METHOD FOR QUALITY ASSURANCE OF AN EXAMPLE-BASED SYSTEM

Description

[0001] The invention relates to a method for quality assurance of an example-based system.

[0002] Example-based systems, such as artificial neural networks, are generally well-known. They are typically used in areas where a direct algorithmic solution does not exist or cannot be adequately created using conventional software methods. Example-based systems make it possible to create and train a task based on a set of examples. The learned task can then be applied to a further set of examples.

[0003] The dissertation "Quality-assured efficient development of forward-directed artificial neural networks with supervised learning (QUEEN)" by Thomas Waschulzik describes the development of forward-directed artificial neural networks with supervised learning (hereinafter: WASCHULZIK).

[0004] The publication entitled "Were not in Kansas Anymore: Detecting Domain Changes in Streams", in Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, October 1, 2010, pages 585-595, describes the adaptation of a system in natural language processing in one domain when that system has been trained in another domain.

[0005] The publication entitled "Boosting Few-Shot Image Recognition Via Domain Alignment Prototypical Networks", in IEEE 30TH INTERNATIONAL CONFERENCE ON TOOLS WITH ARTIFICIAL INTELLIGENCE (ICTAI), November 5, 2018, pages 260-264, describes the introduction of a module for domain alignment in so-called few-shot learning.

[0006] Against this background, the purpose of the invention is to improve the quality assurance of an example-based system.

[0007] This problem is solved according to the invention by a method for quality assurance of an example-based system, in which the example-based system is created and trained using collected examples that form an example set. Each example in the example set comprises an input value located in an input space. A first example set, comprising a plurality of examples, and a second example set, also comprising a plurality of examples, are collected. A first quality assessment (or quality indicator), representing the coverage of the input space by examples from the first example set, is determined based on the distribution of the input values ​​in the input space. A second quality assessment (or quality indicator), representing the coverage of the input space by examples from the second example set, is determined based on the distribution of the input values ​​in the input space.The first quality assessment and the second quality assessment are compared.

[0008] The invention is based, on the one hand, on the realization that example-based systems, such as neural networks, are often treated as black boxes. In this approach, the internal information processing is not analyzed, and the generation of an understandable model is omitted. Furthermore, the system is not verified through inspection. This leads to reservations about using example-based systems in highly critical tasks.

[0009] The invention is also based on the realization that when capturing examples for the creation and training of the example-based system, it is often unknown how many examples in which areas of the input space need to be captured in order to create a suitable knowledge base.

[0010] Another key insight of the invention is that examples for example-based systems are often captured in several steps. In this iterative approach, it is desirable to ensure that no application-relevant differences occur between the capture of the first and second sets of examples, or at least that these differences can be detected and evaluated.

[0011] An example of a technical situation where differences between the first and second example sets are relevant is the age-related change in the properties of sensors used in capturing the examples: Due to aging, the calibration of the sensors changes, and it must be checked whether the first example set, which was captured in a first stage of sensor aging, and the second example set, which was captured in a second stage of sensor aging, still describe the same task.

[0012] Another example of differing example sentences occurs when one of the example sentences has been deliberately falsified to harm the creator or user of the knowledge base.

[0013] The solution according to the invention solves these problems by determining the first and second quality assessments and comparing them. It may turn out, for example, that one of the sample data sets must be discarded entirely or partially.

[0014] The respective quality assessment represents the coverage of the input space by examples, which is determined based on the distribution of input values ​​within the input space. This results in a mapping of the input space. The mapping can serve as the basis for further example acquisition to create a suitable knowledge base. Thus, example acquisition can be controlled according to the distribution within the input space, even though the specific type of classifier or approximator has not yet been defined. The number of degrees of freedom with which the knowledge base is trained also does not yet need to be determined. By knowing in which areas further examples need to be acquired, the examples can be captured more selectively, and consequently, the costs of example acquisition (since fewer examples need to be captured overall) can be significantly reduced.

[0015] The invention also recognized that a suitable representation and encoding of the features is a prerequisite for using mappings of the input space for example-based systems. The raw data is transformed into a representation adapted to the solution of the task through application-specific transformations. This representation is then transformed using standard procedures so that it can be used as activity for the input neurons of a neural network (so-called encoding). The quality assessment, which represents the coverage of the input space by examples from the example set, can be implemented at both the representation and encoding levels.

[0016] The invention is further based on the understanding that the encoding and / or representation of the input features in the input space preferably has a semantic relationship with the desired output of the example-based system. Preferably, the mapping of the input space is carried out when, for example, features have been determined through preprocessing that have a semantic relationship to the outputs.

[0017] The invention is further based on the realization that the ratio between the number of independent input features that determine the dimension of the spanned state space and the number of examples to be captured for the configuration, training, evaluation and testing of the system is preferably not too large: because the coverage of the input space by examples is insufficient in the case of a large ratio.

[0018] Furthermore, the invention is based on the understanding that the dimensions spanning the state space are preferably semantically independent of one another (i.e., they represent independent aspects of the problem). More preferably, the dimensions are of equal relevance for solving the problem.

[0019] Preferably, for quality assurance purposes, only a single classification or approximation task is considered. For example, in an artificial neural network used as a Single Shot Multibox Detector (SSD), only the classification for a given object size in a so-called default box (i.e., with a given aspect ratio, scale, and position in the image) is considered.

[0020] Preferably, the example-based system is intended for use in a safety-related function. A person skilled in the art understands a "safety-related function" to be a function of a system that is safety-relevant, i.e., whose behavior influences the safety of the system's environment. Here, the term "safety" is to be understood in the sense of the term "safety." In technical terminology, "safety" refers to the goal of protecting the environment of a system from hazards emanating from the system. In contrast, in technical terminology, the goal of protecting the system from hazards emanating from the system's environment is referred to as "security."

[0021] Another example of an application of the method according to the invention is a situation in which the first example set has undergone quality assurance and the corresponding first quality assessment is known. If the quality assessment for the second example set is corresponding or similar, it can be assumed that the second example set is of satisfactory quality.

[0022] According to a preferred embodiment of the method according to the invention, a third example set is formed from the first and second example sets, and a third quality assessment, which represents the coverage of the input space by the examples of the third example data set, is determined based on the distribution of the input values ​​in the input space. The first quality assessment, the second quality assessment, and the third quality assessment are then compared.

[0023] The third example sentence represents, so to speak, the union of the first and second example sentences.

[0024] An example of the application of the third example sentence is a scenario in which the second example sentence is collected based on knowledge gained from the first example sentence. For instance, the second example sentence was collected to fill in the gaps in the input space. The third example sentence can then be used to determine whether the gaps were successfully filled.

[0025] Determining the quality rating involves distributing representatives within the input space and assigning a number of examples from the example set to each representative. The examples assigned to each representative are located within a region of the input space that surrounds the representative. The first quality rating is a local assessment for this region based on the examples from the first set of examples assigned to the representative. The second quality rating is a local assessment for this region based on the examples from the second set of examples assigned to the representative.

[0026] By assigning examples from the example set to representatives, sample data sets are determined within the environment areas that are assigned to the representatives. For each of these sample data sets, the local quality ratings are calculated, i.e., the first quality rating and the second quality rating.

[0027] For example, it may be of interest to further investigate those environmental areas where the relative difference between the first quality assessment and the second quality assessment is comparatively large.

[0028] Dividing the example set into several subset areas offers advantages typically associated with the divide-and-conquer approach from computer science. For instance, a developer of the example-based system can focus on those parts of the input space where certain quality criteria are not met. The quality in these parts can then be checked and improved as needed. This significantly reduces the effort required to evaluate the entire example set.

[0029] Preferably, a representative example is distributed. The distribution is preferably equal. For example, a grid is chosen in the input space to arrange the representative examples. The grid can be defined individually for each dimension of the input space. A criterion for defining the grid, for example with categorical variables, can be a model of the desired properties of the example distribution in the input space, which is based on the requirements of the example-based system. The grid can be hierarchically structured, for example to represent hierarchical codings. When using a grid to arrange the representative examples, one representative example is distributed in each hypercube in the input space of the grid. With a hierarchical grid structure, one representative example is distributed per hierarchy level.

[0030] Alternatively, the representative is the center of a cluster, which is determined using a clustering method. The clustering method is preferably used to determine the position and extent of each cluster within the input space. More preferably, the clustering method is performed taking into account the output values ​​of the examples located in an output space. The clusters can be defined based on requirements for the properties of the example-based system or based on a subset of example data. For example, in the application of the example-based system, a set of examples can be collected in an early phase, selected based on knowledge of how they fulfill the requirements. This distribution of example data is then quality-assured. In a subsequent project phase, further examples with the same distribution can be collected.In this case, each example from the quality-assured sample set represents a representative for the subsequent sample collection phase. This ensures that for each initial example, an additional quality-assured set of examples is collected. The representative's position can be determined, for example, by the cluster center. Alternatively, a hierarchical clustering method can be used, in which one representative is inserted per cluster and per hierarchy level, and each example is assigned to a cluster and consequently to a representative at each hierarchy level. The set of examples available for calculating the quality score is then assigned to the clusters and consequently to the representative using a predefined metric. For an example that cannot be assigned to any cluster, a new cluster with a representative is preferably created.Alternatively, this example, along with other examples that could not be assigned to any cluster, will be recorded separately through a quality assessment.

[0031] Preferably, the examples are not fully assigned to a representative, but only to a predetermined proportion. This can be achieved, for example, by using a clustering algorithm that provides a partial assignment of the examples to the example datasets (e.g., a percentage assignment to several environments, where the sum of the proportions equals 1). When determining the quality ratings based on this partial assignment, the respective example is considered according to its corresponding proportion.

[0032] The quality assessment is determined based on the number of examples assigned to each representative. This is particularly advantageous if the specific examples are not used again later. A benchmark for comparing the first and second quality assessments is the difference between the number of examples assigned to the representative in the first example set and the number assigned to the representative in the second example set.

[0033] Alternatively or additionally, the specific examples or a reference to the examples are stored in the representative (transformation of the example data set into a structure oriented to the topography of the input space). This is advantageous if the specific examples are needed later on.

[0034] The storage space required for processing is preferably reduced by storing representatives only if at least one example exists in the respective environment area. When determining the coverage of the input space, environment areas in which no representative was created are marked as "no example found." Nevertheless, a histogram of the number of examples per representative can be generated, since the number of environment areas in which no example was captured can be determined with minimal effort (sum of expected representatives - created representatives = number of fields without captured examples).

[0035] In the case of preferred further training, the third quality assessment is a local quality assessment for the surrounding area, which is determined based on the examples assigned to the representative in the third example set.

[0036] Preferably, the representatives are assigned to the same positions in the input space when determining the first, second, and / or third quality rating. In other words, the position of the representatives is preferably chosen to be the same for all example sentences. This allows the representatives of the first, second, and / or third example sentence to be assigned to each other.

[0037] According to a further preferred embodiment of the method according to the invention, the quality assessment comprises a statistical mean which is determined on the basis of the set of examples and / or the examples assigned to a respective representative.

[0038] In this way, quality assessments can be defined based on the information assigned to the representatives, for example using methods of descriptive statistics (as described in one of the following textbooks: "Statistics: The Path to Data Analysis" (Springer Textbook) Paperback - September 15, 2016 by Ludwig Fahrmeir (Author), Christian Heumann (Author), Rita Künstler (Author), Iris Pigeot (Author), Gerhard Tutz (Author); "Statistics for Dummies" Paperback - December 4, 2019 by Deborah J. Rumsey (Author), Beate Majetschak (Translator), Reinhard Engel (Translator); "Workbook on Descriptive and Inductive Statistics" (Springer Textbook) Paperback - February 27, 2009 by Helge Toutenburg (Author), Michael Schomaker (Contributor), Malte Wißmann (Contributor), Christian Heumann (Contributor)).

[0039] The determined local quality ratings can be sorted and processed using a histogram. For example, for each environment, the relative difference between the first quality rating (of the first example set) and the second quality rating (of the second example set) is calculated. A histogram is then created based on this relative difference. The histogram display further allows, for instance, the selection of environments assigned to a specific bin by clicking on it. One of these environments can then be selected to display additional information, such as the detailed distribution of the examples.

[0040] According to a further preferred method, a statistical measure, in particular a mean, median, minimum, maximum and / or quantiles of the number of examples assigned to a representative, is determined as the statistical means.

[0041] Following further preferred processing, neighboring environmental areas are identified in the input space, each of whose representatives is assigned a number of examples that meet a predefined quality criterion for quality assessment.

[0042] Preferably, the specified quality criterion is met if the number of examples assigned to a respective representative falls below, exceeds, or lies within a specified quality threshold of the quality assessment.

[0043] When determining whether two regions are adjacent, various neighborhood relationships can be used, such as the von Neumann neighborhood (also called the 4-neighborhood), the Moore neighborhood (also called the 8-neighborhood), or the neighborhood from graph theory. These defined neighborhood relationships must be adapted for higher-dimensional spaces: In three-dimensional space, for example, the 6-neighborhood is considered for cuboids with common faces, the 18-neighborhood for cuboids with common edges, and the 26-neighborhood for cuboids with common vertices. Neighborhood is defined by the number of dimensions in which two lattice points may differ and still be considered adjacent.

[0044] In a preferred training program, a related area within the input space is identified, consisting of adjacent surrounding areas. A number of examples are assigned to each representative of these surrounding areas. This number of examples fulfills a predefined quality criterion.

[0045] Preferably, the specified quality criterion is met if the number of examples assigned to a given representative falls below, exceeds, or lies within a specified quality threshold of the quality assessment. Further preferably, additional examples are recorded in the respective area if the quality assessment determined for that area is less than a specified quality threshold. Alternatively or additionally, examples are removed from a given area if the quality assessment determined for that area is greater than a specified quality threshold.

[0046] When the quality criterion is met by falling below a predefined quality threshold, the location and size of areas in the input space where too few examples have been recorded (so to speak, "holes in the input space") can be determined in a particularly advantageous way. In other words, a particular advantage of this embodiment lies in the identification of sub-areas of the input space where the example values ​​do not provide a sufficient basis for a safety-critical application. This, in turn, has the advantage that corrective action can be taken, for example, by recording further examples (e.g., as the second set of examples described above) or by restricting the knowledge base in the application to the related areas of high quality.

[0047] In particular, identifying areas where too few examples have been recorded has the advantage of enabling proactive countermeasures against attacks using adversarial examples. This is because the probability of success for an attack using an adversarial example is comparatively high in these areas. This probability can be reduced by recording more examples in these areas or by limiting the knowledge base to related areas of high quality.

[0048] Based on the identified areas of correlation, quality assessments can be calculated. For example, the number of representatives in an area of ​​correlation can be determined. Histograms can be generated showing the size or other properties of an area of ​​correlation. Furthermore, statistical measures such as the mean, median, quantiles, or standard deviations of properties of the areas of correlation can be calculated. Additionally, the extent of the areas of correlation in the dimensions of the input space can be determined. The dimensions can be ordered according to the largest extent of the area of ​​correlation.

[0049] Identifying a related area is particularly advantageous in the example described above for applying the third example sentence, where the second example sentence is collected based on knowledge gained from the first example sentence. In this example, the second example sentence was collected to fill the gaps in the input space. These gaps can be identified and characterized using the related areas described above. The third example sentence can then be used to determine whether the gaps have been successfully filled.

[0050] According to a particularly preferred embodiment of the method according to the invention, each example comprises an output value located in an output space. A local complexity rating is determined for the respective environment, representing the complexity of a task of the example-based system as defined by the examples in the environment. The local complexity rating is determined by the relative positions of the examples in the environment to each other in the input and output spaces.

[0051] The person skilled in the art understands the phrase "relative position of the examples of the environment to each other in the input space and output space" preferably to mean that the complexity assessment is defined based on considering the similarity of the distances of the examples in the input space to the distances in the output space. For example, the task of the example-based system has a comparatively low complexity if the distances in the input space (apart from scaling) correspond approximately to the distances in the output space.

[0052] This offers the advantage of effectively capturing examples. Based on the complexity assessment, areas are identified where, due to the high complexity of the task in the example-based system, a comparatively large number of examples must be captured. Preferably, in areas of the input space with higher complexity, the density of representatives is dynamically increased until a homogeneous complexity is reached and a sufficient number of examples lie in the vicinity of the representatives.

[0053] The complexity assessment corresponds, for example, to the quality indicators described in Section 4 (QUEEN Quality Indicators) of WASCHULZIK. These quality indicators can be defined and applied for both representing and coding the characteristics (see Section 4.5 of WASCHULZIK).

[0054] In a particularly preferred embodiment of the method according to the invention, a first complexity assessment is determined for the examples of the first example set, a second complexity assessment for the examples of the second example set, and a third quality assessment for the examples of the third example set. The third local complexity assessment is compared with the first and / or second local complexity assessment.

[0055] Particular attention should be paid to areas where the third complexity rating indicates a significantly higher complexity of the third example set than the first and / or second complexity ratings. In these areas, there is either a problem with the sample collection process, or the first, second, or both example datasets have been falsified to harm the creator or user of the knowledge base. This application example is particularly important because the high complexity cannot be detected based on the first and second example sets alone (but only on the third example set).

[0056] For the scenario described above, one possible measure is to remove examples from one of the two example sets and, based on the recalculated complexity rating, determine whether an "improvement" (i.e., reduced complexity) has been achieved. If an improvement is observed, further examples can be removed. If the third complexity rating shows a drift toward lower complexity, this approach can be continued.

[0057] To handle the respective differences between the first, second, and third complexity assessments, an acceptance range or measure can be defined. In other words, if the agreement lies within a certain range, the collected examples are accepted. The acceptance measure can be chosen depending on the size of the dataset. For comparatively large datasets, the acceptance range is preferably chosen to be narrower, and for comparatively small datasets, it is chosen to be wider.

[0058] According to a further preferred embodiment of the method according to the invention, a complexity distribution is determined by means of a histogram representation of the complexity assessment.

[0059] Preferably, the range of values ​​for the complexity ratings is binned (i.e., divided into ranges) for the histogram representation.

[0060] In a preferred further development, the complexity distribution over k nearest neighbors of an example in the input space is determined. This reveals how the complexity is distributed in the local environment of an example. Specifically, the complexity characteristics in the local environment of the example are determined, essentially creating a complexity fingerprint of the local environment.

[0061] For example, the "binned" values ​​are plotted on the y-axis and the representation of the increasing k (the k-nearest neighbors) is plotted on the x-axis.

[0062] To reduce the computational resources required to determine the complexity distribution, the step size for values ​​of k > 1 is chosen. For example, with a step size of 5, a distribution of the complexity rating is determined for values ​​of k = 5, 10, 15, 20, etc. Furthermore, preferably, the step size of k is chosen to be small only in areas of particular interest. Thus, the distribution of the complexity rating is initially calculated with a comparatively large step size of k, and then, in an area of ​​particular interest, with a small step size of k.

[0063] Furthermore, preferably for the calculated histogram field (Complexity assessment binnt, k) The number of values ​​for the complexity rating is stored. Furthermore, preferably, identification information (e.g., a number) is also stored, indicating the example in whose environment the complexity distribution was determined.

[0064] According to a preferred embodiment of the method according to the invention, the integrated quality indicator QI 2< according to section 4.6 of WASCHULZIK is used as a quality indicator for the representations, which can be defined on the basis of formula 4.21 as follows: QI 2 P = 1 P 2 ∑ x i ∈ P 2 d NRE x i − d NRA x i 2 where, according to formula 4.18 by WASCHULZIK: d NRE x = d RE x ∑ y ∈ P 2 d RE y P 2 the normalized distance of the represented inputs (NRE) and d NRA x = d RA x ∑ y ∈ P 2 d RA y P 2 The normalized distance of the represented expenditures (NRA) is used. This is... x the couple ( x 1 , x 2 ,) consisting of the two examples x 1 and x 2. x 1 and x 2 are examples from the example set P. P = { p 1 , p 1 , ... , p | P |} is the set of elements of BAG P, where | P 2< | the number of elements of the BAG P is. At Federal OfficeThis is a multiset (also called a bag) as defined in specification 21.5 on page 27 of the WASCHULZIK appendix. The task QAG is defined in Definition 3.1 on page 23 by WASCHULZIK and is referred to there as the QUEEN task.

[0065] d RE ( x ) is an abbreviation for the distance in the input space d re ( vep x 1 , vep x 2) and d RA ( x ) is an abbreviation for the distance in the output space d ra ( vap x 1 , vap x 2).

[0066] The definition of the distance between the representations of two examples according to WASCHULZIK is based on the Euclidean norm. Thus, the distance in the input space is defined as (see formula 4.3 of WASCHULZIK): d re p k 1 p k 2 = ∑ i = 1 aem vemp i , k 1 − vemp i , k 2 2 with p k 1 , p k 2 as examples from the setP , where p k = vep k vap k = vemp 1 , k , vemp 2 , k , … , vemp AnzahlEingabeMerkmale , k , vamp 1 , k , vamp 2 , k , … , vemp AnzahlAusgabeMerkmale , k with i Running index across all variants; vemp i,kx Characteristic of the input feature i of the example kx with kx ∈ R (R is the set of real numbers); and aem Number of input features the task QAG.

[0067] The person skilled in the art understands the formulation "on the basis of the following definition" or "on the basis of formula 4.21 can be defined as follows" preferably to mean that modifications and a function of the quality indicator F(QI 2< ) are also included in the idea of ​​this definition.

[0068] For preferred further training, an aggregated complexity assessment is determined by aggregating the local complexity assessments.

[0069] The aggregated complexity assessment has the advantage that a developer of the example-based system can easily perform their quality assurance.

[0070] For example, an aggregated complexity assessment is generated as a histogram of the complexity in the different environments of the input space. For this purpose, the range of values ​​for the complexity assessments is binned (i.e., divided into bins). Preferably, only the number of environments with the corresponding complexity is included in the bins if the positions of the environments are no longer needed. Preferably, this histogram is combined with information about the number of examples, for example, also in a histogram showing the number of examples assigned to the representative. Furthermore, preferably, information about the representatives is stored in the histogram so that this information can be accessed during detailed analyses.

[0071] According to another preferred further development approach, the aggregated complexity assessment identifies environmental areas whose complexity rating falls below a predefined complexity threshold. Within these identified environmental areas, the task of the example-based system is implemented using an algorithmic solution. This is particularly advantageous for applications with high quality requirements, such as safety-related functions.

[0072] This preferred further training approach is based on the understanding that the exact functioning of the system (i.e., semantic relationships) is often known for areas with low task complexity. In this case, the task can be implemented as a conventional algorithm (rather than as an example-based system). This is particularly advantageous because demonstrating sufficient safety of the safety-related function within an approval process is generally easier for the simple algorithmic solution.

[0073] This further training also has the advantage that no further examples need to be recorded in areas of low complexity.

[0074] Preferably, when searching for simple areas, data collection artifacts are also sought that reveal a relationship between input and output, which arises from specific circumstances of data collection but does not represent a practically applicable relationship (such as the so-called Kluger-Hans effect: https: / / de.wikipedia.org / wiki / Kluger_Hans). In areas of particularly high complexity, the examples are analyzed to determine whether, for example, problems occurred during the collection and recording of the examples.

[0075] According to a further preferred embodiment of the method according to the invention, the input space is hierarchically divided based on the quality assessment.

[0076] Preferably, a hierarchical mapping of the input space is achieved through its hierarchical division. The hierarchy is further preferably derived from the representation or encoding of the input feature and / or from the analysis of the complexity of the task.

[0077] By introducing an additional hierarchy in the input space analysis, the density of representatives can be dynamically increased (until homogeneous complexity is reached) or a new hierarchy level can be introduced in areas of high complexity. A new hierarchy level is introduced by adding a new subdivision with a higher resolution to the representative's domain. This process can be iterated by adding another hierarchy level to the high-resolution domain as local complexity increases again. This allows the resolution to be dynamically adapted to the specific task.

[0078] According to a further preferred embodiment of the method according to the invention, the example-based system is provided for use in a safety-related function, wherein the safety-related function comprises object recognition based on image recognition, in which the object is recognized using the example-based system.

[0079] In a preferred further training, object recognition is used in the automated operation of a vehicle, in particular a track-bound vehicle, a motor vehicle, an aircraft, a watercraft and / or a spacecraft.

[0080] Object detection in automated vehicle operation is a particularly useful implementation of a safety-related function. Object detection is necessary, for example, to identify obstacles in the roadway or to analyze traffic situations with regard to the right-of-way of other road users.

[0081] A motor vehicle is, for example, a motor vehicle, e.g. a passenger car (PKW), a truck (LKW) or a tracked vehicle.

[0082] The watercraft is, for example, a ship or submarine.

[0083] The vehicle can be manned or unmanned.

[0084] One example of an application area is the autonomous or automated operation of a rail vehicle. Object recognition systems are used to solve the tasks involved, analyzing scenes digitized by sensors. This scene analysis is necessary, for example, to detect obstacles on the track or to analyze traffic situations with regard to right-of-way. Currently, systems based on the use of examples to train the parameters of the pattern recognition system are particularly successful for object recognition. Examples include neural networks, such as those using deep learning algorithms.

[0085] According to a further preferred embodiment of the method according to the invention, the example-based system is intended for use in a safety-related function, wherein the safety-related function comprises a classification based on sensor data from organisms.

[0086] Tissue classification of animal or human tissue is a particularly useful implementation of a safety-related function in the field of medical image processing. The organisms include, for example, Archaea (ancient bacteria), Bacteria (true bacteria), and Eukarya (nucleated organisms), or tissues of Protista (also Protoctista, founder organisms), Plantae (plants), Fungi (fungi, chitinous fungi), and Animalia (animals).

[0087] Other areas of application include the safe control of industrial plants (e.g. synthesis in chemistry, the control of production processes, e.g. rolling mills), a classification of chemical substances (e.g. environmental toxins, warfare agents), a classification of vehicle signatures (e.g. radar or ultrasonic signatures) and / or control in the field of industrial automation (e.g. production of machines).

[0088] According to a further preferred embodiment of the method according to the invention, the example-based system comprises a supervised learning system, a system built using statistical methods, preferably an artificial neural network with one or more layers of neurons that are neither input nor output neurons and are trained using backpropagation, in particular a convolutional neural network, especially a single-shot multibox detector network.

[0089] The use of artificial neural networks often enables an improvement in classification or approximation performance.

[0090] The one or more layers of neurons that are neither input nor output neurons are often referred to as hidden neurons. Training neural networks with many layers of hidden neurons is also commonly known as deep learning. A specific type of deep learning network for pattern recognition is the convolutional neural network (CNN). A special case of CNNs is the so-called SSD network (Single Shot Multibox Detector network). The term "Single Shot Multibox Detector" refers to a deep learning-based object detection method based on a convolutional neural network, as described in: Liu, Wei (October 2016). SSD: Single shot multibox detector. European Conference on Computer Vision. Lecture Notes in Computer Science. 9905. pp. 21-37. arXiv:1512.02325

[0091] The invention further relates to a computer program comprising instructions which, when the program is executed by a computing unit, cause it to carry out the method of the type described above.

[0092] The invention further relates to a computer-readable storage medium comprising instructions which, when executed by a computing unit, cause it to carry out the method of the type described above.

[0093] For advantages, embodiments and details of the features of the computer program and computer-readable storage medium according to the invention, reference can be made to the above description of the corresponding features of the method according to the invention.

[0094] An embodiment of the invention is explained with reference to the drawings. The drawings show: Figure 1 schematically shows the process of an embodiment of a method according to the invention, Figure 2 schematically shows the structure of an example-based system with unsupervised learning, Figure 3 schematically shows the structure of an example-based system with supervised learning according to the embodiment of the method according to the invention, Figure 4 schematically shows a two-dimensional input space according to the embodiment of the method according to the invention, Figure 5 schematically shows the in Figure 4Figure 6 shows a two-dimensional input space in a second state, Figure 6 a schematic side view of a track-bound vehicle located on a driving track, Figure 7 schematically another example of a two-dimensional input space according to a further embodiment of the method according to the invention, Figure 8 two axis diagrams which represent the application of the complexity assessment to a first synthetic function, Figure 9 two axis diagrams which represent the application of the complexity assessment to a second synthetic function, and Figure 10 two axis diagrams which represent the application of the complexity assessment to a third synthetic function.

[0095] Figure 1Figure 1 shows a schematic flowchart representing the process of an exemplary embodiment of a method according to the invention for quality assurance of an example-based system. The method can be applied to example-based systems with supervised and unsupervised learning.

[0096] In supervised learning, the goal is to learn a function that maps data x (as input values) to a label y. An example of supervised learning is classification, where, for instance, image data x is mapped to a class y (e.g., cat). Other examples of supervised learning include regression, object recognition, image labeling, etc.

[0097] In unsupervised learning, the goal is to learn a structure of data x (without using a label y). Clustering is an example of unsupervised learning, where groups within the data are identified that exhibit similarities in a specific metric. Other examples of unsupervised learning include dimensionality reduction and feature learning (also known as representation learning), etc.

[0098] The Figures 2 and 3 show examples of implementation of example-based systems 1. Figure 2 Figure 1 schematically shows the structure of an exemplary implementation of an example-based system 1, which is implemented as an autoencoder. Autoencoders are a type of artificial neural network 2 that can be used for efficient data coding and learn this ability unsupervised. The autoencoder maps the input values ​​x to a feature vector Z.

[0099] Figure 3Figure 1 schematically shows the structure of an exemplary implementation of an example-based system 1 with supervised learning, which is implemented as a multilayer perceptron. Further examples of supervised learning systems include a recurrent neural network, a convolutional neural network, or, in particular, a so-called single-shot multibox detector network.

[0100] The example-based system 1 is formed by an artificial neural network 2, which has a layer 4 of input neurons 5 and a layer 6 of output neurons 7.

[0101] The in Figure 3 The artificial neural network 2 shown has several layers 8 of neurons 9 that are not input neurons 5 or output neurons 7.

[0102] The example-based system and the method according to the invention are implemented by means of one or more computer programs. The computer program comprises instructions which, when the program is executed by a computing unit, cause it to execute the method according to the invention as described in the [document / section]. Figure 1 The illustrated embodiment is to be carried out. The computer program is stored on a computer-readable storage medium.

[0103] The example-based system is used in a safety-related function of a system. The behavior of the function therefore influences the safety of the system's environment. An example of a safety-related function is object recognition based on image recognition, in which the object is recognized using the example-based system 1 (with supervised learning). Object recognition is used, for example, in the automated operation of a vehicle, especially one in Figure 4track-bound vehicle 40 shown, a motor vehicle, an aircraft, a watercraft or a spacecraft, used.

[0104] Another example of a safety-related function is classification based on sensor data from organisms, e.g., Archaea (ancient bacteria), Bacteria (true bacteria), and Eukarya (nucleated organisms), or from tissues of Protista (also Protoctista, founders), Plantae (plants), Fungi (fungi, chitinous fungi), and Animalia (animals); safe control of industrial plants; classification of chemical substances; classification of vehicle signatures; or control in the field of industrial automation.

[0105] In process step A, the examples to be collected are determined. In step B, the examples are collected: The collected examples form an example set. Each example has an input value 12, which lies in an input space, and an output value 14, which lies in an output space. This applies to object recognition (as one of several possible examples of a safety-related function in supervised learning) for the automated operation of the [system / device]. Figure 6The examples are collected for the track-bound vehicle 40 shown by equipping the track-bound vehicle 40 with a camera unit 42 for capturing images. The camera unit 42 is oriented in the direction of travel 41 such that a spatial area 43 ahead in the direction of travel 41 is captured by the camera unit. The track-bound vehicle 40 travels with the camera unit 42 in the direction of travel 41 along a route 44. To capture the examples, scenes relevant for the creation and training of the example-based system 1 for object recognition are recreated. For example, cardboard cutouts, crash test dummies, or actors 45 are used to represent people on the route 44 who are to be recognized by the example-based system 1 to be created and trained. Alternatively, scenes can be recreated using virtual reality.

[0106] Figure 4Figure 20 shows an example of a two-dimensional input space. In the actual application of the method according to the invention, the input space and output space will often have a higher dimensionality.

[0107] According to the inventive method, in a process step B1 a first example set 21, which comprises a plurality of examples 22, is collected. The examples 22 of the example set 21 are represented as crosshairs 23 in Figure 4 depicted.

[0108] In a process step C1, an initial quality assessment is determined, representing the coverage of the input space by examples from the first example set 21. In a process step C2, during the determination of the quality assessment C1, representatives are distributed within the input space. The representatives 24 are uniformly distributed and are represented as intersection points 25 of the grid 26 shown.

[0109] In process step C3, a number of examples 29 from the example set are assigned to each representative 28. The examples 29 assigned to the representative 28 lie in an environment 30 of the input space 20, which surrounds the respective representative 28. The environment 30 is exemplified in Figure 4 represented as a dotted area. In process step C4, an initial local quality assessment for the surrounding area 30 is determined as part of the quality evaluation. The local quality assessment is also carried out for further areas in Figure 4 The surrounding areas shown were determined.

[0110] In process step C5, adjacent environmental areas 32-36 are determined in the input space, and a number of examples are assigned to each representative of these areas, all of which fall below a predefined quality threshold. Figure 4These surrounding areas 32-36 are represented as surfaces with diagonal stripes. This is in the Figure 4 The example shown in the surrounding areas 32-36 identifies areas where no example is located. Furthermore, in a process step C6, a related area 38 within the input space 20 is determined, consisting of the adjacent surrounding areas 32-36, each of whose representatives is assigned a number of examples that falls below a predefined quality threshold. This determines the location and size of areas within the input space 20 where too few examples have been recorded. In other words, sub-areas of the input space 20 are identified where the example values ​​do not provide a sufficient basis for a safety-critical application.

[0111] Corrective action can be taken based on the identification: For example, in a further procedural step D1, examples of a second example sentence 27 are collected, and the knowledge about the first example sentence is taken into account during the collection. The examples of the second example sentence 27 are in Figure 5 represented as stars.

[0112] Corresponding to the determination of the first local quality assessment, a second local quality assessment is determined for each of the surrounding areas in a process step D2.

[0113] The first and second example sentences together form a third example set when combined. In a process step D3, a third local quality assessment is determined for each of the surrounding areas, representing the coverage of the input space by the examples of the third example set.

[0114] The first, second, and third quality assessments are compared in a process step E. For example, in a process step E1, a difference between the third quality assessment and the first or second quality assessment is determined. If the quality of the example set has improved in the union set compared to the first example set 21, it can be assumed that the knowledge base has been improved by collecting the second example set 27.

[0115] For example, the quality rating can be determined based on the number of examples assigned to each representative. A comparative measure for the first and second quality ratings can be the difference between the number of examples from the first set of examples assigned to the representative and the number of examples from the second set of examples assigned to the representative.

[0116] In process step F1, an initial local complexity assessment is determined for the respective environment. This assessment represents the complexity of a task in the example-based system, as defined by the first example set 22 of the environment. According to process step F11, this local complexity assessment is determined by the relative positions of the examples in the environment to each other in the input space 20 and the output space. In other words, the complexity assessment is defined based on the similarity of the distances between the examples in the input space 20 and the distances in the output space.

[0117] For example, the task of the example-based system has a comparatively low complexity if the distances in the input space 20 (apart from the scaling) roughly correspond to the distances in the output space. Based on the complexity assessment, areas are identified where, due to the high complexity of the task of the example-based system, a comparatively large number of examples must be recorded. For example, in areas of the input space 20 where higher complexity exists, the density of representatives is dynamically increased until a homogeneous complexity is reached.

[0118] The complexity assessment corresponds to the quality indicators described in Section 4 (QUEEN Quality Indicators) of WASCHULZIK. These quality indicators can be defined and applied for both the representation and coding of the features (see Section 4.5 of WASCHULZIK). An example of this quality indicator for representations is the integrated quality indicator QI 2< according to Section 4.6 of WASCHULZIK.

[0119] In process step F2, a second local complexity assessment is determined for the respective environment, representing the complexity of a task in the example-based system as defined by the second example set. A third local complexity assessment is determined for the examples in the third example set (union set) in process step F3.

[0120] In process step G, the first, second, and third complexity assessments are compared. Particular attention should be paid to areas where the third local complexity assessment indicates a significantly higher complexity of the third example set than the first and / or second complexity assessments. In these areas, there is either a problem with the collection of the examples, or the first, second, or both example datasets 21 and 27 have been falsified to harm the creator or user of the knowledge base. This application example is particularly important because the high complexity cannot be detected based solely on the first and second example sets 21 and 27 (but only on the third example set).

[0121] Alternatively to the one relating to the Figures 4 and 5 The described embodiment, in which representatives are equally distributed in the input space, shows Figure 7An embodiment of an input space 220, in which the representatives each form a center of a cluster, which is determined by means of a clustering method. The examples 222 of the example set are in Figure 7 depicted as crosshairs 223.

[0122] Figure 7 The diagram shows four example clusters, 230, 232, 234, and 236, each containing several examples. These examples lie within a dashed boundary line in the diagram, which, however, does not represent an actual cluster boundary but is merely drawn for illustration. Clusters 230, 232, 234, and 236 each have an associated cluster center, 240, 242, 244, and 246 (shown with a plus sign). The cluster centers 240, 242, 244, and 246 are each located in the center of their respective clusters and are assigned to a cluster independently of the boundaries of the input space grid.

[0123] The clusters according to Figure 7They have the advantage of representing the data topology particularly well. The grid according to Figures 4 and 5 This has the advantage that the uncovered areas are represented more appropriately. For example, the coverage of the input space (according to process steps C1-C6) can be calculated using the grid, and the complexity assessment (according to process steps F1-F3) can be calculated using the cluster center as well as the grid. Which approach is more suitable can also depend on the neural network method. If the coding neurons can move within the input space, then the cluster approach is preferred, or the cluster centers are equated with the positions of the coding neurons in the input space.

[0124] To gain an understanding of the properties and behavior of the quality indicators described in WASCHULZIK as examples of complexity assessment, it is helpful to apply them to synthetic functions (e.g., y=x). From this, it can be deduced how these quality indicators can be applied to example-based systems.

[0125] The Figures 8 to 10 Each figure shows a histogram of the distribution of the complexity rating across the k nearest neighbors of a pre-selected example for a synthetic function. The example could be, for instance, a representative example or the center of a cluster (as described above). Alternatively, the example could be selected from the vicinity of a representative example, chosen for a more in-depth investigation regarding the complexity of the task.

[0126] Figure 8Figure 4.1 on the left and Figure 4.4 on the right are from WASCHULZIK. Figure 8 on the left shows the synthetic function y=x as an axis diagram (the entries in the axis diagram are shown as "+"). The axis diagram on the right shows a histogram. SHLQ 2< from QI 2< about the k-nearest neighbors of an example for the function y=x. It turns out that for arbitrary local neighborhoods k of an example, the histogram shown SHLQ 2< has the value zero.

[0127] Figure 9 shows the left Figure 4 .17 and on the right, Figure 4.20 by WASCHULZIK. As a synthetic function, in Figure 9 left y=ru(seed,300)*300 The data is represented as an axis chart. It is a uniformly distributed random variable with values ​​between 0 and 300. The axis chart on the right shows the histogram. SHLQ 2< from QI 2< about the k-nearest neighbors of an example of the function y=ru(seed,300)*300. The axis diagram in Figure 9The scale on the right is such that 40 represents the value 1.

[0128] Figure 10 shows the left Figure 4 .41 and on the right, Figure 4.44 by WASCHULZIK. As a synthetic function, in Figure 10 left y=sin(8*pi*x / 300)+br(seed,300) It is represented as an axis diagram. It is a sine function that ranges from 0 to 1000. 0 < x ≤ 50 and 100 < x ≤ 200 It has stochastic noise. The axis diagram on the right shows the histogram. SHLQ 2< from QI 2< about the k-nearest neighbors of an example of the function y=sin(8*pi*x / 300)+br(seed,300). The axis diagram in Figure 10The graph is scaled such that 40 represents the value 1. A person skilled in the art can see from this representation that there are several k-neighborhoods up to a size of approximately 45 where the value of QI 2 < 0 is almost 0 (recognizable by the dark gray shading of the bins with small numbers plotted on the V-axis), thus indicating an almost linear mapping of the input and output space. If a person skilled in the art now analyzes the information in the histogram to determine in which neighborhoods the examples exhibit low complexity, they will find the example with x = 75 in whose neighborhood k = 45 the complexity is very low. The same applies to x = 225 or x = 275 for k = 45. Thus, without prior knowledge of how the examples are distributed in the input space, a person skilled in the art can easily, quickly, and reliably identify the areas where the complexity is particularly low or high.By reading the bins with high values, even in large environments, he can identify areas of high complexity (e.g., bin number 80 in K=20). This identification of areas of high or low complexity can be performed independently of the dimension of the input and output spaces, since the distance between the k-nearest neighbors can be determined in spaces of arbitrary dimensionality. Using the same procedure, the person skilled in the art can also identify, from the histograms and the size of the connected regions, the representatives that contain, for example, very few examples. The position in the input space where further examples need to be recorded can then be determined from these representatives.

Claims

1. Computer-implemented method for quality assurance of an example-based system (1), wherein - the example-based system (1) is created and trained on the basis of examples captured by means of sensors of an object recognition system and forming an example set, - the respective example (22) of the example set comprises an input value (12) which lies in an input space (20), - a first example set (21) comprising a plurality of examples captured by means of the sensors of the object recognition system and a second example set (27) comprising a plurality of examples captured by means of the sensors of the object recognition system are collected (B1, D1), - a first local quality rating which represents coverage of the input space by the examples (29) of the first example set (21) is determined on the basis of the distribution of input values in the input space (C1), wherein representatives (28) are distributed in the input space and a number of examples is assigned to these representatives in each case, wherein the examples assigned to this representative are located in a surrounding area (30) of the input space which surrounds the representative and the quality rating is determined on the basis of the number of examples (29) assigned to the representative (28), - a second local quality rating which represents coverage of the input space by examples of the second example set (27) is determined on the basis of the distribution of input values in the input space (D2), wherein representatives are distributed in the input space and a number of examples is assigned to these representatives in each case, wherein the examples assigned to this representative are located in a surrounding area of the input space which surrounds this representative and the quality rating is determined on the basis of the number of examples assigned to this representative and - the first quality rating and second quality rating are compared with one another, wherein a comparative variable is determined which corresponds to the absolute value of the difference between the number of examples of the first example set assigned to the representative and the number of examples of the second example set assigned to the representative and wherein as a function of the determined comparative variable one of the example datasets is wholly or partially rejected, wherein the object recognition system is designed to identify objects using the example-based system and is used in an automated operation of a vehicle, in particular a trackbound vehicle, a motor vehicle, an aircraft, a watercraft and / or a spacecraft.

2. Computer-implemented method according to claim 1, wherein - a third example set is formed from the first and second example set, and - a third quality rating which represents coverage of the input space by the examples of the third example set is determined on the basis of the distribution of the input values in the input space (D3), and - the first quality rating, the second quality rating and the third quality rating are compared (E).

3. Computer-implemented method according to claim 2, wherein the third quality rating is a local quality rating for the surrounding area (30) and said third quality rating is determined on the basis of the examples of the third example set which are assigned to the representative.

4. Computer-implemented method according to at least one of the preceding claims, wherein the quality rating comprises a statistical mean which is determined on the basis of - the set of examples and / or - the examples which are assigned to a respective representative (28) according to claim 2, wherein a statistical measure, in particular a mean, median, minimum, and / or quantiles of the number of examples which are assigned to a representative, is determined as a statistical mean.

5. Computer-implemented method according to at least one of claims 3 to 4, wherein adjacent surrounding areas (32-36) are determined (C5) in the input space (20), the respective representative of which is assigned a number of examples which fulfils a predetermined quality criterion of the quality rating.

6. Computer-implemented method according to claim 5, wherein a connection area (38), which consists of adjacent surrounding areas (32-36), is determined (C6) within the input space (20) and the representatives of said surrounding areas are each assigned a number of examples which fulfils a predetermined quality criterion of the quality rating.

7. Computer-implemented method according to at least one of the preceding claims 3 to 6, wherein - the respective example comprises an output value (14) that is located in an output space, - for the respective surrounding area, a local complexity rating is determined which represents a complexity of a task of the example-based system (1), said complexity being defined by the examples of the surrounding area, and - the local complexity rating is determined by the relative position of the examples of the surrounding area with respect to one another in the input space (20) and the output space.

8. Computer-implemented method according to claim 2 and 7, wherein a first local complexity rating is determined for the examples of the first example set (Fl), a second local complexity rating is determined for the examples of the second example set (F2) and a third local complexity rating is determined for the examples of the third example set (F3) and wherein the third local complexity rating is compared with the first and / or second local complexity rating (G).

9. Computer-implemented method according to at least one of claims 7 to 8, wherein a complexity distribution is determined by means of a histogram representation of the complexity rating.

10. Computer-implemented method according to claim 9, wherein the complexity distribution over k-nearest neighbours of an example in the input space is determined.

11. Computer-implemented method according to at least one of claims 7 to 10, wherein the complexity rating is an integrated quality indicator QI2, - wherein the quality indicator is determined on the basis of the following definition: QI 2 P = 1 P 2 ∑ x i ∈ P 2 d NRE x i − d NRA x i 2 - wherein: d NRE x = d RE x ∑ y ∈ P 2 d RE y P 2 is the normalised distance of the represented inputs and d NRA x = d RA x ∑ y ∈ P 2 d RA y P 2 is the normalised distance of the represented outputs, - wherein x is the pair (x1, x2,) consisting of the two examples x1 and x2, - wherein x1 and x2 are examples from the example set P, - wherein P = {p1,p1, ... , p|P|} is the number of elements of the multiset BAG P and - wherein |P2| is the number of elements of the multiset BAG P.

12. Computer program, comprising commands which, when the program is executed by a computing unit, cause the computing unit to perform the method according to at least one of claims 1 to 11.

13. Computer-readable storage medium, comprising commands which, when executed by a computing unit, cause the computing unit to perform the method according to at least one of claims 1 to 11.