 1 is a schematic representation of an embodiment of the device for

Providing a database for robustness assessment of at least one AI-based information processing system;

Fig. 2 is a schematic flowchart to illustrate an embodiment of the

Method for providing a database for evaluating the robustness of at least one AI-based information processing system.

1 shows a schematic representation of an embodiment of the device 1 for providing a database for the robustness assessment of at least one AI-based information processing system 10, which is embodied as a trained neural network, for example. In principle, however, the device 1 can also be used for other AI-based information processing systems 10 .

Provision is made for the at least one AI-based information processing system 10 to provide a function for automated driving of a motor vehicle and/or for driver assistance of the motor vehicle and/or for surroundings detection and/or surroundings perception or for another application.

The device 1 comprises a data processing device 2. The

Data processing device 2 includes a computing device 3 and a memory 4. The device 1 is set up in particular to carry out the method described in this disclosure.

At least one data processing system 10 (i.e. at least one neural network in the example), at least one data set 11, at least one data augmentation definition 12 and at least one difference measure definition 13 are supplied to the data processing device 2 as input parameters 9, these are received by the data processing device 2.

A AI-based information processing system 10 includes in particular a structural description and parameters (e.g. weightings, activation functions, filter parameters, etc.) of the AI-based information processing system 10. The data set 11 includes data, for example two-dimensional camera images and/or other one- or multi-dimensional sensor data from at least one Sensor (camera, lidar, radar, ultrasound, etc.). The at least one data augmentation definition 12 includes in particular a description of at least one data augmentation method, ie a description of how data is to be augmented (eg disturbed) within the framework of the method described in this disclosure. The at least one difference measure definition includes in particular a description of at least one difference measure, i.e. a description of how the non-augmented data is to be compared with the augmented data within the scope of the method described in this disclosure or how a difference value 22 is to be determined .

Based on the input parameters 9, a multidimensional data structure 20 is generated by the data processing device 2, in particular by the computing device 3, and stored in the memory 4, which is shown schematically as a cube in FIG. The dimensions and value ranges of the dimensions of the multidimensional data structure 20 are or are defined by the received input parameters 10, 11, 12, 13 (in the example described, the multidimensional data structure 20 includes four dimensions in particular). To generate the multidimensional data structure 20, a data point 21 is generated or determined for each combination of the input parameters 10, 11, 12, 13. Each data point 21 of the multidimensional data structure 20 comprises a difference value 22 determined by means of the difference measure defined by the at least one difference measure definition 13. The difference value 22 is determined by means of the data processing device 2, in particular by means of the computing device 3, in that the at least one defined difference measure between output data is formed is generated by the at least one AI-based information processing system 10 for data of the at least one data set 11 and for the same data augmented by means of the data augmentation defined by the at least one data augmentation definition 12 . For this purpose, the data processing device 2, in particular the computing device 3, executes the AI-based information processing system 10 on the data and on the augmented data. In a simple example, in which the AI-based information processing system 10 supplies a number as output data, for example, a difference measure can be a difference between the supplied numbers, for example. If vectors are output, differences between the vectors can be formed, for example using a scalar product.

If the difference values 22 are determined for all data points 21 of the multidimensional data structure 20, then the generated multidimensional data structure 20 is provided. In particular, the multidimensional data structure 20 is output. With the help of the provided multidimensional data structure 20, a robustness of at least one Kl-based Information processing system 10 can be assessed based on the difference values 22 comprised by the multidimensional data structure 20 .

With the help of the device and the method, difference values 22 can be determined for different data sets, different data augmentation methods and different difference measures. It is also possible to compare different AI-based information processing systems 10 (e.g. different trained neural networks) with one another. Every possible combination of the input parameters 9 corresponds to a data point 21 for which a difference value 22 is determined. The device 1 and the method therefore allow the provision of an extensive and flexibly expandable database for the robustness assessment of at least one AI-based information processing system 10.

Provision can in particular be made for the robustness of the at least one AI-based information processing system 10 to be evaluated on the basis of the multidimensional data structure 20 provided. For this purpose, the difference values 22 of the individual data points 21 are evaluated and a robustness is determined from this. For example, the difference values 22 can be compared with at least one robustness requirement. Such a robustness requirement can be, for example, a threshold value that is predefined as a function of the respective data augmentation method and that, for example, must not be exceeded by (averaged) difference values 22 .

In a further development, it can be provided that the at least one AI-based information processing system 10 is discarded based on the evaluation, the method is repeated with changed input parameters 10, 11, 12, 13, or the AI-based information processing system 10 is certified as robust.

Provision can be made here for the AI-based information processing system 10 to be loaded into a memory of at least one control device 30 after certification.

It can be provided that the providing includes providing an interface 5 for the targeted retrieval of data points 21 of the multidimensional data structure 20 . The interface 5 can be in the form of hardware and/or software.

It can be provided that providing a transmission of the multidimensional data structure 20 to a certification service provider and / or a user of the class based information processing system 10 and / or loading the multidimensional data structure 20 in a memory of at least one control unit 30 comprises.

Provision can be made for a set of sub-parameters 14 for the at least one data augmentation definition 12 to be additionally received as an input parameter 9, with the multidimensional data structure 20 being generated and/or the data being augmented taking into account the received set of sub-parameters 14. The sub-parameters 14 include, for example, value ranges as input parameters 9 for a data augmentation function.

It can be provided that a set of filter criteria 15 for individual ones of the input parameters 9 is additionally received as the input parameter 9 , the multidimensional data structure 20 being generated taking into account the received set of filter criteria 15 .

It can be provided that a selection of statistical distribution functions 16 for parameter distributions for the at least one data augmentation definition 12 is additionally received as an input parameter 9, with the multidimensional data structure 20 being generated taking into account the selection of statistical distribution functions 16 for parameter distributions.

It can be provided that a selection of distributions 17 for combinations of parameters of data augmentation definitions 12 and data of the data set 11 is additionally received as an input parameter 9, with the generation of the multidimensional data structure 20 taking into account the received quantity of distributions 17 for the combinations of the parameters he follows.

It can be provided that a selection for a relevance 18 of individual data points 21 is received as the input parameter 9 , the multidimensional data structure 20 being generated taking into account the relevance 18 received.

It can be provided that the results 23 generated by the at least one AI-based information processing system 10 are additionally stored in the multidimensional data structure 20 for each data point 21 . Provision can be made for the multidimensional data structure 20 to be expanded after it has been provided by inserting at least one additional dimension and/or by expanding a value range of at least one dimension, with the expanded multidimensional data structure 20+ being provided.

FIG. 2 shows a schematic flowchart to illustrate an embodiment of the method for providing a database for robustness assessment of at least one AI-based information processing system 10 . Provision is made for the at least one AI-based information processing system 10 to provide a function for automated driving of a motor vehicle and/or for driver assistance of the motor vehicle and/or for surroundings detection and/or surroundings perception or another application.

As an input parameter 9, at least one Kl-based

Information processing system 10 (e.g. a trained neural network), at least one data record 11, at least one data augmentation definition 12 and at least one difference measure definition 13 are specified. Based on this, at least one AI-based information processing system 10, data 11a from at least one data set 11, at least one data augmentation method 12a and at least one differential measure 13a are specified.

In a method step 100, the data 11a is augmented by means of the at least one data augmentation method 12a, e.g. the data 11a is disturbed by adding noise or at least one adversarial disturbance. Provision can be made here for parameters for the data augmentation to be selected at random in a preceding method step 90 .

In a method step 101, the at least one AI-based information processing system 10 is applied to the non-augmented data 11a and the respectively associated augmented data. The results obtained in each case are compared with one another by date in a method step 102 and a differential value is determined for each date by means of the at least one differential measure 13a. In the case of several K1-based information processing systems 10, several data sets 11, several data augmentation methods 12a and/or several difference measures 13a, this is carried out for every possible combination, so that a difference value is determined for every combination. A multidimensional data structure 20 is generated from the difference values, the Difference values are assigned to individual data points, each of which is assigned to a possible combination of the input parameters 9 .

The multidimensional data structure 20 generated in this way is provided, for example in the form of an interface 5, by means of which the difference values for any combination of the input parameters 9 can be queried or retrieved, so that the at least one Kl-based information processing system 10 is robust based on the difference values comprised by the multidimensional data structure 20 can be assessed.

The determination of the robustness is shown schematically in method steps 200 to 202. In this case, in a method step 200, difference values are retrieved according to various filter criteria, with the filter criteria resulting from a predetermined degree of robustness. The retrieved difference values are aggregated in a method step 201, for example by statistically or otherwise evaluating the retrieved difference values, so that (statistical) key figures (e.g. in the form of key performance indicators 40, for example in the form of aggregated difference measures, e.g. "average error" = 0.5 or 0.4 etc.) can be calculated and made available (distribution parameters etc.). The aggregation takes place here, for example, across one or more dimension axis(s) of the multidimensional data structure 20, e.g. across a complete data set 11 and/or across all data augmentation methods 12a and/or all difference measures 13a. Then, for example, maximum values and/or unweighted or weighted average values can be formed and provided.

In a method step 202, the key figures and/or aggregated values generated in this way can be visualized, for example by displaying the results as a graph 41, forming histograms 43, displaying the scalar values 44,

Heatmaps 45 are generated and/or a cake graphic 46 is displayed. This allows a robustness 25 of the at least one Kl-based

Assess and certify information processing system 10 improved. Furthermore, a report 42 can be created and output, which contains the results of a robustness assessment and/or certification.

It is particularly possible that the provision of the multidimensional data structure 20 provides an interface 5 for the targeted retrieval of data points multidimensional data structure 20 includes. A query can include the following parameters in particular: desired filter criteria 50, desired axis or dimension selection 51, desired aggregation method(s) 52, desired visualization method 53.

In this way, in particular, a metric generator 300 can be created with which AI-based information processing systems 10 (e.g. trained neural networks) can be evaluated and certified in terms of robustness 25 in a comparable and repeatable manner.

The provision of the multidimensional data structure 20 makes it possible in particular to be able to assess the robustness 25 of a AI-based information processing system 10 even without the data 11a and without the AI-based information processing system 10 itself. This is particularly useful for sensitive data.

reference list

contraption

data processing facility

computing device

Storage

interface

input parameters

AI-based information processing system

record a data

Data augmentation definition a data augmentation method

Difference dimension definition a Difference dimension

sub-parameters

filter criteria

distribution function

distributions

Relevance multidimensional data structure + extended multidimensional data structure

data point

difference value inferred result

robustness

control unit

Key Performance Indicator

graph

report

Histogram scalar value Heatmap cake graphic Desired filter criterion Desired axis or dimension selection Desired aggregation method Desired visualization method Process step -102 Process steps -202 Process steps Metric generator

Claims

patent claims

1. A method for providing a database for the robustness assessment of at least one AI-based information processing system (10), the AI-based information processing system (10) having a function for automated driving of a motor vehicle and/or for driver assistance of the motor vehicle and/or for environmental detection and/or environmental perception or for another application, with at least one AI-based input parameter (9).

Information processing system (10), at least one data record (11), at least one data augmentation definition (12) and at least one difference dimension definition (13) are received, a multidimensional data structure (20) being generated based on the input parameters (9), the dimensions and value ranges of the dimensions of the multidimensional data structure (20) are defined by the received input parameters (9), and wherein each data point (21) of the multidimensional data structure (20) comprises a difference value (22) determined by means of the at least one defined difference measure (13a), which determines is formed in that the at least one defined difference measure (13a) is formed between output data that is generated by the at least one AI-based information processing system (10) for data (11a) of the at least one data set (11) and for the same by means of the at least one defined data augmentation (12a) augmented data was generated n, and wherein the generated multidimensional data structure (20) is provided so that a robustness of the at least one AI-based information processing system (10) can be assessed based on the difference values (22) comprised by the multidimensional data structure (20).

2. The method according to claim 1, characterized in that the providing includes providing an interface (5) for the targeted retrieval of data points (21) of the multidimensional data structure (20).

3. The method according to claim 1 or 2, characterized in that providing a transmission of the multidimensional data structure (20) to a certification service provider and / or a user of the AI-based information processing system (10) and / or loading the multidimensional data structure (20) in a memory of at least one control unit (30).

4. The method according to any one of the preceding claims, characterized in that a set of sub-parameters (14) for the at least one data augmentation definition (12) is additionally received as an input parameter (9), with the generation of the multidimensional data structure (20) and/or the Augmenting the data (11a) taking into account the received set of sub-parameters (14).

5. The method according to any one of the preceding claims, characterized in that a selection of statistical distribution functions (16) for parameter distributions for the at least one data augmentation definition (12) is additionally received as input parameters (9), the generation of the multidimensional data structure (20) being Taking into account the selection of statistical distribution functions (16) for parameter distributions.

6. The method according to any one of the preceding claims, characterized in that a selection for a relevance (18) of individual data points (21) is received as the input parameter (9), the generation of the multidimensional data structure (20) taking into account the received relevance (18 ) he follows.

7. The method as claimed in one of the preceding claims, characterized in that the results (23) obtained from the at least one AI-based information processing system (10) are additionally stored in the multidimensional data structure (20) for each data point (21).

8. The method according to any one of the preceding claims, characterized in that the multidimensional data structure (20) is expanded after it has been provided by inserting at least one further dimension and/or by expanding a value range of at least one dimension, the expanded multidimensional data structure (20+ ) provided.

9. The method as claimed in one of the preceding claims, characterized in that the at least one AI-based information processing system (10) is a neural network and/or comprises at least one neural network.

10. Device (1) for providing a database for the robustness assessment of at least one AI-based information processing system (10), wherein the AI-based information processing system (10) has a function for automated driving of a motor vehicle and/or for driver assistance of the motor vehicle and/or for an environment detection and / or environment perception or for another application, comprising a data processing device (2), wherein the data processing device (2) is set up as an input parameter (9) at least one Kl-based

Information processing system (10) to receive at least one data record (11), at least one data augmentation definition (12) and at least one difference measure definition (13), and to generate a multidimensional data structure (20) based on the input parameters (9), the dimensions and value ranges of the dimensions of the multidimensional data structure (20) are defined by the received input parameters (9), and wherein each data point (21) of the multidimensional data structure (20) comprises a difference value (22) determined by means of the at least one defined difference measure (13a), which determines is formed in that the at least one defined difference measure (13a) is formed between output data that is generated by the AI-based information processing system (10) for data (11a) of the at least one data set (11) and for the same by means of the at least one defined data augmentation (12a ) augmented data were generated, and the generated m provide an ultimate data structure (20), so that a robustness of the at least one AI-based information processing system (10) can be assessed based on the difference values (22) comprised by the multidimensional data structure (20).

11. Control unit (30), comprising a memory, wherein a multidimensional data structure (20) is stored in the memory, which is suitable as a database to enable a robustness assessment for at least one AI-based information processing system (10).

12. Use of the control unit (30) according to claim 11 for evaluating and certifying at least one AI-based information processing system (10).