Method for generating graph structures for training a graph neural network

Abstract

The invention relates to a method for generating a graph structure for training a graph neural network, wherein the method comprises the steps of obtaining data representing a computation graph, wherein the computation graph comprises a plurality of nodes (2) connected by edges; and generating a graph structure for training a graph neural network by removing edges from the computation graph, wherein the edges are removed in such a way that an environment in the computation graph corresponds to an environment in the graph structure (3).

Description

Technical Field

[0001] This invention relates to a method for generating graphical structures for training graphical neural networks, which improves the training and implementation of graphical neural networks and saves computer resources. Background Technology

[0002] Graph Neural Networks (GNNs) are a class of deep learning methods designed to perform inference on data described by graphs, where each graph consists of multiple nodes connected by edges. Such GNNs have become very popular for tasks involving graph data. GNNs typically employ message-passing architectures, such as graph convolutional networks. In such architectures, data or information propagates from nodes in the graph through the neighborhood of the corresponding nodes to generate representations that depend on the broad graph context. Furthermore, each node can represent a corresponding operation performed by the neural network, as part of determining the neural network's output from its input. This makes GNNs suitable for learning many real-world tasks.

[0003] However, real-world problems often involve very large graphs, requiring substantial computing resources to train such graph neural networks. There are scenarios where training these networks may require far more computing resources than a typical system can provide. If the graph neural network is not properly regularized, the size of the graph can often lead to overfitting. Furthermore, graph convolutional networks or graph attention networks are computationally intensive tasks due to their neighborhood aggregation mechanisms, which utilize self-attention computation to assign weights to neighboring nodes.

[0004] To address these issues, it is known to base the training of such graph neural networks on simplified architectures or graph structures. For example, since very large graphs often involve many redundant components, it is possible to remove these redundant components to simplify the training of such graph neural networks.

[0005] Document US 2017 / 0132513 A1 discloses a method for training a neural network represented as a computational graph, wherein data representing the computational graph is obtained. The computational graph is then augmented to generate a training computational graph for training the neural network using a machine learning training algorithm, the machine learning training algorithm including calculating the gradient of an objective function with respect to each parameter of the neural network. Augmenting the computational graph includes inserting multiple gradient nodes and training edges into the computational graph to generate a backward path through the computational graph, the backward path representing operations for calculating the gradient of the objective function with respect to the parameters of the neural network. The neural network is trained using the machine learning training algorithm by executing the training computational graph. Summary of the Invention

[0006] The purpose of this invention is to provide an improved method for generating graphical structures for training graphical neural networks. This method can improve the training and implementation of graphical neural networks, and at the same time, it can save computer resources.

[0007] This objective is achieved through the method of the present invention for generating graphical structures for training graphical neural networks.

[0008] Furthermore, this objective is achieved through the control unit of the present invention for generating graphical structures for training graphical neural networks.

[0009] According to one embodiment of the present invention, this objective is achieved by a method for generating a graphical structure for training a graphical neural network, wherein data representing a computational graph is obtained, wherein the computational graph includes a plurality of nodes connected by edges, and wherein the graphical structure for training the graphical neural network is generated by removing edges from the computational graph, wherein the edges are removed in such a manner that the environment in the computational graph corresponds to the environment in the graphical structure.

[0010] Here, the environment in a graph, or the environment in a computational graph or the generated graph structure, is the structure or composition of nodes that form a part of the graph. The fact that the environment in a computational graph corresponds to the environment in the generated graph structure means that environments similar to those in the computational graph are also included in the generated graph structure.

[0011] Generating simplified structures for training graph neural networks has the advantage of significantly reducing the amount of computing resources required to train graph neural networks based on the generated graph structures, such as storage space and / or processing time. This ensures that graph neural networks can also be trained by ordinary systems with limited resources. Furthermore, training graph neural networks on such simplified graph structures also greatly simplifies the subsequent implementation of graph neural networks.

[0012] Furthermore, the fact that the environment included in the computational graph is also included in the generated graph structure has the added advantage that training a graph neural network based on the generated simplified graph structure does not significantly degrade the performance of the trained graph neural network.

[0013] Therefore, an improved method is provided for generating graphical structures for training graphical neural networks. This method can improve the training and implementation of graphical neural networks, especially those utilizing message passing, while simultaneously saving computer resources.

[0014] In one embodiment, at least one edge attribute is assigned to each edge of the computational graph, wherein the step of generating a graph structure for training a graph neural network by removing edges from the computational graph includes: removing edges from the computational graph based on the edge attributes assigned to the edges of the computational graph.

[0015] Here, edge attributes refer to data associated with edges. Specifically, edge attributes define the relationships between nodes of a computational graph, particularly those nodes connected by corresponding edges. For example, edge attributes can define how many edges are connected to a specific node of the computational graph, or how many edges are connected separately between two specific nodes of the graph.

[0016] By considering the edge properties or relationships between nodes in a computational graph when removing edges, a simplified architecture or graph structure for training a graph neural network can be generated. Simultaneously, it can be ensured that the environment, structure, or composition of nodes in a portion of the computational graph corresponds to the environment in the generated graph structure, or that a similar environment is also included in the generated graph structure.

[0017] The step of generating a graph structure for training the graph neural network by removing edges from the computational graph may include: assigning a signature to each edge of the computational graph based on at least one edge attribute assigned to the corresponding edge, such that similar edges are assigned the same signature; applying a hash function to convert the signatures assigned to the edges of the computational graph into numerical values; and, for each node of the computational graph, retaining the edges of the computational graph that are connected to the nodes of the computational graph and whose signatures are converted into the minimum value among all numerical values, and removing all other edges connected to the corresponding nodes from the computational graph.

[0018] Here, "signature" refers to a graph signature, and more specifically, a mathematical signature that defines graph similarity or similarity between edges. Therefore, based on the graph signature assigned to an edge, it is possible to deduce whether the edge satisfies a given characteristic or whether a subset of the computational graph including the edge satisfies at least one graph property. For example, based on the signature, it is possible to deduce which nodes the corresponding edge connects to.

[0019] Similar edges are the edges of a computational graph that are assigned at least one of the same edge attributes.

[0020] Furthermore, a hash function is any function that can be used to map data of any size to values ​​of a fixed size.

[0021] Therefore, the generated graphical structure can be easily obtained through known mathematical operations, which can even be implemented on ordinary systems with limited computer resources.

[0022] In one embodiment, edge attributes assigned to the edges of the computational graph are generated based on at least one node attribute of at least one node connected to the corresponding edge. Thus, it is conceivable that many computational graphs do not include edge attributes, but include node attributes that define data associated with nodes, such as type, label, style attributes (such as width, if available), or arbitrary data values ​​that can be used with nodes, wherein edge attributes can be derived from node attributes.

[0023] According to another embodiment of the present invention, a method for training a graph neural network is provided, wherein the graph structure for training the graph neural network is generated by the method described above for generating a graph structure for training the graph neural network, wherein training data for training the graph neural network is provided, and wherein the graph neural network is trained based on the generated graph structure and the training data.

[0024] Therefore, a method for training graph neural networks based on improved graph structures is provided. Specifically, training the graph neural network based on a simplified architecture has the advantage of significantly reducing the amount of computing resources required to train the graph neural network based on the generated graph structure, such as storage space and / or processing time. This ensures that the graph neural network can also be trained by ordinary systems with limited resources. Furthermore, training the graph neural network on such a simplified graph structure greatly simplifies the subsequent implementation of the graph neural network. Additionally, the fact that the environment included in the computational graph is also included in the generated graph structure has the added advantage that training the graph neural network based on the generated simplified graph structure does not significantly degrade the performance of the trained graph neural network.

[0025] In addition, training data can include sensor data.

[0026] Sensor data is the output of a device that detects and responds to some type of input from the physical environment. The output can be used to provide information or input to another system, or to guide a process.

[0027] Therefore, the method can be based on an environment outside of the data processing device on which the graph neural network is trained.

[0028] According to another embodiment of the present invention, a method for classifying image data by means of a graph neural network is provided, wherein the graph neural network has been trained by the method for training a graph neural network as described above.

[0029] One method for classifying image data using a graph neural network is one that classifies digital image data based on low-level features of the image (such as edges or pixel attributes).

[0030] Therefore, a method for classifying image data is provided, which is based on a graph neural network trained on an improved graph structure. In particular, the fact that the graph neural network is trained on a simplified architecture has the advantage that the amount of computing resources required to train the graph neural network based on the generated graph structure, such as storage space and / or processing time, can be significantly reduced.

[0031] This ensures that graph neural networks can also be trained by ordinary systems with limited resources. Furthermore, training graph neural networks on this simplified graph structure greatly simplifies subsequent implementations of the graph neural network or subsequent classification of image data. Additionally, the environment included in the computational graph also has the added advantage of including the generated graph structure, meaning that training the graph neural network based on the generated simplified graph structure does not significantly degrade the performance of the trained graph neural network.

[0032] According to another embodiment of the present invention, a control unit for generating a graphical structure for training a graphical neural network is provided, wherein the control unit includes: an obtaining unit configured to obtain data representing a computational graph, wherein the computational graph includes a plurality of nodes connected by edges; and a first generating unit configured to generate a graphical structure for training a graphical neural network by removing edges from the computational graph, wherein the first generating unit is configured to remove edges from the computational graph in such a way that the environment in the computational graph corresponds to the environment in the graphical structure.

[0033] Therefore, a control unit for generating improved graph structures for training graph neural networks is provided. Specifically, training the graph neural network based on a simplified architecture has the advantage of significantly reducing the amount of computing resources required to train the graph neural network based on the generated graph structure, such as storage space and / or processing time. This ensures that the graph neural network can also be trained by ordinary systems with limited resources. Furthermore, training the graph neural network on this simplified graph structure greatly simplifies the subsequent implementation of the graph neural network. Additionally, the environment included in the computational graph also includes the generated graph structure, which has the added advantage that training the graph neural network based on the generated simplified graph structure does not significantly degrade the performance of the trained graph neural network.

[0034] In one embodiment, at least one edge attribute is assigned to each edge of the computational graph, wherein a first generation unit is configured to remove edges from the computational graph based on the edge attributes assigned to the edges. By considering these edge attributes or the relationships between nodes of the computational graph when removing edges, a simplified architecture or graph structure for training a graph neural network can be generated. Simultaneously, it can be ensured that the environment or structure or composition of nodes in a portion of the computational graph corresponds to the environment in the generated graph structure, or a similar environment is also included in the generated graph structure.

[0035] The first generation unit may further include: an allocation unit configured to assign signatures to the edges of the computational graph based on at least one edge attribute assigned to the corresponding edge, such that similar edges are assigned the same signature; a calculation unit configured to apply a hash function to convert the signatures assigned to the edges of the computational graph into numerical values; and a removal unit configured to retain the edges of the computational graph that are connected to the nodes of the computational graph and whose signatures are converted to the minimum value among all the numerical values, and to remove all other edges connected to the corresponding nodes from the computational graph. Therefore, the generated graph structure can be easily obtained through known mathematical operations, which can even be implemented on ordinary systems with limited computer resources.

[0036] The control unit may further include a second generation unit configured to generate at least one edge attribute for each edge of the computational graph, based on at least one node attribute of at least one node connected to the corresponding edge. Thus, many computational graphs may be considered that do not include edge attributes but include node attributes that define data associated with nodes, such as type, label, style attributes (e.g., width, if available), or arbitrary data values ​​that can be used for nodes, wherein edge attributes can be derived from node attributes.

[0037] According to another embodiment of the present invention, a control unit for training a graph neural network is provided, wherein the control unit includes a first receiver for receiving a graph structure for training a graph neural network generated by the control unit for generating a graph structure for training a graph neural network as described above, a second receiver for receiving training data for training a graph neural network, and a training unit configured to train a graph neural network based on the graph structure and the training data.

[0038] Therefore, a control unit for training graph neural networks based on an improved graph structure is provided. Specifically, training the graph neural network based on a simplified architecture has the advantage of significantly reducing the amount of computing resources required, such as storage space and / or processing time, for training the graph neural network based on the generated graph structure. This ensures that the graph neural network can also be trained by ordinary systems with limited resources. Furthermore, training the graph neural network on this simplified graph structure greatly simplifies the subsequent implementation of the graph neural network. Additionally, the environment included in the computational graph also includes the advantage of the generated graph structure, namely, that training the graph neural network based on the generated simplified graph structure does not significantly degrade the performance of the trained graph neural network.

[0039] Furthermore, training data can include sensor data. Therefore, the training of a graphical neural network by a control unit can be based on an environment outside the data processing device on which the graphical neural network is trained.

[0040] According to another embodiment of the present invention, an image classifier for classifying image data is provided, wherein the image classifier includes a receiver for receiving a graphical neural network trained by a control unit for training a graphical neural network as described above, and a classification unit configured to classify the image data through the graphical neural network.

[0041] The image classifier can be configured to classify digital image data based on low-level features of the image, such as edges or pixel attributes.

[0042] Therefore, an image classifier for classifying image data is provided, which is based on a graph neural network trained on an improved graph structure. Specifically, training the graph neural network on a simplified architecture has the advantage of significantly reducing the amount of computing resources required to train the graph neural network based on the generated graph structure, such as storage space and / or processing time. This ensures that the graph neural network can also be trained by ordinary systems with limited resources. Furthermore, training the graph neural network on this simplified graph structure greatly simplifies the subsequent implementation of the graph neural network or the subsequent classification of image data. In addition, the environment included in the computational graph also includes the added advantage of the generated graph structure, namely, that training the graph neural network based on the generated simplified graph structure does not significantly degrade the performance of the trained graph neural network.

[0043] The invention will now be described in more detail with reference to the accompanying drawings. Attached Figure Description

[0044] Figure 1 A flowchart illustrating a method for training a graphical neural network according to an embodiment of the present invention is shown;

[0045] Figure 2A -B illustrates a portion of a method for training a graphical neural network according to an embodiment of the present invention; and

[0046] Figure 3 A system for training a graphical neural network according to an embodiment of the present invention is shown. Detailed Implementation

[0047] Figure 1 A flowchart of a method 1 for training a graph neural network according to an embodiment of the present invention is shown.

[0048] Graph Neural Networks (GNNs) are a class of deep learning methods designed to perform inference on data described by graphs, where each graph consists of multiple nodes connected by edges. Such GNNs have become very popular for tasks involving graph data. GNNs often employ message-passing architectures, such as graph convolutional networks. In such architectures, data or information propagates from nodes in the graph through the neighborhood of the corresponding node to generate representations that depend on the broad graph context. Furthermore, each node can represent a corresponding operation performed by the neural network, as part of determining the neural network's output from its input. This makes GNNs suitable for learning many real-world tasks.

[0049] However, real-world problems often involve very large graphs, requiring substantial computing resources to train such graph neural networks. There are scenarios where training these networks may require far more computing resources than a typical system can provide. If the graph neural network is not properly regularized, the size of the graph can often lead to overfitting. Furthermore, graph convolutional networks or graph attention networks are computationally intensive tasks due to their neighborhood aggregation mechanisms, which utilize self-attention computation to assign weights to neighboring nodes.

[0050] To address these issues, it is known to base the training of such graph neural networks on simplified architectures or graph structures. For example, since very large graphs often involve many redundant components, it is possible to remove these redundant components to simplify the training of such graph neural networks.

[0051] Figure 1 A method 1 for generating a graph structure for training a graph neural network is shown, wherein, in a first step 2, data representing a computation graph is obtained, wherein the computation graph includes multiple nodes connected by edges, and wherein, in a further step 3, a graph structure for training a graph neural network is generated by removing edges from the computation graph, wherein the edges are removed in such a way that the environment in the computation graph corresponds to the environment in the graph structure.

[0052] In particular, Figure 1 A method for generating easily processed graphical structures is shown, which can significantly accelerate the training and use or implementation of graphical neural networks based on the generated graphical structures.

[0053] Generating simplified architectures for training graph neural networks has the advantage of significantly reducing the amount of computing resources required to train these networks, such as storage space and / or processing time. This ensures that graph neural networks can also be trained by ordinary systems with limited resources. Furthermore, training graph neural networks on such simplified architectures greatly simplifies subsequent implementations and uses of the networks.

[0054] Furthermore, the environment included in computational graphics also includes the additional advantage of the generated graphics structure: training a graph neural network based on the generated simplified graphics structure does not significantly degrade the performance of the trained graph neural network.

[0055] therefore, Figure 1 An improved method for generating graphical structures for training graphical neural networks is shown, which can improve the training and implementation of graphical neural networks, especially those utilizing message passing, and at the same time, can save computer resources.

[0056] according to Figure 1 In one embodiment, at least one edge attribute is assigned to each edge of the computational graph, wherein step 3 of training the graph neural network to generate the graph structure by removing edges from the computational graph includes: removing edges from the computational graph based on the edge attributes assigned to the edges of the computational graph.

[0057] In this context, edge attributes refer to data associated with edges. Specifically, edge attributes define the relationships between nodes in a computational graph. For example, edge attributes can define how many edges are connected to specific nodes in the computational graph, or how many edges are connected between two specific nodes in the graph. Edge attributes can be in vector or list form, for example.

[0058] according to Figure 1 In one embodiment, step 3, which generates a graph structure for training a graph neural network by removing edges from the computational graph, further includes: step 4, wherein for all edges of the computational graph, a signature is assigned to each edge of the computational graph based on at least one edge attribute assigned to the corresponding edge, such that similar edges are assigned the same signature or the same signature is assigned to similar edges; step 5, wherein a hash function is applied to convert the signatures assigned to the edges of the computational graph into numerical values; and step 6, wherein for all nodes of the computational graph, the edges of the computational graph that are connected to the nodes of the computational graph and whose signatures are converted into the minimum value of all numerical values ​​are retained, and all other edges connected to the corresponding nodes are removed from the computational graph.

[0059] In step 6, for example, the Minimal Independent Permutation Local Sensitive Hash (MinHash-LSH) can be used, which is a technique for efficiently finding similar items in large databases.

[0060] Furthermore, the hash function can be a pseudo-random hash function. This pseudo-random hash function can then be used across all graph regions or environments of the computational graph, wherein the method produces a graph structure whose regions resemble the corresponding regions in the computational graph, and wherein edges whose signatures have not yet been converted to the minimum of all values ​​have been removed.

[0061] like Figure 1 As shown, Method 1 further includes step 7, wherein for all edges of the computation graph, at least one edge attribute is generated for each edge of the computation graph based on at least one node attribute of at least one node connected to the corresponding edge.

[0062] For example, the edge properties of a graph can be calculated as a vector consisting of a vector representing the node properties of the first node connected to the edge and a vector representing the node properties of the second node connected to the edge.

[0063] Method 1 also includes step 8: providing training data for training the graph neural network, and step 9: training the graph neural network based on the generated graph structure and the training data.

[0064] according to Figure 1 In one embodiment, the training data includes sensor data, which may be acquired, for example, by a video camera, RADAR, LiDAR, or ultrasonic sensor.

[0065] The trained graph neural network can then be implemented or used for image classification. For example, graph neural networks can be used to detect traffic signs, road surfaces, pedestrians, or vehicles in digital images.

[0066] However, depending on the type of training data, a trained graphical neural network can also be used to determine one or more continuous values, such as distance, speed, or acceleration, or to control the functions of an electronic control unit (such as the electronic control unit of a car).

[0067] Figure 2A and 2B A portion of a method for training a graph neural network according to an embodiment of the present invention is shown.

[0068] In particular, Figure 2A and Figure 2B The steps for generating a graphical structure for training a graphical neural network by removing edges from a computational graph are shown.

[0069] also, Figure 2A An area of ​​the computational graph is shown, which defines the composition of nodes in the environment 10 or the computational graph, wherein these nodes are connected by edges. Figure 2B The same region is shown in the generated graph structure, and thus, the edges are removed after the step of generating the graph structure for training the graph neural network by removing edges from the computational graph.

[0070] like Figure 2B As can be seen, the generated graphical structure also includes... Figure 2A The same node composition or environment 10 shown. Specifically, the step of generating the graph structure for training the graph neural network involves removing edges from the computation graph in such a way that environment 10 in the computation graph corresponds to environment 10 in the graph structure. This can be achieved, for example, by assigning signatures to the edges of the computation graph respectively and applying hash functions to these signatures to further remove edges.

[0071] Figure 3 A system 20 for training a graph neural network according to an embodiment of the present invention is shown.

[0072] like Figure 3 As shown, system 20 includes: a control unit 21 for generating a graphical structure for training a graphical neural network, and a control unit 22 for training a graphical neural network, wherein the control unit 22 for training the graphical neural network is configured to train the graphical neural network based on the graphical structure generated by the control unit 21 for generating the graphical structure for training the graphical neural network.

[0073] according to Figure 3 In one embodiment, a control unit 21 for generating a graph structure for training a graph neural network includes: an obtaining unit 23 configured to obtain data representing a computational graph, wherein the computational graph includes a plurality of nodes connected by edges; and a first generating unit 24 configured to generate a graph structure for training the graph neural network by removing edges from the computational graph, wherein the first generating unit 24 is configured to remove edges from the computational graph in such a way that the environment in the computational graph corresponds to the environment in the graph structure.

[0074] The obtaining unit may, for example, include a receiver or input device configured to receive data representing a computational graph, wherein the data representing the computational graph can be input via the input device. The first generating unit may also be implemented by code stored in memory and processable by a processor.

[0075] according to Figure 3In one embodiment, at least one edge attribute is assigned to each edge of the computational graph, and the first generation unit 24 is configured to remove edges from the computational graph based on the edge attributes assigned to the edges of the computational graph.

[0076] like Figure 3 As shown, the first generation unit further includes: an allocation unit 25, configured to assign signatures to the edges of the computational graph based on at least one edge attribute assigned to the corresponding edge for all edges of the computational graph, such that similar edges are assigned the same signature; a calculation unit 26, configured to apply a hash function to convert the signatures assigned to the edges of the computational graph into numerical values; and a removal unit 27, configured to retain the edges of the computational graph that are connected to the nodes of the computational graph and whose signatures are converted into the minimum value of all numerical values ​​for all nodes of the computational graph, and remove all other edges connected to the corresponding nodes from the computational graph.

[0077] Furthermore, the allocation unit, calculation unit, and removal unit can be implemented, for example, by code stored in memory and processed by a processor.

[0078] The control unit 21 shown also includes a second generation unit 28, which is configured to generate at least one edge attribute for all edges of the computed graph, based on at least one node attribute of at least one node connected to the corresponding edge.

[0079] The second generation unit can again be implemented by code stored in memory and processed by a processor.

[0080] according to Figure 3 In one embodiment, the control unit 22 for training a graph neural network includes: a first receiver 29 for receiving a graph structure for training a graph neural network generated by a control unit 21 for generating a graph structure for training a graph neural network; a second receiver 30 for receiving training data for training a graph neural network; and a training unit 31 configured to train a graph neural network based on the graph structure and the training data.

[0081] Training units can, for example, be implemented by code stored in memory and processed by a processor.

[0082] In addition, according to Figure 3 In some embodiments, the training data includes sensor data, particularly sensor data acquired by one or more optical sensors.

Claims

1. A method for generating a graphical structure for training a graphical neural network, wherein the trained graphical neural network is used for image classification, wherein, The method includes the following steps: - Obtain data representing a computational graph, wherein the computational graph includes a plurality of nodes connected by edges (2); and - The graph structure (3) for training the graph neural network is generated by removing edges from the computational graph, wherein the edges are removed in such a way that the environment in the computational graph corresponds to the environment in the graph structure. Wherein, at least one edge attribute is assigned to each edge of the computational graph, and wherein the step of generating the graph structure (3) for training the graph neural network by removing edges from the computational graph includes: - Remove edges from the computational graph based on the edge attributes assigned to the edges of the computational graph; - For all edges of the computation graph, based on at least one edge attribute assigned to the corresponding edge, assign a signature to the edges of the computation graph such that similar edges are assigned the same signature (4). - Apply hash functions to convert the signatures assigned to the edges of the computation graph into numerical values ​​(5); and - For all nodes of the computation graph, retain the edges of the computation graph that are connected to the nodes of the computation graph and whose signatures are converted to the minimum value of all values ​​by the hash function, and remove all other edges connected to the corresponding nodes from the computation graph (6).

2. The method according to claim 1, wherein, The method further includes the following steps: - For all edges of the computation graph, generate at least one edge attribute assigned to the edges of the computation graph based on at least one node attribute of at least one node connected to the corresponding edge (7).

3. A method for training a graphical neural network, wherein, The method (1) includes the following steps: - Generate a graphical structure for training the graphical neural network using the method according to any one of claims 1 to 2; - Provide training data (8) for training the graph neural network; and - The graph neural network is trained based on the generated graph structure and the training data (9).

4. The method according to claim 3, wherein, The training data includes sensor data.

5. A method for classifying image data using a graph neural network, wherein, The graph neural network has been trained by the method (1) for training a graph neural network according to claim 3 or 4.

6. A control device for generating a graphical structure for training a graphical neural network, wherein the trained graphical neural network is used for image classification, the control device (21) comprising: - The obtaining unit (23) is configured to obtain data representing a computational graph, wherein the computational graph includes a plurality of nodes connected by edges; and - A first generation unit (24) is configured to generate the graph structure for training the graph neural network by removing edges from the computational graph, wherein the first generation unit (24) is configured to remove the edges from the computational graph in such a way that the environment in the computational graph corresponds to the environment in the graph structure. In this process, at least one edge attribute is assigned to each edge of the computational graph, and the first generation unit (24) is configured to remove edges from the computational graph based on the edge attributes assigned to the edges of the computational graph. The first generation unit (24) further includes: The allocation unit (25) is configured to assign a signature to the edges of the computation graph based on at least one edge attribute assigned to the corresponding edge for all edges of the computation graph, such that similar edges are assigned the same signature; The computation unit (26) is configured to apply a hash function to convert the signatures assigned to the edges of the computation graph into numerical values, respectively; and The removal unit (27) is configured to, for all nodes of the computation graph, retain the edges of the computation graph that are connected to the nodes of the computation graph and whose signatures are converted to the minimum value of all values ​​by the hash function, and remove all other edges connected to the corresponding nodes from the computation graph.

7. The control device according to claim 6, wherein, The control device (21) further includes a second generation unit (28) configured to generate at least one edge attribute for all edges of the computational graph based on at least one node attribute of at least one node connected to the corresponding edge.

8. A control device for training a graphical neural network, wherein, The control device includes: The first receiver (29) is used to receive the graphical structure for training the graphical neural network generated by the control device (21) for generating the graphical structure for training the graphical neural network according to any one of claims 6 to 7; A second receiver (30) is configured to receive training data for training the graphical neural network; and The training unit (31) is configured to train the graph neural network based on the graph structure and the training data.

9. The control device according to claim 8, wherein, The training data includes sensor data.

10. An image classification apparatus for classifying image data, wherein, The image classification device includes: A receiver for receiving a graphical neural network trained by a control device for training a graphical neural network according to claim 8 or 9; and The classification unit is configured to classify image data using the graphical neural network.

Images