Method and device for enabling an explanation and / or verification of a behaviour of a neural network trained with a training database, and vehicle with ai

Abstract

The invention relates to a method for explaining and / or verifying a behaviour of a neural network trained with a training database, said method having the following steps: - determining sensor information by means of a sensor, so that sensor information is available, - applying the neural network to the sensor information, so that target information is available, - determining similarity information, wherein a comparison is made of the sensor information or part of the sensor information with an information database, so that the similarity information is available, linking the similarity information with the target information, so that a target information tuple is available, on the basis of which the explanation and / or verification of the behaviour of the neural network can be realised.

Description

Method and device for enabling explanation and / or verification of a behavior of a neural network trained with a training database, and vehicle with AI

[0001] The invention relates to a method for enabling an explanation and / or verification of a behavior of a neural network trained with a training database, a device for enabling an explanation and / or verification of a behavior of a neural network trained with a training database, and a vehicle which carries out the corresponding method or has such a device.

[0002] Neural networks, which are the main component of artificial intelligence (AI), often result in results that are incomprehensible. For example, a neural network trained on images of ships can identify a church in a harbor area as a military ship, even though it is clearly a building. Therefore, there is a need for the results obtained with neural networks to be either explainable, explainable, or verifiable.

[0003] DE 10 2020 204 758 A1 discloses a method for checking a result determined by means of artificial intelligence using a symmetry comparison. Symmetries in assigned images are used to check the reliability of the assignment.

[0004] DE 10 2020 203 705 A1 describes a method for verifying the plausibility of the output of neural classifier networks using additional information. For example, information about a locally detected match and its spatial assignment is output.

[0005] DE 10 2022 205 413 A1 shows a method for recognizing road signs in the environment of a motor vehicle.

[0006] The object of the invention is to improve the state of the art.

[0007] The task is solved by a method for enabling an explanation and / or verification of the behavior of a neural network trained with a training database, comprising the following steps: - Determining sensor information by means of a sensor so that sensor information is available, - Applying the neural network to the sensor information so that target information generated by the neural network is available, - Determining a similarity information, whereby a comparison with the sensor information or a part of the sensor information with an information database is carried out by a comparison unit, so that the Comparison unit generated similarity in format! on is present, - Linking the similarity information with the target information by a linking unit, so that the linking results in a target information tuple, so that the behavior of the neural network can be explained and / or verified using the target information tuple.

[0008] This method can provide a means of explaining, checking, and / or verifying the behavior of the neural network using the available target information tuple. For example, by analyzing the target information tuple, an operator, observer, or even a computer system can use this target information tuple to check the probability of the accuracy of the target information determined by the neural network.

[0009] For example, false-positive detection cases (i.e., misrecognition or misallocation of the sensor information or misclassification of the sensor information as specific target information) can be directly detected and resolved, so that the rate of false-positive target information can be significantly reduced by using the similarity information.

[0010] In addition, this method allows the reliability of the neural network or any need for further training to be derived or identified quickly and directly.

[0011] The following terms are explained in more detail in connection with the invention:

[0012] "Enabling an explanation and / or verification" of a neural network's behavior describes a procedure that, through targeted queries, comparisons, or other logical actions, enables the explanation or verification of behavior that cannot be directly retrieved due to the design of the neural network through indirect actions or observations. "Explaining" in this case refers to a logical understanding of the behavior or the displayed result, while "verifying" refers to a plausibility check or testing of a result of the neural network's activity, particularly by checking an expected result or an expected quality of the result.

[0013] The "behavior" of the neural network refers in particular to the way in which the neural network operates, although this can typically only be described indirectly, in particular by logically following the corresponding results or result patterns of the neural network.

[0014] A "training database" serves as the basis for the behavior of the neural network, whereby images, data sets or other information are stored in this training database and give the neural network the opportunity to practice and train corresponding behavior patterns or determination of results. For example, a training database is provided with 1000 images of street signs, whereby these 1000 images (number is exemplary) are made available to the neural network in order to develop its own behavior for identifying street signs in images. Naturally, it is often not possible to understand the criteria according to which a neural network makes a corresponding identification. The same applies if, instead of the example images, other data sets are provided in which patterns or rules are to be recognized.

[0015] A "neural network" in this context refers in particular to an artificial neural network, i.e. a neural network implemented within a machine or a computer system, which is formed from artificial neurons and was inspired by networks that biological neurons form, for example, in a brain. Such a neural network, which is used in particular in so-called machine learning, is, under certain conditions, capable of solving a problem whose description with rules or an algorithm is too complex or even impossible. It should be noted that the terms "network" and "network" can be used synonymously here. For example, a Neural networks thus enable the analysis of large amounts of data, automated image analysis, or even the detection of invisible regularities or structures in data sets. The present invention is particularly directed toward automated image recognition, with the neural network being used to detect objects or signals in the images.

[0016] "Determining sensor information" describes the recording of sensor information, i.e., information determined in particular over a distance, using a sensor. The sensor information itself can comprise an electronic data set, which, for example, provides information about a target object recorded by the sensor.

[0017] The "sensor" is the actual detector, which is also referred to as a pickup or sensor. This is a technical component that can record physical or chemical properties and convert them, for example, into an electronic signal. The electronic signal generated in this way represents the recorded physical or chemical property in such a way that a conclusion, for example, a numerical value for this respective property is possible from the electronic signal. Examples of such sensors include temperature sensors, radiation sensors, long-range optical sensors or acoustic sensors.

[0018] The "application of the neural network" describes the activation or use of the neural network on the sensor information in such a way that, for example, the sensor information is examined, classified, analyzed, or searched by the neural network, for example to obtain knowledge from the sensor data regarding a specific object or a specific formation, provided that the sensor data corresponds, for example, to an image. Analogously, a data set can be searched by the neural network for systematics, patterns, or the like.

[0019] A “target information” describes in particular information about a “target” obtained from the sensor information, i.e. for example information sought from the set of sensor information or an assignment of a certain element or a certain sub-area of ​​the sensor information to a property or its provision with an attribute. For example, target information can consist of a street sign or even a building or a vehicle being recognized, marked or specially highlighted as such within an image available as sensor information. The target information is therefore, for example, the label “street sign”, “building” or “vehicle” for a certain image section, i.e. in particular with the labeling of the location or position in the image.

[0020] For example, “determining similarity information” describes finding a similarity between the sensor information and an information database. For example, a comparison is made with the information database for an image captured with the sensor, which leads to the identification of the most similar image from the information database. This most similar image, or a reference to the most similar image, then represents the similarity information. This applies analogously, for example, to the analysis of a large data set, where, for example, the most similar data set to a set of data sets from the information database is identified as similarity information.

[0021] The "comparison" in this case represents in particular the process in which the similarity information is determined or assigned. For an image as a sensor signal, for example, a minimal (color) deviation, a contour detection, a brightness or a color analysis as well as a pixel comparison or the like can be carried out as a comparison, whereby this applies to both the black / white and the color spaces of corresponding compared images. Likewise, a comparison can also lead to several, for example hierarchically sorted, items of similarity information from "very similar" to "less similar" or a comparable quality scale. In this case, the similarity information can also contain several pieces of information.

[0022] The "information database" is in particular a Database in which a large number of data records are stored, which are assigned, for example, to a specific class of target information. This way, for the already For example, for a road sign, an information database may contain comparison images for a stop sign, an information sign, a speed limit sign, or a prohibition sign, wherein the comparison of an image recognized by the neural network as a “stop sign” is compared with all or at least one or some of the reference images stored as a “stop sign” in the information database. As a result, at least one image of the stop sign recognized as being particularly similar is then kept ready as similarity information for the target information. The “comparison unit” is a technical device set up to compare the information, whereby a computer or a processing unit is usually used here, the comparison process being carried out using image analysis software, for example. It should be mentioned that an analogous procedure is also possible for any other data sets.For example, to analyze a large amount of bank data from a large number of customers with suspected money laundering, the neural network processes it and identifies a data set as "suspicious" (target information), with data sets from previously solved money laundering cases being kept in the information database. This would, for example, identify a data set as "most similar" in terms of due booking dates, denominations of amounts, or typical amounts (similarity information). An operator can then, for example, perform a visual comparison. so that an assessment of the validity of the neural network's assessment becomes possible.

[0023] The combination or joint provision of the target information with the similarity information represents "linking," whereby a link is created, for example, by storing it with a common marker, in a common memory area, or with mutual assignment. A "linking unit" represents the computing device that carries out this process. A computer or a processing unit that runs appropriately configured software can also be used for this purpose.

[0024] As a result, a "target information tuple" is available, where this tuple represents the respective associated link between the target information and the similarity in format! on.

[0025] In order to be able to make a better assessment of the behavior of the neural network, a match value is determined within the target information tuple, whereby the match value represents a match between the similarity information and the target information.

[0026] This approach can, for example, provide a qualitative and quantitative assessment of the similarity between the target information and the similarity information.

[0027] The "match value" or a degree of "match" can be specified, for example, as a percentage, as a value on a previously determined scale, or in some other way. For the example of image analysis already mentioned, it can therefore be stated, for example, that the image obtained from the information database has a "similarity of 85%" to the image of the target information, whereby this numerical value naturally depends on the type of comparison and the respective data quality. The match value can, for example, be carried out by a device similar to the comparison unit, generally by a computer or a processing unit.Likewise, a common computer or a common computing unit with respective software can be used for the functions mentioned, which can take over the tasks of the neural network, the comparison unit, the linking unit and / or a unit used to determine the match partially or entirely with respective software.

[0028] According to one embodiment, the similarity information can be displayed together with the target information and in particular together with the match value by means of a display unit so that it can be read by a viewer.

[0029] This allows the viewer to view the target information together with the similarity information and / or the match information and thus make a personal assessment of the quality of the activity of the neural network. For example, in the image recognition example mentioned above, the image captured by the sensor (target information) is displayed on the display unit with the most similar image from the information database (similarity information) and, for example, the match value of 85%, so that the viewer can estimate both the most similar image and the confidence level (the comparison value here is 85%) of the recognition.

[0030] The "display unit" used here can be, for example, a monitor or screen, which can also have additional filters or overlay modes, for example, to minimize glare or adjust brightness information or, for example, to display additional information to the target information. The "viewer" in this case can be an operator, i.e., a person, but also, for example, a computer or another analysis tool.

[0031] According to one embodiment, the sensor information is image information that can be determined or is determined using a camera. Thus, for example, as described above, a camera can be used to record images as sensor information.

[0032] The “image information” can usually be in the form of digital image information, for example as a storable data set, whereby the “camera” can be designed as a digital camera in accordance with the usual technical design.

[0033] According to one embodiment, the training database is a labeled image database, wherein the labeled image database comprises in particular traffic signs, in particular navigation marks, and / or vehicles, in particular watercraft.

[0034] This makes it possible, for example, to provide previously marked, or "labeled," images to facilitate the determination of similarity information. This way, the similarity information can be extracted only from a set of data in the information database that, for example, carries a label matching the target information.

[0035] A "labeled image database" refers to a database of images where an image, several images, or even each image contains a "label," i.e., additional information with attributes for this image or a specific group of images. For example, a specific group of images known to depict buildings can be labeled "house" or "building."

[0036] According to a further embodiment, the target information comprises information of a target object determined by means of the neural network.

[0037] Thus, it is possible that concrete information about the target object is determined by means of the neural network and an assignment and in particular the explanation and / or verification of the target information by the observer is facilitated.

[0038] The "information of a target object" can, for example, be a naming of the target object, but properties of the target object can also be named in order to provide the information of a target object.

[0039] In particular, according to a further embodiment, the information database can comprise the training database or a portion of the training database. This allows the comparison to be performed based on the trained images. This enables an effective assessment of the detected object, such as a ship or a navigation mark. Even just portions of the training database can be usefully used for this purpose, as this may allow a comparison to be performed more quickly or with less computational effort.

[0040] According to a further embodiment, the similarity information comprises a similarity numerical value and / or one or more information database entries, in particular one or more images. As already explained above, a direct comparison of a similar image with the target information can therefore be provided, for example, wherein the similarity numerical value can indicate the specific similarity analogous to the match value with a specific numerical value or, for example, a grouping within a numerical value of a scale.

[0041] According to a further embodiment, the target information comprises an image determined by means of the sensor, in particular a camera, sonar, lidar or radar image.

[0042] A "camera image" refers to an optically acquired image, a "sonar image" to an image generated by sound waves, a "lidar image" to an image generated by a multidimensional laser scanner, and a "radar image" to an image of an environment or viewing area generated by radar beams. Likewise, it can refer, in particular, to an electronically captured and / or stored representation of the respective image for further electronic processing.

[0043] In particular, the target information tuple can be an n-tuple, where n has a value of 2, 3, 4, 5, 6, 7, 8, 9, or 10.

[0044] This means that, for example, several entries of the information database can be kept or provided in the respective target information tuple in order to provide, for example, an order of entries of the information database in descending order of similarity, in particular with a respective similarity number value.

[0045] According to one embodiment, the target object and the target information tuple and / or in particular also the match value can be displayed by means of a display device, in particular for a viewer.

[0046] According to a further aspect, the object is achieved by a device for enabling an explanation and / or verifying a behavior of a neural network trained with a training database, comprising a neural network trained with a training database, a sensor, a comparison unit and a linking unit, configured to carry out a method according to one or more of the previously explained embodiments.

[0047] Such an "enabling device" can, for example, be a computer with corresponding display capabilities, which, as described above, carries out the corresponding method and thus enables the explanation and / or verification of the behavior of the neural network.

[0048] According to one embodiment, the device has a display unit for displaying the similarity information and the target information and in particular the match value for a viewer.

[0049] According to a further aspect, the object is achieved by a vehicle, in particular a train, a motor vehicle, a ship or an aircraft, which has a sensor for determining sensor information, a neural network trained by means of a training database and an information database and / or a device according to one of the previously described embodiments, wherein the vehicle and / or the device is set up and designed in such a way that a Method according to one of the preceding claims is feasible.

[0050] The method according to the invention and / or the device according to the invention can therefore be advantageously used to, for example, carry out a recognition of target objects such as road signs, buildings, ships, other vehicles or even pedestrians on a corresponding vehicle and to provide corresponding data in a processed form, for example in a control station of a locomotive, in the cockpit of an aircraft or helicopter, on a bridge of a ship or in the cockpit of a car, in order to be able to check, for example, the quality of the recognition of railway signals, markings on a runway, ships and navigation marks or traffic signs and road signs.

[0051] According to one embodiment, the vehicle can have a display unit, wherein the display unit is visible to a viewer and the display unit is configured to display the similarity information within the target information tuple and the target information and in particular the match value, so that an explanation and / or verification of the behavior of the neural network trained with a training database is possible by the viewer.

[0052] The invention is explained again by way of example:

[0053] By comparing or evaluating results and information from different sources, the result of a neural network can be explained. This is particularly the case with image data which, for example, must be used to interpret traffic signs, vehicles and the like, for example in a maritime environment or in rail traffic. Especially when visibility is poor, as is the case during storms or fog, traffic signs and watermarks can certainly be misinterpreted by the neural network. The invention in the present case provides additional information which is generated by linking the AI ​​information (target information) with further information (e.g. an image from an image database).

[0054] A core idea of ​​the invention is that target information is determined by means of a neural network and can be displayed to a user, for example on a monitor, in conjunction with other information.

[0055] In general, the neural network was trained using a training database. For example, corresponding traffic signs in different lighting scenarios, daytime conditions, weather conditions, and other environmental conditions were input (trained) into the neural network so that it could subsequently evaluate unknown images and recognize corresponding target objects, such as traffic signs or watermarks.

[0056] The input signal for the neural network is generally sensor information, such as information or image data from a CCD camera or other optical camera operating in the TR, UV, or visible range. The sensor is therefore particularly a camera. Using the sensor information, the neural network determines the target information in particular. This target information is, for example, an object recognized in an image that has been marked (labeled) accordingly. For example, a "STOP" traffic sign with its distinctive octagonal shape can be identified as such and displayed to a user as a STOP sign.

[0057] The similarity information then serves in particular as image information found in a different way than by means of the specially trained neural network, which can be compared with the target information. For example, the target object found in the camera image (target information) is compared with an image database. This comparison can, as mentioned, be carried out pixel by pixel, by comparing the individual color values ​​of the pixels with each other. On the other hand, the comparison can also be carried out using another neural network, for which significantly fewer or more images were used for training and / or which has more or fewer neural nodes. Therefore, part of the sensor information (target information) matches only one image or value of the comparison database. This Similarity comparison can therefore, for example, lead to a numerical value which is an indicator of reliability and / or the most similar image is taken from the comparison database and displayed parallel to the actual target information so that a viewer can see both the target information and the “most similar” image from the image database. In this sense, the target information and, for example, the similarity numerical value form a (2-) tuple which has two elements. This target information tuple can, of course, also contain further elements such as the most similar image or the next three most similar images, so that, for example, the target information, the similarity value and one similarity image each form a 5-tuple.

[0058] If the information database matches the training database, it is even possible to deduce why the neural network chose a particular label or recognition, since the most similar image is compared to the target information. If, for example, the image from the training database corresponds to a warship, and the supposedly recognized image corresponds to a cloud formation or a church building, it is clearly evident to an observer that the neural network has identified a false positive.

[0059] In particular, the image database contains labelled data, for example of traffic signs, so that each traffic and watermark is given a specific name.

[0060] In the following, the invention is explained on the basis of Explained in more detail using examples. It shows Figure 1 is a schematic representation of a Computer system with neural network and a monitor on which images captured by a camera are displayed with additional information, as well as Figure 2 is a schematic representation of a Monitor display for comparing a picture taken with a camera Image section with a labeling by a neural network with a representation of a similarity image in a display field.

[0061] A computer system 101, which can be installed, for example, on a ship or in a vehicle, has a trained neural network 103, a CCD camera 111, a comparison computer 109, and a monitor 113 with a processing unit. The CCD camera 111 is connected to the neural network 103 via a K1 connection 119. The output signal of the neural network 103 is connected to the processing unit of the monitor 113 via a K1 data line 117. Furthermore, the trained neural network 103 is also connected to the comparison computer 109 via an image connection 121. and the output signal of the comparison computer 109 is also connected to the processing unit of the monitor 113 via the image data line 115.

[0062] The neural network 103 was trained using a training database 105 to recognize traffic signs or, alternatively, for example, to recognize navigational aids and ships. Traffic sign recognition will be explained below using an example. Thus, images captured with the CCD camera 111 are analyzed for corresponding traffic signs, and if a corresponding traffic sign is detected, it is displayed and identified as a recognized traffic sign 127 on the monitor 113. The detected traffic sign is fed to the comparison computer 109 via the image connection 121.

[0063] In a first alternative, this comparison computer 109 accesses an image database 107, which, among other things, contains a wide variety of images of respective traffic signs. These images are compared with the traffic sign found in the camera image by means of a pixel comparison. The image with the highest similarity value, or if several images have the same value, one of them, is or are sent to the processing unit of the monitor 113 via the image data line. In addition, the similarity value, the probability of match 131, can also be transmitted as a numerical value to the processing unit of the monitor 113 via the image data line. The processing unit then displays the Monitor 113 displays the actual image of the CCD camera 111, the labeled Kl object (traffic sign found here), the most similar image of the traffic sign as a comparison object 129, and the probability of match 131 in the image in such a way that the essential image content can be perceived by a user.

[0064] In a second alternative, the comparison computer 109 uses the training database 105 and searches for similar images to the detected traffic sign. The rest of the display is identical.

[0065] A monitor display 201 is described as an application on a ship as follows:

[0066] The monitor display 201 is used on a ship's bridge (not shown) to detect possible obstacles using a neural network and to provide evasive recommendations. This is intended to relieve the ship's captain or officer by supporting the lookout. The system that generates the monitor display 201 is a computer that uses a camera (not shown) to record a panoramic view of the moving ship and then uses a neural network, i.e., artificial intelligence, to detect possible collision partners and provide evasive recommendations. The neural network was trained using images from a database, each containing ships and navigational aids (buoys, buoys). The display of the monitor display 201 is for example on a computer screen and is shown in Fig. 2 only for reasons of clarity.

[0067] The monitor display has a display field 202, which points to an image section 203. The image section 203 corresponds to a zoomed portion of the camera image. Within the image section 203, a cloud formation 211, a cloud 215, and a cloud 217 are visible. Furthermore, the sky 218 is visible as a background, as is water 219 at the lower edge of the image. Between the water 219 and the cloud 211, fog 221 obscures the image.

[0068] Below the image section 203, at the lower edge of the image, there is a course scale 204 which shows the visual course 205 (here 253°) in a display field.

[0069] A detection field 231 shows an area that the neural network has identified as a ship. The detection field has a dashed border surrounding the supposed ship, as well as the text "SHIP" at the top of the frame. Cloud 211 and cloud 215 are enclosed within the frame.

[0070] In the course scale 204, a course recommendation 206 is displayed as a flashing arrow, which was given due to the detection of the suspected vessel (cloud 211) and a presumed course (aligned to the left) in order to avoid a collision with the suspected vessel. The flashing arrow of the course recommendation 206 indicates indicates that the alternative recommendation has been recognized and accepted by the system, but has not yet been implemented.

[0071] On a control console (not shown), the captain or officer can now confirm the course recommendation 206 (also by not taking action) or acknowledge it negatively and thus reject it.

[0072] A display field 252 serves to assist the captain or officer in making this decision. For this purpose, the image section 203 is compared for similarity against an image database, which, for example, contains a large number of images of ships. This can also be done using a neural network; however, for this example, a pixel comparison in the black / white color space is assumed. For this purpose, it is assumed that the image section 203 represents a black / white infrared image, in which misidentification by the neural network is more likely than, for example, with a daylight color image.

[0073] The pixel comparison has now shown that an image displayed in display field 252 has the greatest similarity to the image section displayed in detection field 231. The image in display field 252 shows a ship 261 on water 269 with an exhaust plume 265 against a sky 268.

[0074] Furthermore, a display 271 at the top left of the screen shows a match value of 80%, which indicates that the pixel comparison has resulted in a match of 80% for the applied parameters. For example, during automatic detection in Detection field 231 by the neural network of the fog 221 resulted in the lack of contact between the cloud 211 and the water 219 not being recognized, so that the neural network incorrectly assigned the image section as a ship.

[0075] The captain can now visually check in the joint display of display field 202 and display field 252 which formation is marked as "SHIP" and visually compare this with the similar image in display field 252. If it is recognized that this is a misassignment, the automatic course correction can be acknowledged and prevented. Furthermore, the relatively high match value of 80% can be used to determine or question the quality of the image recognition by the neural network and, for example, initiate an extension of the training of the neural network. Reference symbol list 101 Computer System 103 Neural Network 105 Training database 107 Image database 109 comparison calculators 111 CCD camera 113 Monitor with processing unit 115 Image data line 117 Kl data line 119 Kl-Connection 121 Image connection 123 Training database linking 125 Image database linking 127 Kl-Obj ect 129 Comparison object 131 Practice mood probability 133 information tuples 201 Monitor display 202 Display field 203 Image detail 204 Course scale 205 sight course 206 Course recommendation 207 Horizontal grid 209 vertical grids 211 Cloud formation 215 Cloud 217 Cloud 218 Heaven 219 Water 221 Fog 231 Detection field 252 Ad! eld 261 ship 265 Exhaust plume 268 Sky 269 Water 271 ad

Claims

Patent claims:

1. A method for enabling explanation and / or verification of a behavior of a neural network (103) trained with a training database (105), comprising the following steps: - determining sensor information (203) by means of a sensor, so that sensor information (203) is available, - applying the neural network to the sensor information (203) so that target information (127, 231) generated by the neural network is available, - Determining similarity information, where a Comparison with the sensor information (203) or a part of the sensor information (203) with an information database (107) by a Comparison unit (109) is carried out, so that the signal generated by the comparison unit (109) Similarity information (252) is available, - Linking the similarity information (252) with the target information (127, 231) by a Linking unit (113) such that the linking results in a target information tuple (133, 231, 252) such that the explanation and / or verification of the behavior of the neural network can be explained using the target information tuple (133, 231, 252).

2. The method according to claim 1, wherein determining a match value (131, 271) within the Target information tuple (133, 231, 252), wherein the match value (131, 271) indicates a match of the similarity information (252) with the target information (127, 231) represents.

3. The method according to claim 1 or 2, wherein the Similarity information (252) and the target information (127, 231) and in particular the agreement value (131, 271) are displayed jointly for a viewer to read by means of a display unit (113, 201).

4. Method according to one of the preceding claims, wherein the sensor information (203) is image information which can be determined with a camera (111).

5. Method according to one of the preceding claims, wherein the training database is a fabled image database, wherein the fabled image database comprises in particular traffic signs and / or vehicles (261).

6. Method according to one of the preceding claims, wherein the Target information (127, 231) by means of the neuronal Network comprises information of a target object (211).

7. Method according to one of the preceding claims, wherein the information database comprises the training database or a part of the training database.

8. Method according to one of the preceding claims, wherein the Similarity information (252) a Similarity number value (131, 271) and / or a Information database entry or several Information database entries (129), in particular one or more images.

9. Method according to one of the preceding claims, wherein the target information (127, 231) comprises an image determined by means of the sensor (111), in particular a camera, sonar, lidar or radar image.

10. Method according to one of the preceding claims, wherein the Target information tuple (133, 231, 252) is an n-tuple, where n has a value of 2, 3, 4, 5, 6, 7, 8, 9 or 10 .

11. Method according to one of the preceding claims, wherein the target object (211) and the target information tuple (133, 231, 252) and / or in particular the match value (131, 271) are displayed by means of a display device (113, 201).

12. Device (101) for enabling an explanation and / or verification of a behavior of a neural network trained with a training database (105), comprising a neural network trained with a training database (105), a sensor (111), a comparison unit (109) and a Linking unit (113) configured to carry out a method according to one of the preceding claims.

13. Device according to claim 12, characterized by a Display unit (113) for displaying the Similarity information (252) and the target information (127, 231) and in particular the agreement value (131, 271) for a viewer.

14. Vehicle, in particular train, motor vehicle, ship or Aircraft, which has a sensor (111) for determining sensor information, a neural network (103) trained by means of a training database (105) and an information database (107) and / or a device according to one of claims 12 or 13, characterized in that the vehicle or the device is set up and designed such that a Method according to one of the preceding claims is feasible.

15. Vehicle according to claim 14, characterized by a display unit (113, 201), wherein the display unit (113, 201) is visible to a viewer and the display unit (113, 201) is configured to display the similarity information (252) within the target information tuple (133, 231, 252) and the target information (127, 231) and in particular the match value (131, 271), so that an explanation and / or verification of the behavior of the neural network (103) trained with a training database (105) is made possible by the viewer.

Images