Method for qualifying machine learning model

Abstract

A general aspect of the present disclosure relates to a method for qualifying a trained machine learning model. The method includes receiving a trained machine learning model, determining one or more model behavioral features, evaluating a test data set based on one or more test data criteria, and determining a qualified result based on the one or more model behavioral features and the evaluation of the test data set.

Description

Background Art

[0001] The development and use of machine learning methods and associated models are gaining increasing attention across a wide range of technological fields. Their use extends to numerous application areas, including autonomous driving, robotics, medical diagnostics, speech recognition, tools, household products, and more. The trustworthiness and quality of such models are crucial, especially when used in safety-critical applications where errors in model functionality could have serious consequences.

[0002] The evaluation of machine learning models (ML models) and their performance assessment are core aspects of the development and application of such systems. Qualitative evaluation of ML models is based on validation and test datasets. These datasets are typically obtained by randomly partitioning the overall dataset into training, validation, and test datasets, often using a heuristic partitioning ratio. The resulting test datasets are often checked for suitability by experts.

[0003] After a model has been trained on the training data, it is typically tested on the validation data to see how well it generalizes. This allows, for example, to identify overfitting or identify the best combination of hyperparameters (e.g., learning rate, number of layers). Test data is a separate, independent dataset used for a final evaluation of model performance after the model has been optimized using the training and validation data. Test data is not presented to the model during the training process, but is used to evaluate the model's ability to generalize to new, unknown data.

[0004] In addition to the quantitative performance criteria used to evaluate machine learning models, experts also play an important role. Their experience and knowledge provide deep insights into the trustworthiness and quality of the models. Experts use a checklist to assess qualitative criteria that are critical to the trustworthiness of ML models.

[0005] To fully evaluate an ML model, statements must be made about its expected behavior and the conditions expected during operation. For ML models, only incomplete specifications are often available for this purpose. Assumptions about the expected behavior and boundary conditions during operation are often indirectly confirmed through visual inspection of samples and expert approval.

[0006] Given the increasing importance of ML models and their diverse applications, reliable evaluation methods are urgently needed to ensure that these models correspond to the required quality and credibility standards. Summary of the Invention

[0007] A first general aspect of the present disclosure relates to a method for qualifying a trained machine learning model. The method includes receiving the trained machine learning model, determining one or more model behavior characteristics, evaluating a test data set based on one or more test data criteria, and determining a qualification result based on the one or more model behavior characteristics and the evaluation of the test data set.

[0008] A second general aspect of the present disclosure relates to a computer system designed to perform the method for qualifying a trained machine learning model according to the first general aspect (or its embodiments).

[0009] A third general aspect of the present disclosure relates to a computer program designed to perform the method for qualifying a trained machine learning model according to the first general aspect (or its embodiments).

[0010] A fourth general aspect of the present disclosure relates to a computer-readable medium or signal storing and / or containing a computer program according to the third general aspect (or an embodiment thereof).

[0011] The methods proposed in the present disclosure according to the first general aspect (or its embodiments) can be used to provide a method for qualifying trained machine learning models (ML models). These methods can help minimize possible risks and uncertainties in the application of ML models, while improving the efficiency of development and implementation in shortened development cycles. In an example, the method for qualification can help implement the licensing of ML models in terms of safety goals, fulfillment of legal requirements and / or usage risks. Another advantage may be that the method can be used within the scope of various technical functions and systems, such as autonomous driving functions, control device functions in vehicles (e.g. control device functions that replace physical sensors), calculation of correction values or calculation of abstract system states, (cloud-based) fleet monitoring, and / or ML-based systems in the field of tools and / or household products.

[0012] The techniques disclosed herein can contribute to the standardization of ML model qualification. The disclosed techniques can be used to qualify various ML model architectures. Another advantage may be that the standardized method allows for automated ML model qualification, for example in a computer-implemented manner. In one example, industrial self-certification can be achieved. This can include enabling ML models that are integrated into complex and specialized system environments and can only be replicated at an external certification body at great technical and time expense to be self-certified in an existing (target) system environment. Furthermore, the comparability and reproducibility of ML models across various products and / or product generations can be improved. The unified definition of test data standards allows for the unified use of reference models for qualifying test data sets. The techniques disclosed herein allow for a modular implementation of the method, and additional model behavior characteristics and / or test data standards related to new developments and domain properties can be added. The method can contribute to reducing (safety) technical risks when using ML models in technical systems.

[0013] Some terms are used in this disclosure as follows:

[0014] A "machine learning model" may include any file that is trained to learn specific patterns in a data set and to derive patterns, insights, and predictions from these patterns. Here, for example, supervised learning, unsupervised learning, reinforcement learning, semi-supervised learning, and / or transfer learning may be used to train a machine learning model. Machine learning models may include, for example, linear regression models, logistic regression models, support vector machines (SVMs), k-nearest neighbor classifiers, decision trees and random forests, k-means clustering models, neural networks (e.g., recursive neural networks (RNNs)), long short-term memory networks (LSTMs), and / or transformer models. In addition, machine learning models may include algorithms for Q-learning, linear interpolation, nearest neighbor interpolation, and / or principal component analysis.

[0015] A "vehicle" is any device that transports passengers and / or cargo. A vehicle can be a motor vehicle (e.g., a passenger car or truck) or a rail vehicle. A vehicle can also be a motorized two-wheeled or three-wheeled vehicle. However, floating and flying devices are also vehicles. Vehicles can operate at least partially autonomously, either autonomously or with assistance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1-A to Figure 1-BA method for qualifying a trained machine learning model is schematically illustrated.

[0017] Figure 2 An exemplary eligibility matrix according to one or more embodiments of the present disclosure is schematically illustrated. DETAILED DESCRIPTION

[0018] Figure 1-A and Figure 1-B 1 is a flow chart including possible method steps for a method 100 for qualifying a trained machine learning model. The method 100 for qualifying a trained machine learning model includes receiving 110 a trained machine learning model. The method also includes determining 120 one or more model behavior characteristics 13. The method also includes performing 130 an evaluation of a test data set based on one or more test data criteria 14. The method also includes determining 150 a qualification result based on the one or more model behavior characteristics 13 and the evaluation of the test data set.

[0019] In an example, determining 120 one or more model behavior characteristics 13 may include setting one or more model behavior characteristics 13 to be checked in the context of the ML model being qualified. To this end, one or more test data criteria 14 may then be set, by means of which the test data set is evaluated. Figure 2 As shown, in the example, one or more test data criteria 14 can be based on one or more model behavior characteristics 13. It may be advantageous to evaluate not only the output of the machine learning model when qualifying the machine learning model, but also to evaluate the test data set with respect to the model behavior characteristics 13 being examined.

[0020] In some examples, method 100 for qualifying a trained machine learning model can be understood as a chain of reasoning. Figure 2An exemplary qualification matrix 10 is shown. In the example, a quality target 11 for the entire system should be achieved. The entire system may include one or more system components. The machine learning model may be a component of one or more system components. Quality targets 12 may be set specifically for the machine learning model. In order to qualify the machine learning model, the following method 100 of an embodiment of the present disclosure may be performed. At a first level, it may be determined with the aid of one or more model behavior characteristics 13 whether the machine learning model achieves the relevant (expected) functional target. This may be checked based on a qualified test data set. At a second level, the test data set is evaluated based on one or more test data criteria 14. For this purpose, one or more reference models may be defined, which may be applied to the test data set to obtain a quantification of one or more test data criteria 14. At a third level, inaccuracies in the reference model are analyzed by comparison with operational data. If necessary, the test data set may be recreated or adapted. In the example, a new test data set may be generated using the adapted reference model. This method may facilitate the implementation of a continuous chain of reasoning for qualifying the machine learning model.

[0021] In an example, one or more test data criteria 14 may be determined for each of the one or more model behavior characteristics 13. In an example, the method may include determining 120 or more model behavior characteristics 13. In this case, a different test data criterion 14 may be determined for each of the plurality of model behavior characteristics 13. In any case, different numbers and types of one or more test data criteria 14 may be determined for one or more model behavior characteristics 13. For example, for a first model behavior characteristic 13, 1, 2, ..., 10 or more test data criteria 14 may be determined, and for a second model behavior characteristic 13, 1, 2, ..., 10 or more test data criteria 14 may be determined. In an example, the one or more model behavior characteristics may include properties of the machine learning model that are oriented toward (desired) functionality of the machine learning model. In an example, the one or more model behavior characteristics may include at least one of an objective function, domain robustness, generalization behavior of the network, input context, or output context. For example, the objective function may be described by a previously defined function of the machine learning model. For example, the objective function may be described by physical model knowledge, a simulation model, and / or a semantic functional relationship (e.g., related to the entire system). For example, domain robustness can be described by the relationship between the variance of the input variables of the machine learning model and the invariance of the output variables of the machine learning model. For example, the generalization behavior of the network can be described by the acceptable maximum deviation relative to the (test) dataset and / or by the maximum allowed interpolation behavior of the machine learning model. In an example, the input context can be described by a previously defined operating range of the machine learning model, a safety-critical boundary, and / or a distance to a known dataset. In an example, the output context can be described by a physical boundary, a previously defined operating range of the machine learning model, a safety-critical boundary, and / or a distance to a known dataset.

[0022] In an example, the one or more test data criteria 14 may include at least one of coverage, validity of input context, input distribution, validity of output context, output distribution, functional accuracy, concept drift, or dataset dependency.

[0023] In an example, one or more test data criteria can be evaluated by comparing the quantification of the one or more test data criteria based on the test data set with the quantification of the same one or more test data criteria based on a data set of a model used as a reference / model. For example, the data distribution of the test data set can be evaluated as a test data criterion based on the data distribution of the reference / model data set.

[0024] In an example, the degree of coverage may include a criterion for describing the completeness of the content of the test dataset with respect to the relevant attributes. For example, the degree of coverage may be evaluated based on classification, such as the maximum allowed variance of the nearest neighbor points of the output value in the test dataset. In an example, the degree of coverage may be evaluated based on percentages. For example, the degree of coverage may be evaluated based on a percentage of the content of the test dataset, such as relative to a dataset of a model used as a pattern / reference. An example of this may be a traffic scenario (e.g., weather, road, lighting, pedestrians) during autonomous driving.

[0025] In an example, the validity of the input context may include criteria for describing the validity of the test data-input data in the context of the corresponding system. For example, the validity of the input data of the test data set may be evaluated with the help of known attributes in the (input) data set (e.g., from resampled or unlabeled data). For example, the validity of the input data of the test data set may be evaluated by accepting an example model / reference data set with allowed and / or prohibited data points, and / or by means of interpolation and extrapolation rules about existing data points. In an example, physical model knowledge (e.g., from simulations related to the entire system) may also be used to evaluate the validity of the input data in the test data set in the context of the corresponding system. In another example, the validity of the input data in the test data set may be evaluated with the help of a set of semantic rules that can set possible or impossible input data combinations (scenarios). In an example, the validity of the input context may be evaluated based on a percentage. For example, the validity of the input context may be evaluated based on the context distribution of the test data set, e.g., relative to the data set of the context model used as a model / reference. An example of this may be the operating range of an internal combustion engine.

[0026] In an example, the input distribution may comprise a criterion for describing the distribution of the test data - input data. For example, the input distribution may be evaluated based on the acceptance of an exemplary distributed data set (e.g. from measurements with a known data generating process), by means of statistical analysis (e.g. external temperature distribution), by means of established field standards (e.g. operating points of a standardized driving cycle) and / or by means of legal provisions. In an example, the input distribution may be evaluated based on the Kernel-Stein-Discrepancy. For example, the input distribution may be evaluated based on the data distribution of the test data set relative to a data set of a model used as a pattern / reference. An example of this may be the frequency distribution of wall materials when measured using a power tool positioning device (a so-called "wall scanner").

[0027] In an example, the validity of the output context may include criteria for describing the admissibility of the output data of the machine learning function in the context of the corresponding system. In an example, the validity of the output context may include criteria for describing the admissibility of the test data-output data (target value) in the corresponding context of the entire system. For example, the validity of the output context may be evaluated with the help of an allow / disallow threshold from a safety requirement (e.g., based on a physical model). For example, the validity of the output context may be evaluated with the help of an example data set with allowed / disallowed output values of a model accepted as a reference / model, interpolation rules and / or extrapolation rules about existing data points. In another example, the validity of the output data may be evaluated with the help of a set of semantic rules that can set possible or impossible output data combinations (scenarios).

[0028] In an example, the output distribution may include a criterion for describing the distribution of the output data. In an example, the output data may include a target value in a test data set to be evaluated. For example, the output distribution may be evaluated with the aid of a distributed data set of acceptance examples (e.g., from measurements with a known data generating process). In an example, the output distribution may be evaluated based on a kernel-stem difference. For example, the output distribution may be evaluated based on the data distribution of the test data set relative to a data set of a model used as a pattern / reference. An example of this may be the frequency distribution of wall materials when measured using a power tool positioning device (a so-called "wall scanner").

[0029] In an example, functional accuracy can include criteria for corresponding model behavior characteristics that describe the content correctness of output data with respect to the assumed "ground truth." For example, functional accuracy can be evaluated based on the objective function, domain robustness, generalization behavior of the network, input context, and / or output context.

[0030] In an example, concept drift may include criteria for describing the transferability of the domain of test data and / or training data to the domain of operational data. For example, concept drift may be evaluated by interpolation based on a test data set and comparing the interpolation with labeled operational data. In an example, concept drift may be evaluated based on kernel-stem differences. For example, concept drift may be evaluated based on the data distribution of the test data set relative to a data set of a model used as a pattern / reference. An example of this may be the data distribution of a test data set in a vehicle relative to a test bench test data set of a temperature estimation model for the stator temperature of an electric vehicle.

[0031] In an example, dataset dependencies may include criteria for describing the dependency of a test dataset on other datasets (particularly a training dataset).

[0032] In an example, the method may include determining 140 one or more evaluation indicators based on one or more model behavior characteristics 13. In an example, determining 150 a qualified result may also be based on one or more evaluation indicators. For example, the one or more evaluation indicators may include a misclassification rate. In an example, the one or more evaluation indicators may include a confidence interval, for example based on an application-specific indicator. In an example, the one or more evaluation indicators may include a deviation frequency. For example, the one or more evaluation indicators may include a deviation frequency from a robustness requirement of a model used as a reference, for example, robustness with respect to image rotation for optical good-bad part analysis during a manufacturing process. In an example, the one or more evaluation indicators may include a deviation. For example, the one or more evaluation indicators may include a deviation from a functional requirement from a model used as a reference. An example of this is the cumulative fuel consumption deviation of a control device function for estimating fuel consumption in a vehicle. In an example, the one or more evaluation indicators may include a mean squared error (MSE).

[0033] In an example, the method may include using a machine learning model when a qualified result falls within a permitted range. In an example, the method may include using a machine learning model when one or more evaluation metrics fall within a permitted range. In an example, a qualified result may include a separate result for one or more model behavior characteristics 13 and a separate result for the evaluation of a test data set. In an example, a permitted range may be set separately for one or more model behavior characteristics 13 and for the evaluation of a test data set. In an example, using a machine learning model may include switching from a conventional method to a method based on a machine learning model. In an example, the switching may be performed in a computer-implemented manner in an operating environment. In an example, the permitted range may include one or more first thresholds for one or more model behavior characteristics 13 and / or one or more second thresholds for one or more test data criteria 14. For example, the first threshold for one or more model behavior characteristics 13 may be a threshold for the aforementioned misclassification rate.

[0034] In an example, evaluating 130 a test data set can include defining 131 one or more reference models for each of the one or more test data standards 14; and applying 132 the one or more reference models to the test data set to obtain test data quantification for at least one of the one or more test data standards 14. In an example, the one or more reference models can include one or more of a data density estimate and / or a data density function, a semantic model, principal component analysis, linear interpolation, modeling of a frequency distribution, k-nearest neighbor classification, and / or nearest neighbor interpolation. For example, data density estimation can be performed based on given data, according to a kernel-based method, based on a histogram, and / or with respect to neighboring points using k-nearest neighbors. For example, the semantic model can be mapped in the form of an ontology and / or a rule set.

[0035] For example, for the test data criterion "coverage", the one or more reference models may include a (test) data density estimate and / or a semantic model.

[0036] For example, for the test data criterion "validity of input context", one or more reference models may include linear interpolation. In an example, linear interpolation may be applied to the subspace produced by principal component analysis. For example, for the test data criterion "validity of input context", one or more reference models may include a semantic model. For example, for the test data criterion "input distribution", one or more reference models may include modeling of frequency distribution. In an example, this may include partitioning the input data of the machine learning model. For example, for the test data criterion "validity of output context", one or more reference models may include a semantic model.

[0037] For example, for a test data standard "output distribution," one or more reference models may include a data density function.

[0038] For example, for the test data criterion "concept drift", the one or more reference models may include nearest neighbor interpolation and / or local linear interpolation based on the test data set and comparison of the interpolation with labeled operational data.

[0039] In an example, one or more test data criteria may be quantified with the aid of one or more reference models. In an example, the quantization of one or more test data criteria may be summarized in a test data quantization.

[0040] In some embodiments, the test data set may include a first original test data set and / or a generated test data set. In an example, the generated test data set may be generated by applying at least one reference model of the one or more reference models to the first original test data set. In an example, the generated test data set may be generated by a selection set from a plurality of second original test data sets. For example, the reference model may include generating the test data set with the aid of a manifold model and / or an approximate k-nearest neighbor classifier. For example, a reference model for the test criterion "validity of input context" may include generating the test data set with the aid of a manifold model and / or an approximate k-nearest neighbor classifier.

[0041] In some embodiments, method 100 may further include applying 160 one or more reference models to the operational data set to obtain operational data quantifications 15 for at least one of the one or more test data standards 14. In some examples, the method may further include comparing 170 the test data quantifications with operational data quantifications for at least one of the one or more test data standards 14 to obtain a comparison result 15. In an example, if the comparison result 15 is outside a defined acceptance range, method 100 may include generating 180 a new test data set and / or modifying the original test data set. Furthermore, method 100 may include performing 190 an evaluation of the new test data set based on the one or more test data standards 14. In an example, method 100 may further include determining 200 a new qualified result based on the one or more evaluation metrics and the evaluation of the new test data set. For example, if comparison result 15 indicates an unacceptable deviation between the operational data and the test data set, determining 200 a new qualified result may be necessary to ensure that the machine learning model provides correct results during operation. In an example, method 100 may include switching from a machine learning model-based approach to a conventional approach if comparison result 15 is outside a defined acceptance range. For example, this can represent the inverse of the previously mentioned method steps of switching from a conventional method to a method based on a machine learning model. In an example, one or more reference models can be applied 160 to an operational dataset and / or the test data quantification can be compared 170 to the operational data quantification in real time. For example, this can be performed continuously, at specific time intervals, multiple times per day, and / or with the aid of a trigger. In an example, applying 160 one or more reference models to an operational dataset can be part of monitoring and / or monitoring operational data and / or model behavior in operational mode.

[0042] In an example, at least a subset of the operational dataset may include labeled data. For example, for the test criteria "functional accuracy" and / or "concept drift," the operational dataset may include labeled data. For example, for the test criteria "coverage," "input context validity," and / or "input distribution," labeled data may not be necessary.

[0043] In an example, the method 100 for qualifying a trained machine learning model and / or the machine learning model (qualified and / or unqualified) can be designed to be executed in a vehicle, a robot, a building, a power tool, and / or a home appliance, and / or can be designed to control and / or monitor a vehicle function, a robot function, a building automation function, a power tool automation function, and / or a home appliance automation function. In an example, the method 100 can include implementing the qualified machine learning model on a computer system in the vehicle, the robot, the building, the power tool, and / or the home appliance.

[0044] In an example, the robot may include a vehicle (autonomously driven or assisted driving). For example, the vehicle function may be a function for autonomous and / or assisted driving. In some examples, the method 100 for qualifying a trained machine learning model and / or the machine learning model (qualified and / or unqualified) may be designed to be executed on a computer system of a vehicle (e.g., an autonomous, highly automated, or assisted driving vehicle). For example, the computer system may be implemented locally in the vehicle or (at least partially) in a backend that is communicatively connected to the vehicle. For example, the computer system may include a control device on which the method 100 for qualifying a trained machine learning model and / or the machine learning model may be executed. In some examples, the vehicle may include a computer system having a communication interface that enables communication with a backend. For example, the method 100 for qualifying a trained machine learning model and / or the machine learning model may be executed in the backend. In an example, a control device function in the vehicle may include or access the method 100 for qualifying a machine learning model and / or the machine learning model, for example when the control device function is remotely executed in the cloud. For example, a control device function that replaces physical sensors, calculates correction values and / or abstracts system states (e.g., aging) may include or access method 100 and / or a machine learning model for qualifying a machine learning model, for example when the control device function is executed remotely in the cloud.

[0045] In other examples and as described above, the method 100 for qualifying a trained machine learning model and / or the machine learning model (qualified and / or unqualified) can be designed to be executed in a robot, and / or to control and / or monitor robot functions (particularly to control and / or monitor the robot's motion functions). In some examples, the method 100 for qualifying a trained machine learning model and / or the machine learning model can be executed on a computer system of the robot. For example, the computer system can be implemented locally in the robot, or (at least partially) in a backend that is communicatively connected to the robot.

[0046] In one example, the method 100 for qualifying a trained machine learning model and / or the machine learning model (qualified and / or unqualified) can be designed to be executed in a building and / or used to control a building function (particularly a building automation function). For example, the building function can be a function for adjusting room temperature, lighting, and / or security devices. In some examples, the method 100 for qualifying a trained machine learning model and / or the machine learning model can be designed to be executed on a computer system within the building. For example, the computer system can be implemented locally in the building or (at least partially) implemented in a backend that is communicatively connected to the building. For example, the computer system can include a control system or building automation control device, on which the method 100 for qualifying a trained machine learning model and / or the machine learning model can be executed. In an example, the building can have a computer system with a communication interface that enables communication with an external backend. For example, the method 100 for qualifying a trained machine learning model and / or the machine learning model can be executed in the backend. In an example, the operating data and / or the input data for the machine learning model can be based on information such as room temperature, brightness, or the presence of people. In some cases, operational data and / or input to machine learning models can include relative temperature differences, illuminance, or distances to specific locations or objects in a building. In some examples, this information can come from a network, such as sensor data or settings from other buildings or building components. This information can be provided through communication between buildings or building components or via an external backend.

[0047] In other examples, the method 100 for qualifying a trained machine learning model and / or the machine learning model (qualified and / or unqualified) can be designed to be executed in the power tool and / or can be designed to control and / or monitor the power tool functions (particularly control and / or monitor the working functions of the power tool). In some examples, the method 100 for qualifying a trained machine learning model and / or the machine learning model can be executed on a computer system of the power tool. For example, the computer system can be implemented locally in the power tool or (at least partially) in a back end that is communicatively connected to the power tool.

[0048] In other examples, the method 100 for qualifying a trained machine learning model and / or the machine learning model (qualified and / or unqualified) can be designed for execution in a household appliance and / or can be designed for controlling and / or monitoring household appliance functions (particularly controlling and / or monitoring operating functions of the household appliance). In some examples, the method 100 for qualifying a trained machine learning model and / or the machine learning model can be executed on a computer system of the household appliance. For example, the computer system can be implemented locally in the household appliance or (at least partially) in a backend that is communicatively connected to the household appliance.

[0049] In another example, the method 100 for qualifying a trained machine learning model and / or the machine learning models (qualified and / or unqualified) can be designed to be executed in a machine tool, a personal assistant, an access control system, and / or a medical device (e.g., for medical imaging and / or accessible via a network). In an example, the method 100 can include implementing the qualified machine learning model on a computer system of the machine tool, the personal assistant, the access control system, and / or the medical device.

[0050] In an example, a machine learning model can be used to analyze audio data and / or video data. In an example, a machine learning model can be used to classify sensor data, identify objects, or perform semantic segmentation on sensor data, for example, with respect to traffic signs, road surfaces, pedestrians, vehicles, power lines, water pipes, air pipes, biochemical reactions, or physical obstructions, for example, in path planning. In an example, a machine learning model can be used to determine continuous values, for example, by means of regression related to distance, speed, acceleration, gas concentration, azimuth, altitude, fuel consumption, emissions, temperature, current, voltage, aging, yaw rate, humidity, pressure, vibration, and / or cloud cover. In an example, a machine learning model can be used for object tracking, for example, based on pixel attributes. In another example, a machine learning model can be used for anomaly recognition, optionally with the aid of an autoencoder. In an example, a machine learning model can be used to estimate sensor signals, in particular sensor signals from sensors for measuring speed, rotational speed, current, voltage, temperature, pressure, air pressure, weight, deformation, flow of liquid or gas, or gas composition.

[0051] For example, the above-mentioned labeled operating data can be acquired by means of a sensor system. For example, this may mean using a sensor system to detect an output signal of a controller so that a labeled operating data set can be acquired from the input data and the target data.

[0052] Also disclosed is a computer system designed to perform the method 100 for qualifying a machine learning model. The computer system may include at least one processor and / or at least one working memory. The computer system may also include a (non-volatile) memory.

[0053] Also disclosed is a computer program designed to perform the method 100 for qualifying a machine learning model. For example, the computer program can exist in an interpretable form or a compiled form. The computer program can be loaded (even partially) into the RAM of a computer for execution, for example as a sequence of bits or bytes.

[0054] Also disclosed is a computer readable medium or signal storing and / or containing the computer program or at least a portion thereof. The medium may comprise, for example, one of RAM, ROM, EPROM, HDD, SDD, etc., on which / in which the signal is stored.

Claims

1. A method (100) for qualifying a trained machine learning model, comprising: - receiving (110) a trained machine learning model, - determining (120) one or more model behavior characteristics (13), - evaluating (130) the test data set based on one or more test data criteria (14), and - determining (150) a qualified result based on the one or more model behavior characteristics (13) and an evaluation of the test data set.

2. The method according to claim 1, wherein the method comprises - determining (140) one or more evaluation metrics based on the one or more model behavior characteristics (13), and wherein Determining (150) a passing result is also based on the one or more evaluation indicators.

3. The method according to claim 1 or 2, wherein the method further comprises: -When the qualified result falls within the permitted range, using the machine learning model.

4. The method according to claim 3, wherein the permission range includes one or more first threshold values for the one or more model behavior characteristics (13) and / or one or more second threshold values for the one or more test data criteria (14).

5. A method according to claim 3 or 4, wherein using the machine learning model includes switching from a conventional method to a method based on the machine learning model.

6. The method according to any of the preceding claims, wherein the evaluation (130) of the test data set comprises - defining (131) one or more reference models for each of said one or more test data standards (14), - applying (132) the one or more reference models to the test data set to obtain test data quantification of at least one test standard of the one or more test data standards (14).

7. The method of claim 6 , wherein the test dataset comprises a first original test dataset and / or a generated test dataset, wherein the generated test dataset is generated by applying at least one reference model of the one or more reference models to the first original test dataset, and / or wherein the generated test dataset is generated by selecting a set from a plurality of second original test datasets.

8. The method according to claim 6 or 7, wherein the method (100) further comprises - applying (160) the one or more reference models to an operational data set to obtain an operational data quantification (15) of at least one of the one or more test data standards (14), - comparing (170) the test data quantification with the operational data quantification of at least one test standard of the one or more test data standards (14) to obtain a comparison result (15), and if the comparison result (15) is outside a defined acceptance range, - generating (180) a new test data set and / or modifying said original test data set, - evaluating (190) the new test data set based on the one or more test data criteria (14), - determining (200) a new qualified result based on the one or more evaluation metrics and an evaluation of the new test data set.

9. The method of claim 8, wherein at least a subset of the operational data set comprises labeled data.

10. The method of any of the preceding claims, wherein the one or more model behavior features include at least one of an objective function, domain robustness, generalization behavior of the network, input context, or output context.

11. The method according to any of the preceding claims, wherein the one or more test data criteria (14) include at least one of degree of coverage, validity of input context, input distribution, validity of output context, output distribution, functional accuracy, concept drift or dataset dependency.

12. The method (100) according to any of the preceding claims, wherein the method (100) for qualifying a trained machine learning model and / or the machine learning model is designed for execution in a vehicle, a robot, a building, a power tool and / or a household appliance, and / or is designed for controlling and / or monitoring vehicle functions, robot functions, building automation functions, power tool automation functions and / or household appliance automation functions.

13. A computer system designed to perform the method (100) for qualifying a machine learning model according to any one of the preceding claims 1 to 12.

14. A computer program comprising instructions which, when executed by a computer system, cause the computer system to perform the method (100) for qualifying a machine learning model according to any one of the preceding claims 1 to 12.

15. A computer-readable medium or signal storing and / or containing a computer program according to claim 14.

Images