METHOD FOR TRAINING AND OPERATING A MULTITASKING ARTIFICIAL NEURAL NETWORK, MULTITASKING ARTIFICIAL NEURAL NETWORK AND DEVICE

Description

[0001] The present invention generally relates to the field of artificial intelligence and machine learning. In particular, the invention relates to a method for training a multitasking artificial neural network, a multitasking artificial neural network, and a device comprising such an artificial neural network. Furthermore, the invention also relates to a computer program and a storage medium. State of the art

[0002] Artificial neural networks (ANNs) are frequently used in the field of artificial intelligence, and especially in machine learning. The technical tasks that such an ANN can perform are numerous and extend across a wide range of technical fields, including automation technology, robotics, and their subfields such as image recognition and similar areas.

[0003] Multitasking-capable ANNs can also be used, which are configured to execute several, i.e. at least two, different tasks in parallel, in particular simultaneously, concurrently, alternately, or similarly, based on input data supplied to the ANN. In comparison to the operation of several parallel (only) single-task-capable ANNs, which can each execute only one specific task individually based on each input data supplied, multitasking-capable ANNs can share at least some of the functions, intermediate calculations, etc. inherent in the ANN for several tasks. This can, for example, save computing resources such as computing time, storage space, memory bandwidth, etc. compared to the use of several ANNs, each only capable of a single task.Even during the training of such a multitasking ANN, shared information can be used to train for the best possible overall performance for the majority of tasks to be performed. However, it has proven disadvantageous that the shared use of functions, intermediate calculations, information or similar of the ANN can lead to dependencies between the different tasks. This can complicate task-specific adaptations of the multitasking ANN, for example by requiring (re-)validation of all tasks. For example, an adaptation that is actually only intended to affect a single task could result in an undesirable adaptation of another task, meaning that not only the task or task that was actually adapted but also the undesirably adapted task or task must be re-validated.

[0004] From the publication PENGFEI LIU ET AL, "Recurrent Neural Network for

[0005] Text Classification with Multi-Task Learning", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, May 17, 2016 (2016-05-17), (available online: https: / / arxiv.org / abs / 1605.05101) is a multi-task learning framework for learning collaboratively across multiple related tasks.

[0006] From the publication HUNG KIM-HAN ET AL, "Multi-stage Diagnosis of Alzheimer's Disease with Incomplete Multimodal Data via Multi-task Deep Learning", September 9, 2017 (2017-09-09), INTERNATIONAL CONFERENCE ON FINANCIAL CRYPTOGRAPHY AND DATA SECURITY, LECTURE NOTES IN COMPUTER SCIENCE, SPRINGER, BERLIN, HEIDELBERG, PAGE(S) 160 - 168, ISBN: 978-3-642-17318-9, a multi-task deep learning for incomplete data is known, in which prediction tasks associated with different modality combinations are learned jointly to improve the performance of each task. Disclosure of the invention

[0007] Embodiments of the invention provide an improved possibility for applying a multitask-capable artificial neural network (ANN). Advantageous developments of the invention emerge from the dependent claims, the description, and the accompanying figures.

[0008] A first aspect of the invention relates to a method for training a multitasking artificial neural network, ANN, according to independent claim 1. The method provides the following steps: A first path for a first information flow through the ANN is provided. This first path couples an input layer (e.g. input layer) of the ANN with at least one cross-task intermediate layer (e.g. hidden layer) of the ANN, wherein the at least one cross-task intermediate layer can also be referred to as a cross-task ANN section. The intermediate layer is common to a plurality, at least two, of mutually different tasks of the ANN or is shared by or for these tasks, so that it can be further specified from the hidden layer to a shared layer.In addition, the first path couples the at least one cross-task intermediate layer to a respective task-specific ANN section from the plurality of mutually different tasks. The respective ANN section is thus assigned to exactly one single task, whereby several, in particular hierarchically arranged, ANN sections can be provided for each task. The at least one ANN section can be or comprise a task-specific output layer (e.g., output layer) of the ANN for executing a specific task. Accordingly, the task-specific output layer can generate or provide output data and / or an output signal.

[0009] Initial training data for training cross-task parameters is fed via the input layer and the first path. These cross-task parameters are common to the majority of different tasks of the ANN, meaning they can be shared between tasks, reused, etc. Cross-task parameters can include, in particular, weights, biases, activation functions, etc. of the ANN.

[0010] In addition to the first path, at least one task-specific, second path is provided by the ANN for a second information flow that is different from the first information flow. The second path couples the input layer of the ANN with only a portion of the task-specific ANN sections from the plurality of mutually different tasks. In other words, the information flow effected via the second path does not flow into all of the different tasks, but only into a portion of them. For example, in an ANN that is intended to execute a total of two different tasks, the second path can only allow information flow to a single task, whereas the other task is decoupled from the second path.

[0011] Second training data for training task-specific parameters are supplied via the second path. The second training data are preferably at least task-specific, i.e., are assigned to at least one of the plurality of different tasks.

[0012] This method can improve the application of a multitask-capable ANN by separately modifying and / or improving only a single task or a subset of several specific tasks. Modification and / or improvement in this context can be understood in particular to mean that this single task or this subset of several specific tasks can be retrained, adapted, corrected, fine-tuned, etc. However, the proposed method ensures that the remaining tasks are not affected, in particular the task-specific parameters assigned to them, such as their weights, biases, activation functions, etc. It is understood that this retraining, adaptation, correction, fine-tuning, etc., is carried out using the second training data via the second path.

[0013] The proposed method, on the other hand, makes it possible, on the one hand, to advantageously utilize desired commonalities or dependencies between the different tasks in order to save computing resources through the shared use of functions, intermediate calculations, information, or similar, while, on the other hand, achieving a certain degree of decoupling of the different tasks in order to adapt individual tasks individually. Should the ANN need to be validated for its application, for example, to obtain approval for productive operation in, for example, automation technology, robotics, etc., the proposed method can significantly reduce the required effort. This is because, in principle, only the linked task needs to be validated using the second path, since the other tasks have not changed since the last validation.

[0014] An ANN can be understood, in particular, as a directed graph whose nodes represent a number of neurons arranged in the layers described above, and whose edges represent connections between them. The parameters to be trained can, in particular, influence the intensity of the information flow along these connections. Accordingly, the first path can allow information along all nodes and / or edges, while the second path can only allow information along a subset of these nodes and / or edges. In this context, a task can be understood, in particular, as a possibly self-contained task, represented, for example, by at least part of a computer program.

[0015] The different tasks can each perform tasks in automation technology, robotics, etc., such as for an at least partially automated, i.e., semi-autonomous robot. Examples of applications for such an ANN can also be found in automotive engineering, particularly for at least partially automated vehicles.

[0016] In addition, other applications in automation technology in general and, for example, in vehicle technology in particular are conceivable.

[0017] A further development provides that a flow of information from the at least one cross-task intermediate layer to the second path can be permitted, but a counter-flow of information from the second path to the cross-task intermediate layer can be prevented. In other words, the shared parameters, etc., can continue to be used for the tasks coupled with the second path, while these cannot be changed by the second path and can thus remain unchanged for the remaining tasks. The information flow of the second path is unidirectional with respect to the cross-task intermediate layer. This allows for a particularly resource-efficient adaptation of individual tasks.

[0018] Another development provides that the at least one task-specific ANN section can be multi-layered, i.e., with a corresponding number of hierarchically arranged, task-specific task (intermediate) layers (i.e., possibly task-specific hidden layers). It should be noted that the number of task-specific task layers of one task can differ from the number of task-specific task layers of another task. Based on a representation of the ANN as a graph, the at least one task-specific ANN section can therefore have a number of task-specific nodes that are connected to one another via corresponding task-specific edges only within the same task-specific ANN section.

[0019] In a further development, the respective task-specific ANN section can be changed or adapted via the second path. In particular, a number of task-specific task layers, etc., of this ANN section can be changed. For example, further task-specific task layers can be added to this task-specific ANN section, or existing task-specific layers can be removed. Additionally or alternatively, a number of neurons, for example, in one or more task-specific layers of the task-specific ANN section can be changed, for example, reduced or increased. This enables particularly good adaptation of individual tasks without changing the other tasks.According to a further development, the second training data supplied via the second path can be combined from input data supplied to the input layer and from intermediate layer data derived from the at least one cross-task intermediate layer. This allows for a particularly computationally resource-efficient adaptation of individual tasks.

[0020] In a further development, from a plurality of cross-task intermediate layers, those that support the training of task-specific parameters can be selected for linking to the second path. For example, it is possible that from the plurality of cross-task intermediate layers, especially those arranged hierarchically, only those that are relevant, helpful, useful, or similar for adapting the task linked to the second path are connected to the second path. This allows the task-specific quality of the ANN to be further improved.

[0021] According to a further development, a validation of at least some of the tasks executable by the ANN can be performed between the supply of the first training data and the supply of the second training data. The supply of the second training data for adapting at least one specific task can be performed without adapting at least one other, different specific task. Since the other specific task has already been validated after the supply of the training data and, if applicable, a test run, and has not been changed by the supply of the second training data due to the decoupling from the second path, further validation of the other specific task can be dispensed with. This allows the ANN to be adapted with little validation effort.

[0022] A second aspect of the invention relates to a computer-implemented multitasking artificial neural network, ANN, according to independent claim 6.

[0023] As described above, the ANN can be understood in particular as a directed graph whose nodes represent a number of neurons arranged in the layers described above, and whose edges represent connections between them. The proposed ANN has an input layer, a plurality of task-specific ANN sections, namely a number corresponding to the number of tasks assigned to a plurality of mutually different tasks of the ANN, and at least one cross-task intermediate layer arranged between the input layer and the plurality of task-specific ANN sections and comprising a number of parameters that can be used across tasks.

[0024] In addition, the proposed ANN has a first path that couples the input layer via the at least one cross-task intermediate layer to the plurality of task-specific ANN sections for a first information flow through the ANN.

[0025] In addition, the proposed ANN comprises at least one task-specific second path coupling the input layer to only a portion of the plurality of task-specific ANN sections for a task-specific second information flow through the ANN that is different from the first information flow.

[0026] This ANN enables the computationally resource-efficient adaptation of one or more specific ANN tasks. Advantageously, parameters such as weights, biases, activation functions, etc., that are common to all (or at least several) tasks can be trained via the first path. However, if it should become apparent during validation or productive operation of the ANN that a specific task or its execution needs to be modified in any way, this can be done without affecting the other specific tasks. This allows the advantages of a multitask-capable ANN with regard to the shared use of certain parts of the ANN to be combined with the advantages of simple adaptation of specific tasks, similar to a single-task ANN.

[0027] According to a further development, a number of layers of the second path can differ from a number of cross-task intermediate layers. In other words, the second path can have a different number of nodes and / or edges than the first path, particularly with regard to the cross-task intermediate layers. A node and / or an edge of the second path can be connected to one or more nodes and / or edges of the first path. This enables free configuration of the ANN for improving and / or modifying one or more specific tasks with the lowest possible computing resources.

[0028] In a further development, the ANN can have a plurality of second paths. Each second path can be set up for a task-specific information flow to only a subset of the plurality of task-specific ANN sections. In other words, the ANN can have a plurality of second paths, each of which is assigned (only) to one specific task, such that the different second paths are decoupled from one another with regard to their respective information flow. As a result, the advantages described above for changing and / or improving individual specific tasks can also be achieved if the ANN is to execute three or more mutually different tasks and, for example, only two of three or at least three of at least four different tasks, etc., are to be changed and / or improved.

[0029] According to a further development, the ANN can have at least one recurrent cross-task intermediate layer configured for information flow toward the second path. A recurrent cross-task intermediate layer can be understood as having direct feedback, in which the node uses its output as a new input, indirect feedback, in which the output of a node is used as the input of a node in an upstream cross-task intermediate layer, or lateral feedback, in which the output of a node is used as the input of a node in the same cross-task intermediate layer.

[0030] In another development, the second path can also have at least one recurrent layer. As described above, this recurrent layer can be one of possibly several nodes of the second path. In this context, a recurrent layer can be understood as having direct feedback, in which the node uses its output as a further input, indirect feedback, in which the output of a node is used as the input of a node in an upstream cross-task intermediate layer, or lateral feedback, in which the output of a node is used as the input of a node in the same cross-task intermediate layer.

[0031] A third aspect of the invention provides a device comprising at least one computer-implemented multitasking artificial neural network (ANN), as described above. The ANN can be implemented as software, hardware, or a hybrid of software and hardware.

[0032] The device can be designed, for example, as a computer, electronic control unit, or control unit network, or the like. In addition, the device can have at least one processor, a memory, such as a volatile and a non-volatile memory, one or more data interfaces to detection devices, such as sensors, actuators of a machine, robot, or the like, a communication interface, etc. The device can also be configured to receive input data in the form of, for example, signals, feed them to the input layer of the ANN described above, and provide its output data. Accordingly, the device can be used, for example, in automation technology, robotics, or the like, wherein specific tasks, as described above, can be modified and / or improved.

[0033] A fourth aspect of the invention provides a method for operating a computer-implemented multitasking artificial neural network (ANN). The ANN may, for example, be implemented in the device described above, such as a computer, electronic control unit, etc.

[0034] The proposed method provides the following steps: In a first phase, a plurality of mutually different tasks of the ANN are jointly trained by supplying first training data via a first path that allows a first flow of information through the ANN. This can be done, in particular, using the method described above in one or more of the described embodiments.

[0035] In a second phase, one or more of the tasks of the trained ANN can be executed. These could be, for example, various tasks in automation technology, robotics, etc.

[0036] In a third phase, at least one of the tasks of the ANN can be trained and / or corrected independently of at least one other of the plurality of tasks by supplying second training data via a second path that is different from the first path and allows a second information flow through the ANN that is different from the first information flow. This can be done, in particular, by the method described above in one or more of the described embodiments.

[0037] This allows at least one individual task of the ANN to be subsequently modified and / or improved without undesirably changing other tasks.

[0038] A fourth aspect of the invention relates to a computer program comprising instructions which, when the computer program is executed by a computer, cause the computer to execute one of the methods described above, i.e. a method according to the first or third aspect, in one or more of the respectively described embodiment variants.

[0039] A fifth aspect of the invention relates to a machine-readable storage medium on which a computer program according to the fourth aspect is stored.

[0040] The computer-readable storage medium may in particular, but not necessarily, be a non-transitory medium, particularly suitable for storing and / or distributing a computer program. The computer-readable storage medium may be a CD-ROM, a DVD-ROM, an optical storage medium, a solid-state medium, or the like, supplied together with or as part of other hardware. Additionally or alternatively, the computer-readable storage medium may also be distributed or distributed in another form, for example via a data network, such as the Internet or other wired or wireless telecommunications systems. For this purpose, the computer-readable storage medium may, for example, be embodied as one or more data packets.

[0041] Further measures improving the invention are presented in more detail below together with the description of the preferred embodiments of the invention with reference to figures. Short description of the characters

[0042] Advantageous embodiments of the invention are described in detail below with reference to the accompanying figures. They show: Figure 1 shows a block diagram of a device with an artificial neural network according to an embodiment of the invention, Figure 2 shows a block diagram of a device with an artificial neural network according to an embodiment of the invention, Figure 3 shows a block diagram of a device with an artificial neural network according to an embodiment of the invention, Figure 4 shows a block diagram of a device with an artificial neural network according to an embodiment of the invention and Figure 5 shows the embodiment according to Figure 4 for better illustration as a directed acyclic graph.

[0043] The figures are merely schematic and not to scale. In the figures, identical, equivalent, or similar elements are provided with the same reference numerals throughout. Embodiments of the invention

[0044] Figure 1shows a block diagram of a device 100 which can be implemented, for example, in an at least partially automated robot (not shown). The device 100 is, by way of example, a computer in the form of an electronic control unit which can be designed, for example, as an embedded system in order to carry out automation and / or control tasks of the robot, for example. By way of example only, the robot can be an at least partially automated vehicle, such as an at least partially automated motor vehicle. The device 100 is further configured - at least in productive operation - to receive input signals from, for example, sensors of the robot, to process them and to generate output signals itself in response to the input signals and to provide them, for example, to an actuator of the robot.With reference to the example of a vehicle, the device 100 can control at least partial functions of the vehicle based on incoming sensor signals by controlling an actuator and / or vehicle drive, etc.

[0045] The device 100 comprises an artificial neural network 110, which is abbreviated to ANN below and is represented here for clarity as a directed graph with a number of nodes and edges. The nodes represent layers or a number of neurons arranged in the layers. The edges represent connections between the layers or the neurons arranged in the layers.

[0046] The ANN 110 is multitasking capable, i.e., configured to execute multiple tasks in parallel, e.g., simultaneously, alternately, etc. In this embodiment, the ANN 110 is configured to execute at least two different tasks, which are Figure 1and are distinguished from each other below by the reference symbols A and B.

[0047] The ANN 110 is configured for image processing. Task A is traffic sign recognition, and task B is semantic scene segmentation.

[0048] As in Figure 1 As shown, the ANN 110 is multi-layered, with an input layer 120, at least one task-spanning intermediate layer 130 and several, namely the number of tasks A, B corresponding to the number of task-specific ANN sections 140. In Figure 1 The majority of KNN sections 140 thus relate to the two tasks A and B, so that the KNN 110 here has, as an example, two KNN sections 140. In principle, each of the KNN sections 140 can represent an output layer of the KNN 110. As in Figure 1As indicated, each of the ANN sections 140 can also be configured as a multi-layered structure with a corresponding number of hierarchically arranged, task-specific task layers. Three task-specific task layers are shown here for each of tasks A and B only as an example, and the number of these can vary upwards and downwards. The number of intermediate layers 130 is five here, for example, although more or fewer intermediate layers 130 can also be provided.

[0049] The cross-task intermediate layers 130 are, as in Figure 1indicated by the corresponding nodes and edges, are common to both tasks A and B. These are therefore layers that are shared between tasks A and B (shared layers). In particular, parameters that can be used across tasks, such as weights of the individual nodes, i.e., layers or neurons, activation functions, etc., can be shared. For example, the complexity of the intermediate layers 130 can increase hierarchically.

[0050] According to Figure 1 The input layer 120 is coupled via an edge to the intermediate layers 130, and these, in turn, are each coupled via an edge to the task-specific ANN sections 140, so that an information flow from the input layer 120 via the intermediate layers 130 to the ANN sections 140 assigned to tasks A and B, which function as the respective output layer, is possible. Figure 1An arrow indicates an input signal S that causes the corresponding information flow. In a training phase, the input signal S can represent training data, while in productive operation of the device 100, for example, sensor data or the like is fed in.

[0051] As in Figure 1As shown, the ANN 110 has a first path P1 that couples the input layer 120 via the cross-task intermediate layers 130 to the task-specific ANN sections 140 for a first information flow in the direction indicated by the edge arrows through the ANN 110. Accordingly, an edge of the last intermediate layer 130 in the direction of the information flow couples it to the first layer of the respective task-specific ANN section 140 in the direction of the information flow. This means that the information flow of the first path P1 branches from the last of the intermediate layers 130 towards the first task A and towards the second task B.

[0052] How to continue in Figure 1As shown, the ANN 110 additionally has a second path P2, whose second information flow differs from the first information flow described above. Thus, the second path P2 itself comprises further nodes 170, which are connected to one another via further edges. The nodes 170 can in turn be understood as layers or as neurons arranged in layers. The second path P2 only leads to one of the two tasks, namely the first task A in this exemplary embodiment. Thus, the path P2 represents a task-specific side path which, although it allows the second information flow to a single one of the tasks, in this case the first task A, is decoupled from the at least one other task, in this case task B. Accordingly, the second information flow along the second path P2 only influences the task assigned to it, in this case the first task A, while the other task, in this case task B, is not influenced.

[0053] In the embodiment according to Figure 1 the second path P2 merely exemplarily has a number of nodes that corresponds to the number of cross-task intermediate layers 130. However, this number can also differ. The last of the nodes or layers 170 of the second path P2 in the direction of the information flow is coupled to the ANN section 140 assigned to the first task A. Merely exemplarily, the ANN section of, for example, task A here has three task-specific task layers, of which the first in the direction of the information flow is coupled to the last of the nodes or layers 170 of the second path P2. It should be noted that the information flow of the second path P2 towards the first task A can be carried out with at least one operator, so that the information flow of the second path P2 towards the first task A can be, for example, additive, subtractive, multiplicative, concatenating, etc. As in Figure 1As shown, individual or all layers 170 of the second path P2 can be combined with intermediate layer data, e.g., the above-described cross-task usable parameters, which are derived from the cross-task intermediate layers 130. In other words, there can be an information flow along an edge that couples one or more of the cross-task intermediate layers 130 to the second path P2, e.g., to an edge or a node thereof. Along this edge, for example, cross-task usable data, in particular parameters, can then be fed to the second path P2. It should be noted that this information flow is preferably unidirectional (i.e., exclusively from the cross-task intermediate layers 130 to the second path P2) so that the information flow via the second path P2 cannot change the parameters, data, etc. of the cross-task intermediate layers 130.

[0054] Figure 2 shows a block diagram of another embodiment of the device 100 or the KNN 110. This embodiment differs substantially from that shown in Figure 1 shown in that the second path P2 has a number of layers 170 that differs from the number of cross-task intermediate layers 130 of the first path P1. Merely by way of example, the number of layers 170 of the second path P2 is less than the number of cross-task intermediate layers 130 of the first path P1, although the reverse is also possible. Nevertheless, all parameters that can be used across tasks can still be supplied to the second path P2, since corresponding edges from the first path P1 to the second path P2 are provided.

[0055] Figure 3shows a block diagram of a further embodiment of the device 100 or of the ANN 110. This embodiment differs from the embodiments described above in that a further second path P2' is also provided for the second task B for a second information flow through the ANN 110. In principle, the further second path P2' is the same as that described above for the second path P2. It should be noted that the second path P2' - as clearly shown in Figure 3 shown - can only influence the second task B, but cannot influence the first task A.

[0056] The Figure 4 and 5show, in a block diagram, a further embodiment of the device 100 and the ANN 110, respectively. According to this embodiment, some or possibly all of the cross-task intermediate layers 130 can be implemented as recurrent layers. Accordingly, the information flow between the cross-task intermediate layers 130 and / or towards the second path P2 can be more complex than in the embodiments described above. In order to illustrate a correspondingly more complex, exemplary information flow through the ANN 110, the ANN 110 according to this embodiment is Figure 5 represented as a directed acyclic graph at different times t for the input signal S with t=0, t=1 and t=2.

[0057] Based on Figure 1, which shows a block diagram, it is explained below by way of example how the device 100 or the KNN 110 can be trained to carry out the different tasks A, B.

[0058] As described above, the ANN 110 has the first path P1 for the first information flow through the ANN 110. In one phase, initial training data is supplied via this first path P1, which is designed to train the cross-task parameters common to tasks A and B in the cross-task intermediate layers 130. During this training, for example, weights of the individual layers or the neurons arranged therein are adjusted in order to then be able to execute the respective task A or B with a desired quality. In this way, the different tasks A and B can also be trained gradually. After training, the ANN 110 is then, in principle, adapted or trained for productive operation. It should be noted that this training data is represented here by the input signal 120 for better clarity.

[0059] Usually, the KNN 110 is validated before going into production. However, if during validation it turns out that a task, e.g. the first task A according to Figure 1 , does not deliver the desired performance, the architecture of the ANN 110 described above enables retraining, fine-tuning, correction, etc., or generally speaking, adaptation, of the specific task A by means of the second path P2 for the second information flow through the ANN 110. For this purpose, task-specific, second training data are supplied via the task-specific second path P2, which enable adaptation of one task, here the first task A, without influencing the other task, here the second task B. It should be noted that this training data is represented here by the input signal 120 for better clarity.

[0060] Since only the parameters, settings, etc. of task A change by supplying the task-specific training data via the second path P2 (this also applies to the further second path P2'), i.e., neither the parameters of the cross-task intermediate layers 130 nor the task-specific parameters, settings, etc. of task B assigned to task B change, the subsequent validation effort is significantly reduced. It is understood that if task B remains the same, only task A, modified via the second path P2 or P2' by supplying the second training data, needs to be validated, since the parameters, settings, etc. common to tasks A and B have not changed.

[0061] The principle of training described above, in particular task-specific training, can be transferred accordingly to all embodiments of this application.

Claims

1. Computer-implemented method for training an artificial neural network (110), ANN, capable of multitasking, wherein the ANN is configured to process images supplied to the ANN for the execution of a plurality of different tasks by means of the ANN, wherein the plurality of different tasks comprises at least a first task (A) and a second task (B), wherein the first task (A) is traffic sign recognition and the second task (B) is a semantic scene segmentation, wherein the ANN comprises an input layer (120), at least one cross-task intermediate layer (130) and a plurality of corresponding task-specific ANN sections, wherein each task-specific ANN section functions as the output layer of the ANN for the task corresponding to this ANN section, comprising the following steps: - providing a first path (P1) for a first information flow through the ANN (110), wherein the first path (P1) couples the input layer (120) of the ANN (110) to the least one cross-task intermediate layer (130) of the ANN (110), and the first path (P1) couples the at least one cross-task intermediate layer (130) to each task-specific ANN section (140) from the plurality of corresponding task-specific ANN sections, - supplying first training data for training cross-task parameters, which are common to the plurality of mutually different tasks of the ANN (110), via the input layer (120) and the first path (P1), - providing at least one task-specific, second path (P2) for a second information flow, different from the first information flow, through the ANN (110), wherein the second path (P2) couples the input layer (120) of the ANN (110) to only one task-specific ANN section (140) from the plurality of task-specific ANN sections, wherein the task for which one task-specific ANN section functions as the output layer of the ANN is the task (A) or the task (B), and - supplying second training data and training only task-specific parameters via the second path (P2) on the basis of the second training data for training the ANN for that task for which one task-specific ANN section functions as the output layer of the ANN.

2. Method according to Claim 1, wherein an information flow from the at least one cross-task intermediate layer to the second path (P2) is permitted, but an opposite information flow from the second path (P2) to the cross-task intermediate layer (130) is prevented.

3. Method according to Claim 1 or 2, wherein the second training data supplied via the second path (P2) are combined from input data supplied to the input layer (120) and from intermediate layer data derived from the at least one cross-task intermediate layer (130).

4. Method according to any of the preceding claims, wherein from a plurality of cross-task intermediate layers (130) those which assist the training of the task-specific parameters are selected for a linkage with the second path (P2).

5. Method according to any of the preceding claims, wherein a validation of at least a portion of the tasks (A, B) executable by the ANN (110) is carried out between supplying the first training data and supplying the second training data, and supplying the second training data, for the purpose of adapting at least one specific task (A, B), is carried out with the exclusion of an adaptation of at least one further specific task (A, B) which is different therefrom.

6. Computer-implemented artificial neural network (110), ANN, capable of multitasking, and trained by a method according to any of the preceding claims, wherein a first task (A) is traffic sign recognition and a second task (B) is a semantic scene segmentation, comprising - an input layer (120), - a plurality of task-specific ANN sections (140) assigned to a plurality of mutually different tasks of the ANN (110), - at least one cross-task intermediate layer (130) arranged between the input layer (120) and the plurality of task-specific ANN sections (140) and comprising a number of parameters usable in a cross-task manner, - a first path (P1), which couples the input layer (120) via the at least one cross-task intermediate layer (130) to the plurality of task-specific ANN sections (140) for a first information flow through the ANN (110), and at least one task-specific, second path (P2), which couples the input layer (120) to only a portion of the plurality of task-specific ANN sections (140) for a task-specific second information flow, different from the first information flow, through the ANN (110).

7. ANN capable of multitasking according to Claim 6, wherein a number of layers of the second path (P2) is different from a number of the cross-task intermediate layers.

8. ANN capable of multitasking according to Claim 6 or 7, which has a plurality of second paths (P2) and each second path (P2) is configured for a task-specific information flow to only a partial number of the plurality of task-specific ANN sections (140).

9. ANN capable of multitasking according to any of Claims 6 to 8, which has at least one recurrent cross-task intermediate layer (130) configured for an information flow to the second path (P2).

10. Device, comprising at least one artificial neural network (110), ANN, capable of multitasking according to any of Claims 6 to 9.

11. Computer-implemented method for operating an artificial neural network (110), ANN, capable of multitasking according to any of Claims 6-9.

12. Computer program, comprising instructions which, when the computer program is executed by a computer, cause the latter to carry out a method according to any of Claims 1 to 5 or 11.

13. Machine-readable storage medium on which a computer program according to Claim 12 is stored.