DE102023100839A1 - Computer-implemented, iterative method for adapting a machine learning model, data processing device, computer program and computer-readable storage medium - Google Patents

Abstract

The invention relates to a computer-implemented, iterative method for adapting a machine learning model for a starting scenario, wherein the machine learning model is trained on the basis of a training data set with a plurality of data points with an annotation for training the machine learning model for the starting scenario to predict an annotation from the plurality of classes for an input data point.

Description

The invention relates to a computer-implemented, iterative method for adapting a machine learning model. The invention further relates to a data processing device, a computer program and a computer-readable storage medium.

The starting point of the invention is a machine learning (ML) application that has been developed, for example, for use in a motor vehicle. The ML application can be an application for computer-based vision for a driver assistance function or for autonomous driving. The ML application processes input data. The input data can be data from a sensor, for example images or sequences of images (videos). The task of the ML model can be, for example, object recognition (classification or semantic segmentation).

• - Szenario A ist die Verwendung in Land 1 und Szenario B ist die Verwendung in Land 2;
• - Szenario A ist die Verwendung in einer städtischen Umgebung und Szenario B ist die Verwendung in einer ländlichen Umgebung;
• - Szenario A ist die Verwendung auf kleinen Feldwegen und Szenario B ist die Verwendung auf Fernverkehrsstraßen; oder
• - Szenario A ist die Verwendung während des Tages und Szenario B ist die Verwendung während der Nacht.

Assume that the ML application was developed for a scenario A (start scenario) and is to be adapted to a scenario B (target scenario). The following cases are examples:

• - Scenario A is use in country 1 and Scenario B is use in country 2;
• - Scenario A is use in an urban environment and Scenario B is use in a rural environment;
• - Scenario A is use on small country lanes and Scenario B is use on main roads; or
• - Scenario A is use during the day and Scenario B is use during the night.

The ML model was trained using a training dataset collected for Scenario A. The challenge is to collect suitable training data to train the ML model in order to extend the use of the ML model to Scenario B.

ML models can have the property that they already generalize to different situations "around the training data". Therefore, it is not clear for which aspects of scenario B the ML model works and for which aspects in scenario B additional training data must be collected. This is particularly challenging in an early preparation stage of a rollout for scenario B, since little or no information exists about which data should be collected.

Collecting and processing (including classifying) new training data can be extensive. Therefore, a goal of preferred embodiments of the invention is to gain an understanding of what training data is required for scenario B. Alternatively, one can collect all possible data for scenario B, but this is not satisfactory because a large amount of data may not contribute much to supplementing the existing data for scenario A. Such a process can be inefficient and computationally intensive and would take a very long time.

With an understanding of what data is required, the appropriate data can be collected, added to the existing training dataset, and finally the ML model can be additionally trained for scenario B.

The US 2016 / 0 070 986 A1 , the US 2020 / 0 364 520 A1 and the US 2022 / 0 219 698 A1 are state of the art.

The invention aims to provide an efficient, resource-saving and fast method for extending a machine learning model for use in a given target scenario.

1. a) In einem Iterationsschritt einer Iteration, Sammeln eines Adaptionsdatensatzes mit einer Mehrzahl von (dem Zielszenario entstammenden) Datenpunkten zum Trainieren des maschinellen Lernmodells für das Zielszenario;
2. b) In dem Iterationsschritt der Iteration, Zuweisen einer Annotierung zu den jeweiligen Datenpunkten des gesammelten Adaptionsdatensatzes;
3. c) In dem Iterationsschritt der Iteration, Testen des auf Grundlage des Trainingsdatensatzes (für das Startszenario) trainierten, maschinellen Lernmodells mit dem in dem Iterationsschritt gesammelten Adaptionsdatensatz oder einer Teilmenge davon, wobei die Datenpunkte des in dem Iterationsschritt gesammelten Adaptionsdatensatzes oder der Teilmenge davon jeweils als Eingangsdatenpunkt für das maschinelle Lernmodell dienen;
4. d) Feststellen für jeden Eingangsdatenpunkt, ob die durch das in Schritt c) getestete, maschinelle Lernmodell vorhergesagte Annotierung zugewiesenen Annotierung entspricht;
5. e) In einem weiteren Iterationsschritt der Iteration, Sammeln eines weiteren Adaptionsdatensatzes mit einer Mehrzahl von Datenpunkten zum Trainieren des maschinellen Lernmodells für das Zielszenario auf Grundlage der Feststellung in Schritt d) und Fortfahren mit Schritt b) in dem weiteren Iterationsschritt der Iteration mit dem weiteren Adaptionsdatensatz; und
6. f) Beenden der Iteration durch Trainieren des maschinellen Lernmodells mit den Datenpunkten des Trainingsdatensatzes und den in den Iterationsschritten gesammelten Adaptionsdatensätzen zum Adaptieren des für das Startszenario vorgesehenen maschinellen Lernmodells an das Zielszenario.

According to one aspect, the invention provides a computer-implemented, iterative method for adapting a machine learning model, which is intended for a starting scenario or even already trained, to a target scenario, wherein the machine learning model can be trained on the basis of a training data set with a plurality of data points (originating from the starting scenario) with an annotation (ie they are so-called labeled data points) for training the machine learning model for the starting scenario to predict or infer an annotation for an input data point, the method comprising:

1. a) In an iteration step of an iteration, collecting an adaptation dataset with a plurality of data points (originating from the target scenario) for training the machine learning model for the target scenario;
2. b) In the iteration step of the iteration, assigning an annotation to the respective data points of the collected adaptation dataset;
3. c) In the iteration step of the iteration, testing the machine learning model trained on the basis of the training data set (for the starting scenario) with the adaptation data set collected in the iteration step or a subset thereof, wherein the data points of the adaptation data set collected in the iteration step tion dataset or subset thereof serve as input data points for the machine learning model;
4. d) determining, for each input data point, whether the annotation predicted by the machine learning model tested in step c) corresponds to the assigned annotation;
5. e) In a further iteration step of the iteration, collecting a further adaptation data set with a plurality of data points for training the machine learning model for the target scenario based on the determination in step d) and continuing with step b) in the further iteration step of the iteration with the further adaptation data set; and
6. f) Terminating the iteration by training the machine learning model with the data points of the training dataset and the adaptation datasets collected in the iteration steps to adapt the machine learning model intended for the start scenario to the target scenario.

One advantage of the method can be that only adaptation data sets are collected for training that are necessary and suitable for adapting the machine learning model from the starting scenario to the target scenario. Adaptation data sets with data points whose annotation is already sufficiently predicted by the existing, trained machine learning model can be disregarded. This keeps the amount of data in the collected adaptation data sets to a minimum. The method thus enables efficient adaptation of the machine learning model. The collection can be carried out by operating the motor vehicle or at least another test vehicle in environmental scenarios that correspond to the data points with false detection and recording the other data points.

A data point can correspond to an image, a sequence of images (video) or other data. The data points can have been recorded by any sensor. For example, the data points can have been recorded by a sensor in the motor vehicle.

Each data point has an annotation. If the data point corresponds to an image, for example, each pixel can be assigned a class as an annotation of the data point (as in semantic segmentation). Alternatively, a section (bounding box) can be defined as an annotation of the data point and a class can be assigned to this section. An annotation can thus be the assignment of the data point or a part to a class from a plurality of classes and/or the selection of a part of the data point. Annotations are generally used in the field of "supervised machine learning". The invention can be applied in particular to machine learning models in this field, but also to any machine learning model.

The machine learning model can be trained to predict or infer an annotation for an input data point. If the input data point corresponds to an image, for example, the machine learning model can predict the class or section and the respective class for each pixel. The machine learning model thus predicts (infers) a pixel-by-pixel classification and/or object recognition (i.e. it predicts bounding boxes, for example, to recognize where an object is in the image). The machine learning model can thus be used, for example, for semantic segmentation and/or object recognition.

• e1) falls in Schritt d) die vorhergesagte Annotierung von der zugewiesenen Annotierung für wenigstens einen oder einer vordefinierten Anzahl von Eingangsdatenpunkten abweicht, Sammeln des weiteren Adaptionsdatensatzes mit einer Mehrzahl von Datenpunkten, die der zugewiesenen Annotierung des wenigstens einen oder der vordefinierten Anzahl von abweichenden Eingangsdatenpunkten zuweisbar sind.

It is preferred that step e) comprises:

• e1) if in step d) the predicted annotation deviates from the assigned annotation for at least one or a predefined number of input data points, collecting the further adaptation data set with a plurality of data points that are assignable to the assigned annotation of the at least one or the predefined number of deviating input data points.

Advantageously, additional adaptation data sets are collected if the existing, trained machine learning model makes poor predictions for the adaptation data sets. One measure of how well or poorly the machine learning model works for an adaptation data set is, for example, the number of data points for which the annotation was predicted correctly or incorrectly compared to the assigned annotation. Here, a predefined number can define a threshold, and if this is exceeded or not met, another adaptation data set is collected with a plurality of data points that can be assigned to the assigned annotation of the input data points.

• c1) Aufteilen des gesammelten Adaptionsdatensatzes in eine erste Teilmenge und eine zweite Teilmenge;
• c2) Trainieren des maschinellen Lernmodells auf Grundlage des Trainingsdatensatzes und der ersten Teilmenge des gesammelten Adaptionsdatensatzes; und
• c3) Testen des in Schritt c2) trainierten, maschinellen Lernmodells mit der zweiten Teilmenge des gesammelten Adaptionsdatensatzes.

It is preferred that step c) comprises:

• c1) dividing the collected adaptation data set into a first subset and a second subset;
• c2) training the machine learning model based on the training dataset and the first subset of the collected adaptation dataset; and
• c3) Testing the machine learning model trained in step c2) with the second subset of the collected adaptation dataset.

By dividing the adaptation dataset into a subset for training and a subset for testing, the procedure can be made even more efficient.

• c4) Trainieren des maschinellen Lernmodells auf Grundlage des Trainingsdatensatzes und eines Adaptionsdatensatzes oder einer Teilmenge davon eines dem Iterationsschritt vorhergehenden Iterationsschritts.

It is preferred that step c) comprises:

• c4) Training the machine learning model based on the training data set and an adaptation data set or a subset thereof from an iteration step preceding the iteration step.

For use cases or application situations that may arise during the method and which are not explicitly described here, it may be provided that, in accordance with the method, an error message and/or a request to enter user feedback is issued and/or a default setting and/or a predetermined initial state is set.

According to a further aspect, the invention provides a use of the machine learning model, which is intended for the start-up scenario or even already trained, in a motor vehicle.

It is preferred that the use comprises adapting the machine learning model according to one of the preceding embodiments of the method.

It is preferred that the use comprises collecting the adaptation data set by means of the motor vehicle in the target scenario.

It is preferred that the use comprises transmitting the adapted machine learning model to the motor vehicle.

It is preferred that the use comprises applying the machine learning model in the target scenario, in particular for object detection of objects (e.g. road users) in an environment of the vehicle (so-called computer vision application).

According to a further aspect, the invention provides a data processing device comprising means for carrying out the method according to one of the preceding embodiments.

The invention also includes the data processing device or the processor device, which is set up to carry out an embodiment of the method according to the invention. For this purpose, the processor device can have at least one microprocessor and/or at least one microcontroller and/or at least one FPGA (Field Programmable Gate Array) and/or at least one DSP (Digital Signal Processor). Furthermore, the processor device can have program code that is set up to carry out the embodiment of the method according to the invention when executed by the processor device. The program code can be stored in a data memory of the processor device. The processor circuit of the processor device can, for example, have at least one circuit board and/or at least one SoC (System on Chip).

The invention also includes further developments of the data processing device according to the invention, which have features as have already been described in connection with the further developments of the method according to the invention. For this reason, the corresponding further developments of the data processing device according to the invention are not described again here.

According to a further aspect, the invention provides a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method according to one of the preceding embodiments.

According to a further aspect, the invention provides a computer-readable storage medium on which the computer program is stored.

As a further solution, the invention also includes a computer-readable storage medium, comprising program code which, when executed by a processor circuit of a computer or a computer network, causes it to carry out an embodiment of the method according to the invention. The storage medium can, for example, be provided at least partially as a non-volatile data memory (e.g. as a flash memory and/or as an SSD - solid state drive) and/or at least partially as a volatile data memory (e.g. as a RAM - random access memory). The storage medium can be arranged in the processor circuit in its data memory. The storage medium can also, for example, be operated as a so-called app store server on the Internet. The computer or computer network can provide a processor circuit with at least one microprocessor. The program code can be provided as binary code or assembler and/or as source code of a programming language (e.g. C) and/or as a program script (e.g. Python).

The invention also includes combinations of the features of the described embodiments. The invention therefore also includes implementations that each have a combination of the features of several of the described embodiments, provided that the embodiments have not been described as mutually exclusive.

• 1 eine Ausführungsform eines computer-implementierten, iterativen Verfahrens zum Adaptieren eines maschinellen Lernmodells für ein Startszenario an ein Zielszenario; und
• 2 eine Ausführungsform eines Schrittes des Verfahrens gemäß 1.

Embodiments of the invention are described below.

• 1 an embodiment of a computer-implemented, iterative method for adapting a machine learning model for a starting scenario to a target scenario; and
• 2 an embodiment of a step of the method according to 1 .

The exemplary embodiments explained below are preferred embodiments of the invention. In the exemplary embodiments, the components of the embodiments described each represent individual features of the invention that are to be considered independently of one another and which also develop the invention independently of one another. Therefore, the disclosure should also include combinations of the features of the embodiments other than those shown. Furthermore, the described embodiments can also be supplemented by other features of the invention already described.

In the figures, identical reference symbols designate functionally identical elements.

1 shows an embodiment of a computer-implemented, iterative method for adapting a machine learning model for a starting scenario to a target scenario.

The machine learning model may include different, trainable learning algorithms. The machine learning model may be used in a motor vehicle. For example, the machine learning model may be an image analysis application in a motor vehicle to enable driver assistance system functions or autonomous driving functions.

The machine learning model is trainable based on a training dataset with a plurality of annotated data points for training the machine learning model for the starting scenario.

A data point can correspond to an image, a sequence of images (video) or other data. The data points can have been recorded by any sensor. For example, the data points can have been recorded by a sensor in the motor vehicle.

Each data point has an annotation. An annotation can be a feature or label of a data point. For example, an annotation can identify an object in a data point.

If the data point corresponds to an image, for example, each pixel can be assigned a class as an annotation of the data point (as in semantic segmentation). Alternatively, a section (bounding box) can be defined as an annotation of the data point and a class can be assigned to this section. An annotation can thus be the assignment of the data point or a part to a class and/or the selection of a part of the data point. Annotations are generally used in the field of "supervised machine learning". The invention can be applied in particular to machine learning models in this field, but also to any machine learning model.

The machine learning model can be used, for example, for semantic segmentation and/or object recognition.

The launch scenario may, for example, be a use of the machine learning model in a launch country. The launch scenario may also be a use of the machine learning model in a launch environment, such as an urban environment, on a dirt road, in daylight, or the like.

If the machine learning model is trained with the training data set for the starting scenario, the machine learning model can predict an annotation for an input data point. The better the machine learning model is trained for the starting scenario, the better the prediction can be. The functioning of the machine learning model therefore depends on the training data set for the starting scenario.

An aim of preferred embodiments of the invention is to adapt the machine learning model trained for the start scenario to a target scenario.

The target scenario may, for example, be a use of the machine learning model in a target country that is different from the starting country. The target scenario may also be the use of the machine learning model in a target environment that is different from the starting environment. For example, the target environment may It could be a rural area rather than an urban area, a main road rather than a track, a night-time area rather than a daytime area, or something similar.

To adapt the machine learning model trained for the start scenario to the target scenario, at least one adaptation dataset with data points for training for the target scenario is required.

• - In einem Iterationsschritt einer Iteration, Sammeln eines Adaptionsdatensatzes mit einer Mehrzahl von Datenpunkten zum Trainieren des maschinellen Lernmodells für das Zielszenario.

In a first step S11, the method comprises:

• - In an iteration step of an iteration, collecting an adaptation dataset with a plurality of data points for training the machine learning model for the target scenario.

In step S11, an adaptation data set is collected with data points that are suitable for training the machine learning model for the target scenario. The adaptation data set preferably comprises a small number of data points compared to the training data set. The respective adaptation data set can be collected, for example, by a motor vehicle in the target scenario.

• - In dem Iterationsschritt der Iteration, Zuweisen einer Annotierung zu den jeweiligen Datenpunkten des gesammelten Adaptionsdatensatzes.

In a further step S12, the method comprises:

• - In the iteration step of the iteration, assigning an annotation to the respective data points of the collected adaptation dataset.

The annotation can be assigned manually or by computer. In step S12, the selected annotation is assigned to the data points of the collected adaptation data set. The assignment can be made by the motor vehicle.

• - In dem Iterationsschritt der Iteration, Testen des auf Grundlage des Trainingsdatensatzes trainierten, maschinellen Lernmodells mit dem in dem Iterationsschritt gesammelten Adaptionsdatensatz oder einer Teilmenge davon, wobei die Datenpunkte des in dem Iterationsschritt gesammelten Adaptionsdatensatzes oder der Teilmenge davon jeweils als Eingangsdatenpunkt für das maschinelle Lernmodell dienen.

In a further step S13, the method comprises:

• - In the iteration step of the iteration, testing the machine learning model trained on the basis of the training data set with the adaptation data set collected in the iteration step or a subset thereof, wherein the data points of the adaptation data set collected in the iteration step or the subset thereof each serve as input data point for the machine learning model.

The machine learning model is trained on the basis of the training data set. In addition, the machine learning model can be trained on the basis of an adaptation data set or a subset thereof from an iteration step preceding the iteration step. The machine learning model is tested with the adaptation data set collected in the iteration step. The testing can be carried out by the motor vehicle.

• - Feststellen für jeden Eingangsdatenpunkt, ob die durch das in Schritt S13 getestete, maschinelle Lernmodell vorhergesagte Annotierung der zugewiesenen Annotierung entspricht.

In a further step S14, the method comprises:

• - Determining for each input data point whether the annotation predicted by the machine learning model tested in step S13 corresponds to the assigned annotation.

• - In einem weiteren Iterationsschritt der Iteration, Sammeln eines weiteren Adaptionsdatensatzes mit einer Mehrzahl von Datenpunkten zum Trainieren des maschinellen Lernmodells für das Zielszenario auf Grundlage der Feststellung in Schritt S14 und Fortfahren mit Schritt S12 in dem weiteren Iterationsschritt der Iteration mit dem weiteren Adaptionsdatensatz.

In a further step S15, the method comprises:

• - In a further iteration step of the iteration, collecting a further adaptation data set with a plurality of data points for training the machine learning model for the target scenario based on the determination in step S14 and continuing with step S12 in the further iteration step of the iteration with the further adaptation data set.

The determination made in step S14 is thus used in step S15 to collect another adaptation data set.

For example, in step S14 it can be determined that the predicted annotations of the input data points correspond to the assigned annotations. This can indicate that the annotations of the data points of the adaptation data set collected in the iteration step are sufficiently predicted by the existing, trained machine learning model. Thus, in step S15, a further adaptation data set can be collected which is, for example, disjoint to the adaptation data set collected in the iteration step.

For example, in step S14 it can also be determined that the predicted annotation of the input data points deviates from the assigned annotation. This can indicate that the annotation of the data points of the adaptation data set collected in the iteration step is not sufficiently predicted by the existing, trained, machine learning model. Then, in step S15, another adaptation data set can be collected, which for example comprises a plurality of data points that can be assigned to the assigned annotation.

Steps S14 and S15 can be performed by the motor vehicle.

• - Beenden der Iteration durch Trainieren des maschinellen Lernmodells mit den Datenpunkten des Trainingsdatensatzes und den in den Iterationsschritten gesammelten Adaptionsdatensätzen zum Adaptieren des maschinellen Lernmodells für das Startszenario an das Zielszenario.

In a further step S16, the method comprises:

• - End the iteration by training the machine learning model with the data points of the training dataset and the adaptation datasets collected in the iteration steps to adapt the machine learning model for the start scenario to the target scenario.

For example, the iteration can be terminated when the collected adaptation sets sufficiently cover the target scenario and/or when the classes of data points from further collected adaptation data sets are sufficiently predicted by the existing trained machine learning model, i.e. when collecting new adaptation data sets has no further effect.

In one or more intermediate steps, the machine learning model can be trained, for example, with the training data set and the adaptation data sets that have already been collected. When training the machine learning model, coefficients of the machine learning model can be adjusted. A criterion for ending the iteration can also be the situation when, for example, no significant change in the coefficients can be detected during training in a further intermediate step.

Step S16 can be performed by the motor vehicle. The adapted machine learning model can finally be transferred to another motor vehicle for use in the target scenario.

2 shows an embodiment of a step of the method according to 1 .

• - Aufteilen des gesammelten Adaptionsdatensatzes in eine erste Teilmenge und eine zweite Teilmenge (Schritt S21);
• - Trainieren des maschinellen Lernmodells auf Grundlage des Trainingsdatensatzes und der ersten Teilmenge des gesammelten Adaptionsdatensatzes (Schritt S22); und
• - Testen des in Schritt S22 trainierten, maschinellen Lernmodells mit der zweiten Teilmenge des gesammelten Adaptionsdatensatzes (Schritt S23).

Step S13 may include the following steps S21, S22 and S23:

• - dividing the collected adaptation data set into a first subset and a second subset (step S21);
• - training the machine learning model based on the training data set and the first subset of the collected adaptation data set (step S22); and
• - Testing the machine learning model trained in step S22 with the second subset of the collected adaptation dataset (step S23).

The invention also includes a data processing device comprising means for carrying out the method. The invention also includes a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method. Furthermore, the invention also includes a computer-readable storage medium on which the computer program is stored.

• Die Daten für Szenario B (Zielszenario) werden iterativ gesammelt. Dabei wird nur ein kleiner Datensatz gesammelt und klassifiziert. Dieser Datensatz wird dann unter Verwendung des bestehenden ML-Modells für Szenario A (Startszenario) bewertet.

A solution principle of preferred embodiments of the invention can thus be summarized as follows:

• The data for Scenario B (target scenario) is collected iteratively. Only a small dataset is collected and classified. This dataset is then evaluated using the existing ML model for Scenario A (start scenario).

The collected and classified data for Scenario B is evaluated using the existing ML model for Scenario A. In other words, all collected data is used as an evaluation set. This allows identifying data points and cases of Scenario B for which the existing model does not work sufficiently. Such cases can be positions, locations, objects or environments. The findings are then used for further data collection (iterative approach).

Finally, the collected data is used to train the ML model. This extends the existing ML model and its functionality for scenario A to scenario B.

• Die Daten für Szenario B werden iterativ gesammelt. Dabei wird nur ein kleiner Datensatz gesammelt und klassifiziert. Dieser Datensatz wird dann unter Verwendung des bestehenden ML-Modells für Szenario A bewertet.

A further solution principle of preferred embodiments of the invention can be summarized as follows:

• The data for Scenario B is collected iteratively, collecting and classifying only a small dataset. This dataset is then evaluated using the existing ML model for Scenario A.

The collected data is split into two subsets that are used for training and evaluation (in both cases together with the existing training data for scenario A). This results in a model that has been used to train new data. The evaluation data is used to evaluate the new model for scenario B. This allows to identify data points and cases of scenario B for which the existing model does not work sufficiently. Such cases can be positions, places, objects or environments. The findings are then used for further data collection (iterative approach).

Finally, the collected data is used to train the ML model. This improves the existing ML model and its Functionality for scenario A extended to scenario B.

Overall, the examples show how machine learning domain adaptation can be provided while collecting and classifying data.

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions.

Cited patent literature

• US 20160070986 A1 [0008]
• US 20200364520 A1 [0008]
• US 20220219698 A1 [0008]

Claims (7)

Computer-implemented, iterative method for adapting a machine learning model for a starting scenario to a target scenario, wherein the machine learning model can be trained or is already trained to predict an annotation for an input data point based on a training data set with a plurality of data points with a respective annotation for training the machine learning model for the starting scenario, the method comprising: a) In an iteration step of an iteration, collecting an adaptation data set with a plurality of data points for training the machine learning model for the target scenario; b) In the iteration step of the iteration, assigning an annotation to the respective data points of the collected adaptation data set; c) In the iteration step of the iteration, testing the machine learning model trained on the basis of the training data set with the adaptation data set collected in the iteration step or a subset thereof, wherein the data points of the adaptation data set collected in the iteration step or the subset thereof each serve as an input data point for the machine learning model; d) Determining for each input data point whether the annotation predicted by the machine learning model tested in step c) corresponds to the assigned annotation; e) In a further iteration step of the iteration, collecting a further adaptation data set with a plurality of data points for training the machine learning model for the target scenario based on the determination in step d) and continuing with step b) in the further iteration step of the iteration with the further adaptation data set; and f) terminating the iteration by training the machine learning model with the data points of the training data set and the adaptation data sets collected in the iteration steps to adapt the machine learning model for the starting scenario to the target scenario. Procedure according to Claim 1 , characterized in that step e) comprises: e1) if in step d) the predicted annotation deviates from the assigned annotation for at least one or for a predefined number of input data points, collecting the further adaptation data set with a plurality of data points that are assignable to the assigned annotation of the at least one or the predefined number of deviating input data points. Method according to one of the preceding claims, characterized in that step c) comprises: c1) dividing the collected adaptation data set into a first subset and a second subset; c2) training the machine learning model based on the training data set and the first subset of the collected adaptation data set; and c3) testing the machine learning model trained in step c2) with the second subset of the collected adaptation data set. Method according to one of the preceding claims, characterized in that step c) comprises: c4) training the machine learning model on the basis of the training data set and an adaptation data set or a subset thereof of an iteration step preceding the iteration step. Data processing device comprising means for carrying out the method according to one of the preceding claims. Computer program comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method according to one of the Claims 1 until 4 to execute. Computer-readable storage medium on which the computer program is Claim 6 is stored.

Images