Federated learning with improved coordination between participating clients

Abstract

Method (100) for generating a training contribution for a neural network (1) on a client node C1, ..., CK for federated training of this neural network (1), comprising the steps: • an original set of parameters W is received from a server node Q as a starting point for generating a new version W' of these parameters (110); • a set of anchor examples a is received from the server node Q (120); • training examples x from a given set D are processed by the neural network (1) parameterized with the new version W' of the parameters W to produce outputs y; • anchor examples a are processed under new parameters W' and under original parameters W, and a processing product A' is provided to the neural network (1) at the same position in each case.A taken from (140);• with a given cost function L, both deviations &Dgr;(y, y*) of the outputs y from the respective target outputs y* and results ϕ(A', A),ϕ(R(A'), R(A)) of a comparison of processing products A', A, or of derived quantities R(A'), R(A), are evaluated with a comparison metric ϕ (150);• the new version W' of the parameters W is optimized to the goal (160) that, with further processing of training examples x, the evaluation by the cost function L is improved; and• the result W* of the optimization of the new version W' of the parameters W is transmitted as a training contribution for the federated training to the server node Q (170).

Description

The present invention relates to federated neural network training in which a plurality of client nodes C1,..., CK cooperate in a coordinated manner by a server node Q.Prior ArtThe training of neural networks, as used for example for classification and / or semantic segmentation of images, requires a large number of training examples with sufficient variability. The effort required for storing these training examples and the actual training can be too great for an individual entity. Also, for legal reasons, it is not always possible to merge all training examples at an entity performing the training. For example, an image classifier for monitoring the environment of a vehicle driving at least partially in an automated manner requires images that can contain identifiers, faces and other personal data as training examples. If this image classifier is to be trained to work equally well not only in North America and Europe, but also in other world regions, merging all training examples for performing training on privacy laws, such as European Basic Privacy Regulation (DSGVO), may fail.One way is federated learning, in which a plurality of client nodes C 1,..., CK train the neural network in each case on a local dataset of training examples and the respective work results are collected at a server node Q. The training contributions made by the individual client nodes C1,..., CK are collected by the server node Q and aggregated to form an overall training result.Disclosure of the InventionThe invention provides a method for generating a training contribution for a neural network on a client node C1,..., CK for federated training of the neural network.This method assumes that the client node C1,..., CK receives from a server node Q an original set of parameters W characterizing the behavior of the neural network as the starting point for the development of a new version W' of these parameters. These parameters can in particular comprise, for example, weights with which inputs which are supplied to each neuron of the neural network are summed up to activate this neuron. Since the optimization described below also refers to the original parameters W, the client node C1,..., CK can, in particular, for example, create a working copy W' of the original parameters W in order to keep access to the original parameters W at any time. However, it is also possible, for example, to initially perform all calculations for which the original parameters W are required at the beginning of the optimization and then to overwrite these with the new version W' in situ in the memory. In either case, the new version W' to be created is initialized with the original version W of the parameters.The client node C1,..., CK further receives a set of anchor examples a from the server node Q. These anchor examples a should advantageously cover that domain of the input data on which the fully trained neural network is to operate.Training examples x from a predefined set D are fed to the neural network parameterized with the new version W' of the parameters W. These training examples x are each labeled with desired outputs y* that the neural network should deliver upon input of the respective training example x. The neural network processes the training examples into actual outputs y. Deviations Δ(y,y*) of the outputs y from the respective target outputs y* provide feedback for the optimization of the parameters W'. In particular, in federated training, in which many client nodes C 1,..., CK cooperate, these client nodes C 1,..., CK may each operate with different sets D 1,..., DK of training examples x.The anchor examples a are fed on the one hand to the neural network parameterized with the new version W' of the parameters W and on the other hand to the neural network parameterized with the original parameters W. In this case, "originally" is to be understood as meaning in particular that version W of the parameters which was last communicated from the server node Q to the client node C1,..., CK and thus specifies the last state of the overall trained model M trained in a federated manner overall. Typically, training the overall model M requires several passes, in which the current state W is communicated to the client nodes C1,..., CK in each case and, starting from this, each client node C1,..., CK processes its own new version W'.A processing product A' or A is taken from the neural network parameterized with the new version W' of the parameters or with the original version W of the parameters at the same point in its architecture. By comparing these processing products A' and A, and / or variables R(A') and R(A) derived from these processing products A' and A, respectively, with one another with a comparison metric φ, further feedback for the optimization of the parameters W' can be obtained. As explained above, the original parameters W for the calculation of the processing products A or the derived variables R(A) can be stored in the memory, so that the new version W' is optimized with a working copy. However, it is also possible, for example at the beginning of the optimization, to perform all calculations in which the original parameters W occur in advance and to keep the results, while the original parameters W themselves are overwritten by the new version W'. It depends on the particular application, which is overall more efficient with respect to the memory requirement.In order to combine the feedback from both said sources, a predetermined cost function L is used• Both deviations Δ(y, y') of the outputs y from the respective desired outputs y*• and results φ(A', A), φ(R(A'), R(A)) of a comparison of processing products A', A or of variables R(A'), R(A) derived therefrom with the comparison metric φevaluated. In this case, these two contributions to the cost function L can be weighted with one another in any desired manner. For example, both contributions may add additively to the cost function L, and a contribution may be multiplied by a factor λ.The new version W' of the parameters W is optimized to the aim that, when training examples x are further processed, the evaluation by the cost function L is improved. The result W* of the optimization of the new version W' of the parameters W is transmitted to the server node Q as a training contribution for the federated training.It has been recognized that, by the additional feedback on the basis of the anchor examples a, the training of a plurality of client nodes C1,..., CK can be coordinated such that the changes of the parameters W' proposed by all client nodes C1,..., CK point substantially in the same general direction. Different data sets D1,..., DK of training examples x on which the client nodes C1,..., CK operate may be very heterogeneous, in particular with regard to the difficulty of training. If, for example, images of traffic situations are to be analyzed in the context of at least partially automated driving, this task is easier to learn on a data record of clearly structured traffic situations with marked roads, traffic lights and traffic signs than on a data record of traffic situations in which unpaved roads and / or a fallen infrastructure prevail. Training on higher quality images is also simpler than on poorer quality images, for example.Different degrees of difficulty in training may in turn result in different client nodes C1,..., CK proposing changes to certain parameters W' pointing in widely different or even opposite directions. If these different suggestions are then aggregated by the central server node Q, for example by averaging, they can cancel each other out. Thus, for example, the mean value of "avoid the obstacle to the left" and "avoid the obstacle to the right" is to continue straight ahead and collide with the obstacle.This can be illustrated by a simple example in which an image classifier for the classification of black and white images is to be trained in a federated manner. For a first client node C1, given its set D1 of training examples x, it may be advantageous to consider these training examples x as representations of white features on a black background. For a second client node C 2, on the other hand, it may be advantageous in view of its set D 2 of training examples x to consider these training examples x as representations of black features on a white background. Thus, one and the same new parameters W' have completely different meanings in the context of both client nodes C1 and C2. It is then not very meaningful to aggregate these parameters W' in a meaningful manner, as one cannot reasonably average quantities relating to completely different measurement units (for example length in meters and electrical voltage in volts).Precisely this tendency is suppressed by the use of the anchor examples a. By all the client nodes C1,..., CK using the same anchor examples a and the cost function favouring when these anchor examples a are processed by all the client nodes C1,..., CK to form similar processing products A' or variables R(A') derived therefrom, a tendency is introduced that the individual client nodes C1,..., CK can be guided during their optimization by the basic structures that the totality of all the client nodes C1,..., CK has already learned. These basic structures are embodied in the original parameters W that the server node Q has made available to the client nodes C1,..., CK as starting point for its current training step. Thus, with respect to the cost function L, it is unfavorable for the client nodes C1,..., CK to learn dynamics that are fundamentally different from the dynamics learned by all the client nodes C1,..., CK so far. The probability is thus increased that the training contributions of all client nodes C 1,..., CK can be aggregated in a meaningful manner to form an overall result.In a particularly advantageous embodiment, the variables R(A'), R(A) derived from the processing products A', A additionally depend on the respective training example x. In this way, the different sets D1,..., DK of training examples x with which the different client nodes C1,..., CK operate can enter into the check as to whether the common anchor examples a are processed uniformly. In particular, the cost function is then no longer only entered at points by the behavior of the neural network on a small set of anchor examples a. Rather, in each case a planar behavior on the respective set D1,..., DK of training examples x can be examined, which is parameterized with the anchor examples a.In particular, for example, the derived variables R(A'), R(A) can contain at least one similarity measure ψ(X, A'), ψ(X, A) of the processing product A', A with the processing product X determined in the same way from the respective training example x. This supports the fact that all client nodes C 1,..., CK map their respective training examples x into spaces of processing products X that have a uniform coarse structure.In a particularly advantageous embodiment, the derived variables R(A'), R(A) contain similarity measures ψ(X,A'(a 1)),..., ψ(X,A'(a N)) or ψ(X,A(a 1)),...,( X,A(a N)), which compare one and the same processing product X of the training example x with processing products A',A of all N anchor examples a 1,..., a N respectively. Thus, the derived variable R(A') determined with the new parameters W' of a client node i becomes a relative representation R i of the respective training example x with respect to this client node i. At the same time, the derived variable R(A) determined with the original parameters W becomes a relative representation R M( x) of the respective training example x with respect to the overall model M to be trained in a federated manner overall.R M( x) express relations of the respective training example x to the anchor examples a 1,..., a N in the context of the model currently trained by the client node i or in the context of the overall model M. These relative representations R i( x) and R M( x) can then be compared with the comparison metric φ. The cost function (loss function) L i for the client node i can then be written, for example, as wherein L y,i( x) is the loss relating to the comparison of the output y obtained for the respective training example x with the respective desired output y*.The similarity measure 0 can in particular include, for example, a cosine similarity. Such a similarity measure has good isometric invariance properties.The comparison metric φ can in particular include, for example, a norm, and / or another distance measure, and / or a power (for example a square) thereof. In this way, in particular relative representations R i( x) and R M( x) in a space spanned by the N anchor examples a 1,..., a N can be reasonably compared with one another.In a further particularly advantageous embodiment, a neural network is selected which is capable of being transformed into a neural network.• a feature extractor for converting the input of the neural network into a representation in a workspace, and• a task header for determining the output of the neural network from this representation.The representation in the workspace is then used as the processing product A' and A, respectively. The processing product A' or A is then something which was provided from the outset as a finished product in the design of the neural network.For example, a convolutional neural network (CNN) may be provided as a feature extractor. Such a network produces successive feature maps that first extract fundamental features from the input and then combine these fundamental features into more complex features. The last feature map in the chain is then the representation in the workspace that can be evaluated by any task head. For example, a classifier can map the representation, and thus also the original input, to classification scores with respect to one or more classes of a predefined classification. It is also possible, for example, for a regressive device to ascertain from the representation, and thus from the original input, a value of interest of one or more variables.In a further particularly advantageous embodiment, a neural network is selected which is capable of being transformed into a neural network.• an encoder for converting the input of the neural network into a representation with reduced dimensionality in a workspace, and• a decoder for converting this representation into an output of the neural network with again increased dimensionality. The representation in the workspace is then used as the processing product A' and A, respectively. Here too, the representation in the workspace has already been provided as a finished product during the design of the neural network.In particular, in such workstations, it is easier to find a well-instantiated measure of similarity between processing products. Since the representation with reduced dimensionality forms a bottleneck, as it were, by means of which the encoder forces the information into the decoder, the neural network learns to take over the information that is most important for the respective application into the representation. Accordingly, the similarity of processing products may also be attributed a meaning in terms of application.In a further particularly advantageous embodiment, the server node Q transmits the original set of parameters W to a plurality of client nodes C1,..., CK which operate with different sets D1,..., DK of training examples x. The training contributions received from these client nodes C1,..., CK are aggregated to form an overall change of the parameters W of the neural network 1. As explained above, synchronizing the processing products A', A formed by the individual client nodes C1,..., CK from anchor examples a ensures that the client nodes C1,..., CK essentially all operate according to the same basic structure. Thus, for example, a client node C 1,..., CK whose set D 1,..., DK contains particularly simple training examples x does not take an "abbreviation" in the processing, which is easier to learn, but whose working results are poorly aggregated with the working results of the other client nodes C 1,..., CK. Conversely, a special solution found by a client node C1,..., CK with a particularly difficult set D1,..., DK of training examples x can also be particularly good from the perspective of this client node C1,..., CK with respect to the predefined application, but can also stand again in the field of the other client nodes C1,..., CK alone, so that the aggregating also falls again with difficulty. A solution which is not quite perfect with regard to the actual application, but which also provides a training contribution usable for cooperation with the other client nodes C1,..., CK, would be better.This is roughly comparable to the fact that almost all client nodes C1,..., CK come to training contributions that can be represented on a DIN A4 sheet. The server node Q wants to drag all these contributions through the scanner and summarize them into a final result. If a client node C1,..., CK now issues a particularly simple solution on a DIN A5 sheet or even a DIN A6 sheet in order to save paper, the single sheet feeding of the scanner cannot capture this sheet and it is lost. If another client node C1,..., CK dispenses a particularly difficult solution on two stapled DIN A4 sheets or on one DIN A3 sheet, this clogs the single sheet feeder.Aggregating can in particular comprise, for example, the formation of an average value and / or median. These forms of aggregating would be susceptible to interfering with or even cancelling the training contributions from each other in training contributions that do not point in uniform directions.In a further particularly advantageous embodiment, measurement data that have been recorded using at least one sensor are supplied to the trained neural network. A control signal is formed from the output subsequently supplied by the neural network. A vehicle, a robot, a driving assistance system, a system for quality control, a system for monitoring regions, and / or a system for medical imaging, is actuated with this control signal. Since the training contributions of all client nodes C 1,..., CK can be aggregated better, the overall accuracy achieved by the federated trained neural network is better. Accordingly, the probability is increased that the reaction carried out by the respective controlled technical system in response to the control signal is appropriate for the situation embodied by the measurement data.As explained above, the neural network can be designed in particular as an image classifier. Image classifiers, which can be used, for example, for monitoring the environment of vehicles and / or robots, must be trained very frequently in a federated manner. On the one hand, the data volume of the images is overall too large to collect all images at a central entity. On the other hand, such a collection may fail because images are currently considered to be person-related data and the transfer across the border between countries with different levels of data protection is not permitted.The method can be fully or partially computer-implemented in particular. Therefore, the invention also relates to a computer program having machine-readable instructions which, when executed on one or more computers and / or compute instances, cause the computer or computers and / or compute instances to carry out the described method. In this sense, control units for vehicles and embedded systems for technical devices, which are also capable of executing machine-readable instructions, are also to be regarded as computers. Compute instances may be, for example, virtual machines, containers, or server-less execution environments, which may be provided in particular in a cloud.The invention also relates to a machine-readable data carrier and / or to a download product with the computer program. A download product is a digital product that can be transmitted via a data network, i.e. downloaded by a user of the data network, and can be stored in an online shop for immediate download, for example.Furthermore, one or more computers and / or computer instances can be equipped with the computer program, with the machine-readable data carrier or with the download product.Further measures which improve the invention are described in more detail below together with the description of the preferred exemplary embodiments of the invention on the basis of figures.Exemplary EmbodimentsIt shows: FIG. 1 shows an exemplary embodiment of the method 100 for generating a training contribution for federated training of a neural network 1; FIG. 2 illustrates the effect of unifying the operation of client nodes C1,..., CK.FIG. 1 is a schematic flow diagram of an embodiment of the method 100 for generating a training contribution for federated training of a neural network 1.In step 110, an original set of parameters W characterizing the behavior of the neural network 1 is received from a server node Q as the starting point for the development of a new version W' of these parameters. That is, this new version W' is initialized with the original version W regardless of whether this original version W is maintained or overwritten with the new version W'.In step 120, a set of anchor examples a is received from the server node Q.In step 130, training examples x from a predefined set D are supplied to the neural network 1 parameterized with the new version W', so that the neural network 1 supplies outputs y in each case. In this case, the training examples x are each labeled with desired outputs y*.In step 140, anchor examples a are fed on the one hand to neural network 1 parameterized with new version W' of parameters W and on the other hand to neural network 1 parameterized with original parameters W. In this case, a processing product A' (if the network 1 is parameterized with the new parameters W') or A (if the network 1 is parameterized with the original parameters) is respectively taken from the neural network 1 at the same point.According to block 105, a neural network 1 may be selected, which• a feature extractor 1 afor converting the input 2 of the neural network 1 into a representation 3 in a workspace, and• a task header 1 bfor ascertaining the output 4 of the neural network 1 from this representation 3has. According to block 141, this representation 3 can then be used in the workspace as a processing product A' or A, respectively.According to block 106, a neural network 1 may be selected, which• an encoder 1c for converting the input 2 of the neural network 1 into a representation 3 with reduced dimensionality in a workspace, and• a decoder 1d for converting this representation 3 into an output 4 of the neural network 1 with again increased dimensionalityhas. According to block 142, this representation 3 can then be used in the workspace as a processing product A' or A, respectively.In step 150, deviations Δ(y, y*) of the outputs y from the respective desired outputs y* and results φ(A', A), φ(R(A'), R(A)) of a comparison of processing products A', A or of variables R(A'), R(A) derived therefrom are evaluated using a predefined cost function L using a comparison metric φ.According to block 151, the quantities R(A'), R(A) derived from the processing products A', A may additionally depend on the respective training example x.According to block 152, the derived quantities R(A'), R(A) may include at least one similarity measure ψ(X,A'), ψ(X,A) of the processing product A',A with the processing product X similarly determined from the respective training example x.The derived variables R(A'), R(A) can, according to block 152 a, in particular include, for example, similarity measures ψ(X,A'(a 1)),..., ψ(X,A'(a N)) or ψ(X.A(a 1)),..., ψ(X,A(a N)) which compare one and the same processing product X of the training example x with processing products A', A of all N anchor examples a 1,..., a N respectively.According to block 152 b, the similarity measure ψ may in particular include, for example, a cosine similarity.According to block 153, the comparison metric φ may in particular include, for example, a norm and / or another distance measure.In step 160 the new version W' of the parameters W is optimized to the aim that, when training examples x are further processed, the evaluation by the cost function L is improved.In step 170, the result W* of the optimization of the new version W' of the parameters W is transmitted to the server node Q as a training contribution for the federated training.According to block 111, the server node Q may transmit the original set of parameters W to a plurality of client nodes C 1,..., CK operating with different sets D 1,..., DK of training examples x. It can then aggregate the training contributions received from these client nodes C1,..., CK to form an overall change of the parameters W of the neural network 1.In the example shown in FIG. 1, measurement data 2 that were recorded using at least one sensor 5 are fed to the trained neural network 1* in step 190. From the output 4 subsequently supplied by the neural network 1*, a control signal 6 is formed in step 200. In step 210, a vehicle 50, a robot 51, a driving assistance system 60, a system 70 for quality control, a system 80 for monitoring regions, and / or a system 90 for medical imaging, is actuated with the actuation signal 6.FIG. 2 illustrates the effect of considering anchor points a on federated training of a neural network using multiple client nodes C 1,..., CK.In the example shown in FIG. 2, there are four anchor examples a 1, a 2, a 3, a 4. It is desirable that the internal processing chain of the network trained locally on each client node C1,..., CK takes a form outlined rectangularly in Fig. 2. Derived variables R(A') of processing products A' of the anchor points, which can also depend, for example, on the specifically used training examples x, are to be seated in the corners of the rectangle. For the client nodes C2,..., CK, the state shown in FIG. 2 is also already the case because the rectangular structure also already corresponds quite well to the shape that the processing chain would have to have in order to reproduce the desired outputs y* identically for all training examples x with the outputs y of the respective network.However, the client node C1 has allocated a set D1 of training examples x which can be most easily solved with a processing chain in triangular form, in the sense that the outputs y of the network trained by the client node C1 reproduce the target outputs y* identically. This may be, but a training contribution determined for the triangular processing chain then necessarily has a completely different (inter alia semantic) meaning than a training contribution determined for a rectangular processing chain. The training contribution would thus be poorly aggregated with the training contributions provided by rectangular processing chains. In this situation, the additional consideration of the anchor examples a causes the processing chain on the client node C1 to still assume the rectangular shape and the derived quantities R(A') of processing products A' corresponding to the anchor examples a 1, a 2, a 3, a 4 to migrate into the corners of this rectangle. With regard to the achievable accuracy in the task of the neural network 1 that is actually to be achieved, victims naturally have to be brought about here, since the shape of the processing chain now no longer corresponds to the optimum shape. For federated training with all client nodes C1,..., CK, a training contribution of the client node C1, which, although not quite exactly meeting the content optimum, can be aggregated meaningfully with the training contributions of the client nodes C2,..., CK, is more valuable than a training contribution that speaks quite a different language than the other training contributions and thus interferes more easily during aggregating. At least this disturbance would still be avoidable, in that this training contribution is completely ignored. Then, however, all training examples x from the set D1 for which the client node C1 was responsible are ignored in the training of the neural network 1. This, in turn, is a significantly greater sacrifice in the overall achievable accuracy in the context of the given application than to make the processing chain of the client node C1 rectangular, degrading the accuracy on the set D1, and being able to fully introduce the training contribution of the client node C1 into the federated training.

Claims

Method (100) for generating a training contribution for a neural network (1) on a client node C1,..., CK for federated training of this neural network (1), comprising the steps: • an original set of parameters W characterizing the behaviour of the neural network (1) is received (110) from a server node Q as a starting point for the development of a new version W' of these parameters; • a set of anchor examples a is received (120) from the server node Q; • Training examples x from a predefined set D are fed (130) to the neural network (1) parameterized with the new version W' of the parameters W, so that the neural network (1) supplies outputs y in each case, the training examples x being labeled with setpoint outputs y* in each case; • Anchor examples a are fed on the one hand to the neural network (1) parameterized with the new version W' of the parameters W and on the other hand to the neural network (1) parameterized with the original parameters W, and a processing product A' or A is taken (140) from the neural network (1) in each case at the same point; • Deviations Δ(y, y*) of the outputs y from the respective desired outputs y* and results φ(A', A), φ(R(A'), R(A)) of a comparison of processing products A', A or of variables R(A'), R(A) derived therefrom are evaluated (150) using a comparison metric φ; • the new version W' of the parameters W is optimized (160) to the aim that, when training examples x are further processed, the evaluation is improved by the cost function L; and • the result W* of the optimization of the new version W' of the parameters W is transmitted (170) to the server node Q as a training contribution for the federated training.Method (100) according to claim 1, wherein the quantities R(A'), R(A) derived from the processing products A', A additionally depend (151) on the respective training example x.Method (100) according to one of Claims 1 to 2, wherein the derived variables R(A'), R(A) comprise (152) at least one similarity measure ψ(X, A'), ψ(X, A) of the processing product A', A with the processing product X determined in the same way from the respective training example x.The method (100) of claim 3, wherein the derived quantities R(A'), R(A) include similarity measures ψ(X,A'(a 1)),..., ( X,A'(a N)) and ψ(X,A(a 1)),..., ( X,A(a N)) respectively, that compare (152a) one and the same processing product X of the training example x with processing products A', A of all N anchor examples a 1,..., a N respectively.The method (100) of any of claims 3 to 4, wherein the similarity measure ψ includes a cosine similarity (152b).The method (100) of any of claims 1 to 5, wherein the comparison metric φ includes (153) a norm, and / or another distance measure.Method (100) according to one of Claims 1 to 6, wherein • a neural network (1) is selected (105) which ◯ has a feature extractor (1a) for converting the input (2) of the neural network (1) into a representation (3) in a working space and ◯ has a task header (1b) for ascertaining the output (4) of the neural network (1) from this representation (3), and • the representation (3) in the working space is used (141) as processing product A' or A, respectively.Method (100) according to one of Claims 1 to 7, wherein • a neural network (1) is selected (106) which ◯ has an encoder (1c) for converting the input (2) of the neural network (1) into a representation (3) with reduced dimensionality in a working space and ◯ has a decoder (1d) for converting this representation (3) into an output (4) of the neural network (1) with increased dimensionality again, and • the representation (3) in the working space is used (142) as processing product A' or A, respectively.Method (100) according to any one of claims 1 to 8, wherein the server node Q • transmits (111) the original set of parameters W to a plurality of client nodes C1,..., CK operating with different sets D1,..., DK of training examples x and • aggregates (180) the training contributions received from these client nodes C1,..., CK to an overall change of the parameters W of the neural network (1).The method (180) of claim 9, wherein the aggregating comprises forming an average and / or median (181).Method (100) according to one of Claims 1 to 10, wherein • measurement data (2) which have been recorded using at least one sensor (5) are fed (190) to the trained neural network (1*); • a control signal (6) is formed (200) from the output (4) subsequently supplied by the neural network (1*); and • a vehicle (50), a robot (51), a driver assistance system (60), a system (70) for quality control, a system (80) for monitoring regions, and / or a system (90) for medical imaging, with which control signal (6) is controlled (210).Method (100) according to one of Claims 1 to 11, wherein a neural network (1) designed as an image classifier is selected.A computer program comprising machine readable instructions which, when executed on one or more computers and / or compute instances, cause the computer or computers and / or compute instances to perform a method (100) according to any one of claims 1 to 12.Machine-readable data carrier comprising the computer program according to claim 13.One or more computers and / or compute instances equipped with the computer program according to claim 13 and / or with the machine readable medium according to claim 14.

Images