Method and system for optimized training of machine learning algorithm

Abstract

To provide a method for optimized training of a machine learning algorithm.SOLUTION: A method for optimized training of a machine learning algorithm, includes: providing a domain model that has domain parameters P1 to P3 and / or domain values v11 to v32 for a domain 306; providing a data model 302 that has a training dataset 304 including training data therefor; removing, masking 314 and modifying training data from the training dataset depending on the domain parameters and / or domain values; training a neural network 318 based on the reduced training data set 312; comparing 322 a determined model output 319 with a model output 205 associated with the training dataset and determining a model output deviation; selecting training data from the training dataset depending on the model output deviation; training a machine learning algorithm based the selected training data; and providing the trained machine learning algorithm.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The invention relates to a method and a system for training machine learning algorithms in an optimized manner. The invention further relates to a computer program comprising a program code and to a computer readable data carrier comprising the program code of the computer program. [Background technology]

[0002] Background technology A domain model in the form of an ontology is a structured semantic picture of a particular knowledge domain. In the field of machine learning models, the development of such domain models plays a crucial role to improve the understanding and interpretation of the models, and it allows for effective communication between experts, developers and users.

[0003] The construction of a domain model in the form of an ontology of machine learning models is usually an iterative process that requires a systematic discovery and categorization of relevant knowledge. First, the various concepts and entities related to each other in the domain are identified, e.g., data sources, models, algorithms, metrics, and evaluation methods. Based on this, the relationships between the concepts are defined, so that their dependencies and relationships are captured. In this case, classification, aggregation, association, and hierarchization can be used to form the structure of the ontology. Defining the attributes and properties of the concepts allows a detailed description and characterization of the individual elements. The attributes here can include, for example, information about the parameters, properties, or application domain of the machine learning model.

[0004] It is preferable to involve experts and practitioners to validate and improve the ontology. Their expertise and experience contribute to ensuring the completeness and accuracy of the model. The use of existing ontologies and standards can also be beneficial to ensure consistency and compatibility.

[0005] A well-developed domain model in the form of an ontology allows for a unified and structured representation of knowledge in the domain of machine learning models, which aids understanding, interpretability and reusability of the models and promotes effective collaboration between experts, developers and users.

[0006] The publication “Using ontologies for dataset engineering in automotive AI applications” (http: / / dx.doi.org / 10.23919 / DATE54114.2022.9774675) describes a method for building a domain model in the form of an ontology, which describes the effects and / or properties from the domain or area of ​​use of the machine learning model (the so-called “dimensions” or domain parameters) and the possible specificities in which these may occur (the so-called “options” or domain values).

[0007] The publication “Datamodels: Predicting Predictions from Training Data” (https: / / arxiv.org / abs / 2202.00622) describes the use of data models, i.e. machine learning models, to consider the impact of data points (in class extension) on the performance of a trained machine learning model.

[0008] Furthermore, methods are known that use combinatorial testing developed for software testing, where an approach for incremental construction of a data set using a domain model is described, where it is ensured that the dimensions and options of the domain model are covered, but the performance or capabilities of the trained machine learning model are not considered.

[0009] So far, the performance of the trained machine learning models has been poorly considered, as well as the importance of the dimensions and options considered. [Prior art documents] [Non-patent literature]

[0010] [Non-Patent Document 1] “Using ontologies for dataset engineering in automotive AI applications”(http: / / dx.doi.org / 10.23919 / DATE54114.2022.9774675) [Non-Patent Document 2] “Datamodels: Predicting Predictions from Training Data”(https: / / arxiv.org / abs / 2202.00622) Summary of the Invention [Problem to be solved by the invention]

[0011] The problem on which the invention is based is therefore to provide a method for training a machine learning algorithm, which improves the performance of the machine learning algorithm or model, in particular with respect to the dimensions and options of the domain model, and furthermore, which takes into account the importance of the dimensions and options of the observed domain model. [Means for solving the problem]

[0012] The above problem is solved by a method for optimized training of a machine learning algorithm according to the features of claim 1. Furthermore, the above problem is solved by a system for optimized training of a machine learning algorithm according to the features of claim 12.

[0013] Disclosure of the Invention Here, a method for training a machine learning algorithm in an optimized manner is presented, which in this case comprises at least the following steps: - providing a domain model having domain parameters and / or domain values ​​for at least one domain; - providing a data model having a training dataset comprising training data for at least one domain; - removing and / or masking and / or modifying at least one training data from the training data set in dependence on the at least one domain parameter and / or domain value in order to provide a reduced training data set, in particular in order to evaluate the sensitivity of the at least one removed and / or masked domain parameter and / or domain value; - training a neural network on a reduced training data set to determine a model output in dependence on the reduced data set; - comparing the determined model output with model outputs associated with a training dataset; - selecting training data from the training dataset in dependence on the comparison of the model outputs; - training a machine learning algorithm based on the selected training data; - Providing a trained machine learning algorithm; Includes.

[0014] The data model preferably comprises image data and / or video data, by which preferably different combinations of domain parameters and / or domain values ​​are mapped into at least one domain. In principle, it is not necessary for the entire combinations of domain parameters and / or domain values ​​to be mapped by the data model. Absence can be found by the method according to the invention, for example, when at least one domain parameter and / or domain value is at least temporarily masked and / or removed, for which image data and / or video data is not present.

[0015] The model outputs associated with a training dataset preferably describe model outputs calculated based on the entire training dataset, without necessarily taking into account the domain model.

[0016] In this document, the data model is combined with the domain model, thus providing a means for deducing which domain parameters and which domain values ​​are significant and therefore required in the training phase of the machine learning algorithm. Furthermore, this in particular derives what degree of interaction between the domain parameters and the domain values ​​must be used, further depending on the domain parameters and / or the domain values. With the data model thus obtained, a (training) data set is derived, with which the machine learning algorithm or machine learning model is trained. The model thus trained is preferably robust to the variability of the data model. Here, it is always preferable to keep the training data set as small as possible, in particular preferably based on the computational power. In this case, it is desirable, for example, that only little redundancy is present. However, there are also economic reasons for keeping the training data set small, in particular in the case of (partially) supervised training methods, since with the size of the training data set, the costs of labeling, which is often performed manually, also increase. However, the training data set should preferably not be selected too small in order to include sufficient variability regarding the characteristics of the domain.

[0017] In the field of computer vision, a "domain model" preferably relates to a structured and / or semantic picture of a knowledge domain related to the processing and / or analysis of visual information. Thus, by using computer vision techniques, an improved understanding of images, videos or other visual data is obtained. A domain model may include many different aspects, such as object recognition, face recognition, image segmentation, motion analysis, image classification, etc. A domain model preferably comprises a collection of concepts, entities and / or relationships that represent relevant visual properties, characteristics and / or structures within the respective domain. Such a model may for example define classes of objects, such as cars, people or animals, their features and / or their interrelationships. This may also include algorithms and techniques used for processing and analyzing visual data, such as convolutional neural networks (CNN), feature extraction methods or clustering methods. The development of a domain model is preferably performed taking into account the knowledge of experts in the fields of image processing, pattern recognition and / or artificial intelligence, where the experts preferably identify relevant concepts and define their relationships to achieve a comprehensive understanding of the domain. The domain model advantageously aids in the development of high performance computer vision applications, such as object recognition systems, autonomous vehicles, medical imaging systems and / or video surveillance systems, etc. It may also be useful for integrating computer vision techniques into other application domains such as robotics, safety and / or augmented reality.

[0018] A domain model preferably describes effects and / or properties from the domain or area of ​​use of the machine learning model, where the effects and / or properties are also referred to as "dimensions" and are e.g. indicated as domain parameters. The possible features in which the effects and / or properties may occur are preferably referred to as "options" and are e.g. indicated as domain values.

[0019] The method allows achieving that during the training phase and / or during the computation, sufficient data is present in the training dataset for the individual dimensions and options of the domain model as well as for their interactions. According to the invention, the need for dense sampling of the domain, which is still required in some known machine learning algorithms for sensitivity analysis, and the need to generate simulated or synthetic (training) data for the iterative improvement of the (training) dataset is no longer necessary.

[0020] The method allows for a "discrete" statement of whether a particular effect can potentially improve model performance on different data. Additionally, it allows for the detection of the effects of interactions between various effects, and furthermore, it allows for the evaluation of how these interactions affect model performance. The method increases the efficiency or quality of training, which makes the training phase of machine learning algorithms more convenient overall. That is, the method purposefully utilizes a special data model, so that the training data does not require a complete combination and / or complete sampling of one dimension. Rather, a subset or sub-set of data can be used to obtain an already well-founded estimate. Overall, therefore, the present invention significantly reduces training costs.

[0021] The method is basically applicable to all image processing based applications where no or insufficient image data is present specifically marked or labeled for the target domain. The method is basically applicable for analyzing and / or processing sensor data. This is especially true in driver assistance systems and / or fully automated driving and / or surveillance cameras and / or automation systems and / or other image processing fields, and is particularly advantageous when there is a large amount of data for training and / or applying assistance functions. Furthermore, the invention also applies to applications of multimodal systems based on image data formed, for example, by cameras, LiDAR sensors and / or radar sensors or any combination thereof. Here, the data used by the machine learning algorithms may come from at least one sensor. The sensor is, for example, a sensor that has calculated measurements of the surroundings in the form of a sensor signal that may come from the following sources: digital images, for example videos, radar, LiDAR, ultrasound, motion, thermal images and / or audio signals. Based on the sensor signal, information can be obtained about the element in which this sensor signal is encoded (i.e. an indirect measurement can be performed based on the sensor signal that is used as a direct measurement).

[0022] The method for training machine learning algorithms finds particular application in the field of active learning and / or testing and / or evaluation or data curation. In this case, the method can in particular be used for actively selecting (training) data that any technical system, in particular an autonomous vehicle and / or a robotic system and / or an industrial machine, transmits to a back-end computer. This allows reducing data traffic. The freed bandwidth can in particular be used for efficiently and effectively using the information for training machine learning systems.

[0023] Based on the improved domain model, the method allows for better and / or more effective curation of training and / or test data sets of machine learning algorithms. Furthermore, splits can be set. Similarly, the method allows for the selection of unlabeled data for labeling or labeling, which can then be used, for example, for supervised learning or training of machine learning algorithms. Furthermore, the domain model here is particularly realizable and / or executable in autonomous vehicles and / or robotic systems and / or industrial machines, which allows for effective recording of (training) data based on at least one activation or trigger provided by the domain model here. The method for training machine learning algorithms, in technical terms, is directed to the technical use of mathematical methods, in particular to be executed as computationally efficient and effective as possible on a computer and / or control device. Here, in particular in the method, the functional scheme of the internal technology of the computer and / or processor is useful in designing the implementation of the method for optimizing the internal functioning of the computer and / or control device. Such optimization is achieved by using the data model in order to computationally avoid a combinational "explosion" in particular linked to combinatorial testing.

[0024] The training model achieves at least the following technical effects, which are particularly important when deriving a (training) data set from a domain model: On the one hand, it allows a sensitivity analysis that allows to calculate which "dimensions" and / or "options" in which dimensions are decisive for a good or effectively usable training data set. Furthermore, it allows to calculate how much interaction is required between the different dimensions, thereby providing the required combination of interactions. Additionally, the results of the method can support the selection of previously unlabeled or uncharacterized (training) data, whereby labeling is "rewarded" based on, in particular, a sensitivity and / or performance analysis. This can be advantageous, for example, when the performance or capacity of the trained machine learning algorithm or machine learning model is not yet saturated with respect to the effects contained in the (training) data. Thus, in summary, a training method is described herein for selecting the most important effects or dimensions and / or their interactions or options from a description of a domain, such that the performance of the machine learning model can be optimized, in particular with respect to a selectable metric. By using a data model it is possible to calculate which effects are important in the (training) data and / or which of these effects should be considered specifically and with what strength in the (training) data set.

[0025] The steps according to the invention and further optional steps do not necessarily have to be performed in the order shown, but may be performed in other orders, further intermediate steps may be provided, and individual steps may include one or more sub-steps without departing from the scope of the method according to the invention.

[0026] In a preferred embodiment, the training data for at least one domain are each characterized depending on at least one combination of domain parameters and / or domain values, such that each training data preferably has a label corresponding to a particular combination of domain value and corresponding domain parameter.

[0027] In a preferred embodiment, the training data set of the data model comprises at least one predicted domain identifier, which preferably corresponds to a "standard label" from the prior art, by means of which the machine learning algorithm to be trained receives a description of which domain the training data is contained in. After the algorithm has evaluated and / or classified and / or segmented the image data and / or video data, the result here is compared with the predicted domain identifier in order to check whether the algorithm has correctly detected the semantic content contained in each image data and / or video data.

[0028] In a preferred embodiment, the step of removing and / or masking and / or modifying at least one training data from the training data set in dependence on at least one domain parameter and / or domain value is performed continuously and / or iteratively by varying at least one domain parameter and / or domain value. The method comprises varying at least one domain parameter and / or domain value to perform the removing, masking or modifying of the training data. The exact type of variants and / or step formats of the procedures herein are not specified and various formats can be used.

[0029] In a preferred embodiment, the training of each neural network, the determination of each model output and the comparison of the model outputs are performed for each successively and / or iteratively reduced training data set, a subset is selected from the training data set and a machine learning algorithm is trained on the basis of said subset. For the removed and / or masked and / or modified domain parameters and / or domain values, the neural network is thus trained so that a model output can be determined and compared with the comparative model output. The individual neural networks do not yet correspond to the machine learning algorithm to be trained, which is only trained when the training data set is selected in the manner of the method and / or reduced starting from the original larger training data set.

[0030] In a preferred embodiment, the subset from the training data set comprises training data having a predefined influence on the model output, in particular depending on at least one boundary value. In contrast, preferably, training data from the training data set that has no influence on the model output in the comparison of the model output here is masked, since it does not positively influence the quality and / or performance of the training, but only burdens the computational output during training.

[0031] In a preferred embodiment, based on a comparison of the determined model output with the model output associated with the training data set, an interaction between at least two domain parameters and / or domain values ​​can be calculated. Based on the data model, an embedding in the data model can be preferably calculated for the concrete instance. In case of a linear data model, this is an embedding in a transformed space described by the coefficients of the data model. The embedding or embedding can preferably identify commonalities and / or interactions between the training data. Here, it can be calculated which combinations of domain parameters and / or domain values ​​should preferably be considered in common in order to achieve a predefined model output.

[0032] In a preferred embodiment, the machine learning algorithm comprises a neural network, in particular a deep neural network. A deep neural network is a type of artificial neural network architecture that preferably consists of multiple layers of neurons. Each layer processes input data and forwards it to the next layer, which can learn complex patterns and relationships. Deep neural networks are preferably used for tasks such as image and speech recognition, machine translation and other complex data processing tasks. The machine learning algorithm in the preferred embodiment preferably uses such a deep neural network.

[0033] In a preferred embodiment, there is further provided a manufacturing line including an equipment configuration for producing a configurable product. The manufacturing line is preferably a series of production stations and / or work areas arranged to cooperate to produce at least one product. The manufacturing line may include various devices, machines and / or equipment configured for the production of the configured product. In the present embodiment, it is emphasized that in a preferred embodiment, a specific equipment configuration is present in the manufacturing line. Such an equipment configuration may include, for example, machines, robots, automated assembly lines, tools and / or other devices required for the production of the configured product.

[0034] In a preferred embodiment, the method further comprises the step of producing at least one configurable product using the equipment configuration after the production line has been provided. After the production line has been provided, according to the embodiment, a method is carried out with the goal of producing at least one configurable product, the steps being carried out using the existing equipment configuration in the production line. The method does not specify in detail what type of product or what the exact flow of the production process looks like. The emphasis is on the fact that the method in a preferred embodiment has the goal of producing at least one configurable product in the production line using the supplied equipment configuration.

[0035] According to the invention, a control device included in an autonomous vehicle and / or a robotic system and / or an industrial machine is also claimed, on which a machine learning algorithm trained according to the method in one of its embodiments can be executed.

[0036] According to the invention, a system for training a machine learning algorithm in an optimized state is also claimed. The system comprises a providing device configured to provide a domain model having domain parameters and / or domain values ​​for at least one domain and a data model having a training dataset including training data for the at least one domain. The system further comprises an evaluation computing device configured to remove and / or mask and / or modify at least one training data from the training dataset in dependence on the at least one domain parameter and / or domain value to provide a reduced training dataset, thereby determining a model output in dependence on the reduced dataset, comparing the determined model output with a model output associated with the training dataset, selecting training data from the training dataset in dependence on the comparison of the model outputs, and training the machine learning algorithm in dependence on the selected training data, wherein the providing device is further configured to provide the trained machine learning algorithm.

[0037] According to the invention, a computer program is also claimed, comprising a program code for carrying out at least part of the method according to the invention in one of its embodiments, when the program is executed on a computer. In other words, according to the invention, a computer program product is claimed, comprising instructions for causing a computer to carry out the processes or steps of the method according to the invention in one of its embodiments, when the program is executed by a computer.

[0038] According to the invention, a computer readable data carrier is proposed, comprising a program code of a computer program for carrying out at least a part of the method according to the invention in one of its embodiments, when the program code is executed on a computer. In other words, the invention relates to a computer readable medium (memory) medium, comprising instructions for carrying out the process or steps of the method according to the invention in one of its embodiments, when the program code is executed by a computer.

[0039] The described configurations and developments can be combined with one another in any desired manner.

[0040] Other possible configurations, developments and realizations of the invention also include not explicitly mentioned combinations of the features of the invention described above or below with reference to the exemplary embodiments.

[0041] The accompanying drawings are provided for a better understanding of the embodiments of the present invention, and are used in the context of the specification, explanation of the basic schemes, and concepts of the present invention, to illustrate the embodiments.

[0042] Other embodiments and many of the above-mentioned advantages will become apparent with reference to the drawings, in which: The illustrated elements are not necessarily drawn to scale relative to each other. [Brief description of the drawings]

[0043] [Figure 1] 1 is a schematic flow chart illustrating an embodiment of the present method for training a machine learning algorithm in an optimized state. [Diagram 2] FIG. 1 is a schematic block diagram showing a comparison between a conventional training method (a) and the present method (b) for training a machine learning algorithm in an optimized state. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0044] In the various figures of the drawings, elements, modules or components having the same or similar functions are labeled with like reference numerals unless otherwise stated.

[0045] FIG. 1 shows a schematic flow chart of a method for optimally training a machine learning algorithm.

[0046] The method may in any embodiment be at least partially executed by a system 1, which for this purpose may include several components not shown in detail, such as one or more supplying devices and / or at least one rating calculation device. It will be understood that the supplying device may be configured together with the rating calculation device or may be different from the rating calculation device. The system may further include a memory device and / or an output device and / or a display device and / or an input device.

[0047] A computer-implemented method for training a machine learning algorithm in an optimized manner, as described herein, includes at least the following steps.

[0048] In step S1, a domain model having domain parameters and / or domain values ​​for at least one domain is provided.

[0049] In step S2, a data model is provided having a training data set that includes training data for at least one domain.

[0050] In step S3, at least one training data is removed and / or masked and / or modified from the training data set in dependence on the at least one domain parameter and / or domain value to provide a reduced training data set.

[0051] In step S4, a neural network is trained on the reduced training data set to determine a model output dependent on the reduced data set.

[0052] In step S5, the determined model outputs are compared to model outputs associated with the training data set.

[0053] In step S6, training data is selected from the training data set depending on the comparison of the model outputs.

[0054] In step S7, the machine learning algorithm is trained based on the selected training data.

[0055] In step S8, the trained machine learning algorithm is provided.

[0056] The step S3 of removing and / or masking and / or modifying at least one training data from the training data set in dependence on the at least one domain parameter and / or domain value is particularly preferably performed continuously and / or iteratively by varying the at least one domain parameter and / or domain value. The training of the respective neural network, the determination of the respective model output and the comparison of the model outputs are particularly preferably performed for each reduced training data set continuously and / or iteratively, a subset is selected from the training data set and a machine learning algorithm is trained on the basis of said subset.

[0057] FIG. 2 shows a comparison between the conventional training method (FIG. 2(a)) and the optimized training method of the present application (FIG. 2(b)).

[0058] According to Fig. 2a, a neural network 200 is trained using a training dataset 202 consisting of training data. The model thus trained has a determined model output 205, which is shown diagrammatically in graph 204. Optionally, the training data of the training dataset 202 can be characterized or labelled depending on a predefined category 206, e.g. a predicted class. The label is indicated by reference sign 208.

[0059] 2(b) shows an embodiment of a method for training a machine learning algorithm in an optimized state as claimed herein. Here, a domain model 300 is provided, which can be defined by a number of domain parameters P1, P2, P3, e.g. with corresponding domain values ​​v11, v12, v13, v21, v22 and v31, v32, v33, respectively. Furthermore, a data model 302 is provided with a training dataset 304, which includes training data for at least one domain. Optionally, the training data of the training dataset 304 can be characterized or labelled depending on a predefined category or domain 306, e.g. a predictive class. The label is indicated by reference sign 308. In the method, exemplarily, at least one training data 310 is at least temporarily removed and / or masked from the training dataset 304 depending on at least one domain parameter P3 and / or domain value v33 to provide a reduced training dataset 312. The removal and / or masking is indicated by reference number 314. This allows to determine the influence on the model output, in particular depending on the separated domain parameters and / or domain values. In the present method, the individual training data of the training data set 304 are labeled depending on the domain parameters P1, P2, P3 and the corresponding domain values ​​v11, v12, v13; v21, v22 and v31, v32, v33. Thus, for example, each training data of the training data set 304 is associated with at least one identifier for a specific configuration of the domain parameters and domain values. The identifier is indicated by reference number 316. In the present method, for example, at least one training data having or characterized as the domain value v33 for the domain parameter is removed.As a result, on the basis of the reduced training data set 312, respectively, a neural network 318 is trained, whereby a model output is determined depending on the reduced data set 312. The determined model output 319 for the neural network 318 trained with the reduced data set 312 is shown diagrammatically in a graph 320. The model output 319 thus determined or shown by the neural network 318 trained on the basis of the reduced data set 312 is compared with the model output 205 associated with the (full) training data set. The comparison is indicated by reference sign 322 and corresponds to step S5.

[0060] Now, based on said comparison, training data can be selected from the training data set, which is no longer shown in Fig. 2. Preferably, by means of the data model 302 enriching the semantic domain model, a subset or set of training data and / or combination cases can be determined from the domain parameters and domain values ​​from which a global output behavior can be computed, preferably by in particular successively and / or partially removing and / or adding data points X' in X with special domain values ​​from the domain model.

[0061] A machine learning algorithm can then be trained based on the selected training data.

[0062] That is, in particular in the method, a domain model, preferably a semantic domain model, is enriched by the learned data model, where the domain model S is preferably enriched by a number of input variables D N It has S = {D1,…,D N} where D i =(D i1 ,…,D ik).

[0063] We can define X, Y, M as (training) data, where X is the input data, Y is an arbitrary label for prediction, and M is a description of x in the domain model S.

[0064] A data model can be defined as f(X) → y'. The cost function L(y') is sometimes L(y,y').

[0065] In this method, it is particularly preferred to use a data model D M =g(s)=g(D 11 ,…,D Nk ) → R is calculated and / or presented.

[0066] For the comprehensive combination of the domain models, a prediction of the model output resulting from the neural network is preferably formed. The calculation of the data model is preferably carried out according to a method known per se, but in this case, unlike the prior art, the data model is calculated for a concrete instance s in S. In other words, a mapping to the domain model is performed for the calculation of the data model.

[0067] As model output S' with |S'|<|S|, a subset of training data can be calculated and / or output, which can be effectively used for training machine learning algorithms. The subset of training data preferably forms scenarios for domain parameters and / or domain values ​​that have a significant influence on the model output. Furthermore, in the method, an estimate of the interaction degree k can be output, in particular for combination tests. Here, for example, a check can be made on important (or unimportant) combinations of domain parameters and / or domain values. With the thus reduced data model and / or a particular interaction degree, preferably (training) data X' is selected from X, which meets a predetermined degree of quality of the domain model. Optionally, a label for supervised learning of the reduced (training) data set can be generated otherwise. Then, a machine learning algorithm robust to the data model is trained with the obtained (training) data set.

Claims

1. 1. A method for optimally training a machine learning algorithm, comprising: The method comprises: - providing a domain model (S1) comprising domain parameters and / or domain values ​​for at least one domain; - providing (S2) a data model having a training data set comprising training data for said at least one domain; - removing and / or masking and / or modifying at least one training data from said training data set in dependence on the at least one domain parameter and / or domain value in order to provide a reduced training data set, in particular in order to evaluate the sensitivity of the at least one removed and / or masked domain parameter and / or domain value (S3); - training (S4) a neural network on the basis of said reduced training data set in order to determine a model output in dependence on said reduced data set; - a step (S5) of comparing the determined model output with the model outputs associated with said training data set; - selecting training data from said training data set depending on the comparison of said model outputs (S6); - training (S7) said machine learning algorithm on the basis of selected training data; - providing (S8) said trained machine learning algorithm; A method comprising:

2. The method of claim 1 , wherein the training data for the at least one domain is characterized depending on at least one combination of domain parameters and / or domain values, respectively.

3. The method of claim 1 or 2, wherein the training data set for the data model comprises at least one predicted domain identifier.

4. 4. The method according to claim 1 , wherein the step (S3) of removing and / or masking and / or modifying at least one training data from the training data set in dependence on the at least one domain parameter and / or domain value is performed continuously and / or iteratively by varying the at least one domain parameter and / or domain value.

5. 5. The method of claim 4, wherein the training of each neural network, the determination of each model output and the comparison of the model outputs are performed for each training data set thus successively and / or iteratively reduced, in such a way that a subset is selected from the training data set and the machine learning algorithm is trained on the basis of the subset.

6. The method according to claim 5 , wherein the subset from the training data set comprises training data that has no predefined influence on the model output, in particular in dependence on at least one boundary value.

7. 7. The method according to claim 1, wherein an interaction between at least two domain parameters and / or domain values ​​is calculable based on a comparison of the determined model output with model outputs associated with the training data set.

8. The method according to claim 1 , wherein the machine learning algorithm comprises a neural network, in particular a deep neural network.

9. 9. The method of any one of claims 1 to 8, further comprising providing a manufacturing line including an equipment configuration for producing a configurable product.

10. 10. The method of claim 9, wherein the method further comprises, after the manufacturing line is provided, producing at least one configurable product using the equipment configuration.

11. A control device included in an autonomous vehicle and / or a robotic system and / or an industrial machine, on which a machine learning algorithm trained according to the method of any one of claims 1 to 10 can be executed.

12. A system (1) for training a machine learning algorithm in an optimized state, comprising: The system (1) comprises: a providing device configured to provide a domain model having domain parameters and / or domain values ​​for at least one domain and a data model having a training data set comprising training data for said at least one domain; an evaluation calculation device configured to remove and / or mask and / or modify at least one training data from the training dataset in dependence on at least one domain parameter and / or domain value to provide a reduced training dataset, to train a neural network on the reduced training dataset in order to determine a model output in dependence on the reduced dataset, to compare the determined model output with a model output associated with the training dataset, to select training data from the training dataset in dependence on the comparison of the model outputs, and to train the machine learning algorithm on the selected training data; Equipped with The supplying device is further configured to supply the trained machine learning algorithm, the system (1).

13. A computer program comprising a program code for performing at least a part of the method according to any one of claims 1 to 10, when the computer program is run on a computer.

14. A computer readable data carrier comprising a program code for performing at least a part of the method according to any one of claims 1 to 10, when said computer program is run on a computer.

Images