Method and device for training a neural network

Abstract

In summary, the present invention relates to a method and a device for training a neural network, in particular a neural network for use in an automated system. For this purpose, it is provided to generate further training data elements from training data elements, each with an at least two-dimensional matrix, by modifying the matrices in the training data elements. An extended training data set can thus be generated from the original training data elements and the additional training data elements. The neural network can then be trained with this extended training data set.

Description

State of the art

[0001] Although the present invention is preferably described below in connection with a fully or partially automated vehicle, the present invention is not limited thereto. Rather, the basic principle of the invention can also be applied to any other fully or at least partially automated systems, such as robotic systems.

[0002] Systems for the fully or at least partially automated operation of a vehicle are becoming increasingly important. Neural networks are increasingly being used to implement functionalities for the automated operation of such vehicles. These neural networks must generally be trained before an application (inference). For this purpose, training data sets can be provided to the neural network during a training phase in order to train the neural network. A training data set can, for example, comprise several training data elements, with each training data element being able to be assigned a target output. For example, a training data set can comprise image data. For example, each training data element can comprise an image, with each image being assigned at least one characteristic property (e.g.a recognized object in an object recognition task) that the neural network is supposed to perform.

[0003] The publication DE 10 2019 209 560 A1 describes a method for training a neural network using image data from a vehicle's perspective. It proposes generating additional training images by modifying traffic signs in the image data. Disclosure of the invention

[0004] The present invention provides a method, a device, and a computer program product for training a neural network with the features of the independent patent claims. Further advantageous embodiments are the subject of the dependent patent claims.

[0005] The solution to the problem is described below using the method and the device. Features, advantages, or alternative embodiments mentioned therein are also to be applied to the other claimed subject matter, and vice versa. In other words, the subject claims (which are directed, for example, to a device or a computer program product) can also be developed with the features described or claimed in connection with the method, and vice versa. The corresponding functional features of the method are implemented by corresponding subject modules, in particular by hardware modules or microprocessor modules, of the system, and vice versa.The alternatives or embodiments of the invention described in connection with the method are not explicitly repeated for the device, but can also be applied within the scope of the device and vice versa. In general, in computer science, a software implementation and a corresponding hardware implementation (e.g., as an embedded system) are equivalent. For example, a method step for "storing" data can be carried out using a memory unit and corresponding instructions for writing data to the memory. To avoid redundancy, features of the device are therefore not explicitly described again, even though they can also be used in the alternative embodiments described with reference to the method. In principle, the claimed device is designed to carry out the claimed method.

[0006] According to a first aspect, the present invention provides a method for training a neural network, in particular a neural network for an automated system. The automated system can, for example, be a fully or at least partially autonomously driving vehicle. In principle, however, other at least partially autonomously operating devices such as robots, in particular used in production plants, etc., are also possible. The method comprises a step of providing an original training data set. The original training data set comprises a plurality of training data elements. Each training data element comprises an at least two-dimensional input matrix. Furthermore, the method comprises a step of generating at least one modified input matrix.Each modified input matrix can be formed from an at least two-dimensional input matrix of a training data element from the original training data set. Each modified input matrix is ​​formed by applying a predeterminable modification rule from the at least two-dimensional input matrix of a respective training data element. A further training data element is generated for each modified input matrix. The method further comprises a step for forming an extended training data set. The extended training data set comprises the training data elements from the original training data set as well as the generated additional training data elements. Finally, the method comprises a step for training the neural network. The neural network is trained using the generated extended training data set.In particular, the neural network can be trained to a target output using the generated extended training dataset.

[0007] According to a further aspect, the present invention provides a device for training a neural network. The device comprises at least one input device, an augmentation device, and a training device. The input device is designed to receive an original training data set. The original training data set can comprise a plurality of training data elements. Each training data element comprises an at least two-dimensional input matrix. The augmentation device is designed to generate a modified input matrix from the at least two-dimensional input matrix of a training data element. In particular, the augmentation device can generate the modified input matrix from the at least two-dimensional input matrix of a training data element by applying a predeterminable modification rule.This creates an additional training data element for each changed input matrix.

[0008] Furthermore, the augmentation device is designed to create an augmented training data set comprising the training data elements from the original training data set and the additional training data elements. The training device is designed to train the neural network. In particular, the training device is designed to train the neural network using the generated augmented training data set. The training device can train the neural network to a target output using the generated augmented training data set.

[0009] According to yet another aspect, the present invention provides a computer-readable storage medium. The storage medium comprises instructions that, when executed by a computer, cause the computer to carry out the method according to the invention.

[0010] According to yet another aspect, the present invention provides a computer program product. The computer program product comprises a computer program with instructions that, when executed by a computer, cause the computer to execute the method according to the invention.

[0011] According to yet another aspect, the present invention provides a computer program comprising instructions which, when the computer program is executed by a computer, cause the computer to carry out the method according to one of the preceding method claims.

[0012] The present invention is based on the recognition that training data sets are of crucial importance for training a neural network. Firstly, training the neural network generally requires training data sets with a very large number of training data elements. Furthermore, the quality of the training data elements contained in a training data set is also crucial. In particular, it is desirable for a training data set for a target output to be trained to preferably comprise several at least slightly different training data elements. This can increase the robustness of the property to be trained or the target output of the neural network.

[0013] Based on this finding, one idea of ​​the present invention is to create a concept for the automated generation of training data elements for a neural network. In particular, one goal of the inventive concept for generating additional training data elements for a training data set is to increase the robustness of the trained neural network.

[0014] For this purpose, one or more additional training data elements are generated from a training data element of a training data set for a neural network using a predeterminable modification rule. These additional training data elements can be combined with the training data elements of the original training data set to form an extended training data set. In the simplest case (a combination), the additional training data elements can be added to form the extended training data set. The neural network can then be trained using the extended training data set. By expanding the training data set to the training data elements according to the predeterminable modification rule, additional training data elements are obtained that differ (slightly) from the training data elements of the original training data set.Thus, the neural network can also be trained to the desired target output using these modified training data elements.

[0015] It is intended that further training data elements are generated by modifying the at least two-dimensional input matrix while maintaining the corresponding target output or for the same task of the neural network.

[0016] The modification rule is preferably applied to the at least two-dimensional input matrix. Since the original training data set may comprise multiple training data elements and / or multiple at least two-dimensional matrices, the same modification rule can be applied to all of the multiple at least two-dimensional matrices. Alternatively, different modification rules can be applied to the different at least two-dimensional matrices or to different groups of the at least two-dimensional matrices. This can be configured in a configuration phase.

[0017] A training data element can, for example, be in the form of a two-dimensional or multi-dimensional matrix. In particular, image data can, for example, be provided in the form of such a matrix. Each data element of such a matrix can contain numbers within a predetermined value range. For example, the matrix can be a binary matrix containing only the values ​​0 and 1 (e.g., black, white). Alternatively, however, matrices with a different value range, for example between 0 and 255, or any other value range are also possible. In particular, for image data, for example, multiple color channels can also be provided. In addition to image data, however, any other data can of course also be used as the basis for the training data elements for training the neural network.

[0018] The generation of further, additional training data elements for a training data set can be done automatically, i.e., without user intervention. The expanded training data set can increase the robustness of the neural network trained on the basis of the expanded training data set. In particular, the reliable, stable operation of any technical system using such a trained neural network can also be improved. This is very advantageous for fully or at least partially autonomous systems, such as fully or at least partially autonomous vehicles.

[0019] According to one embodiment, each training data element comprises a label. Such a label represents a target output of the neural network to be trained, assigned to each input according to the task. In principle, each training data element of a training data set can comprise an individual label. If further training data elements are generated on the basis of a training data element, in particular the at least two-dimensional matrix of a training data element, these further training data elements can be assigned the same label that corresponds to the underlying training data element. Depending on the application, however, modifications to the label for the further training data elements may also be possible. For example, information regarding the modification that can be taken into account during the training process can also be specified in the label.

[0020] According to one embodiment, the predeterminable modification rule comprises modifying one or more elements of the input matrix. In particular, the elements in the input matrix can be modified using a predefined probability distribution. Such a probability distribution can, for example, specify a value for an element of the matrix, elements of a predeterminable region of the matrix, or for all elements of the matrix, on the basis of which a suitable modification of one or more elements in the matrix is ​​carried out according to a mathematical function.

[0021] According to one embodiment, the predeterminable modification rule comprises a random modification of at least one element in the input matrix. For example, a modification can be performed for one element or a definable group of multiple elements of the input matrix based on a (pseudo-)random function. For example, it is also possible to randomly select a definable number of one or more elements in the input matrix and modify these randomly selected elements. Likewise, the value by which one or more elements in the input matrix are modified can also be randomly determined.

[0022] According to one embodiment, the predeterminable modification rule is applied to a predeterminable, configurable, or predeterminable region of the input matrix. For this purpose, any suitable approaches for determining the region in the input matrix to be modified can be used. For example, the region of the input matrix to be modified can be specified manually by a user, for example, using a suitable user interface or the like, for example, using a segmentation function or selection function. Furthermore, any suitable fully or at least partially automated methods for selecting a region are also possible. This can be done, for example, based on the detection of structures, objects, or any other properties in the data of the input matrix.Alternatively or additionally, the range of the input matrix to be modified can also be determined automatically. The automatic determination method can be configured, in particular, through appropriate configurations made via a user interface.

[0023] According to one embodiment, the predeterminable modification rule comprises modifying an element of the input matrix depending on values ​​of neighboring elements of the input matrix. For example, it is possible to change one or more elements in an input matrix such that a difference to one or more and / or selected neighboring elements is increased or decreased. In principle, however, any other approaches for modifying one or more elements in the input matrix depending on one or more neighboring elements are also possible. In principle, any elements that have a suitable relationship, in particular a neighborhood relationship, to the element to be modified can be regarded as neighboring elements. In particular, for example, in a two-dimensional matrix, directly vertically and / or horizontally adjacent elements can be regarded as neighboring elements.Depending on the application, however, other approaches are also possible, in which, for example, only a single element or a predeterminable group of elements are considered as neighboring elements.

[0024] According to one embodiment, the method comprises a step for identifying the predeterminable region using the input matrix. For example, the input matrix can be analyzed to identify a region to be modified. For this purpose, edges, structures, objects, and / or any other properties in the data of the input matrix can be detected and / or analyzed. For example, an area can be identified in the input matrix that corresponds to an (automatically) detected object. Likewise, for example, an area that corresponds to an outline of an object is possible. Furthermore, any other approaches for automatically identifying a suitable area are fundamentally possible. Additionally or alternatively, it is also possible to modify the predeterminable region using the data of the input matrix. For example, an initial area can first be determined manually or automatically.In a subsequent step, this range can then be adjusted according to the data in the input matrix. For example, if it is determined that an object, edge, or other structure in the input matrix changes with respect to the initially specified range, such as increasing or decreasing in size, the range can be adjusted accordingly. This enables dynamic modification of the specified range.

[0025] According to one embodiment, identifying the predeterminable region comprises detecting object boundaries in the input matrix. Any suitable detection method can be used for this purpose. For example, an object boundary can be determined based on a difference in the values ​​between adjacent elements of the input matrix.

[0026] According to one embodiment, a training data element comprises a sequential sequence of several consecutive training data points. A training data point is to be understood as a group of data for a position in the sequential sequence. Such a training data point can also comprise a two- or multi-dimensional matrix. For example, the sequential sequence can comprise a temporal sequential sequence in which each training data point specifies training data for a point in time of this sequential sequence. The sequential sequence can be provided in the training data elements as a further dimension of the input matrix. In this way, for example, temporal sequences can also be trained. In this case, the modification in the data of the input matrix can be carried out either uniformly, i.e. identically, for the individual positions in the sequential sequence.Alternatively, the modification in the data of the input matrix for the individual positions in the sequential sequence can also be carried out according to a suitable modification rule.

[0027] According to one embodiment, the input matrix comprises image data. The image data can, for example, comprise image data from one or more cameras. In particular, the image data can be in the form of monochrome or black-and-white image data, grayscale image data, or color images. The image data can be in the form of image data for a specific point in time or as moving images or video sequences. In principle, any suitable image format is possible.

[0028] According to one embodiment, the input matrix comprises multiple color channels. For example, the individual color channels, such as red, green, and blue, can be provided as separate color channels, for example, in the form of two-dimensional matrices. The individual color channels can, for example, form a further dimension of the input matrix. The predeterminable modification rule can either be applied to only one color channel, applied to all color channels in the same way, or a separate modification rule can be applied to each color channel. Furthermore, any other suitable concepts for modifying color images are of course also possible.

[0029] According to one embodiment, the input matrix comprises binary elements. In such a case, the individual elements of the input matrix have only binary states, for example, 0 and 1. Modifying an element in the input matrix thus corresponds to inverting the corresponding element in the input matrix. Such input matrices represent a particularly simple form of input matrices for training neural networks.

[0030] The above embodiments and further developments can be combined with one another as desired, where appropriate. Further embodiments, further developments, and implementations of the invention also include combinations of features of the invention not explicitly mentioned above or described below with respect to the exemplary embodiments. In particular, those skilled in the art will also add individual aspects as improvements or additions to the respective basic forms of the invention. Short description of the drawings

[0031] Further features and advantages of the invention are explained below with reference to the figures. These show: Fig. 1: a schematic diagram illustrating an extended training data set as may form the basis of an embodiment; Fig. 2: a schematic representation of a flowchart as may be the basis of a method for training a neural network according to an embodiment; and Fig. 3: a schematic representation of a block diagram for an apparatus for training a neural network according to an embodiment. Description of embodiments

[0032] Fig. 1 shows a schematic diagram illustrating the concept for generating an extended training data set for training a neural network NN according to one embodiment. As shown in the upper part of Fig. As shown in Figure 1, to train a neural network NN, an original training data set T can first be provided. Such an original training data set T can contain several training data elements. Each training data element can, for example, contain an input matrix Mi. Fig. The number of three input matrices Mi shown in Figure 1 is to be understood merely as an example and does not represent a limitation of the present invention.

[0033] The individual input matrices Mi can generally be two-dimensional or multi-dimensional matrices. For example, if the input matrices Mi are image data, a grayscale image can be represented as a two-dimensional matrix Mi. The individual elements of such a matrix Mi can have values ​​within a defined value range, for example between 0 and 255. However, other value ranges are also possible. In particular, binary matrices Mi are also possible, for example, in which the individual elements can only have binary values, for example 0 and 1. For color images, the image information can also be provided in multiple channels, for example multiple color channels. For example, individual two-dimensional matrices can be provided for red, green, and blue (R, G, B).In principle, however, any other concepts for providing color information are also possible. In such a case, the individual channels can, for example, form an additional dimension of the input matrix Mi. It should be understood, however, that the input matrix Mi is not limited to image data.

[0034] In addition to static image data, which represent a scene, particularly a traffic scene, at a fixed point in time, dynamic image data, such as image sequences or similar, are also possible. For such sequences, multiple images at different points in time can be represented, for example, in the form of an additional dimension of the matrices Mi. In principle, any other suitable information or data suitable for training a neural network is possible.

[0035] Each training data element can also include a label Li or similar. Such a label Li can, for example, specify a desired output of the neural network NN for the respective input matrix Mi according to the neural network's task. The label Li can be specified in any suitable manner and assigned to the respective input matrix Mi. A neural network NN can be trained based on such an original training data set T.

[0036] The concept of the present invention further provides for expanding or modifying the training data set T. For this purpose, additional training data elements are created for one or more data elements of the training data set T. By combining these additional training data elements with the training data elements of the original training data set T, an expanded training data set Te can then be created. The neural network NN can then be trained with this expanded training data set Te, as shown exemplarily in the lower part of the Fig. 1 is shown.

[0037] To generate the additional training data elements, an input matrix Mi is modified according to a predeterminable modification rule. This generates a new, modified input matrix M-ia, M-ib. For each additional input matrix M-ia, M-ib generated in this way, an additional training data element is generated. A further training data element generated by modifying an input matrix Mi from the original training data set T can, for example, be assigned the same label Li that is also assigned to the underlying training data element Mi. Depending on the application, however, it is also possible to modify the label Li in a suitable manner if necessary. In particular, the label L-ia, L-ib of a further matrix M-ia, M-ib can, for example, include information relating to the applied modification or modification rule.

[0038] Fig. 2 shows a flowchart of how a method for training a neural network (NN), in particular a neural network (NN) for an automated system, can be based according to one embodiment. The automated system that uses such a neural network (NN) can, for example, comprise a fully or at least partially autonomously driving vehicle. Likewise, such a neural network can be used, for example, for a fully or at least partially autonomously operating robot, such as a household appliance (vacuum robot, robot mop, or similar) or an at least partially automated process or production system. In principle, any other applications for such a neural network (NN) are also possible.

[0039] In a step S1, an original training data set T is first provided. This original training data set T can comprise multiple training data elements. Each of these training data elements comprises an at least two-dimensional input matrix Mi, in particular an input matrix Mi as previously described. Accordingly, each input matrix Mi of a training data element can be assigned a corresponding label Li.

[0040] In step S2, a modified input matrix M-ia, M-ib can then be generated from an input matrix Mi of a training data element. To generate such a modified input matrix M-ia, M-ib, a predeterminable modification rule can be applied to the initial input matrix Mi of the training data element from the original training data set T. If a label Li is assigned to the input matrix Mi of a training data element of the original training data set T, such a label can also be assigned to the modified matrix M-ia, M-ib. For example, a generated modified input matrix M-ia, M-ib can be assigned the same label Li that is also assigned to the underlying input matrix Mi. If necessary, however, the label Li can also be modified in a suitable manner and such a modified label L-ia, L-ib can be assigned to a generated modified input matrix M-ia, M-ib.

[0041] From an input matrix Mi of a training data element, either one or possibly several (especially differently) modified input matrices M-ia, M-ib can be generated in this way. Fig. The number of two modified input matrices M-ia, M-ib for an original input matrix Mi shown in Figure 1 is to be understood merely as an example and does not represent a limitation of the present invention. In principle, it is also not absolutely necessary to generate the same number of modified input matrices M-ia, M-ib for each input matrix Mi.

[0042] Alternatively or additionally, it is possible to apply different modification rules to all or selected input matrices in order to generate differently modified input matrices.

[0043] Thus, an additional training data element can be created from each generated modified input matrix M-ia, M-ib. By combining these additional training data elements with the training data elements of the original training data set T, an extended training data set Te can be created in step S3.

[0044] This extended training dataset Te can be used in step S4 to train a neural network (NN). Specifically, the neural network (NN) can be trained to a given target output using the extended training dataset Te.

[0045] A variety of predeterminable modification rules can be applied as modification rules for generating further input matrices M-ia, M-ib from an initial or original input matrix Mi. For example, a partial or global probability distribution can be specified. Such a probability value can, for example, define for individual elements of a matrix Mi the probability with which a modification of the corresponding element in the matrix Mi should occur. If, for example, several further input matrices M-ia, M-ib are generated for an initial input matrix Mi, the individual elements of the further input matrices M-ia, M-ib can be modified depending on such a probability distribution.

[0046] The modification of the elements in a further input matrix M-ia, M-ib can also be based on a random function, for example, based on generated pseudo-random numbers. This allows a very good statistical distribution of the modification of individual elements to be achieved.

[0047] In particular, the modification of an input matrix Mi can also be restricted to a predeterminable range. Such a range can, for example, be specified manually by a user. For this purpose, a user interface can be provided, for example, via which a user can specify the range to be modified. Additionally or alternatively, it is also possible to specify a range to be modified at least partially automatically, in particular on the basis of a predeterminable rule. For example, suitable properties such as structures, areas, edges, or the like can be identified in the data of an initial input matrix Mi. A range to be modified can then be selected based on these identified properties. For example, elements along an edge can, in particular, be modified.In this way, for example, a shift of such an edge or a blurring of such an edge can be simulated. Likewise, it is possible, for example, to modify elements within a region, for example, to increase or decrease a contrast or difference of such a region with respect to its surroundings. However, it is understood that any other suitable approaches for identifying and / or selecting regions of an input matrix Mi are also possible.

[0048] Furthermore, the modification rule can, for example, also comprise changing an element in an input matrix Mi depending on neighboring elements. For example, an element can be modified if at least one neighboring element or a predetermined number or selection of neighboring elements have a predetermined property, for example are greater than a threshold value. Neighboring elements can, for example, be regarded as elements that directly border an element, in particular vertically or horizontally. Likewise, diagonally adjacent elements can also be regarded as neighboring elements. Furthermore, it is possible, for example, to regard elements as neighboring elements that have a predetermined maximum distance from an element.

[0049] In addition, any other suitable rules for modifying elements in an initial input matrix Mi are of course possible in order to obtain further modified input matrices M-ia, M-ib.

[0050] Modifying an element in an input matrix Mi can, for example, comprise increasing or decreasing the value of an element by a predetermined absolute or relative value. In the case of binary matrices, i.e., matrices whose elements can only assume two states, the modification can also comprise inverting such an element. Furthermore, modifying an element can also comprise setting a fixed, predetermined value. Furthermore, an element to be modified can, for example, be described with a random value. However, it is understood that, in principle, any other approaches to modifying an element are also possible.

[0051] Fig.Figure 3 shows a schematic representation of a block diagram for a device 10 for training a neural network (NN). The device 10 can, in principle, comprise any components that may be suitable for implementing the method described above. Similarly, the method described above can, in principle, also comprise any steps that may be suitable for implementing a functionality of the device 10 described below.

[0052] The device 10 comprises an input device 1, an augmentation device 2 and a training device 3. The input device 1 can receive an original training data set T. For this purpose, the input device 1 can, for example, be communicatively coupled to a suitable data source, for example a storage device or the like. Any suitable data transmission concepts, such as a data bus, a network connection or the like, can be provided for the data exchange between the input device 1 and the storage device. The original training data set T received by the input device 1 can, in particular, be a previously described training data set T with a plurality of training data elements, wherein each training data element comprises an at least two-dimensional input matrix Mi.

[0053] The original training data set T received by the input device 1 can be provided to the augmentation device 2. For this purpose, a suitable communication connection, for example a data bus or the like, can be provided. The augmentation device 2 can generate at least one further modified input matrix M-ia, M-ib for a training data element of the original data set T. This can be done in particular according to a modification rule already described in connection with the corresponding method. Furthermore, the augmentation device 2 can generate a further training data element for each modified input matrix M-ia, M-ib and form an expanded training data set Te from the training data elements of the original training data set and the further training data elements.

[0054] This extended training data set Te can be provided to the training device 3. The training device 3 can then train a neural network NN using this extended training data set Te. In particular, the neural network NN can be trained to a predetermined target output using the extended training data set Te.

[0055] The device 10 for training the neural network NN may optionally further comprise an analysis device 4. This analysis device 4 may, for example, analyze the data in an input matrix Mi from a training data element to identify properties, such as objects and / or structures. In this case, the modification rule may be adapted, for example, using these identified properties.

[0056] In summary, the present invention relates to a method and a device for training a neural network, in particular a neural network for use in an automated system. For this purpose, it is provided to generate further training data elements from training data elements, each with an at least two-dimensional matrix, by modifying the matrices in the training data elements. An extended training data set can thus be generated from the original training data elements and the additional training data elements. The neural network can then be trained with this extended training data set. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2019 209 560 A1

[0003]

Claims

[1] Method for training a neural network (NN) for an automated system, comprising the steps: Providing (S1) an original training data set (T), wherein the original training data set (T) comprises a plurality of training data elements and wherein each training data element comprises an at least two-dimensional input matrix (Mi); Generating (S2) at least one modified input matrix (M-ia, M-ib) from the at least two-dimensional input matrix of a training data element by applying a predeterminable modification rule in order to generate a further training data element for each modified input matrix (M-ia, M-ib); Forming (S3) an extended training data set (Te) comprising training data elements from the original training data set (T) and the further training data elements; and Training (S4) the neural network (NN) using the generated extended training data set (Te) to a target output of the neural network (NN). [2] The method of claim 1, wherein each training data element comprises a label (Li) each representing an associated target output. [3] Method according to claim 1 or 2, wherein the predeterminable modification rule comprises modifying elements of the input matrix (Mi) using a predetermined probability distribution. [4] Method according to one of the preceding claims, wherein the predeterminable modification rule comprises a random change of at least one element in the input matrix (Mi). [5] Method according to one of the preceding claims, wherein the predeterminable modification rule is applied to a predeterminable region of the input matrix (Mi). [6] Method according to one of the preceding claims, wherein the predeterminable modification rule comprises modifying an element of the input matrix (Mi) depending on values ​​of neighboring elements of the input matrix (Mi). [7] Method according to one of the preceding claims 5 or 6, comprising a step for identifying and / or modifying the predeterminable range, in particular using the input matrix (Mi). [8] Method according to the immediately preceding claim, wherein identifying the predeterminable region comprises detecting object boundaries in the input matrix (Mi). [9] Method according to one of the preceding claims, wherein a training data element comprises a sequential sequence of several consecutive training data points. [10] Method according to one of the preceding claims, wherein the input matrix (Mi) comprises image data. [11] Method according to the immediately preceding claim, wherein the input matrix (Mi) comprises a plurality of color channels. [12] Method according to one of the preceding claims, wherein the input matrix (Mi) comprises binary elements. [13] A computer program product comprising a computer program comprising instructions which, when the computer program is executed by a computer, cause the computer to carry out the method according to one of claims 1 to 12. [14] Device (10) for training a neural network (NN) for an automated system, comprising: an input device (1) designed to receive an original training data set (T), wherein the original training data set (T) comprises a plurality of training data elements and wherein each training data element comprises an at least two-dimensional input matrix (Mi); an augmentation device (2) which is designed to generate at least one modified input matrix (Mi-a, M-ib) from the at least two-dimensional input matrix (Mi) of a training data element by applying a predeterminable modification rule in order to generate a further training data element for each modified input matrix (M-ia, Mi-b) and to form an extended training data set (Te) which comprises the training data elements from the original training data set (T) and the further training data elements; a training device (3) designed to train the neural network (NN) to a target output of the neural network (N) using the generated extended training data set. [15] Device (10) according to the immediately preceding claim, comprising an analysis device (4) which is designed to identify objects and / or structures in the training data elements of the original training data set (T) and to adapt the modification rule using the identified objects and / or structures.

Images