Federated training of neural networks with improved coordination between the participating client nodes.

Abstract

Method (100) for augmenting a set T of training examples x for a neural network N = D°E, comprising an encoder E for converting the input / of the neural network into a representation R in a workspace Z and a decoder D for determining the output 0 of the neural network N from the representation R, comprising the steps: • From the set T of training examples x, two subsets T1 and T2 are selected (110); • for all training examples x ∈ T1 ∪ T2 from these two subsets T1 and T2, semantic meanings s(x) are procured (120); • the semantic meanings s(x) are transformed into representations R(s(x)) in the workspace Z (130); • for pairs R(s1(x1 ∈ T1)),R(s2 (x2 ∈ T2))) of representations R(s 1,s2) semantic meanings s1, s2, to training examples x1, x2, from different subsets T1, T2, a similarity S(R(s1).R(s2)) is determined according to a given similarity metric S (140); • A selection S*(R(s1),R(s2)) of the best similarities S(R(s1).R(s2)) is determined (150); and • for each selected similarity S*(R(s1) , R(s2)) is used to determine a new representation R' as a training example for the decoder D by taking into account the representation R(x1) - E(x1) of a training example x1 involved in this and the representations R(s1),R(s2) of the semantic meanings s1, s2.

Description

[0001] The present invention relates to the federated training of neural networks in which a plurality of client nodes C1, ..., CN cooperate in a coordinated manner by a server node Q. State of the art

[0002] Training neural networks, such as those used for image classification and / or semantic segmentation, requires a large number of training examples with sufficient variability. The effort required to store these training examples and the actual training can be too great for a single entity. Furthermore, for legal reasons, it is not always possible to consolidate all training examples into one entity that performs the training. For example, an image classifier for monitoring the environment of an at least partially automated vehicle requires images that may contain license plates, faces, and other personal data as training examples.If this image classifier is to be trained so that it works equally well not only in North America and Europe, but also in other regions of the world, merging all training examples for conducting the training may fail due to data protection laws, such as the European General Data Protection Regulation (GDPR).

[0003] One solution is federated learning, in which a large number of client nodes C1, ..., CN each train the neural network on a local dataset of training examples, and the respective training contributions are collected at a server node Q. The training contributions are then aggregated by the server node Q into an overall result and applied to the neural network. Disclosure of the invention

[0004] The invention provides a method for augmenting a set T of training examples x for a neural network N = D°E. The neural network N comprises an encoder E for converting the input I of the neural network into a representation R in a working space Z and a decoder D for determining the output 0 of the neural network N from the representation R. It therefore processes the input I in two steps, first to the representation R and then to the output 0. The method specifically aims to generate further training examples for the decoder D. This means that starting from the set T, new training examples are not immediately created in the space of inputs I, but rather new training examples in the space of representations R, which can be further processed in the same way as if they had originated from new training examples x in the space of inputs I.By using new training examples in the representation space R to better train the decoder D, the neural network N as a whole can also be better trained.

[0005] For this purpose, two subsets T1 and T2 are selected from the set T of training examples x. Semantic meanings s(x) are obtained for all training examples from these two subsets T1 and T2 and converted into representations R(s(x)) in the same workspace Z to which the representations R(x) of training examples x belong. The semantic meanings s(x) can come from any source. For example, the training examples x can be labeled with semantic meanings s(x). However, the semantic meanings s(x) can also be determined, for example, using any trained neural network or other machine learning model.

[0006] A similarity S(R(s1), R(s2)) is now determined for pairs (R(s1(x1 ∈ T1)), R(s2(x2 ∈ T2))) of representations R(s1,s2) of semantic meanings s1, s2 to training examples x1, x2 from different subsets T1, T2, each according to a given similarity metric S. This can also be understood as a similarity between training examples x1, x2 from different subsets T1, T2 being determined via the similarity of the respective associated semantic meanings s1, s2 in the space of representations R. The similarity metric can, in particular, be a distance measure in the working space Z, such as a cosine similarity.

[0007] From these similarities S(R(s1), R(s2)), a selection S*(R(s1), R(s2)) of the best similarities S(R(s1),R(s2)) is determined. This can, for example, be a top-n selection of the numerically largest similarities S(R(s1),R(s2)). For each of these selected similarities S*(R(s1), R(s2)), a new training example is now determined for the decoder D. This new training example is a new representation R* in the working space Z, which is determined using the representation R(x1) = E(x1) of a participating training example x1 as well as the representations R(s1), R(s2) of the semantic meanings s1, s2. The transition to the working space Z makes it possible to combine data types that would not have been meaningfully combined in the original state.

[0008] The newly selected training example R* for the decoder D in the workspace Z starts at R(x1) = E(x1) and, starting from this, is influenced by the semantic meanings s1, s2 of the two participating training examples x1, x2. This new training example R* therefore contains the knowledge that the two training examples x1 and x2, which initially appear to originate from different subsets T1 and T2 of the total set T of training examples, are semantically similar to each other. Thus, in particular, a training session that initially targeted one subset T1 of training examples x can be enriched with knowledge about the second subset T2. This is particularly advantageous, for example, in the distributed training of the neural network N, in which the different subsets T1 and T2 are trained on different participating nodes.

[0009] In such distributed training, a training contribution in the form of parameters Ñ that characterize the behavior of the neural network and / or suggested changes to these parameters Ñ is determined on each participating node using a subset T1, T2. These training contributions are then combined in some form. For example, in decentralized training, the participating nodes exchange their training contributions with each other without a central authority. In federated training, however, there is a central node Q that collects training contributions c1, ..., cK from a large number of client nodes C1, ..., CK. If the participating nodes are unaware of each other, training contributions can arise that are optimal from the perspective of the respective node, but cannot be easily aggregated into a larger whole.For example, the same parameter Ñ can be changed in one direction in one training contribution and in the opposite direction in another training contribution. When averaging, these training contributions can cancel each other out completely or partially. Better coordination between the client nodes C1, ..., CK is therefore desirable. On the other hand, for data protection reasons, the client nodes C1, ..., CK cannot generally freely exchange their local subsets T1, T2 with each other or with the central server node Q. In addition to the effort required for data transfer, this is a main motivation for conducting federated training in the first place, and for exchanging only training contributions that do not allow any conclusions to be drawn about the training examples x.According to the procedure proposed here, something can be exchanged that still does not allow any conclusions to be drawn about the training examples x, but does enable a better coordination of the client nodes C1, ..., CK with each other.

[0010] However, the additional training examples also have a beneficial effect when only one entity trains the entire neural network N. This prevents certain subsets T1 and T2 of training examples x from forming "knowledge islands" without exploiting their connections and commonalities.

[0011] In a particularly advantageous embodiment, the semantic meanings s(x) are obtained by processing the training examples x with a trained foundation model G (also called a “foundation model”). In this way, the knowledge learned in this foundation model G can be used to recognize semantic relationships between training examples. If a foundation model G is trained on a large amount of generally available data, it can, for example, be able to map any input into a continuous working space Z, in which meaningful arithmetic can also be explained. This means that directions, such as those of vectors, in the working space Z also have a meaningful meaning in the context of the current application. For example, the well-known CLIP model transforms textual descriptions of images in a working space Z in which representations can be added and subtracted.For example, if one subtracts the representation for "man" from the representation for "king" and adds the representation for "woman" in return, the result is close to the representation for "queen." Thus, it is understandable that in the space of representations R, starting with an image as a starting point, a direction determined from textual descriptions can be used to find a new representation of an image.

[0012] In another particularly advantageous embodiment, a subnetwork V already contained in the base model G is reused as an encoder E. This allows the training effort for this encoder E to be saved. For example, a base model G, such as the well-known CLIP, can comprise a first encoder for transforming images into a first latent space and a second encoder for transforming textual descriptions of these images into the same or a second latent space. Preferably, the two encoders convert the images, on the one hand, and the textual descriptions, on the other, into one and the same latent space.

[0013] In this context, it is advantageous that the training according to the method proposed here is specifically focused on the decoder D. Firstly, a trained encoder E is easier to obtain from the outset, for example through the aforementioned reuse of a subnetwork V. Secondly, the path taken here for training the decoder D via representations R would not yet be available for training the encoder E, because without a fully trained encoder E, there are no representations R. In order to better interlink the client nodes C1, ..., CK during the federated training of the encoder E, the training examples x themselves would have to be exchanged between the client nodes C1, ..., CK. However, the entire federated training aims to avoid precisely this.

[0014] In contrast to known approaches that determine a new representation of a new training example and derive the new training example from it by back-transformation, here the new representation R* is used directly as a training example for the decoder D. The new training example is thus generated directly in the workspace Z, the space in which it is needed.

[0015] In a particularly advantageous embodiment, at least the decoder D of the neural network N is trained in a federated manner. This training takes place in the interaction of several client nodes C1, ..., CK, which generate training contributions c1, ..., cK, with a server node Q, which aggregates these training contributions c1, ..., cK into changes ΔÑ of parameters Ñ that characterize the behavior of the neural network N. The subsets T1 and T2 are taken from the respective sets T1, ..., TK of the training examples intended for the training of different client nodes C1, ..., CK. As previously explained, in this way the training of the various client nodes C1, ..., CK can be coordinated such that the training contributions c1, ..., cK essentially "speak the same language" and can thus be easily aggregated with one another.

[0016] To determine the additional training examples R* for the decoder D of a client node C1, ..., CK, essentially information is required that results from the training examples x1 used by this node from the subset T1, namely representations R(x1) of these training examples x1 and representations R(s1) of their semantic meanings s1. To adapt the training of the decoder D with the training at another client node C1, ..., CK, no knowledge of the training examples x2 used by this other client node C1, ..., CK from the subset T2 is necessary.

[0017] All that is required are representations R(s2) of the semantic meanings s2 of these training examples x2. These semantic meanings s2 do not allow any conclusions to be drawn about the respective concrete training example x2. Therefore, in terms of data protection, it is significantly less problematic to exchange these semantic meanings s2 between client nodes C1, ..., CK than the original training examples x2. In particular, the client nodes C1, ..., CK can, for example, access a common pool of semantic meanings s2 that is fed by all client nodes C1, ..., CK. For example, instead of an image in which a specific person can be recognized, a textual description of this image can be exchanged. This description, while consistent with the specific image, no longer allows any conclusions to be drawn about this one image and no longer allows the person to be recognized.At the same time, the semantic meanings s2 typically have a significantly smaller volume than the original image.

[0018] Thus, the method proposed here allows the federated training of the decoder D across a large number of client nodes C1, ..., CK to be better integrated, without abandoning the basic idea of ​​federated training that the sets T1, ..., TK of the training examples used by each client node C1, ..., CK remain with each client node C1, ..., CK and are not handed over.

[0019] Thus, in a particularly advantageous embodiment, the semantic meanings s(x) for the training examples x ∈ T1, ...,x ∈ TK are acquired on each client node C1, ..., CK. These semantic meanings s(x) and / or their representations R(s(x)) are shared between the client nodes C1, ..., CK.

[0020] In a further particularly advantageous embodiment, each client node C1, ..., CK applies a generalization function ϕ to the semantic meanings s(x) of the training examples x ∈ T1, ..., x ∈ TK, which removes information related to the specific client node C1, ..., CK from the semantic meanings s(x). The results ϕ(s(x)) thus obtained are shared between the client nodes C1, ..., CK instead of the semantic meanings s(x) in order to fulfill their function during further processing. In this way, an inference from the semantic meanings s2 to personal or otherwise sensitive information from the training examples x2 of external client nodes C1, ..., CK can be prevented even if the semantic meaning s2 is actually "too good", i.e., too detailed.For example, a textual description can be generalized as a semantic meaning s2 by replacing terms with their generic terms. If, for example, luxury sports cars of a certain brand are so rare due to their price that owning them immediately de-anonymizes their owner, this brand can be generalized to "luxury sports car" using the generalization function ϕ. In a not quite so extreme example, shoulder pads with Velcro strips, which are now only found on a few shoulders and whose mention would severely restrict a possible group of people, can be generalized to "accessories" or "haberdashery." Specific measurements of people, such as height or weight, can also be generalized to ranges ("between 180 cm and 200 cm", "between 92 kg and 115 kg").

[0021] In another particularly advantageous embodiment, a direction V is determined based on the representations R(s1), R(s2) of the semantic meanings s1, s2, in which the new representation R* is determined in the working space Z, starting from the representation R(x1) of the training example x1. In this way, this new representation R* can literally approach the subset T2 without this subset T2 having to be known in detail. For example, a gradient in the working space Z of the representations R, which leads to the new training example R*, can be determined from representations R(s1), R(s2) of textual descriptions as semantic meanings s1, s2 and applied to a "reference point" R(x1) that was obtained from an image as the training example x1. The working space Z serves as a common denominator for unifying the very different data types.How far the determined direction ∇ is advanced to determine the new representation R* can be specified, for example, as a hyperparameter λ.

[0022] For example, the new representation R* can be R*=R(x1)+λR(s2)−R(s1)‖R(s2)−R(s1)‖2 be obtained.

[0023] At this point, not only a single representation s2 needs to be considered. Rather, for example, for several semantic meanings s2 corresponding to a top-n selection of similarities S(R(s1), R(s2)), the resulting directions R(s2) - R(s1) / ||R(s2) - R(s1)||2 can be aggregated, for example, by averaging.

[0024] Optionally, those other client nodes C1, ..., CK from which semantic meanings s2, or representations R(s2), are obtained, can be randomly selected. In particular, a predefined small number can be randomly selected from the total available number of other client nodes C1, ..., CK. This reduces the additional effort for improved coordination between the client nodes C1, ..., CK, especially in larger-scale federated training with a large number of client nodes C1, ..., CK.

[0025] The training examples x can, in particular, comprise time series of at least one measured variable and / or images or point clouds that indicate a spatial distribution of at least one measured variable. An image can, in particular, consist of pixels arranged in a regular grid, the pixel values ​​of which represent values ​​of at least one measured variable. A point cloud can assign values ​​of at least one measured variable to any location in space, whereby these points do not have to be regularly arranged and do not have to completely cover a spatial region. Cameras, including those for thermal radiation or other non-visible light, typically deliver images. Lidar and radar sensors, which observe the scene with directed interrogation radiation, usually deliver their measurement data as point clouds. Ultrasonic sensors are an example of a measurement modality that can deliver either images or point clouds.

[0026] As previously described, the semantic meanings s(x) can, in particular, include textual descriptions of the training examples x. In these descriptions, the essential content of the respective training example x can, in particular, be encoded in a form that does not allow any conclusions to be drawn about the specific training example x and, at the same time, is highly compressed compared to the training example x.

[0027] It is particularly advantageous to train or further train the decoder D with one or more new representations R* as training examples. Especially in federated training, this has the effect that the training contributions c1, ..., cK provided by the individual client nodes C1, ..., CK are better coordinated and can be more effectively aggregated.

[0028] This training or further training can be supervised or unsupervised, depending on the application. For supervised training, a target output must be obtained for each new training example R*. For this purpose, manual labeling of the new training example R* can be requested, for example. In this case, it is particularly advantageous not to generate just any new training examples R*, but rather those that specifically advance the coordination of the client nodes C1, ..., CK in the federated training. These are training examples for which the effort of manual labeling is particularly worthwhile. Instead of a manually assigned label, a "pseudo-label" can also be determined from existing labels.

[0029] In this respect, the method described here can also be considered an "active learning" method, since new training examples are specifically selected for a specific purpose. Specifically, starting from a client node C1, ..., CK, one or more other client nodes C1, ..., CK are selected, each of which has representations R(s2) with the greatest similarity to the representation R(s1).

[0030] The output 0 of the neural network N can, in particular, comprise, for example, classification scores relating to one or more classes of a given classification. The neural network N can therefore be a classifier, and here in particular, for example, an image classifier. The classes can, in particular, refer to types of objects whose presence is indicated by images or other inputs I of the neural network N.

[0031] The ultimate goal pursued with the improved training of the neural network N is to obtain outputs 0 for the same inputs I that are better in the context of the specific application at hand. Therefore, in a further particularly advantageous embodiment, measurement data recorded with at least one sensor are fed to the trained neural network N as inputs I. A control signal is determined from the outputs 0 subsequently supplied by the neural network N. A vehicle, a driver assistance system, a robot, an area monitoring system, a quality control system, and / or a medical imaging system is controlled with the control signal. In this way, the probability that the reaction executed by the respective technical system in response to the control signal is appropriate to the situation embodied by the measurement data is advantageously increased.

[0032] The method can, in particular, be fully or partially computer-implemented. Therefore, the invention also relates to a computer program with machine-readable instructions that, when executed on one or more computers and / or compute instances, cause the computer(s) and / or compute instances to execute the described method. In this sense, control units for vehicles and embedded systems for technical devices that are also capable of executing machine-readable instructions are also to be regarded as computers. Compute instances can, for example, be virtual machines, containers, or serverless execution environments, which can be provided in a cloud, in particular.

[0033] The invention also relates to a machine-readable data carrier and / or a downloadable product containing the computer program. A downloadable product is a digital product that can be transmitted over a data network, i.e., downloaded by a user of the data network, and which can be offered for immediate download, for example, in an online shop.

[0034] Furthermore, one or more computers and / or compute instances may be equipped with the computer program, the machine-readable data carrier or the download product.

[0035] 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. Examples of implementation

[0036] It shows: Fig. 1 Embodiment of the method 100 for augmenting a set T of training examples x for a neural network N; Fig. 2 Illustration of the generation of a new representation R* using method 100.

[0037] Fig. 1 is a schematic flow diagram of an embodiment of the method 100 for augmenting a set T of training examples x for a neural network N. The neural network N comprises an encoder E for converting the input I of the neural network into a representation R in a workspace Z and a decoder D for determining the output O of the neural network N from the representation R. The execution of the neural network N is thus a sequential execution D°E of the decoder D after the encoder E.

[0038] In step 110, two subsets T1 and T2 are selected from the set T of training examples x.

[0039] According to block 111, at least the decoder D of the neural network N can be trained in a federated manner. This federated training is an interaction of several client nodes C1, ..., CK, which create training contributions c1, ..., cK, with a server node Q. The server node Q collects the training contributions c1, ..., cK and aggregates them into changes ΔÑ of parameters Ñ that characterize the behavior of the neural network N.

[0040] According to block 112, the subsets T1 and T2 can then be taken from the respective sets T1, ..., TK of the training examples intended for the training of different client nodes C1, ..., CK.

[0041] In step 120, semantic meanings s(x) are obtained for all training examples x E T1 ∪ T2 from the two subsets T1 and T2.

[0042] According to block 121, a trained basic model G can be used for this purpose.

[0043] According to block 121a, a subnetwork V already contained in the basic model G can be reused as encoder E.

[0044] If at least the decoder D is trained federated according to blocks 111 and 112, the semantic meanings s(x) for the training examples x ∈ T1, ..., x ∈ TK can be obtained on each client node C1, ..., CK according to block 122. According to block 123, these semantic meanings s(x) and / or their representations R(s(x)) can then be shared between the client nodes C1, ..., CK.

[0045] According to block 123a, each client node C1, ..., CK can apply a generalization function ϕ to the semantic meanings s(x) of the training examples x ∈ T1, ..., x ∈ TK, which removes information related to the specific client node C1, ..., CK from the semantic meanings s(x). Then, according to block 123b, the results ϕ(s(x)) thus obtained can be shared between the client nodes C1, ..., CK instead of the semantic meanings s(x).

[0046] In step 130, the semantic meanings s(x) are converted into representations R(s(x)) in the workspace Z. As previously explained, they thereby become computable with the representations R(x) of training examples x.

[0047] In step 140, for pairs (R(s1(x1 ∈ T1)),R(s2(x2 ∈ T2))) of representations R(s1, s2) of semantic meanings s1, s2 to training examples x1, x2, which were taken from different subsets T1, T2, a similarity S(R(s1), R(s2)) is determined according to a given similarity metric S.

[0048] In step 150, a selection S*(R(s1),R(s2)) of the best similarities S(R(s1),R(s2)) is determined. As previously explained, this can be, for example, a top-n selection.

[0049] In step 160, for each selected similarity S*(R(s1),R(s2)), a new representation R* is determined as a new training example for the decoder D using the representation R(x1) = E(x1) of a participating training example x1 as well as the representations R(s1),R(s2) of the semantic meanings s1, s2. This augments the existing set T of training examples x with respect to the neural network N as a whole. The new training example is simply not fed to the encoder E, but first to the decoder D.

[0050] In particular, according to block 161, a direction ∇ can be determined based on the representations R(s1), R(s2) of the semantic meanings s1, s2. Then, according to block 162, starting from the representation R(x1) of the training example x1 in this direction ∇, the new representation R* in the workspace Z can be determined.

[0051] In the Fig. In the example shown in Figure 1, the decoder D is trained or further trained in step 170 with one or more new representations R* as training examples. In this process, parameters Ñ that characterize the behavior of the neural network N are optimized. The fully optimized state of these parameters Ñ is denoted by the reference symbol Ñ* and defines the fully trained state N* of the neural network N.

[0052] Then, in step 180, N* measurement data recorded with at least one sensor are fed to the trained neural network as inputs I. This produces outputs of 0.

[0053] In step 190, a control signal 190a is determined from these outputs 0.

[0054] In step 200, a vehicle 50, a driver assistance system 51, a robot 60, an area monitoring system 70, a quality control system 80, and / or a medical imaging system 90 are controlled with the control signal 190a.

[0055] Fig. 2 illustrates the course of processing during the implementation of the method 100. In the Fig. In the example shown in Figure 2, a federated training of the decoder D is carried out with two client nodes C1 and C2, whereby the server node Q is not shown for the sake of clarity.

[0056] The total set T of training examples x is divided into a first subset T1 of training examples x1 and a second subset T2 of training examples x2. The first client node C1 trains on the first subset T1, and the second client node C2 trains on the second subset T2. Each client node C1, C2 applies a base model G to assign semantic meanings s1 and s2 to the respective training examples x1 and x2.

[0057] The semantic meanings s1 and s2 are respectively transformed into representations R(s1) and R(s2) in the working space Z, and similarities S(R(s1), R(s2)) are determined for pairs (R(s1), R(s2)) of these representations. Pairs (R(s1), R(s2)) whose similarities S(R(s1), R(s2)) belong to the selection S*(R(s1),R(s2)) of the best similarities are then used to determine new training examples for the decoder D.

[0058] The selection S*(R(s1),R(s2)) can be determined using any strategy. For example, for each semantic meaning s1 of a training example x1 from the subset T1, the semantic meaning s2 of a training example x2 from the subset T2 can be selected for whose representations R(s1) and R(s2) the best similarity S(R(s1), R(s2)) results. This will then generate as many new training examples for the decoder D as there are training examples x1 in the subset T1. However, it is also possible, for example, to select from a pool of all pairs eR( s1 ),R(s2)), those with the best similarities S(R(s1), R(s2)) are selected. Then, starting from some training examples x1 in the subset T1, several new training examples for the decoder D can be generated, but starting from other training examples x1 in the subset T1, none.

[0059] In the Fig.In the snapshot shown in Figure 2, a representation R(x1) of the training example x1, to which the semantic meaning s1 belongs, has been generated in the working space Z for a selected pair (R(s1), R(s2)). Based on this, a direction ∇ is determined from the representations R(s1) and R(s2), in which the new training example for the decoder D is searched. By progressing in this direction V by the specified amount λ, a new representation R* is obtained as this new training example.

Claims

[1] Method (100) for augmenting a set T of training examples x for a neural network N = D°E, comprising an encoder E for converting the input I of the neural network into a representation R in a workspace Z and a decoder D for determining the output O of the neural network N from the representation R, comprising the steps: • two subsets T1 and T2 are selected from the set T of training examples x (110); • for all training examples x ∈ T1 ∪ T2 from these two subsets T1 and T2, semantic meanings s(x) are obtained (120); • the semantic meanings s(x) are transformed into representations R(s(x)) in the workspace Z (130); • for pairs (R(s1(x1 ∈ T1)),R(s2(x2 ∈ T2))) of representations R(s1,s2) of semantic meanings s1, s2 to training examples x1, x2 from different subsets T1, T2, a similarity S(R(s1), R(s2)) is determined according to a given similarity metric S (140); • a selection S*(R(s1), R(s2)) of the best similarities S(R(s1), R(s2)) is determined (150); and • for each selected similarity S*(R(s1), R(s2)) a new representation R* is determined as a training example for the decoder D using the representation R(x1) = E(x1) of a training example x1 involved in it as well as the representations R(s1),R(s2) of the semantic meanings s1, s2 (160). [2] Method (100) according to claim 1, wherein the semantic meanings s(x) are obtained by processing the training examples x with a trained baseline model G (121). [3] Method (100) according to claim 2, wherein a subnetwork V already contained in the basic model G is reused as encoder E (121a). [4] Method (100) according to one of claims 1 to 3, wherein • at least the decoder D of the neural network N is trained in a federated manner (111) in the interaction of several client nodes C1, ..., CK, which create training contributions c1, ..., cK, with a server node Q, which aggregates these training contributions c1, ..., cK into changes ΔÑ of parameters Ñ that characterize the behavior of the neural network N; and • the subsets T1 and T2 are taken from the respective sets T1, ...,TK of the training examples intended for the training of different client nodes C1, ..., CK (112). [5] Method (100) according to claim 4, wherein • on each client node C1, ..., CK the semantic meanings s(x) for the training examples x ∈ T1, ..., x E TK are obtained (122), and • these semantic meanings s(x), and / or their representations R(s(x)), are shared between the client nodes C1, ..., CK (123). [6] Method (100) according to claim 5, wherein • each client node C1, ..., CK applies a generalization function ϕ (123a) to the semantic meanings s(x) of the training examples x ∈ T1, ..., x ∈ TK, which removes information related to the concrete client node C1, ..., CK from the semantic meanings s(x); and • the results ϕ(s(x)) thus obtained are shared between the client nodes C1, ..., CK instead of the semantic meanings s(x) (123b). [7] Method (100) according to claim 6, wherein at least one textual description is generalized as a semantic meaning s(x) by replacing at least one term in the textual description by a generic term of that term. [8] Method (100) according to one of claims 1 to 7, wherein on the basis of the representations R(s1), R(s2) of the semantic meanings s1, s2 a direction V is determined (161), in which starting from the representation R(x1) of the training example x1 the new representation R* in the workspace Z is determined (162). [9] Method (100) according to one of claims 1 to 8, wherein the training examples comprise x time series of at least one measured variable, and / or images or point clouds indicating a spatial distribution of at least one measured variable. [10] Method (100) according to one of claims 1 to 9, wherein the semantic meanings s(x) comprise textual descriptions of the training examples x. [11] Method (100) according to one of claims 1 to 10, wherein the decoder D is trained or further trained (170) with one or more new representations R* as training examples. [12] Method (100) according to claim 11, wherein • the trained neural network N* is fed with measurement data recorded with at least one sensor as inputs I (180), • a control signal (190a) is determined (190) from the outputs 0 subsequently provided by the neural network N*, and • a vehicle (50), a driver assistance system (51), a robot (60), a system (70) for monitoring areas, a system (80) for quality control, and / or a system (90) for medical imaging, with which the control signal (190a) is controlled (200). [13] A computer program comprising machine-readable instructions which, when executed on one or more computers and / or compute instances, cause the computer or compute instances to execute a method (100) according to any one of claims 1 to 12. [14] Machine-readable data carrier and / or download product with the computer program according to claim 13. [15] One or more computers and / or compute instances with the computer program according to claim 13, and / or with the machine-readable data carrier and / or download product according to claim 14.

Images