Evaluating results of computer-based machine learning system

Abstract

One aspect of the present disclosure relates to a computer-implemented method for evaluating results of a computer-based machine learning system. The method includes receiving a real element dataset, where each real element of the real element dataset corresponds to a real observation. The method includes receiving a first result calculated by a computer-based machine learning system using a real element data set. The method includes generating a synthetic element dataset associated with a real element dataset. The method includes transmitting a synthetic element dataset. The method further includes receiving a second result calculated by the computer-based machine learning system using the synthetic element dataset. The method includes comparing the first result with the second result. The method includes determining, based on the first result and / or a comparison of the first result with the second result, one or more validity values characterizing the calculated confidence of the first result.

Description

Technical Field

[0001] The present invention relates to techniques for evaluating the results of a computer-based machine learning system. Related aspects relate to a computer program, a monitoring module, an apparatus, and a distributed system. Background Art

[0002] Computer-based machine learning systems are increasingly being used in a variety of technological devices. In many cases, computer-based machine learning systems (such as those using artificial neural networks) must be extensively trained and / or validated on known data sets in order to provide reasonable output data for any given input data of the device (i.e., unknown output data). However, the use of such models in safety-critical applications (e.g., related to autonomous or assisted driving) is limited because some existing methods in the prior art may provide unexpected results (e.g., false positive and / or false negative results) if the machine learning system is run on a data set that is different from those used for training and / or validation. In some cases, such as in image processing, if the image is mixed with noise (even if the human eye cannot detect such noise), the artificial neural network may misclassify an object that was originally correctly classified in the same image.

[0003] Thus, some existing computer-based machine learning systems according to the prior art utilize the output data to jointly generate uncertainty values (either directly or using a separate probabilistic model), which quantify the degree of reliability of the calculated output data. However, these uncertainty values provided by the computer-based (e.g., in-vehicle) machine learning systems of the prior art, which may be generated in real time in a moving vehicle, are independent of the current environment / scenario and, perhaps even more problematically, are independent of its evolution over time. Furthermore, some prior art methods do not take into account how these uncertainty values change over time. This may result in an overestimation of the reliability of the results generated by the computer-based machine learning system, which is unacceptable for safety-critical applications.

[0004] Thus, there is a need to develop new techniques for computer-based machine learning systems that can address some or all of the above-mentioned problems. Summary of the Invention

[0005] A first general aspect of the present disclosure relates to a computer-implemented method for evaluating the results of a computer-based machine learning system. The method includes receiving a real element dataset, wherein each real element of the real element dataset corresponds to a real observation. The method also includes receiving a first result, which is calculated by the computer-based machine learning system using the real element dataset. In a next step, the method of the first aspect includes generating a synthetic element dataset related to the real element dataset. The method also includes sending the synthetic element dataset. The method also includes receiving a second result, which is calculated by the computer-based machine learning system using the synthetic element dataset. To this end, the method of the first aspect includes comparing the first result with the second result. Finally, the method includes determining one or more validity values based on the first result and / or the comparison of the first result with the second result, wherein the validity values represent the confidence of the calculated first result.

[0006] A second general aspect of the present disclosure relates to a computer program designed to execute the method according to the first general aspect of the present disclosure.

[0007] A third general aspect of the present disclosure relates to a monitoring module, which is designed to execute the computer-implemented method according to the first general aspect and / or to execute the computer program according to the second general aspect.

[0008] A fourth general aspect of the present disclosure relates to a device comprising the monitoring module according to the third aspect. The device further comprises a module comprising a computer-based machine learning system, wherein the module is configured to receive a real element dataset, the real element dataset being received by the monitoring module according to the third aspect. The module is further configured to calculate a first result using the computer-based machine learning system and transmit the first result to the monitoring module.

[0009] A fifth general aspect of the present disclosure relates to a distributed system comprising the apparatus according to the fourth aspect and an additional module, the additional module comprising a computer-based machine learning system. The additional module is configured to receive a synthetic element dataset, the synthetic element dataset being sent by the monitoring module according to the third aspect. The additional module is further configured to: generate a second result using the synthetic element dataset by the computer-based machine learning system; and send the second result to the monitoring module.

[0010] The techniques of the first to sixth general aspects may have one or more of the following advantages.

[0011] First, the present technology can efficiently determine the confidence level of results generated in real time by a computer-based machine learning system in relation to real-world elements associated with observations from the real world (e.g., when these real-world elements, such as detected image data, are processed in real time by the computer-based machine learning system). In particular, based on these generated results, which in some cases include uncertainty values along with the output data of the computer-based machine learning system, a conclusion can be drawn in real time as to whether these output data are sufficiently trustworthy (or not trustworthy) for a particular application.

[0012] Secondly, the present technology provides the possibility of generating synthetic elements for the real elements analyzed at the current point in time, these synthetic elements can simulate the evolution of the current scene / environment over time (for example, including synthetic elements that simulate future, as-yet-undetected real elements), and these synthetic elements are also processed by the computer-based machine learning system. Here, the results generated by the computer-based machine learning system with respect to the synthetic elements and the real elements can be compared with each other to further specify the confidence level of the results generated with respect to the real-world elements (for example, in real time). Thus, using the present technology, the confidence level of the generated results with respect to the real elements can be assessed more accurately using relatively controllable hardware resources than some techniques in the prior art.

[0013] In this disclosure, some terms are used in the following manner:

[0014] In this disclosure, the term "computer-based machine learning system" is understood to mean any device that can be or has been trained for one or more tasks using machine learning. During training, a training dataset can be provided to the computer-based machine learning system as input data, and the characteristics of the computer-based machine learning system (e.g., corresponding parameters) can be adjusted in response to processing the training dataset (e.g., by analyzing output data) so as to solve the one or more tasks in a defined manner (e.g., with a certain accuracy). The computer-based machine learning system can include a parameterizable model. For example, the characteristics of the machine learning system can be adjusted during learning by an optimization process related to the (unknown) parameters of the machine learning system (e.g., the corresponding model), which can be expressed as minimization of a loss function (within the framework of a predetermined numerical accuracy and / or until a predetermined interruption criterion is reached). The adjusted machine learning system can then be used to predict responses to observations in another dataset, a so-called validation dataset: this validation dataset can thus provide an assessment of the fitness of the machine learning system with respect to the training dataset. Finally, in some cases, a test dataset can be used, which is typically used to evaluate the final machine learning system selected during the validation process.

[0015] In some cases, a "computer-based machine learning system" can perform a classification task or a regression task. In a non-limiting example, a computer-based machine learning system (e.g., its model) can include an artificial neural network having: a specific topology; and a plurality of neurons with corresponding connections. According to some embodiments, the neural network can be a convolutional neural network ("convolutional neural network" or "CNN"), which is defined, for example, by the number of filters, filter size, step size, etc. A convolutional neural network can be used for image classification purposes and perform one or more transformations on a digital image, the one or more transformations being based on, for example, convolution, nonlinearity (ReLU), pooling, or classification operations (e.g., when using a fully connected layer). The neural network can also be designed as a multi-layer feedforward or recurrent network, a direct or indirect feedback neural network, or a multi-layer perceptron. The neural network-based machine learning system can be used in a vehicle computer ("Vehicle Computer") or other vehicle components or in an at least partially autonomous robot (e.g., for evaluating the operating state of a vehicle or robot and / or for controlling the functions of a vehicle or robot based on the vehicle or robot's state data and / or environmental data as input data). For example, a computer-based machine learning system can be implemented in any suitable form, i.e., in software, in dedicated hardware, or in a hybrid of software and dedicated hardware. Thus, a computer-based machine learning system can be a software module (which can also be integrated into a higher-level software system) that can be executed on a general-purpose processor. In other cases, a computer-based machine learning system can be (at least partially) implemented in circuit technology.

[0016] Within the scope of the present disclosure, the term "real element dataset" may refer to a collection of real elements (or in some cases a single real element) that are related to real observations (in other words: observations of the real world, such as measurements). Depending on the nature of the task for which the computer-based machine learning system is used, the elements of the "real element dataset" may contain different information. Thus, for example, each real element of the "real element dataset" may contain a single image (or a fragment of the single image). For example, an image element may contain multiple image pixels (e.g., 1024x2048 pixels), each of which has multiple color values (e.g., three color values with a 16-bit color depth). In other examples, the elements of the "real element dataset" may contain data sequences (e.g., time series). These image data and / or data sequences may be generated with the aid of various sensors (e.g., cameras, radars, lidars, ultrasonic sensors, GPS sensors, or thermal sensors, sensors for vehicle engine control). In some examples, a vehicle's sensors may be used in the context of partially autonomous, autonomous, or assisted driving of a vehicle to provide functionality for the vehicle (e.g., related to safety features, driver information, and implementation of comfort features such as intelligent headlight control and traffic sign information).

[0017] In the present disclosure, a "feature" of an element of a "real element dataset" may refer to a portion of the element that is identifiable in the element (e.g., through corresponding data processing) and that can be characterized by a particular property or assigned to a particular category (e.g., by an image classifier, such as a computer-based machine learning system). If the element is an image, a feature may refer to a spatially confined portion of the scene shown in the image and / or to a plurality of pixels corresponding to the corresponding portion of the image. In the example of a vehicle-integrated sensor system, a "feature" may be an object in the image that is related to a road scene captured (in other words, recorded) thereon (e.g., in real time while the vehicle is traveling). A non-exhaustive list of such objects (i.e., features) may include traffic signs, lane or other road or pedestrian area markings, trees, buildings, and traffic participants such as pedestrians, cyclists, or other vehicles. Within the scope of the present disclosure, it is also contemplated that a "feature" may be a portion of an object: in certain cases where a vehicle is identified in an image, "vehicle color" and "vehicle tires" are examples of such features. Furthermore, "features" such as "weather" or "time of day" may also be extracted from an image. An "element" containing a (measured) data sequence (e.g., a time series) from a sensor may also contain corresponding "features." For example, the change in steering wheel angle over time during driving or the change in injection timing over a pressure profile in a vehicle engine are two non-exhaustive examples of such features. In some cases, features can be extracted (e.g., calculated) from the data sequence of an element: thus, for example, the minimum and / or maximum value of the time series or the maximum and / or minimum derivative (i.e., slope) of the time series or any combination thereof can be calculated and assigned to a corresponding feature.

[0018] Accordingly, the term "synthetic element dataset" can refer to a collection of synthetic elements (or, in some cases, a single synthetic element) that, unlike the elements of a "real element dataset," are not associated with real observations but rather have an artificial origin. For example, a "synthetic element" is not a captured image and, therefore, does not require data processing, such as image processing. A synthetic element can be generated based on one or more real elements (e.g., a captured image) that are altered (in other words, perturbed) according to their properties to generate the synthetic element. The synthetic element may be similar to the original real elements (or elements) from which it was generated after the alteration of these original real elements (or elements), in that it also contains one or more "features" present in these original real elements (or elements). In the example of an image captured (e.g., by a camera), such alterations can include one or more of the following non-exhaustive list: i) pixel-level alterations to model sensor noise; ii) optical flow-based alterations to model pixel-level object motion; iii) alterations based on an object motion model; iv) alterations using a synthetic rendered image; or other alterations.

[0019] Within the scope of the present disclosure, it is contemplated that a "real element" or "synthetic element" is fed as input data into a computer-based machine learning system, which can provide a corresponding "result." The "result" can be the result of the machine learning system's processing of the element, generated by the machine learning system in the form of output data. Furthermore, the "result" can include an "uncertainty value" associated with the generated output data, which can quantify the degree of confidence in the output data (see below for more details). As mentioned above, the "computer-based machine learning system" can be a classifier that can be designed to receive "elements" as input data and classify the "features" of these elements into corresponding categories. In some cases, the classifier can provide (e.g., after identifying these features on the elements) probabilities of classifying the features into different categories. In some examples, the classifier can be an image classifier that performs semantic segmentation on image data on a pixel-by-pixel or region-by-region basis. The image classifier can be designed to receive input data (e.g., "elements") in the form of image data and classify the input data into multiple categories. In some examples, this may include: taking an input vector (R n ) is mapped to the output vector (R) of the second dimension representing the classification result. m) in the form of output data. For example, the components of the input vector may represent image data associated with an image. Each component of the output vector may represent an image classification result, which may be calculated according to the computer-based machine learning system of the present disclosure.

[0020] The term "vehicle" includes any device designed to transport passengers and / or cargo. A vehicle can be a motor vehicle (e.g., an at least partially autonomous / assisted motor vehicle, in particular a passenger car (PKW) or a truck (LKW)). However, the vehicle can also be a ship, train, aircraft, or spacecraft. Other devices are also contemplated within the scope of the present disclosure, as will be discussed further below. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1a is a flow chart illustrating an example of a method for evaluating results of a computer-based machine learning system according to the first aspect. Figure 1b and 1c is a flow chart illustrating other possible method steps according to the first aspect.

[0022] Figure 2a and Figure 2b The figures schematically illustrate examples of real elements 1 and 2 from a real element dataset. These real elements are two detected images at different time points, in which six features are visible: vehicles 10 and 14, pedestrians 11 and 13, and a cyclist 15. These figures also show examples of when portions of these images (the areas within the dashed circles) are considered as real elements 1a and 2a having the feature "vehicle 10."

[0023] Figure 3 An exemplary monitoring module 20 is schematically shown, communicating with: i) module 30, which includes a computer-based machine learning system 70; ii) additional module 40, which also includes the computer-based machine learning system 70; and iii) planner 50. Incoming and outgoing arrows illustrate exemplary communications between modules. 60: Real element dataset; 61 / 62: Output data / uncertainty values for real elements ("first result"); 63: Generated synthetic element dataset; 64 / 65: Output data / uncertainty values for synthetic elements ("second result"). DETAILED DESCRIPTION

[0024] First, based on Figures 1a to 1c to describe techniques for evaluating the results of computer-based machine learning systems. Figure 2a and Figure 2bto discuss exemplary "real elements" of the present technology on which "features" exist. Finally, reference is made to Figure 3 , an exemplary monitoring module 20 for implementing the computer-implemented method of the first aspect is described.

[0025] like Figures 1a to 1c As shown in , a first general aspect relates to a computer-implemented method for evaluating the results of a computer-based machine learning system. As already described above, the computer-based machine learning system can be, for example, an artificial neural network (e.g., CNN). Figures 1a to 1c In the embodiment, the method steps of the corresponding independent claims are shown in boxes drawn by solid lines, while the method steps of some dependent claims are shown in boxes presented by dotted lines.

[0026] The first step of the method comprises receiving 100 a real element dataset 60, wherein each real element 1; 1a; 2; 2a of the real element dataset corresponds to a real observation (in the sense defined above). The real element dataset 60 may be received, for example, by a monitoring module 20, which may be used to evaluate the results of a computer-based machine learning system (see discussion below and Figure 3 , in which a “real element data set 60” is fed to the monitoring module 20). In some cases, the monitoring module 20 may be a vehicle component. As already described above, the real element may comprise a captured image (e.g. a digital image) or a portion thereof, which image is detected, for example, by a sensor such as a camera-based system (e.g. a camera-based system of a vehicle). In this case, the real element data set may comprise an image sequence (or a portion thereof) (e.g., these images may be fed into the monitoring module 20 directly after the images are captured during vehicle travel). In some cases, the “real element data set” may comprise images captured by the camera-based system at predefined points in time (e.g., the “real element data set” comprises consecutive images, which are separated from each other by specified time intervals). In some cases, the real element data set may comprise an image sequence, which is processed by the camera-based system, e.g., by the camera-based system in real time during the image capture. It is also conceivable that the “real element data set” comprises a video recorded by the camera-based system (e.g., the video may first be stored at the time of capture and then transmitted to the monitoring module 20 at a later point in time). In Figure 2a and Figure 2bIn the example of FIG, two real elements 1 and 2 of the "real element dataset" are shown as two images captured at different time points, in which two captured road scenes can be identified. Alternatively, two segments 1a and 2a of these images can be used as two real elements of the "real element dataset" (see below for more details). The real element dataset 60 can include one or more images, five or more images, ten or more images, one hundred or more images, one thousand or more images, or ten thousand or more images.

[0027] Next, the present technology includes receiving 200 (e.g., via the monitoring module 20 described above) a first result 61; 62, the first result being calculated by the computer-based machine learning system 70 using the real element data set. Here, the computer-based machine learning system 70 may be implemented in the module 30 (in other words, the module 30 may include the computer-based machine learning system 70, such as Figure 3 ), or the module 30 may be a machine learning system 70. Figure 3 In the example of FIG, the real element data set 60 (see the previous paragraph for an example of such a data set) supplied to the monitoring module 20 is also fed into the module 30. In some cases, the module 30 may be a vehicle component. In this regard, the computer-based machine learning system 70 may receive images captured by the camera-based system, for example, in real time while the vehicle is traveling, for further processing of these images.

[0028] In the technology of the present disclosure, multiple real elements of the real element dataset can be used as input data for a computer-based machine learning system 70, which generates corresponding output data and generates an uncertainty value for each of the multiple real elements. Returning to the example with images as elements: the multiple real elements of the real element dataset can be multiple images (or fragments thereof) that are captured by a camera-based system at predefined time points (e.g., within a specified time interval). For example, one image out of every "N" captured images may belong to the multiple real elements (e.g., "N" equals one or more, two or more, five or more, ten or more, fifty or more). In addition, the uncertainty value of a real element can represent the confidence level (also see the definition above) of the output data generated by the computer-based machine learning system related to the real element in the multiple real elements. For example, when the generated output data related to the real element is most confident, the uncertainty value of the real element can take a value (e.g., a value of "one") to reflect this, and can take another value (e.g., a value of "zero") when the generated output data is untrustworthy. The uncertainty value of the real element may be between the first value and the second value (e.g., between the values "zero" and "one"); other definitions of the uncertainty value of the real element are also conceivable. In this case, the first results 61; 62 (introduced above) may include multiple output data 61 and multiple uncertainty values 62 for the multiple real elements.

[0029] The next step of the method comprises generating 300 a synthetic element dataset 63 related to the real element dataset. The synthetic element dataset 63 may be generated, for example, by the monitoring module 20. As already explained in detail above, "synthetic elements" have an artificial origin, which distinguishes them from "real elements".

[0030] In the present technology, one or more real elements of a real element dataset may include one or more features 10-15 (see definitions above) that are present in one or more synthetic elements of a synthetic element dataset, the one or more synthetic elements being generated in relation to (in other words, based on) the corresponding one or more real elements of the real element dataset. Figure 2a and Figure 2bIn the example of FIG, two real images at different points in time are shown, in which six features can be seen, namely "vehicles 10, 14", "pedestrians 11, 13", "cyclist 15" (for example in the corresponding processed images). The two images 1; 2 can be two images taken successively (or images whose taking points are relatively close to each other), so that these images have the same features, wherein the corresponding features (for example, vehicles 10; 12 in images 1; 2) are spatially offset from each other. (For example, images 1; 2 were taken using a vehicle-integrated sensor system while the vehicle was driving.) The figures also show examples when parts of these images (areas within the dashed circles) can be regarded as real elements 1a; 2a having features, namely "vehicle 10".

[0031] In one example, generating a synthetic element dataset associated with a real element dataset may include generating 320 synthetic elements of the synthetic element dataset using real elements of the real element dataset. This process can be illustrated using these images as an example as follows: if a real image is an element of the real element dataset, then the image (i.e., the real element in this example) can be used to generate a synthetic image (i.e., the synthetic element in this example). For example, the synthetic image can be generated using one of the methods described above. Image 1 includes six features: vehicles 10, 14, pedestrians 11, 13, and cyclist 15. A synthetic image generated using image 1 may have the same features (thus, the images can be said to be similar in this sense). Another synthetic image generated using image 2 may also have the same six features. In another example, a synthetic image may be generated using portion 1a of image 1. Another synthetic image may be generated using portion 2a of image 2. In this manner, the remaining (e.g., all) synthetic elements of the synthetic element dataset 63 may be generated (i.e., one synthetic element for each corresponding real element).

[0032] In the technology of the present disclosure, the real element dataset may include two or more real elements related to each other, wherein ... Determine the correlation between the two or more real elements. In addition, generating a synthetic element dataset related to the real element dataset may include: generating 340 synthetic elements of the synthetic element dataset using two or more correlated real elements of the real element dataset. In some cases, this overlap between features present on different elements (e.g., images) can be defined by a distance metric that characterizes the spatial offset of the same features at the pixel level on these elements. For example, if the distance metric between features on two considered elements exceeds a predefined value, this may mean that: the overlap is below the predefined overlap value, so that the two elements are classified as unrelated. (If a feature is present on an element and absent on another image, this may also mean that: the overlap is below the predefined overlap value and the two elements are unrelated.) For example, in Figure 2a and Figure 2b Two real elements shown in (i.e. the two images 1; 2) can be declared to be two related elements (in the sense defined above), for example because the overlap of two or more (e.g. all six) of the corresponding (identical) features in the two images exceeds a predefined overlap value.

[0033] In some cases, when spatially confined parts of an image are considered as elements, a predefined value of the distance metric can be used to select corresponding spatially confined parts on different images. In this case, two elements can have the same characteristics, such as Figure 2a and Figure 2b As shown in : From these images, the feature "vehicle 10" can be extracted on the corresponding selected parts 1a; 2a (ie the elements in this example). Figure 2a and Figure 2b In the example of , these elements 1a; 2a can be classified as correlated (in the sense defined above) because these elements contain the following feature, the overlap value of which on the two images exceeds the overlap value. In other cases, the spatially restricted parts of the images (as elements) can have more than one feature. In this way, other (e.g. all) synthetic elements of the synthetic element dataset 63 can be generated (i.e. there is one synthetic element for two or more corresponding real elements). Here, the correlated two or more real elements of the real element dataset can be a sequence of two or more images, which are detected by a sensor system (e.g. a vehicle integrated sensor system) at different time points (see also the discussion above). In some cases, the synthetic elements generated using correlated real elements can simulate future real elements that have not yet been detected.

[0034] Within the scope of the present disclosure, it is conceivable that the "degree of overlap" of one or more features is determined by the module 30. Furthermore, information about these related elements can be forwarded to the monitoring module 20 (in some cases, the "first result" can contain this information). The monitoring module 20 can use this information, for example, to generate one or more synthetic elements with the help of the related real elements (as already described above).

[0035] The composite element dataset 63 may include composite elements generated by the above two techniques. The composite element dataset 63 may include one or more images, five or more images, ten or more images, one hundred or more images, one thousand or more images, or ten thousand or more images.

[0036] The next step of the method comprises sending 400 the synthetic element dataset. The synthetic element dataset 63 may be sent, for example, by the monitoring module 20 to the additional module 40. The additional module 40 may comprise a computer-based machine learning system 70, i.e., the same computer-based machine learning system as may be part of the module 30, see Figure 3 In another case, the synthetic element dataset 63 may be sent, for example, by the monitoring module 20 to a module 30 that includes a computer-based machine learning system 70 (this case is not shown in these figures).

[0037] Next, the present technique includes receiving 500 a second result 64; 65 calculated by the computer-based machine learning system using the synthetic element dataset. The second result 64; 65 may be received, for example, by the monitoring module 20, such as Figure 3 As shown in .

[0038] In the disclosed techniques, multiple synthetic elements of the synthetic element dataset can be used as input data to a computer-based machine learning system 70, which generates corresponding output data and generates an uncertainty value for each synthetic element in the multiple synthetic elements. Furthermore, the uncertainty value for a synthetic element can represent the confidence level of the output data generated by the computer-based machine learning system associated with the synthetic element in the multiple synthetic elements. For example, the uncertainty value for the synthetic element can take a first value (e.g., a value of "one") to reflect when the generated output data associated with the synthetic element is most confident, and can take a second value (e.g., a value of "zero") when the generated output data is untrustworthy. The uncertainty value for the synthetic element can be between the first and second values (e.g., between the values of "zero" and "one"); other definitions of the uncertainty value for the synthetic element are also contemplated. In this case, the second result 64, 65 can include the multiple output data 64 and / or the multiple uncertainty values 65 for the multiple synthetic elements. The multiple synthetic elements can be generated using one or both of the methods described in detail above.

[0039] The next step of the method comprises comparing 600 the first result with the second result (for example, the method step "comparing" can be performed by the monitoring module 20). Finally, the present technology comprises determining 700 one or more validity values 66 based on the first result, said validity values characterizing the confidence level of the calculated first result. Alternatively or additionally, the present technology comprises determining 700 one or more validity values 66 based on the comparison of the first result with the second result, said validity values characterizing the confidence level of the calculated first result. The method step "determining" can be performed by the monitoring module 20. The validity value ("validity bit" in English) can, for example, have two values: one value can classify the first result as credible (in other words, the validity value can be set to "true"), while another value can classify the first result as untrustworthy (in other words, the validity value can be set to "false").

[0040] The method step of "determining" one or more validity values may include setting 710 a first validity value to a value that classifies the first result as untrustworthy (e.g., to "false") if a deviation between an uncertainty value corresponding to a true element observed at a first time point and an uncertainty value corresponding to a true element observed at a second time point satisfies a predetermined criterion, wherein the first time point is a subsequent time point relative to the second time point. Otherwise, if the deviation does not satisfy the predetermined criterion, the method step of "determining" one or more validity values may include setting 720 the first validity value to a value that classifies the first result as trustworthy (e.g., to "true").

[0041] In another case, the method step of "determining" one or more validity values may include: if the deviation between the uncertainty value corresponding to the real element observed at the first time point and the characteristic uncertainty value meets a predetermined criterion, setting 730 the first validity value to a value that classifies the first result as unreliable (for example, to "false"). In addition, the real element data set may include two or more real elements. Here, the characteristic uncertainty value may be a weighted sum of two or more uncertainty values of two or more real elements observed at corresponding time points (for example, an average of the two or more real elements), which are previous time points relative to the first time point. Otherwise, if the deviation does not meet the predetermined criterion, the method step of "determining" one or more validity values may include: setting 740 the first validity value to a value that classifies the first result as reliable (for example, to "true").

[0042] The predetermined criterion can (for both variants of the method step “determining” described above) include that the deviation exceeds a predefined threshold value.

[0043] In yet another embodiment, the method step of "determining" one or more validity values may include setting 750 a first validity value to a value that classifies the first result as untrustworthy (e.g., to "false") if a derivative of the uncertainty value function at a certain point in time satisfies a predetermined continuity criterion. Here, the uncertainty value function may describe the uncertainty value as a function of time, which is determined based on a plurality of uncertainty values of the plurality of real elements observed at different points in time. Otherwise, if the derivative does not meet a predetermined continuity criterion, the method step of "determining" one or more validity values may include setting 760 a first validity value to a value that classifies the first result as trustworthy (e.g., to "true"). The predetermined continuity criterion may include that the derivative exceeds a predefined threshold.

[0044] In the technology of the present disclosure, the comparison of the first result with the second result may include: comparing 620 one or more uncertainty values of the multiple uncertainty values of one or more real elements in the multiple real elements with one or more uncertainty values of the multiple uncertainty values of one or more synthetic elements in the multiple synthetic elements, the one or more synthetic elements being generated in relation to the corresponding one or more real elements (as described above). Alternatively or additionally, the comparison of the first result with the second result may include: comparing 640 output data for one or more real elements in the multiple real elements with output data for one or more synthetic elements in the multiple synthetic elements, the one or more synthetic elements being generated in relation to the corresponding one or more real elements.

[0045] For example, the comparison of one or more uncertainty values of the real elements with the corresponding uncertainty values of the synthetic elements may include calculating one or more deviations between the one or more uncertainty values of the real elements and the corresponding uncertainty values of the synthetic elements. Alternatively or additionally, the comparison of the output data for the real elements with the corresponding output data for the synthetic elements may include determining one or more differences between the output data for the one or more real elements and the corresponding output data for the synthetic elements (e.g., differences related to the classification results of the features on the synthetic elements and the corresponding real elements, as already mentioned above).

[0046] Furthermore, determining the one or more validity values may include setting 770 a second validity value to a value that classifies the first result as untrustworthy (e.g., to "false") if one of the calculated one or more deviations between the uncertainty values of the real element and the corresponding synthetic element meets a predetermined deviation criterion. Otherwise, if the deviation does not meet the predetermined criterion, determining the one or more validity values may include setting 780 the second validity value to a value that classifies the first result as trustworthy (e.g., to "true"). The predetermined deviation criterion may include the deviation exceeding a predefined threshold.

[0047] Alternatively or additionally (alternatively or additionally to the setting of the second validity value), determining the one or more validity values may include setting 790 a third validity value to a value that classifies the first result as untrustworthy (e.g., to "false") if the determined number of differences between the output data for the real element and the corresponding synthetic element meets a predetermined difference criterion. Otherwise, if the determined number of differences does not meet the predetermined difference criterion, determining the one or more validity values may include setting 795 the third validity value to a value that classifies the first result as trustworthy (e.g., to "true"). The predetermined difference criterion may include the determined number of differences exceeding a predefined threshold.

[0048] The method of the first aspect may further include sending 800 one or more validity values (e.g., one or more of the first, second, and third validity values defined above). For example, the monitoring module 20 may send the one or more validity values (e.g., sending the one or more validity values to the planner 50, as described below).

[0049] In the technology of the present disclosure, the computer-based machine learning system can be designed for image processing (as explained above), wherein the real element dataset includes image data. In addition, the computer-based machine learning module can be an image classifier (e.g., an image classifier that performs semantic segmentation on image data pixel by pixel or region by region, also see the definition above). The image classifier is designed to: receive input data in the form of image data (e.g., real and / or synthetic elements, as described in detail above); and classify these input data into multiple categories (e.g., classify the "features" of these elements into corresponding categories, which can then form the generated output data defined above).

[0050] In some examples, the image classification may include semantic segmentation of the image (e.g., region-by-region and / or pixel-by-pixel classification of the image). The image classification may, for example, be object classification. For example, the presence of one or more objects (in other words, "features" in the sense defined above) may be detected in the image data (e.g., traffic participants such as pedestrians, cyclists, or other vehicles in the context of autonomous driving, or traffic signs or lanes in the context of assisted driving). In this case, the computer-based machine learning system may be integrated into the vehicle system (e.g., into an assisted or autonomous driving system) to provide functionality for the vehicle.

[0051] In other examples, computer-based machine learning systems can be used to monitor tasks (e.g., manufacturing processes and / or for quality assurance). For example, a computer-based machine learning system can be designed and used to monitor the operating state and / or environment of an at least partially autonomous robot. In some examples, the at least partially autonomous robot can be an industrial robot. In other examples, a computer-based machine learning system can be designed or used to monitor the operating state and / or environment of a machine (e.g., a machine tool) or a group of machines (e.g., an industrial facility). In these examples, the input data can include state data of the at least partially autonomous robot, the machine or group of machines, and / or its environment, and the output data can include information about the operating state and / or environment of the corresponding equipment.

[0052] In the present disclosure, equipment (eg, vehicles, robots, or industrial facilities) may be monitored and / or controlled based on the classification results.

[0053] In other examples, a computer-based machine learning system may be designed or used to control (or regulate) a device. The device may in turn be one of the devices discussed above (e.g., a vehicle, an at least partially autonomous robot, or a machine). In these examples, the input data (as "real elements" or "synthetic elements") may include state data of the device regarding an internal state of the device (e.g., at least in part, sensor data). Additionally or alternatively, the input data may include state data about the environment of the device (e.g., at least in part, sensor data). The output data of the computer-based machine learning system may characterize the operating state or other internal state of the device (e.g., whether there is a fault, an anomaly, or a critical operating state).

[0054] A second general aspect of the present disclosure relates to a computer program designed to execute the method according to the first general aspect of the present disclosure. The present disclosure also relates to a computer-readable medium (e.g., a machine-readable storage medium, such as an optical storage medium or a read-only memory, such as a flash memory) and signals storing or encoding the computer program of the present disclosure.

[0055] A third general aspect of the present disclosure relates to a monitoring module 20, which is designed to perform a computer-implemented method according to the first general aspect and / or to perform a computer program according to the second general aspect. The monitoring module 20 of the third aspect can be designed to receive a real element dataset 60. The monitoring module 20 can also be designed to receive a first result 61; 62. The monitoring module 20 of the third aspect can also be designed to generate a synthetic element dataset 63. The monitoring module 20 can also be designed to: send the synthetic element dataset 63 and receive a second result 64; 65. The monitoring module 20 can also be designed to: compare the first result with the second result; and based on the first result and / or the comparison of the first result with the second result, determine one or more validity values 66. The monitoring module 20 can also be designed to send the one or more validity values.

[0056] The monitoring module 20 of the third aspect can be integrated into an onboard computer, or can be designed as an independent vehicle internal component (the vehicle internal component, for example, communicates with the onboard computer). The monitoring module 20 can be designed in the same manner for the other devices mentioned above. The monitoring module 20 of the third aspect can communicate with the outside world via a corresponding communication protocol.

[0057] Here, the monitoring module 20 of the third aspect may include one or more processors, controllers, memories, (communication) interfaces, network components, software applications, software architecture, all other software components and / or hardware components, parts of the foregoing, or any combination thereof.

[0058] A fourth general aspect of the present disclosure relates to a device (e.g., a vehicle) comprising a monitoring module 20 according to the third aspect. The device further comprises a module 30 comprising a computer-based machine learning system 70 (see the discussion above for various embodiments of the computer-based machine learning system), wherein the module is configured to receive a real element dataset, which is received by the monitoring module 20 of the third aspect. The module 30 is further configured to calculate a first result using the real element dataset by the computer-based machine learning system and transmit the first result to the monitoring module 20. The device can communicate with the outside world via a corresponding communication protocol.

[0059] Here, the module 30 may include one or more processors, controllers, memories, (communication) interfaces, network components, software applications, software architecture, all other software components and / or hardware components, parts of the foregoing, or any combination thereof.

[0060] The device of the fourth aspect may further include a planner 50, which may be configured to receive one or more validity values from the monitoring module 20 according to the third aspect. The planner 50 may also be configured to control the operation of the device (e.g., the operation of a vehicle or other device discussed above) based on the received validity values.

[0061] In the example of an autonomous or partially autonomous vehicle, if the first validity value classifies the first result as untrustworthy (e.g., when the first validity value is set to "false"), the planner 50 may send an instruction (e.g., to an onboard computer or to a control unit of the vehicle) that triggers a hard maneuver of the vehicle (the hard maneuver may, for example, include emergency braking, stopping the vehicle at a nearby location, turning off the autonomous driving mode, or any combination thereof). On the other hand, if the second and / or third validity value classifies the first result as untrustworthy (e.g., when the second and / or third validity value is set to "false") and the first validity value classifies the first result as trustworthy (e.g., when the first validity value is set to "true"), the planner 50 may send an instruction (e.g., to an onboard computer or to a control unit of the vehicle) that triggers a soft maneuver of the vehicle (the soft maneuver may, for example, include reducing the vehicle speed).

[0062] Here, the planner 50 may include one or more processors, controllers, memories, (communication) interfaces, network components, software applications, software architecture, all other software components and / or hardware components, portions of the foregoing, or any combination thereof.

[0063] A fifth general aspect of the present disclosure relates to a distributed system (e.g., in the form of a cloud system) comprising a device according to the fourth aspect and an additional module 40. The additional module 40 can be designed as an external component of the device (e.g., an external component of a vehicle) (the external component of the device, for example, communicates with an on-board computer). The additional module 40 includes a computer-based machine learning system 70. Here, the computer-based machine learning system of the additional module 40 can be the same computer-based machine learning system as the computer-based machine learning system of the module 30 belonging to the device (e.g., the two machine learning systems can include models described by the same parameters). The additional module 40 can be designed to receive a synthetic element dataset sent by the monitoring module 20 according to the third aspect. The additional module 40 can also be designed to: generate a second result by the computer-based machine learning system using the synthetic element dataset; and send the second result to the monitoring module 20.

[0064] Here, the additional module 40 may include one or more processors, controllers, memories, (communication) interfaces, network components, software applications, software architectures, all other software components and / or hardware components, parts of the aforementioned, or any combination thereof. The additional module 40 may communicate with the outside world via a corresponding communication protocol.

Claims

1. A computer-implemented method for evaluating the results of a computer-based machine learning system, wherein: The method comprises the following steps: Receiving (100) a real element data set (60), wherein each real element (1; 1a; 2; 2a) of the real element data set corresponds to a real observation; receiving (200) a first result (61; 62), the first result being calculated by the computer-based machine learning system (70) using the real element data set; generating (300) a synthetic elemental dataset (63) related to the real elemental dataset; sending (400) the synthetic element dataset; receiving (500) a second result (64; 65), the second result being computed by the computer-based machine learning system using the synthetic element dataset; comparing the first result with the second result (600); Based on the first result and / or a comparison of the first result with the second result, one or more validity values (66) are determined (700), the one or more validity values characterizing a confidence level of the calculated first result.

2. The computer-implemented method of claim 1 , wherein: One or more real elements of the real element dataset include one or more features (10-15), and the one or more features are present in one or more synthetic elements of the synthetic element dataset, and the one or more synthetic elements are generated in relation to the corresponding one or more real elements of the real element dataset.

3. The computer-implemented method of claim 2, wherein: Generating the synthetic element dataset related to the real element dataset includes generating (320) synthetic elements of the synthetic element dataset using real elements of the real element dataset.

4. The computer-implemented method according to any one of claims 1 to 3, wherein: The real element dataset includes two or more real elements that are related to each other, wherein the correlation between the two or more real elements is determined based on an overlap of one or more features present on the two or more real elements, and Wherein, generating the synthetic element dataset related to the real element dataset includes: generating (340) synthetic elements of the synthetic element dataset using two or more related real elements of the real element dataset.

5. The computer-implemented method according to any one of claims 1 to 4, wherein: using a plurality of real elements of the real element dataset as input data to the computer-based machine learning system (70), the computer-based machine learning system generating corresponding output data and generating an uncertainty value for each real element of the plurality of real elements, wherein the uncertainty value of the real element represents a confidence level of output data generated by the computer-based machine learning system in relation to the real element of the plurality of real elements, The first result (61; 62) includes a plurality of output data (61) and a plurality of uncertainty values (62) for the plurality of real elements.

6. The computer-implemented method of claim 5, wherein: Determining the one or more validity values includes setting (710) a first validity value to a value that classifies the first result as untrustworthy if a deviation between an uncertainty value corresponding to a true element observed at a first time point and an uncertainty value corresponding to a true element observed at a second time point satisfies a predetermined criterion, wherein the first time point is a subsequent time point relative to the second time point, and otherwise, if the deviation does not meet the predetermined criterion, setting (720) the first validity value to a value that classifies the first result as credible; or Determining the one or more validity values includes setting (730) a first validity value to a value that classifies the first result as untrustworthy if a deviation between an uncertainty value corresponding to a real element observed at a first point in time and a characteristic uncertainty value satisfies a predetermined criterion, Wherein, the real element data set includes two or more real elements, wherein the characteristic uncertainty value is a weighted sum of two or more uncertainty values of two or more real elements observed at corresponding time points, the time points being previous time points relative to the first time point, And otherwise, if the deviation does not meet the predetermined criterion, setting (740) the first validity value to a value that classifies the first result as credible, Optionally, the predetermined criterion includes: the deviation exceeds a predefined threshold.

7. The computer-implemented method of claim 5, wherein: Determining the one or more validity values includes setting (750) a first validity value to a value that classifies the first result as untrustworthy if a derivative of the uncertainty value function at a point in time satisfies a predetermined continuity criterion, The uncertainty value function describes the uncertainty value as a function of time, and the function is determined according to multiple uncertainty values of the multiple real elements observed at different time points. and otherwise, if the derivative does not satisfy the predetermined continuity criterion, setting (760) the first validity value to a value that classifies the first result as credible, Optionally, the predetermined continuity criterion includes: the derivative exceeds a predefined threshold.

8. The computer-implemented method according to any one of claims 5 to 7, wherein: using a plurality of synthetic elements of the synthetic element dataset as input data to the computer-based machine learning system (70), the computer-based machine learning system generating corresponding output data and generating an uncertainty value for each of the plurality of synthetic elements, wherein the uncertainty value of a synthetic element represents a confidence level of output data generated by the computer-based machine learning system in relation to the synthetic element of the plurality of synthetic elements, The second result (64; 65) includes a plurality of output data (64) and / or a plurality of uncertainty values (65) for the plurality of synthetic elements.

9. The computer-implemented method of claim 8, wherein: The comparison between the first result and the second result includes: comparing (620) one or more of the plurality of uncertainty values for one or more of the plurality of real elements with one or more of the plurality of uncertainty values for one or more of the plurality of synthetic elements, the one or more synthetic elements being generated in relation to the corresponding one or more real elements; and / or Output data for one or more of the plurality of real elements is compared (640) with output data for one or more of the plurality of synthetic elements, the one or more synthetic elements being generated in relation to the corresponding one or more real elements.

10. The computer-implemented method of claim 9, wherein: The comparison of the one or more uncertainty values of the real element with the corresponding uncertainty values of the synthetic element includes calculating one or more deviations between the one or more uncertainty values of the real element and the corresponding uncertainty values of the synthetic element.

11. The computer-implemented method of claim 10, wherein: Determining the one or more validity values includes setting (770) a second validity value to a value that classifies the first result as untrustworthy if one of the calculated one or more deviations between uncertainty values for the real element and the corresponding synthetic element meets a predetermined deviation criterion, And otherwise, if the deviation does not meet the predetermined deviation criterion, setting (780) the second validity value to a value that classifies the first result as credible, Optionally, the predetermined deviation standard includes: the deviation exceeds a predefined threshold.

12. The computer-implemented method according to any one of claims 9 to 11, wherein: The comparison of the output data for the real elements and the corresponding output data for the synthetic elements includes determining one or more differences between the output data for the one or more real elements and the corresponding output data for the synthetic elements.

13. The computer-implemented method of claim 12, wherein: Determining the one or more validity values includes setting (790) a third validity value to a value that classifies the first result as untrustworthy if the determined number of differences between the output data for the real element and the corresponding synthetic element meets a predetermined difference criterion, and otherwise, if the determined number of differences does not meet the predetermined difference criterion, setting (795) the third validity value to a value that classifies the first result as credible, Optionally, the predetermined difference criterion includes: the determined number of differences exceeds a predefined threshold.

14. The computer-implemented method of any one of claims 1 to 13, further comprising: One or more validity values are sent (800).

15. The computer-implemented method of any one of claims 1 to 13, wherein: The computer-based machine learning system is designed for image processing, and wherein the real element dataset comprises image data.

16. The computer-implemented method of claim 15, wherein: The computer-based machine learning module is an image classifier, and optionally, an image classifier that performs semantic segmentation on image data pixel by pixel or region by region. 17 . A computer program designed to execute the method according to claim 1 .

18. A monitoring module (20), designed to execute the computer-implemented method according to any one of the preceding claims 1 to 16 and / or to execute the computer program according to claim 17.

19. An apparatus comprising: The monitoring module (20) according to claim 18; and a module (30) comprising a computer-based machine learning system (70), The module is designed to: receiving a real element data set, the real element data set being received by a monitoring module (20) according to claim 18; calculating, by the computer-based machine learning system, a first result using the real element dataset; and The first result is sent to the monitoring module (20).

20. A distributed system comprising: The apparatus according to claim 19; An additional module (40), comprising the computer-based machine learning system (70), wherein the additional module is configured to: receiving a composite element data set, the composite element data set being sent by a monitoring module (20) according to claim 18; generating, by the computer-based machine learning system, a second result using the synthetic element dataset; The second result is sent to the monitoring module (20).

Images