DE102023201104A1 - Method for generating additional training data for training a machine learning algorithm - Google Patents

Abstract

The invention relates to a method for generating additional training data for training a machine learning algorithm, wherein the method (1) comprises the following steps: acquiring data relating to a system to be modeled (2), generating labeled training data for training the machine learning algorithm based on the acquired data (3), and generating additional training data for training the machine learning algorithm, wherein the generation of additional training data comprises generating uninformative training data based on the generated labeled training data (4).

Description

The invention relates to a method for generating additional training data for training a machine learning algorithm, and in particular to a method for generating additional, uninformative training data for training a machine learning algorithm.

Machine learning algorithms are based on the use of statistical methods to train a data processing system so that it can perform a specific task without it originally being explicitly programmed to do so. The goal of machine learning is to construct algorithms that can learn from data and make predictions. These algorithms create mathematical models that can be used, for example, to classify data and map input data into output data.

A system to be modelled can be recorded, for example, through measurements, whereby an empirical model can be created based on measured values and a machine learning algorithm can be trained accordingly. However, situations can arise in which it is impossible to completely measure a process or system to be modelled. However, this can result in only partial data from a subspace being available for the empirical modeling or the corresponding training of the machine learning algorithm, although process states can also occur during operation that are not captured by this training data.

Augmentation methods, i.e. methods for generating additional training data, have been proposed as a solution to this problem. However, the disadvantage of known augmentation methods is that they are very complex and require a lot of computer resources, especially storage and computing capacity, so that they are difficult to implement with conventional data processing systems.

Furthermore, when controlling a controllable system using a machine learning algorithm, machine learning algorithms often incorrectly assign unknown input data to certain output data, even though the input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability, which can lead to safety-critical situations when controlling a safety-critical application based on the machine learning algorithm, since such situations are usually not taken into account when training the machine learning algorithm.

From the publication EN 10 2019 209 560 A1 A method for training a neural network is known, the method comprising providing a training data set comprising training images showing a vehicle environment from the perspective of a vehicle, wherein a plurality of the training images show traffic signs, generating additional training images by augmenting training images showing traffic signs by augmenting a training image showing a traffic sign by partially covering the traffic sign and/or augmenting a training image showing a variable message sign by changing the lighting state of one or more lighting elements of the variable message sign, and training the neural network based on at least the augmented training images.

The invention is therefore based on the object of specifying an improved method for generating additional training data for training a machine learning algorithm.

The invention is solved by a method for generating additional training data for training a machine learning algorithm according to the features of patent claim 1.

The object is also achieved by a system for generating additional training data for training a machine learning algorithm according to the features of patent claim 8.

Disclosure of the invention

According to one embodiment of the invention, this object is achieved by a method for generating additional training data for training a machine learning algorithm, wherein the method comprises acquiring data relating to a system to be modeled, generating labeled training data for training the machine learning algorithm based on the acquired data, and generating additional training data for training the machine learning algorithm, wherein the generation of additional training data comprises generating uninformative training data based on rend on the generated labeled training data.

The steps of the method can in particular be carried out on a computer. In other words, the method can in particular be a computer-implemented method.

The data relating to a system to be modelled or an application of the machine learning algorithm refers to data which are to be processed by the trained machine learning algorithm in the future, i.e. possible input data of the trained machine learning algorithm.

Labeled training data also refers to data or training data that are annotated or provided with corresponding ground truth information.

Uninformative training data also includes training data that is based on input data or input data that cannot be clearly assigned to any output data or output values of the machine learning algorithm. The assignment of input data to output data by a machine learning algorithm is usually based on probabilistic methods or probability values, with uninformative training data based on input data that cannot be clearly assigned to any output data with a certain probability and can in particular be assigned to several different output values with the same probability.

The uninformative training data therefore represents situations in which input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability. This has the advantage that during training of the machine learning algorithm, it is also possible to train how the machine learning algorithm should react in situations in which an input value cannot be clearly assigned to an output value with a certain probability and, in particular, can be assigned to several different output values with the same probability.

Overall, an improved method for generating additional training data for training a machine learning algorithm is provided.

In one embodiment, the step of generating additional training data further comprises selecting first labeled training data and second labeled training data from the generated labeled training data, determining a midpoint between the input data of the first labeled training data and the input data of the second labeled training data, and then shifting the determined midpoint orthogonal to a connection vector connecting the input data of the first labeled training data and the input data of the second labeled training data by adding a noise signal to the determined midpoint.

The center point is understood here in particular to be an arithmetic mean of the input data, whereby input data represented by the center point cannot usually be clearly assigned to any class or output data.

A noise signal is understood to be a disturbance or an additional deviation.

That the specific center point is shifted orthogonally to the connection vector by adding a noise signal to the specific center point further means that the specific center point is shifted along an orthogonal line by a value predetermined by the noise signal.

Thus, based on simple algebraic operations, additional training data can be generated that not only represents a mixture of the known training data, but systematically uninformative training data that does not belong to any class, without the need for complex and resource-intensive operations or adjustments. Since the additional, uninformative training data is generated based on simple algebraic operations, comparatively few computer resources are required to generate the additional training data, so that the method can also be carried out, for example, on control units of a motor vehicle, which have comparatively low computer resources.

The noise signal can be a random vector.

A random vector is a vector in which N or a certain number of random variables are summarized.

The systematically generated, additional uninformative training data can thus be generated in a simple manner by adding a random vector or based on simple vector operations, without the need for complex and resource-intensive operations.

The step of generating additional training data may also comprise generating additional training data by interpolation between generated uninformative training data and generated labeled training data.

Interpolation is understood to mean the approximate determination of an unknown value using or based on known values at neighboring locations.

Thus, based on the generated additional uninformative training data, additional training data can be generated based on simple mathematical operations without the need for complex and resource-intensive operations.

The data relating to the system to be modelled may also include sensor data.

A sensor, which is also called a (measuring) probe, is a technical component that can detect certain physical or chemical properties and/or the material properties of its environment qualitatively or quantitatively as a measured value.

In this way, conditions outside the actual data processing system on which the additional training data is generated can be easily recorded and taken into account when generating additional training data for training the machine learning algorithm.

A further embodiment of the invention also provides a method for training a machine learning algorithm, the method comprising providing training data for training the machine learning algorithm, generating additional training data for training the machine learning algorithm by a method described above for generating additional training data for training a machine learning algorithm, and training the machine learning algorithm based on the provided training data and the generated additional training data.

Thus, a method for training a machine learning algorithm is specified, which is based on additional training data generated by an improved method for generating additional training data for training a machine learning algorithm. In particular, additional uninformative training data is generated which represents situations in which input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability. This has the advantage that during training of the machine learning algorithm, it can also be trained how the machine learning algorithm should react in situations in which an input value cannot be clearly assigned to an output value with a certain probability and, in particular, can be assigned to several different output values with the same probability.

The training data provided can correspond to the labeled training data generated during the application of the method for generating additional training data for training a machine learning algorithm. Furthermore, the training data can also be provided independently of the application of the method for generating additional training data for training a machine learning algorithm.

A further embodiment of the invention also provides a method for controlling a controllable system based on a machine learning algorithm, the method comprising providing a machine learning algorithm for controlling the controllable system, the machine learning algorithm having been trained by a method for training a machine learning algorithm described above, and controlling the controllable system based on the provided machine learning algorithm.

A controllable system is understood to be a robotic system, for example a driver assistance system of a motor vehicle or a function of an autonomously driving motor vehicle.

Thus, a method is provided for controlling a controllable system which is based on additional training data generated by an improved method for generating additional training data for training a machine learning algorithm. nized machine learning algorithm. In particular, additional uninformative training data is generated which represents situations in which input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability. This has the advantage that during the training of the machine learning algorithm, it can also be trained how the machine learning algorithm should react in such situations in which an input value cannot be clearly assigned to an output value with a certain probability and in particular can be assigned to several different output values with the same probability.

A further embodiment of the invention also provides a system for generating additional training data for training a machine learning algorithm, the system comprising a detection unit designed to detect data relating to a system to be modeled, a first generation unit designed to generate labeled training data for training the machine learning algorithm based on the detected data, and a second generation unit designed to generate additional training data for training the machine learning algorithm, the generation of additional training data comprising generating uninformative training data based on the generated labeled training data.

This provides an improved system for generating additional training data for training a machine learning algorithm. The uninformative training data represents situations in which input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability. This has the advantage that during training of the machine learning algorithm, it is also possible to train how the machine learning algorithm should react in situations in which an input value cannot be clearly assigned to an output value with a certain probability and, in particular, can be assigned to several different output values with the same probability.

In one embodiment, the second generation unit has a selection unit which is designed to select first labeled training data and second labeled training data from the generated labeled training data, a determination unit which is designed to determine a midpoint between the input data of the first labeled training data and the input data of the second labeled training data, and a displacement unit which is designed to shift the determined midpoint orthogonally to a connection vector which connects the input data of the first labeled training data and the input data of the second labeled training data to one another by adding a noise signal to the determined midpoint. Thus, based on simple algebraic operations, additional training data can be generated which not only represent a mixture of the known training data, but systematically uninformative training data which do not belong to any class, without the need for complex and resource-intensive operations. Since the additional, uninformative training data are generated based on simple algebraic operations, comparatively few computer resources are required to generate the additional training data, so that the method can also be carried out, for example, on control units of a motor vehicle, which have comparatively low computer resources.

The noise signal can be a random vector. The systematically generated, additional uninformative training data can thus be generated in a simple manner by adding a random vector or based on simple vector operations, without the need for complex and resource-intensive operations.

The second generation unit can also be designed to generate additional training data by interpolation between generated uninformative training data and generated labeled training data. Thus, based on the generated, additional uninformative training data, further additional training data can be generated based on simple mathematical operations, without the need for complex and resource-intensive operations.

The data relating to an application of the machine learning algorithm can also comprise sensor data. In this way, circumstances outside the actual data processing system on which the additional training data for training the machine learning algorithm are generated can be easily recorded and used when generating additional training data for training. ren of the machine learning algorithm.

A further embodiment of the invention also provides a system for training a machine learning algorithm, the system comprising a first provision unit designed to provide training data for training the machine learning algorithm, a second provision unit designed to provide additional training data for training the machine learning algorithm, the additional training data having been generated by a system described above for generating additional training data for training a machine learning algorithm, and a training unit designed to train the machine learning algorithm based on the provided training data and the provided additional training data.

Thus, a system for training a machine learning algorithm is specified, which is based on additional training data generated by an improved system for generating additional training data for training a machine learning algorithm. In particular, additional uninformative training data is generated which represents situations in which input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability. This has the advantage that, during training of the machine learning algorithm, it can also be trained how the machine learning algorithm should react in situations in which an input value cannot be clearly assigned to an output value with a certain probability and, in particular, can be assigned to several different output values with the same probability.

A further embodiment of the invention also provides a system for controlling a controllable system based on a machine learning algorithm, the system comprising a provision unit which is designed to provide a machine learning algorithm for controlling the controllable system, the machine learning algorithm having been trained by a system for training a machine learning algorithm described above, and a control unit for controlling the controllable system based on the provided machine learning algorithm.

Thus, a system is specified for controlling a controllable system, which is based on a machine learning algorithm trained on the basis of additional training data generated by an improved system for generating additional training data for training a machine learning algorithm. In particular, additional uninformative training data is generated which represents situations in which input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability. This has the advantage that, during training of the machine learning algorithm, it can also be trained how the machine learning algorithm should react in situations in which an input value cannot be clearly assigned to an output value with a certain probability and, in particular, can be assigned to several different output values with the same probability.

In summary, the present invention provides a method for generating additional training data for training a machine learning algorithm, and in particular a method for generating additional, uninformative training data for training a machine learning algorithm.

The described designs and further training courses can be combined as desired.

Further possible embodiments, further developments and implementations of the invention also include combinations of features of the invention described above or below with respect to the embodiments that are not explicitly mentioned.

Short description of the drawings

The accompanying drawings are intended to provide a further understanding of embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain principles and concepts of the invention.

Other embodiments and many of the above advantages will become apparent upon consideration of the drawings. The illustrated elements of the drawings are not necessarily shown to scale with respect to one another.

• 1 zeigt ein Flussdiagramm eines Verfahrens zum Erzeugen von zusätzlichen Trainingsdaten zum Trainieren eines Algorithmus des maschinellen Lernens gemäß Ausführungsformen der Erfindung; und
• 2 zeigt ein schematisches Blockschaltbild eines Systems zum Erzeugen von zusätzlichen Trainingsdaten zum Trainieren eines Algorithmus des maschinellen Lernens gemäß Ausführungsformen der Erfindung.

They show:

• 1 shows a flowchart of a method for generating additional training data for training a machine learning algorithm according to embodiments of the invention; and
• 2 shows a schematic block diagram of a system for generating additional training data for training a machine learning algorithm according to embodiments of the invention.

In the figures of the drawings, identical reference symbols designate identical or functionally identical elements, parts or components, unless otherwise stated.

1 shows a flowchart of a method for generating additional training data for training a machine learning algorithm 1 according to embodiments of the invention.

To train a machine learning algorithm, a system to be modelled can be recorded, for example, through measurements, whereby an empirical model can be created based on measured values and a machine learning algorithm can be trained accordingly. However, situations can arise in which it is impossible to completely measure a process or system to be modelled. However, this can result in only partial data from a subspace being available for the empirical modeling or the corresponding training of the machine learning algorithm, although process states can also occur during operation that are not captured by this training data.

Augmentation methods, i.e. methods for generating additional training data, have been proposed as a solution to this problem. However, the disadvantage of known augmentation methods is that they are very complex and require a lot of computer resources, especially storage and computing capacity, so that they are difficult to implement with conventional data processing systems.

Furthermore, when controlling a controllable system using a machine learning algorithm, machine learning algorithms often incorrectly assign unknown input data to certain output data, even though the input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability, which can lead to safety-critical situations when controlling a safety-critical application based on the machine learning algorithm, since such situations are usually not taken into account when training the machine learning algorithm.

1 shows a method 1 which comprises a step 2 of collecting data relating to a system to be modeled, a step 3 of generating labeled training data for training the machine learning algorithm based on the collected data, and a step 4 of generating additional training data for training the machine learning algorithm, wherein the generation of additional training data comprises generating uninformative training data based on the generated labeled training data.

The uninformative training data represents situations in which input data does not exactly correspond to any class or output data or cannot be clearly assigned to a possible output value with a certain probability. This has the advantage that during training of the machine learning algorithm, it can also be trained how the machine learning algorithm should react in situations in which an input value cannot be clearly assigned to an output value with a certain probability and, in particular, can be assigned to several different output values with the same probability.

Overall, an improved method for generating additional training data for training a machine learning algorithm 1 is provided.

The machine learning algorithm can, for example, be an artificial neural network. The uninformative training data is particularly training data that reinforces a uniform weighting of the network output.

According to the embodiments of the 1 the step 4 of generating additional training data for training the machine learning algorithm comprises a step 5 of selecting first labeled training data and second labeled training data from the generated labeled training data, a step 6 of determining a midpoint between the input data of the first labeled training data and the input data of the second labeled training data, and a subsequent step 7 of shifting the determined midpoint orthogonal to a connection vector which connects the input data of the first labeled training data and the input data of the second labeled training data, by adding a noise signal to the determined center point.

The idea is not only to interpolate between two known data points or to determine the midpoint between the two data points, so that a mixture of the classes or output data of the two data points is created, but to redirect the interpolation via a completely uninformative point.

Selecting the first and second labeled training data may involve randomly selecting training data from the generated training data.

According to the embodiments of the 1 Furthermore, it is a random vector.

The random vector can be added with a variable step size or a random length, whereby the tip of the random vector after addition points to or lies on a completely uninformative point, which represents input data from corresponding uninformative training data, whereby the output data equally represents the output data of the first and the second labeled training data and neither of them can be clearly assigned.

Furthermore, several such completely uniform points can be found by applying the random vector in the high-dimensional space in different, in particular randomly selected directions.

How 1 shows, the method 1 further comprises a step 8 of generating additional training data by interpolation between generated uninformative training data and generated labeled training data, for example between a point to which the tip of the random vector points after addition and the corresponding first labeled training data and/or second labeled training data.

According to the embodiments of the 1 The data relating to an application of the machine learning algorithm also include sensor data.

Based on the generated training data and the generated additional, uninformative training data, a machine learning algorithm can then be trained.

For example, a machine learning algorithm can be trained to control a controllable system based on corresponding labeled training data.

The controllable system can, for example, be a function of an autonomously driving motor vehicle, whereby, if the trained machine learning algorithm cannot clearly assign input data to individual output data, the autonomously driving motor vehicle is transferred to a safe mode, for example to a mode in which it can only drive at a maximum speed of 30 km/h.

2 shows a schematic block diagram of a system for generating additional training data for training a machine learning algorithm 10 according to embodiments of the invention.

How 2 shows, the system 10 has a detection unit 11 which is designed to detect data relating to a system to be modeled, a first generation unit 12 which is designed to generate labeled training data for training the machine learning algorithm based on the detected data, and a second generation unit 13 which is designed to generate additional training data for training the machine learning algorithm, wherein the generation of additional training data comprises generating uninformative training data based on the generated labeled training data.

The detection unit can in particular be a receiver which is designed to receive corresponding data, for example sensor data. The first and second generation units can also each be implemented, for example, based on a code stored in a memory and executable by a processor.

How 2 As further shows, the second generating unit 13 according to the embodiments of the 2 a selection unit 14 which is designed to select first labeled training data and second labeled training data from the generated labeled training data, a determination unit 15 which is designed to determine a midpoint between the input data of the first labeled training data and the input data of the second labeled training data, and a shift unit 16 which is designed is to shift the determined center point orthogonally to a connection vector which connects the input data of the first labeled training data and the input data of the second labeled training data, by adding a noise signal to the determined center point.

The selection unit, the determination unit and the displacement unit can each be implemented based on a code stored in a memory and executable by a processor.

According to the embodiments of the 2 The noise signal is again a random vector.

In addition, the illustrated generation unit 13 is further designed to generate additional training data by interpolation between uninformative training data and provided labeled training data.

According to the embodiments of the 2 The data relating to an application of the machine learning algorithm also include sensor data.

The illustrated system 10 is in particular designed to carry out a method described above for generating additional training data for training a machine learning algorithm.

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

• DE 102019209560 A1 [0006]

Claims (14)

Method for generating additional training data for training a machine learning algorithm, the method (1) comprising the following steps: - collecting data relating to a system to be modeled (2); - generating labeled training data for training the machine learning algorithm based on the collected data (3); and - generating additional training data for training the machine learning algorithm, wherein generating additional training data comprises generating uninformative training data based on the generated labeled training data (4). Procedure (1) according to Claim 1 , wherein the step of generating additional training data (4) comprises the following steps: - selecting first labeled training data and second labeled training data from the generated labeled training data (5); - determining a midpoint between the input data of the first labeled training data and the input data of the second labeled training data (6); and - shifting the determined midpoint orthogonally to a connection vector which connects the input data of the first labeled training data and the input data of the second labeled training data, by adding a noise signal to the determined midpoint (7). Procedure (1) according to Claim 2 , where the noise signal is a random vector. Method (1) according to one of the Claims 1 until 3 , wherein the step of generating additional training data (4) further comprises generating additional training data by interpolation between generated uninformative training data and generated labeled training data (8). Method (1) according to one of the Claims 1 until 4 , wherein the data relating to an application of the machine learning algorithm comprises sensor data. A method for training a machine learning algorithm, the method comprising the following steps: - providing training data for training the machine learning algorithm: - generating additional training data for training the machine learning algorithm by a method for generating additional training data for training a machine learning algorithm according to one of the Claims 1 until 5 ; and - training the machine learning algorithm based on the provided training data and the generated additional training data. Method for controlling a controllable system based on a machine learning algorithm, the method comprising the following steps: - providing a machine learning algorithm for controlling the controllable system, the machine learning algorithm being developed by a method for training a machine learning algorithm according to Claim 6 trained; and - controlling the controllable system based on the provided machine learning algorithm. System for generating additional training data for training a machine learning algorithm, the system (10) comprising an acquisition unit (11) designed to acquire data relating to a system to be modeled, a first generation unit (12) designed to generate labeled training data for training the machine learning algorithm based on the acquired data, and a second generation unit (13) designed to generate additional training data for training the machine learning algorithm, wherein generating additional training data comprises generating uninformative training data from the generated labeled training data. System (10) according to Claim 8 , wherein the second generation unit (13) has a selection unit (14) which is designed to select first labeled training data and second labeled training data from the generated labeled training data, a determination unit (15) which is designed to determine a midpoint between the input data of the first labeled training data and the input data of the second labeled training data, and a displacement unit (16) which is designed to shift the determined midpoint orthogonally to a connection vector which connects the input data of the first labeled training data and the input data of the second labeled training data to one another by adding a noise signal to the determined midpoint. System (10) according to Claim 9 , where the noise signal is a random vector. System (10) according to one of the Claims 8 until 10 , wherein the second generation unit (13) is further configured to generate additional training data by interpolation between generated uninformative training data and generated labeled training data. System (10) according to one of the Claims 8 until 11 , wherein the data relating to an application of the machine learning algorithm comprises sensor data. System for training a machine learning algorithm, wherein the system comprises a first provision unit which is designed to provide training data for training the machine learning algorithm, a second provision unit which is designed to provide additional training data for training the machine learning algorithm, wherein the additional training data is generated by a system for generating additional training data for training a machine learning algorithm according to one of the Claims 8 until 12 and a training unit configured to train the machine learning algorithm based on the provided training data and the provided additional training data. System for controlling a controllable system based on a machine learning algorithm, wherein the system comprises a provision unit which is designed to provide a machine learning algorithm for controlling the controllable system, wherein the machine learning algorithm is implemented by a system for training a machine learning algorithm according to Claim 13 trained, and a control unit for controlling the controllable system based on the provided machine learning algorithm.

Images