Device and method for supporting neural network segmentation

Abstract

The present disclosure relates to a UP entity and a control entity for supporting NN segmentation. The disclosure proposes a UP entity that is configured to: receive an activation message from the control entity, wherein the activation message is associated with a flow and comprises a configuration file and flow-related information, wherein the configuration file indicates a particular segment of a neural network that is allocated to the user plane entity for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers; derive the configuration file from the activation message; and compose the segment of the neural network for execution based on the configuration file. This disclosure further proposes a control entity configured to send the activation message to the UP entity.

Description

[0001] DEVICE AND METHOD FOR SUPPORTING NEURAL NETWORK SEGMENTATION

[0002] TECHNICAL FIELD

[0003] The present disclosure relates generally to the field of telecommunications and computing, specifically to advancements in network architecture and artificial intelligence (Al) applications. It encompasses the integration of In-Network Computing (INC) and Neural Networks within mobile networks. This disclosure introduces innovative methodologies for executing split- AI processes, leveraging the computational capabilities of network devices to enhance the performance and efficiency of Al- driven applications.

[0004] BACKGROUND

[0005] Traditionally, network devices such as routers and switches are designed primarily for forwarding data traffic. However, with the advent of INC, these devices are now capable of performing computations on network packets in addition to their conventional role. This capability is particularly relevant in mobile networks, where Access Nodes (ANs) and User Plane Functions (UPFs) can process packets of a network flow. INC emerges as a critical enabler for native computing within 6G or future networks, integrating computation as a core feature of network infrastructure.

[0006] Neural Networks represent a cornerstone of modem Machine Learning and Al, inspired by the neuronal structure of the human brain. They are extensively used in various domains, including object detection, image recognition, machine translation, and speech recognition. A typical Neural Network (NN) comprises an input layer, an output layer, and one or more hidden layers, each containing neurons that apply specific activation functions to transform inputs into outputs. These functions include tanh (hyperbolic tangent), ReLU (Rectified Linear Unit), sigmoid, softmax, and binary step, with each layer's output serving as the input for subsequent layers, thereby facilitating complex data processing and analysis.

[0007] Split- Al refers to the distributed execution of NN layers across different computational nodes, which traditionally includes user devices and application servers. This concept extends further with INC support, allowing for the NN's execution to be divided among user equipment (UE), servers, and networking equipment such as routers and switches. In mobile networks, this facilitates a distributed Al processing framework, where the computation is shared between UE, servers, ANs, and UPFs, thus enhancing the efficiency and responsiveness of Al applications.

[0008] Existing solutions demonstrate the feasibility of executing NN layers within the User Plane (UP) but fall short in terms of dynamicity and flexibility. They often rely on static configurations and lack efficient mechanisms for signaling and configuring the execution of NN layers across network devices. Furthermore, prior approaches do not fully leverage the potential of INC to enhance split-AI operations, limiting the performance and adaptability of Al applications in mobile networks. Therefore, an advanced solution is desired.

[0009] SUMMARY

[0010] In view of the above-discussed limitations, this disclosure aims to propose a system that not only supports split-AI operations through UP entities but also introduces efficient signaling mechanisms to dynamically configure NN layer execution. One objective is to ensure that the network can adapt to various Al applications and user needs. Another objective is to allow optimizing performance while minimizing overhead and maximizing resource utilization. These and other objectives are achieved by the solution of the present disclosure as provided in the enclosed independent claims. Advantageous implementations are further defined in the dependent claims.

[0011] A first aspect of the disclosure provides a UP entity, configured to receive an activation message from a control entity, wherein the activation message is associated with a flow and comprises a configuration file and flow-related information, wherein the configuration file indicates a particular segment of a neural network that is allocated to the UP entity for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers; derive the configuration file from the activation message; and compose the particular segment of the neural network for execution based on the configuration file.

[0012] This disclosure proposes a UP entity that is enhanced with new functionalities, which can support individually composing sub- NNs for execution (on a per-flow granularity), and the execution of sub-NNs as delegated by the CP (activation by the CP). This disclosure enables the dynamical execution of NN layers tailored to specific Al applications and network conditions. The disclosure further introduces an efficient signaling mechanism that allows the Control Plane (CP) to communicate with UP entities, enabling quick setup and adjustment of neural network configurations for each individual data flow.

[0013] In an implementation form of the first aspect, the configuration file indicates a structure of the particular segment of the neural network utilizing a neural network template.

[0014] Possibly, the UP entity 400 may utilize a unified representation scheme for NNs, in alignment with all relevant entities on the different planes in the mobile communication system (e.g., UP, CP, and Management Plane (MP)). This enables a common understanding and representation of all NN components within the mobile communication system.

[0015] In an implementation form of the first aspect, the neural network template comprises information about the neural network, information about each layer of the neural network, information about each neuron of each layer, and information about each connection between neurons.

[0016] In an implementation form of the first aspect, the configuration file comprises one or more of the following information: positions of all active neurons of the one or more consecutive layers, an activation function of each active neuron, or a reference to the activation function, all incoming connections to each active neuron, each incoming connection being associated with a weight, a bias of each active neuron, and compute-related information for the execution.

[0017] It may be worth mentioning that for keeping the signaling between CP and UP efficient, the configuration file contained in the activation message may utilize pointers or references (instead of the description of the whole computation to carry out), and the pointers may be utilizing the common representation scheme for NNs. This requires the UP to be capable of understanding these pointers, i.e. to be aware of the computation to execute for a given pointer.

[0018] In an implementation form of the first aspect, the UP entity is configured to: receive a configuration message from a management entity, wherein the configuration message configures the UP entity for composition and execution of the particular segment of the neural network, thereby enabling the UP entity to compose and execute the particular segment of the neural network; and send a first acknowledgement message to the management entity, wherein the first acknowledgement message indicates whether the UP entity is successfully configured. Optionally, the UP entity is programmable by an MP entity for split-AI capability. This includes equipping the UP with the capability of dynamically composing a particular segment of the neural network and with the generic capability of executing (arbitrary) composed networks.

[0019] In an implementation form of the first aspect, the configuration message comprises a unified neural network template, and / or information of activation function of each neuron.

[0020] In an implementation form of the first aspect, the unified neural network template employs a grid-based structure, thereby enabling the addressing of neurons by their horizontal and vertical positions.

[0021] For example, the grid-based structure may support a maximum number of NN layers, which can be denoted as layer-ID along the x-dimension / horizontally, and a maximum number of neurons per NN layer, which can be denoted as neuron-position along the y-dimension / vertically.

[0022] In an implementation form of the first aspect, the UP entity is configured to compose the particular segment of the neural network for execution based on the unified neural network template and the configuration file.

[0023] In an implementation form of the first aspect, the configuration file employs a hierarchical representation with a tree-based structure.

[0024] In an implementation form of the first aspect, the UP entity is configured to send a second acknowledgement message to the control entity, wherein the second acknowledgement message indicates whether the segment of the neural network is successfully composed.

[0025] A second aspect of the disclosure provides a control entity, configured to send an activation message to a UP entity, wherein the activation message is associated with a flow and comprises a configuration file and flow-related information, wherein the configuration file indicates a particular segment of a neural network that is allocated to the UP entity for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers.

[0026] This disclosure further proposes a control entity for supporting the NN segmentation. It is determined in the CP how the NN is split, and how the layers are allocated among the compute nodes (including UE, AN, UPF(s), server). Especially, the execution of a particular NN segment at the UP entity is triggered by the activation message provided by the control entity.

[0027] In an implementation form of the second aspect, the configuration file indicates a structure of the particular segment of the neural network utilizing a neural network template.

[0028] In an implementation form of the second aspect, the neural network template comprises information about the neural network, information about each layer of the neural network, information about each neuron of each layer, and information about each connection between neurons.

[0029] In an implementation form of the second aspect, the configuration file comprises one or more of the following information: positions of all active neurons of the one or more consecutive layers, an activation function of each active neuron, or a reference to the activation function, all incoming connections to each active neuron, each incoming connection is associated with a weight, a bias of each active neuron, and compute-related information for the execution.

[0030] In an implementation form of the second aspect, the control entity is configured to receive a second acknowledgement message from the UP entity, wherein the second acknowldegement message indicates whether the particular segment of the neural network is successfully composed in the UP entity.

[0031] In an implementation form of the second aspect, the control entity is configured to determine a segmenting option for the neural network, wherein the segmenting option is associated with the flow; and determine to allocate a first segment of the neural network to a first UP entity and allocate a second segment of the neural network to a second UP entity.

[0032] The operations of determining how to split and how to allocate the layers to compute nodes are not discussed in this application. The important point to note is that the control entity holds information on the NN layers to be executed by which compute nodes and that the control entity provides an allocation of NN layers to UP nodes.

[0033] In an implementation form of the second aspect, the control entity is configured to provide a first configuration file to the first UP entity, wherein the first configuration file indicates the first segment of the neural network that is allocated to the first UP entity for execution, and provide a second configuration file to the second UP entity, wherein the second configuration file indicates the second segment of the neural network that is allocated to the second UP entity for execution, wherein the first configuration file and the second configuration file are different.

[0034] It should be understood that an activation message is sent from the CP to all UP entities along the UP path for the split-AI Communication and Compute (CC) Flow. It contains a different configuration file for each UP entity, the configuration file containing the structure of the sub-NN to be executed by the respective UP entity.

[0035] In an implementation form of the second aspect, the control entity is configured to notify a network entity that is on a path of the flow, that a segment of the neural network is allocated to the network entity for execution.

[0036] In an implementation form of the second aspect, the network entity comprises a user device, an access network device, or an application function (AF).

[0037] In an implementation form of the second aspect, the control entity is configured to receive a third acknowldegement message from the network entity, wherein the third acknowledgement message indicates whether the segment of the neural network is successfully activated in the network entity.

[0038] A third aspect of the disclosure provides a method performed by a UP entity, comprising: receiving an activation message from a control entity, wherein the activation message is associated with a flow and comprises a configuration file and flow-related information, wherein the configuration file indicates a particular segment of a neural network that is allocated to the UP entity for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers; deriving the configuration file from the activation message; and composing the segment of the neural network for execution based on the configuration file.

[0039] Implementation forms of the method of the third aspect may correspond to the implementation forms of the UP entity of the first aspect described above. The method of the third aspect and its implementation forms achieve the same advantages and effects as described above for the UP entity of the first aspect and its implementation forms. A fourth aspect of the disclosure provides a method performed by a control entity, comprising: sending an activation message to a UP entity, wherein the activation message is associated with a flow and comprises a configuration file and flow-related information, wherein the configuration file indicates a particular segment of a neural network that is allocated to the UP entity for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers.

[0040] Implementation forms of the method of the fourth aspect may correspond to the implementation forms of the control entity of the second aspect described above. The method of the fourth aspect and its implementation forms achieve the same advantages and effects as described above for the control entity of the second aspect and its implementation forms.

[0041] A fifth aspect of the disclosure provides a computer program product comprising a program code for carrying out, when implemented on a processor, the method according to the third aspect and any implementation forms of the third aspect, or the fourth aspect and any implementation forms of the fourth aspect.

[0042] A sixth aspect of the disclosure provides a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to cany out, the method according to the third aspect and any implementation forms of the third aspect, or the fourth aspect and any implementation forms of the fourth aspect.

[0043] It has to be noted that all devices, elements, units, and means described in the present application could be implemented in software or hardware elements or any kind of combination thereof. All steps which are performed by the various entities described in the present application as well as the functionalities described to be performed by the various entities are intended to mean that the respective entity is adapted to or configured to perform the respective steps and functionalities. Even if, in the following description of specific embodiments, a specific functionality or step to be performed by external entities is not reflected in the description of a specific detailed element of that entity that performs that specific step or functionality, it should be clear for a skilled person that these methods and functionalities can be implemented in respective software or hardware elements or any kind of combination thereof.

[0044] BRIEF DESCRIPTION OF DRAWINGS

[0045] The above-described aspects and implementation forms of the present disclosure will be explained in the following description of specific embodiments in relation to the enclosed drawings, in which:

[0046] FIG. 1 shows a neuron;

[0047] FIG. 2 shows a NN with different layer characteristics;

[0048] FIG. 3 shows the proposed solution according to an embodiment of this disclosure;

[0049] FIG. 4 shows a UP entity according to an embodiment of this disclosure;

[0050] FIG. 5 shows a mobile communication system with CP and UP according to an embodiment of this disclosure;

[0051] FIG. 6 shows a sequence diagram for dynamic NN layer activation according to an embodiment of this disclosure;

[0052] FIG. 7 shows a hierarchical representation scheme of the NN according to an embodiment of this disclosure; FIG. 8 shows an NN template according to an embodiment of this disclosure;

[0053] FIG. 9 shows a UP Entity providing two services according to an embodiment of this disclosure;

[0054] FIG. 10 shows a control entity according to an embodiment of this disclosure;

[0055] FIG. 11 shows a high-level solution according to an embodiment of this disclosure;

[0056] FIG. 12 shows a method according to an embodiment of this disclosure; and

[0057] FIG. 13 shows a method according to an embodiment of this disclosure.

[0058] DETAILED DESCRIPTION OF EMBODIMENTS

[0059] Illustrative embodiments of a UP entity, a control entity, and corresponding methods for supporting NN segmentation and execution are described in the following with reference to the figures. Although this description provides a detailed example of possible implementations, it should be noted that the details are intended to be exemplary and in no way limit the scope of the application.

[0060] Moreover, an embodiment or example may refer to other embodiments or examples. For example, any description including but not limited to terminology, element, process, explanation, and / or technical advantage mentioned in one embodiment or example may also apply to the other embodiments or examples.

[0061] For ease of understanding, the concepts of NN and split- Al are first introduced. FIG. 1 denotes inputs, weights, internal logic, and output of the neuron of a NN. It shall be used in the following to explain the key terms (weights, bias, activation function). The input values (denoted as x0, x±, x2, ..., xd) are coming from the neurons of the previous layer, i.e. x0, x±, x2, ..., xdare the outputs of the neurons from the preceding layer. To each input, a different weight is given, denoted as iv(l. wt. w2, ..., wd. All weighted inputs (x, * iv are linearly combined and an additional constant is added, referred to as bias (bt). The resulting linear combination is then fed into the activation function (f), the core of the neuron. The output of the neuron is either used as an input for neurons of the next layer or (in case the denoted neuron is the output layer), it is the final output, i.e. the predicted value (inference goal of the NN).

[0062] FIG. 2 illustrates a fully connected NN with different layer characteristics. In particular, this figure denotes an exemplary NN for image recognition. The photo of a tree is the input of the NN. Several intermediate / hidden NN layers follow before it is finally returned what type of object could be recognized from the photo. Notably, the intermediary NN layers differ with respect to the following key characteristics: complexity, and data size.

[0063] Different layers differ with respect to complexity. The complexity is thereby determined by the number of neurons contained in a layer and the complexity of the neurons’ activation functions. Different layers are also characterized by the intermediary data size they produce. Some layers reduce the data size (i.e. they receive more input than they produce output), while other layers increase the data size (i.e. they receive less input than they produce output).

[0064] The integration of Al through NNs within UP entities presents a complex challenge. This challenge is particularly pronounced in environments employing split-AI techniques, where NN layers are distributed across various computational nodes. The dynamic nature of modem network traffic, coupled with diverse Al applications such as image recognition, noise reduction, and object detection, necessitates a highly adaptable NN configuration capable of handling the specific computational demands of different data flows.

[0065] The execution of NN layers by UP entities is influenced by several factors that contribute to the complexity of this task. Firstly, the nature of the neural network itself can vary significantly depending on the Al application for which it is intended. This variability introduces a broad spectrum of computational demands on UP entities, as each application may require different NN architectures and processing capabilities. Furthermore, how a neural network is split - determining which layers are executed by which computational nodes - is typically decided by the CP. This decision process adds another layer of complexity, as it must consider the optimal distribution of computational load and the specific capabilities of each node.

[0066] Additionally, the selection of UP paths and the allocation of specific neural network layers to different UP nodes are critical decisions that impact the efficiency and effectiveness of the split-AI system. These decisions must account for the dynamic requirements of network traffic and the computational capabilities of UP entities, leading to a multitude of potential configurations. This variability underscores the need for a flexible and dynamic approach to configuring UP entities, which traditional static or manual configurations fail to adequately address.

[0067] Given these challenges, there is a pressing need for a novel solution that enhances the flexibility and dynamic configurability of UP entities. Such a solution must efficiently manage the signaling between the CP and UP entities to enable the real-time configuration and execution of neural network layers. This requirement is crucial not only to accommodate the diverse and fluctuating demands of different Al applications but also to minimize the signaling overhead, which can become substantial given the large sizes of contemporary neural network models.

[0068] In response to these identified problems, the present disclosure proposes an innovative approach that enables UP entities to dynamically execute NN layers tailored to specific Al applications and network conditions. The disclosure further introduces an efficient signaling mechanism that allows the CP to communicate with UP entities, enabling quick setup and adjustment of neural network configurations for each individual data flow. This approach not only addresses the need for dynamicity and flexibility in split-AI operations but also significantly reduces the complexity and overhead associated with managing large- scale neural network computations across a distributed network environment.

[0069] FIG. 3 illustrates the proposed solution according to an embodiment of this disclosure. FIG. 3 shows a mobile communication system (e.g. a 6G network or a future network) where the CP has determined in an earlier step, which compute node shall execute which part of the NN. It may be understood that the CP has determined the split of a multiple-layer NN and the allocation of NN layers to compute nodes.

[0070] Then, the UP entities of the mobile communication system are being enhanced by new capabilities: the support of individually composing sub-NNs for execution (on a per-flow granularity), and the execution of sub-NNs as delegated by the CP (activation by the CP).

[0071] The UP entities are configured to receive new signaling information from the CP (NN layer activation) upon split-AI CC flow establishment, containing all relevant information needed for NN layer configuration and execution.

[0072] FIG. 4 shows a UP entity 400 according to an embodiment of the disclosure. The UP entity 400 may be adapted for supporting the execution of segmented neural networks. The UP entity 400 may comprise processing circuitry (not shown) configured to perform, conduct, or initiate the various operations of the UP entity 400 described herein. The processing circuitry may comprise hardware and software. The hardware may comprise analog circuitry digital circuitry, or both analog and digital circuitry. The digital circuitry may comprise components such as application-specific integrated circuits (ASICs), field-programmable arrays (FPGAs), digital signal processors (DSPs), or multi-purpose processors. The UP entity 400 may further comprise memory circuitry, which stores one or more instructions) that can be executed by the processor or by the processing circuitry, in particular under the control of the software. For instance, the memory circuitry may comprise a non-transitory storage medium storing executable software code which, when executed by the processor or the processing circuitry, causes the various operations of the UP entity 400 to be performed. In one embodiment, the processing circuitry comprises one or more processors and a non-transitory memory connected to the one or more processors. The non-transitory memory may carry executable program code which, when executed by the one or more processors, causes the UP entity 400 to perform, conduct, or initiate the operations or methods described herein.

[0073] The UP entity 400 is configured to receive an activation message 401 from a control entity 410, wherein the activation message 401 is associated with a flow and comprises configuration file 402 and flow-related information. The configuration file 402 indicates a particular segment 403 of a neural network that is allocated to the UP entity 400 for execution. Notably, the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers. The UP entity 400 is configured to derive the configuration file 402 from the activation message 401 and compose the particular segment 403 of the neural network for execution based on the configuration file 402.

[0074] This disclosure proposes a UP entity 400 that is enhanced with new capabilities, which include the capability of dynamically composing sub-NNs / NN layers upon activation by the CP, and the capability of executing sub-NNs / NN layers using the contents of packets of flows belonging to Al-based applications.

[0075] It canbe understood that the activation message 401 triggers the dynamic configuration of the UP entities for split- Al execution according to the configuration file 402 and further flow-related information.

[0076] In a particular embodiment, the UP entity 400 may utilize a unified representation scheme for NNs, in alignment with all relevant entities on the different planes in the mobile communication system (UP, CP, and MP). This enables a common understanding and representation of all NN components within the mobile communication system.

[0077] The capability of the UP entity 400 includes to dynamically (i.e. upon receiving an activation message) prepare for NN-specific computations to be carried out on a flow. For keeping the signaling between CP and UP efficient, the configuration file 402 contained in the activation message 401 may utilize pointers (instead of the description of the whole computation to carry out), and the pointers may be utilizing the common representation scheme for NNs. This requires the UP to be capable of understanding these pointers, i.e. to be aware of the computation to execute for a given pointer.

[0078] In one particular embodiment, the UP entity 400 is programmable by a management plane (MP) entity for split-AI capability. This includes equipping the UP with the capability to dynamically prepare for NN-specific computations, including the understanding of pointers to computations and equipping the UP with the computation associated with these pointers.

[0079] FIG. 5 shows a high-level architecture where the proposed solution canbe applied. The Split Control Entity (SCE) canbe any NF in the CP of a mobile network, or it can be composed through a set of NFs and their interactions. The SCE may be the control entity 410 shown in FIG. 4. The UP entities may include the UP entity 400 shown in FIG. 4. The task of the SCE is to determine how the NN - which shall be used for the application flow - must be split, and how the different resulting NN layers are allocated to compute nodes (such as UE, AN, UPF, and server, etc.). Notably, the SCE and its internal logic are not discussed in the embodiments of this disclosure. In this disclosure, the role of the SCE is to provide the allocation of which NN layers should be executed by which compute node. The SCE is not necessarily a single entity but can be composed of a multitude of CP entities.

[0080] The UE intends to establish an application flow to the application server (via the Data Network (DN)), to use INC / split-AI. The flow thereby traversed the UP entities of the mobile network (e.g. AN and UPF). In an initial step, a User Plane Configuration Manager (located in the MP of the mobile communication system), has prepared the UP entities with the new capabilities needed for split-AI operations. During the split-AI CC Flow establishment, the SCE communicates to the involved UP entities the allocated sub-NNs to execute. For instance, the SCE activates the compute nodes in the UP by sending new signaling information, i.e., by providing an activation message 401.

[0081] FIG. 6 denotes a sequence diagram for dynamic NN layer activation in the UP in a 5G environment according to an embodiment of this disclosure. Specifically, the sequence diagram involves steps when dynamically activating the allocated sub-NNs in the UP.

[0082] Step 1 to Step 4: The CP determines based on the operator’s preference, system capabilities, and underlying conditions, the best splitting option and allocation for the NN. These steps are out of the scope of the present disclosure and are only added for the sake of completeness.

[0083] Step 5 a: The SMF, which may be the control entity 401 as shown in FIG. 4 or FIG. 5, sends the layer activation message to the UPF, i.e., it notifies the UPF which layers to execute. The UPF in this embodiment may be the UP entity 400 of FIG. 4 or FIG. 5. b: The UPF prepares for NN layer execution. That is, it processes the activation message 401 received from the SMF and uses the template for dynamically activating the partial NN / NN layer. c: The UPF provides feedback to the SMF, indicating the success of the dynamic layer activation.

[0084] Step 6 a: The SMF sends the layer activation message to the AN, i.e., it notifies the AN which layers to execute. b: The AN prepares for NN layer execution. That is, it processes the activation message received from the SMF and uses the template for dynamically activating the partial NN / NN layer. c: The AN provides feedback to the SMF, indicating the success of the dynamic layer activation.

[0085] Step 7 a: The SMF sends the layer activation message to the UE, i.e., it notifies the UE which layers to execute. b: The UE prepares for NN layer execution. That is, it processes the activation message received from the SMF and uses the template for dynamically activating the partial NN / NN layer. c: The UE provides feedback to the SMF, indicating the success of the dynamic layer activation.

[0086] Step 8

[0087] Via the AF, the application is informed about the parts being executed in the network, and the parts being executed by the application / server. Optionally, the AF may also provide feedback to the SMF, indicating the success of the dynamic layer activation. The CC Flow with split- Al capability can be set up. That is, the UP entities can compute the NN layers on the packet's contents while forwarding them.

[0088] FIG. 7 shows an illustration of a hierarchical representation scheme of the NN according to an embodiment of this disclosure. This embodiment describes a possible realization of a unified representation scheme for NNs.

[0089] The CP communicates to the UP entities, which neurons and connections of the NN template shall be activated and which activation functions are implemented by the neurons. Forthat communication, the present disclosure proposes a specific unified NN representation scheme, which is among others used in the activation message 401 (in the form of a configuration file 402) and utilized to dynamically compose the allocated sub-NNs in the UP entities. The key goal is an efficient encoding of the information, to reduce the signaling load between UP and CP.

[0090] In a particular embodiment, the configuration file 402 indicates a structure of the particular segment 403 of the neural network, which is to be composed and executed by the UP entity 400, utilizing a neural network template.

[0091] Optionally, the neural network template comprises information about the neural network, information about each layer of the neural network, information about each neuron of each layer, and information about each connection between neurons.

[0092] According to an embodiment of this disclosure, the configuration file 402 comprises one or more of the following information: positions of all active neurons of the one or more consecutive layers, an activation function of each active neuron, or a reference to the activation function, all incoming connections to each active neuron, each incoming connection is associated with a weight, a bias of each active neuron, and compute-related information for the execution.

[0093] This specific embodiment is shown in FIG. 7 leverages a hierarchical representation of a NN, which is a tree-based structure, where the root contains information about the NN itself.

[0094] Information relating to the Neural Network:

[0095] • Optional: Name of the NN / NN architecture (can be both, an own defined name, or the name of well-known NNs, such as AlexNet, VGG-16, ResNet-50, etc.)

[0096] • Optional: Purpose of the inference (e.g. image recognition, noise reduction in audio files, time series predictions, etc.)

[0097] • Optional: Any further meta-data relating to the NN

[0098] • Required: Contained layers in the NN : including Number of layers, and Layer IDs

[0099] The next level in the hierarchy - after the root (NN information), are the individual NN layers, represented as branches in the tree.

[0100] Information relating to each Laver:

[0101] • Required: Layer-ID

[0102] • Optional: Layer type (input, hidden, output)

[0103] • Optional: Complexity of the layer (e.g. in terms of the number of FLOPs, compute requirements, expected compute duration) Optional: Layer output size / intermediary output size: (Expected) average output size, and / or (Expected) variation of the output size

[0104] Required: Neurons on that layer: including Number of neurons and Neuron IDs

[0105] After the layers, the next level in the hierarchy is the neurons.

[0106] Information relating to each individual Neuron:

[0107] • Required: The position of the neuron in the layer

[0108] • Required: The neuron ID (in this embodiment, the neuron-ID is a combination of the layer-ID to which it belongs and the neuron’s position within that layer)

[0109] • Required: The bias of the neuron

[0110] • Required: The activation function (in one example, the activation function is referenced by its name: e.g. ReLu, Tanh, sigmoid, ...)

[0111] • Required: Incoming connections: including Number of incoming connections, and Incoming connection IDs

[0112] The connections form the lowest level in the hierarchy, hence, they are the leaves in the tree-based representation scheme.

[0113] Information relating to each Connection:

[0114] • Required: Connection-ID (in this embodiment, the connection-ID is composed of the two neurons-IDs of the neurons it connects)

[0115] • Required: The weight of the connection.

[0116] This hierarchical representation has the following benefits:

[0117] • It supports the usage of well-accepted file formats, such as JSON files, which also support a hierarchical representation. The NN information tree could simply be transformed into a JSON file, as also shown in FIG. 7.

[0118] • The subtrees with the layers being the root, represent how the NN can be split. For instance, a vertical cut in the NN means a cut in the tree-based structure along the layer branch.

[0119] • Allows a simple and intuitive sub-selection of the information. For instance, only those layer branches are selected and sent to those UP entities, for which the information is relevant. For instance, assuming the AN would execute the first layer (layer 0) only the branch / sub-tree with root being layer 0 would be sent to the AN. Assuming that layer k is executed at UPF 1 , only the branch / sub-tree with layer k being the root is sent to UPF 1.

[0120] According to an embodiment of this disclosure, the UP entity 400 may be further configured to receive a configuration message from a management entity, wherein the configuration message configures the UP entity 400 for composition and execution of the particular segment 403 of the neural network, thereby enabling the UP entity 400 to compose and execute the particular segment 403 of the neural network. The UP entity 400 may be further configured to send a first acknowledgement message to the management entity, wherein the first acknowledgement message indicates whether the UP entity 400 is successfully configured.

[0121] This embodiment describes how the UP entities can dynamically compose the allocated sub-NNs, i.e. prepare for the NN layer execution, once the activation message is received. Possibly, the configuration message comprises a unified neural network template, and / or information of activation function of each neuron. For instance, the unified neural network template may employ a grid-based structure, thereby enabling the addressing of neurons by their horizontal and vertical positions.

[0122] FIG. 8 shows a grid-based NN template as a possible embodiment for dynamic NN layer activation. It supports a maximum number of NN layers (denoted as layer-ID along the x-dimension / horizontally in FIG. 8) and a maximum number of neurons per NN layer (denoted as neuron-pos along the y-dimension / vertically in FIG. 8). Notably if the neurons are “blank”, it means that the neurons themselves and their connections do exist, but it is not specified i) whether they are used and ii) if they are used, which activation functions they compute. The grid-based template is fully connected, which means that all neurons of layer n are connected to all neurons of the preceding layer n- 1. In this way, the template provides the highest degree of freedom possible (as there are no limitations in terms of which neurons of the current layer can be connected to which neurons of the following layer).

[0123] In this embodiment, the following schemes are used for referencing NN components:

[0124] • Neurons are addressed by their position (vertical position + horizontal position in the grid-based template). In the example of FIG. 8, the neuron on layer 2 (horizontal position = 2) with vertical position 1 in that layer, would be addressed as “neuron_ID = L2P1”, uniquely describing the position using locations along the x- and y-dimension. Similarly, a second neuron, which is part of NN layer 3 and located in position 2 would be referred to as “neuron_ID=L3P2”

[0125] • Connections are referred to by utilizing the two neurons they connect. In the example of FIG. 8, the connection of the two highlighted neurons could be addressed as “conn_ID = L2P1 - L3P2”.

[0126] To improve the efficiency of dynamically preparing the UP for the execution of sub-NNs, in a particular embodiment, only pointers for representing the NN (instead of the description of the whole computation to cany out), are provided to the UP entities.

[0127] This part of the embodiment describes specifically how the UP can be capable of composing the sub-NN even if only pointers are received (and not the full computation).

[0128] The “execution of NN layers” includes the computation of activation functions. Hence, the UP entities involved in the split- Al operations must be capable of computing / executing activation functions.

[0129] In order to make the neuron activation as efficient as possible, the UP entities are already equipped (e.g., using MP programming) with a mapping of a unique reference to an activation function (e.g. an ID or a specific name) and the specific computation to cany out. Optionally, further meta-data is associated with the unique reference of an activation function.

[0130] That means, when a UP entity “supports an activation function”, it can compute it and the following information is available at the respective entity:

[0131] • Required: A unique reference name or ID, which is aligned with the unified representation scheme. Examples: ReLu, TanH, or Sigmoid - if the reference shall be the activation function’s name.

[0132] • Required: The formula corresponding to the activation function.

[0133] Examples: Sigmoid, where x represents the input value of the activation function, which is composed (as a linear combination) of the output values of the previous layer. • Optional: Meta-data relating to the computation, that can be relevant:

[0134] Examples: Number of floating point operations for executing the neuron, expected utilization of compute resources (RAM / CPU, ...), or expected duration for computing the activation function.

[0135] Notably, even if only referred to by the activation function’s name in the activation message, the UP entity is fully aware of what needs to be executed.

[0136] According to an embodiment of the disclosure, the UP entity 400 is further configured to compose the particular segment 403 of the neural network for execution based on the unified neural network template and the configuration file 402.

[0137] According to an embodiment of the disclosure, the configuration file 402 employs a hierarchical representation with a treebased structure.

[0138] The proposed hierarchical representation scheme shall be used for efficient signaling of the layers to execute (from CP to UP). For the activation (i.e. telling a UP entity 400 which part of the NN it shall execute), an activation message 401 is sent to the respective UP entity. The sub-tree(s) of the representation scheme shall be part of the activation message 401.

[0139] More specifically, an activation contains the one or more NN layers to execute, including the following information:

[0140] • Position of the active neurons (using their ID)

[0141] • Activation functions of the neurons (which can be just in the form of using the reference, e.g. in an ID or the name)

[0142] • Incoming connections to the neurons, including their weights

[0143] • Bias

[0144] • Further compute related information

[0145] All this information is encoded using the hierarchical representation scheme.

[0146] The activation may further contain a communication- and compute-profile, providing compute-related information, that is not limited to the split-AI operation (i.e. this is information provided for any INC case):

[0147] • Resource requirements

[0148] • Flow-ID

[0149] • Computation and communication delay budget

[0150] FIG. 9 shows a UP entity 400 providing two services (used for split-AI operations) according to an embodiment of this disclosure.

[0151] In this embodiment, the UP entity 400 provides two dedicated services. One for the MP (denoted as Service 1 in FIG. 9), which can be used to prepare the UP entity for split-AI support. That is, Service 1 would be used for programming the UP Entity for split-AI support (this includes providing / updating the NN template and equipping it with the capability of executing (a set of) activation functions).

[0152] The second service provided by the UP entity 400 is dedicated to the activation of NN layers by the CP (denoted as Service 2 in FIG. 9). This service processes the activation message received from the CP and performs the layer activation using the received information (including the configuration file), the NN template, and the local information about the activation functions. Further information on the two proposed services (i.e. exchanged messages and procedure) is given in the following.

[0153] Service 1 : Service for the MP to program the UP to provide split- Al support

[0154] • MP sends a configuration message, including: o An NN template allowing the dynamic activation of sub-NNs o Activation functions information (optional, depending on the specific embodiment).

[0155] • Procedure:

[0156] 1) MP sending the configuration message

[0157] 2) The UP entity responds with an ACK / NACK indicating successful configuration.

[0158] Service 2: Service for the CP to activate the UP for dynamic sub-NN activation

[0159] • CP sends a request message, including: o Activation message containing the configuration file for the sub-NN to execute o Flow-ID (the flow associated with the flow-ID does not necessarily must exist, yet) o Flow characteristics.

[0160] • Activation procedure :

[0161] 1) CP to UP request: The CP sends the request message (which includes the activation message) to all UP entities to be involved in the split-AI CC flow.

[0162] 2) After the reception of the request message (which includes the activation message), the UP entity checks if the allocated sub-NN for the given Flow-ID can be activated.

[0163] 3) UP to CP response: The UP entity responds with ACK / NACK to the CP to indicate the possibility of successful activation.

[0164] 4) In case all UP entities involved in the split-AI CC Flow response with ACK to the CP: flow setup or flow reconfiguration a) If the flow with the Flow-ID is not up yet: The CP triggers UP path establishment for indicated Flow- ID with indicated flow characteristics, which includes the sending of an activation message. The UP entities - upon reception of the activation message - activate the allocated sub-NN. b) If the flow with the Flow-ID is already up: Reconfiguration of the flow, which includes the sending of an activation message. The UP entities - upon reception of the activation message - activate the allocated sub-NN.

[0165] FIG. 10 shows a control entity 410 according to an embodiment of the disclosure. The control entity 410 may be adapted for supporting the segmentation of neural networks.

[0166] The control entity 410 may comprise processing circuitry (not shown) configured to perform, conduct, or initiate the various operations of the control entity 410 described herein. The processing circuitry may comprise hardware and software. The hardware may comprise analog circuitry digital circuitry, or both analog and digital circuitry. The digital circuitry may comprise components such as ASICs, FPGAs, DSPs, or multi-purpose processors. The control entity 410 may further comprise memory circuitry, which stores one or more instruction(s) that can be executed by the processor or by the processing circuitry, in particular under the control of the software. For instance, the memory circuitry may comprise a non-transitory storage medium storing executable software code which, when executed by the processor or the processing circuitry, causes the various operations of the control entity 410 to be performed. In one embodiment, the processing circuitry comprises one or more processors and a non-transitory memory connected to the one or more processors. The non-transitory memory may cany executable program code which, when executed by the one or more processors, causes the control entity 410 to perform, conduct, or initiate the operations or methods described herein. The control entity 410 is configured to send an activation message 401 to a UP entity 400, wherein the activation message 401 is associated with a flow and comprises configuration file 402 and flow-related information. The configuration file 402 indicates a particular segment 403 of a neural network that is allocated to the UP entity 400 for execution. The neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers.

[0167] According to an embodiment of this disclosure, the control entity 410 may be configured to determine a segmenting option for the neural network, wherein the segmenting option is associated with the flow; and determine to allocate a first segment of the neural network to a first UP entity 400 and allocate a second segment of the neural network to a second UP entity 400.

[0168] It should be understood that the activation message 401 is sent from the CP to all UP entities along the UP path for the split- AI CC Flow. It contains a different configuration file for each UP entity, the configuration file containing the structure of the sub-NN to be executed by the respective UP entity. The activation message 401 may contain further information relating to the flow characteristics, which may include: compute requirements for the execution (e.g. CPU, RAM requirements), and / or QoS information.

[0169] Optionally, the control entity 410 is configured to provide a first configuration file 402 to the first UP entity 400, wherein the first configuration file 402 indicates the first segment of the neural network that is allocated to the first UP entity 400 for execution, and provide a second configuration file 402 to the second UP entity 400, wherein the second configuration file 402 indicates the second segment of the neural network that is allocated to the second UP entity 400 for execution. The first configuration file 402 and the second configuration file 402 are different.

[0170] FIG. 11 shows a high-level solution according to an embodiment of this disclosure. In the first step, the CP determines how the NN is split, and how the layers are allocated among the compute nodes (UE, AN, UPF(s), server). The step of determining how to split and how to allocate the layers to compute nodes is not discussed here and is only added for the sake of completeness. The important point to note is that the CP holds the NN and that it provides an allocation of NN layers to UP nodes. The allocation may be represented using pointers / unique layer IDs (e.g., l0, l±, ...) and the indication, of which compute node is in charge of which layer to execute. In any form or representation, the CP holds the information of the NN layers to be executed by which compute node.

[0171] Based on this information, the next step is to enforce the split-AI allocation in the UP entities according to the provided allocation. The solution comprises:

[0172] • Efficient signaling: an efficient way for the CP to communicate dynamically to the UP which layers it needs to execute. More specifically, this disclosure proposes the usage of a configuration file containing a sub-structure of an NN. That is, new signaling / information exchange from CP to UP and a specific realization of that.

[0173] • Capability for flexible NN layer execution: on-demand NN layer composition using a template for NN layer activation. That is, it is proposed to equip the UP with novel capabilities.

[0174] The control entity 410 of this embodiment notifies network entities that are on a path of the flow, that a segment of the neural network is allocated to each network entity for execution.

[0175] Possibly, the network entities may comprise a user device, an access network device, or an AF.

[0176] FIG. 12 shows a method 1200 according to an embodiment of the disclosure. In a particular embodiment, the method 1200 is performed by the UP entity 400 shown in FIG. 4 - FIG. 6, FIG. 9, or FIG. 10. The method 1200 comprises a step 1201 of receiving an activation message 401 from a control entity 410, wherein the activation message 401 is associated with a flow and comprises a configuration file 402 and flow-related information, wherein the configuration file 402 indicates a particular segment 403 of a neural network that is allocated to the UP entity 400 for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers. The method 1200 further comprises a step 1202 of deriving the configuration file 402 from the activation message 401, and a step 1203 of composing the segment of the neural network for execution based on the configuration file 402. Possibly, the control entity 410 may be the control entity shown in FIG. 4 - FIG. 6, FIG. 9, or FIG. 10.

[0177] FIG. 13 shows a method 1300 according to an embodiment of the disclosure. In a particular embodiment, the method 2300 is performed by the control entity 410 shown in FIG. 4 - FIG. 6, FIG. 9, or FIG. 10. The method 1300 comprises a step 1301 of sending an activation message 401 to a UP entity 400, wherein the activation message 401 is associated with a flow and comprises a configuration file 402 and flow-related information, wherein the configuration file 402 indicates a particular segment of a neural network that is allocated to the UP entity 400 for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers. Possibly, the UP entity 400 may be the UP entity shown in FIG. 4 - FIG. 6, FIG. 9 or FIG. 10.

[0178] To summarize, embodiments of this disclosure propose a template for NN layer composition in the UP entities, which allows to the composition of practically any sub-NN / NN Layer to be executed and hence provides a high level of flexibility / dynamicity, and can reduce the storage requirements at the UP entities (instead of storing a large number of sub- NNs to execute, they are composed on demand).

[0179] Embodiments of this disclosure also propose a unified representation scheme of NNs within the mobile network. It addresses schemes that can be used throughout all planes of a mobile communication system, presenting a common understanding and reference format. The hierarchical, tree-based representation scheme proposed in a particular embodiment allows for a simple selection of the required subset of information that is needed. It also supports simple updates when something changes, as only the changed information would need to be sent (e.g. a single branch of the tree). In addition, it can directly be mapped to widely used and accepted configuration file formats (e.g. JSON or YAML), allowing an efficient representation of NNs, and hence efficient signaling between CP and UP.

[0180] Embodiments of this disclosure also propose a configuration file containing the sub-structure of an NN being sent from CP to the respective UP entity in charge of executing the sub-NN. This reduces the signaling load from CP to UP, as only that part of the information is sent to each UP entity, that is relevant for this specific entity.

[0181] It is also notable that the proposed solution can be adapted for use in fixed networks, including local network environments such as WiFi and factory network deployments. This adaptability enhances its utility across a broader range of networking scenarios.

[0182] The present disclosure has been described in conjunction with various embodiments as examples as well as implementations. However, other variations can be understood and effected by those persons skilled in the art and practicing the claimed embodiments of the disclosure, from the studies of the drawings, this disclosure, and the independent claims. In the claims as well as in the description the word “comprising” does not exclude other elements or steps and the indefinite article “a” or “an” does not exclude a plurality. A single element or other unit may fulfill the functions of several entities or items recited in the claims. The mere fact that certain measures are recited in the mutually different dependent claims does not indicate that a combination of these measures cannot be used in an advantageous implementation.

[0183] Furthermore, any method according to embodiments of the disclosure may be implemented in a computer program, having code means, which when run by processing means causes the processing means to execute the steps of the method. The computer program is included in a computer-readable medium of a computer program product. The computer-readable medium may comprise essentially any memory, such as a ROM (Read-Only Memory), a PROM (Programmable Read-Only Memory), an EPROM (Erasable PROM), a Flash memory, an EEPROM (Electrically Erasable PROM), or a hard disk drive.

[0184] Moreover, it is realized by the skilled person that embodiments of the UP entity 400 or the control entity 410 comprise the necessary communication capabilities in the form of e.g., functions, means, units, elements, etc., for performing the solution. Examples of other such means, units, elements, and functions are processors, memory, buffers, control logic, encoders, decoders, rate matchers, de-rate matchers, mapping units, multipliers, decision units, selecting units, switches, interleavers, deinterleavers, modulators, demodulators, inputs, outputs, antennas, amplifiers, receiver units, transmitter units, DSPs, trelliscoded modulation (TCM) encoder, TCM decoder, power supply units, power feeders, communication interfaces, communication protocols, etc. which are suitably arranged together for performing the solution.

[0185] Especially, the processor(s) ofthe UP entity 400 orthe control entity 410 may comprise, e.g., one ormore instances ofa Central Processing Unit (CPU), a processing unit, a processing circuit, a processor, an Application Specific Integrated Circuit (ASIC), a microprocessor, or other processing logic that may interpret and execute instructions. The expression "processor" may thus represent a processing circuitry comprising a plurality of processing circuits, such as, e.g., any, some, or all of the ones mentioned above. The processing circuitry may further perform data processing functions for inputting, outputting, and processing of data comprising data buffering and device control functions, such as call processing control, user interface control, orthe like.

Claims

CLAIMS1. A user plane entity (400), configured to : receive an activation message (401) from a control entity (410), wherein the activation message (401) is associated with a flow and comprises a configuration file (402) and flow -related information, wherein the configuration file (402) indicates a particular segment (403) of a neural network that is allocated to the user plane entity (400) for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers; derive the configuration file (402) from the activation message (401); and compose the particular segment (403) of the neural network for execution based on the configuration file (402).

2. The user plane entity (400) according to claim 1, wherein the configuration file (402) indicates a structure of the particular segment (403) of the neural network utilizing a neural network template.

3. The user plane entity (400) according to claim 2, wherein the neural network template comprises information about the neural network, information about each layer of the neural network, information about each neuron of each layer, and information about each connection between neurons.

4. The user plane entity (400) according to one of the claims 1 to 3, wherein the configuration file (402) comprises one or more of the following information: positions of all active neurons of the one or more consecutive layers, an activation function of each active neuron, or a reference to the activation function, all incoming connections to each active neuron, each incoming connection being associated with a weight, a bias of each active neuron, and compute-related information for the execution.

5. The user plane entity (400) according to one of the claims 1 to 4, configured to: receive a configuration message from a management entity, wherein the configuration message configures the user plane entity (400) for composition and execution of the particular segment (403) of the neural network, thereby enabling the user plane entity (400) to compose and execute the particular segment (403) of the neural network; and send a first acknowledgement message to the management entity, wherein the first acknowledgement message indicates whether the user plane entity (400) is successfully configured.

6. The user plane entity (400) according to claim 5, wherein the configuration message comprises a unified neural network template, and / or information of activation function of each neuron.

7. The user plane entity (400) according to claim 6, wherein the unified neural network template employs a grid-based structure, thereby enabling the addressing of neurons by their horizontal and vertical positions.

8. The user plane entity (400) according to claim 6 or 7, configured to: compose the particular segment (403) of the neural network for execution based on the unified neural network template and the configuration file (402).

9. The user plane entity (400) according to one of the claims 1 to 8, wherein the configuration file (402) employs a hierarchical representation with a tree-based structure.

10. The user plane entity (400) according to one of the claims 1 to 9, configured to: send a second acknowledgement message to the control entity (410), wherein the second acknowledgement message indicates whether the segment of the neural network is successfully composed.

11. A control entity (410), configured to: send an activation message (401) to a user plane entity (400), wherein the activation message (401) is associated with a flow and comprises a configuration file (402) and flow-related information, wherein the configuration file (402) indicates a particular segment (403) of a neural network that is allocated to the user plane entity (400) for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers.

12. The control entity (410) according to claim 11, wherein the configuration file (402) indicates a structure of the particular segment (403) of the neural network utilizing a neural network template.

13. The control entity (410) according to claim 12, wherein the neural network template comprises information about the neural network, information about each layer of the neural network, information about each neuron of each layer, and information about each connection between neurons.

14. The control entity (410) according to one of the claims 11 to 13, wherein the configuration file (402) comprises one or more of the following information: positions of all active neurons of the one or more consecutive layers, an activation function of each active neuron, or a reference to the activation function, all incoming connections to each active neuron, each incoming connection is associated with a weight, a bias of each active neuron, and compute-related information for the execution.

15. The control entity (410) according to one of the claims 11 to 14, configured to: receive a second acknowldegement message from the user plane entity (400), wherein the second acknowldegement message indicates whether the particular segment (403) of the neural network is successfully composed in the user plane entity (400).

16. The control entity (410) according to one of the claims 11 to 15, configured to: determine a segmenting option for the neural network, wherein the segmenting option is associated with the flow; and determine to allocate a first segment of the neural network to a first user plane entity (400) and allocate a second segment of the neural network to a second user plane entity (400).

17. The control entity (410) according to claim 16, configured to: provide a first configuration file (402) to the first user plane entity (400), wherein the first configuration file (402) indicates the first segment of the neural network that is allocated to the first user plane entity (400) for execution, and provide a second configuration file (402) to the second user plane entity (400), wherein the second configuration file (402) indicates the second segment of the neural network that is allocated to the second user plane entity (400) for execution, wherein the first configuration file (402) and the second configuration file (402) are different.

18. The control entity (410) according to one of the claims 11 to 17, configured to:notify a network entity that is on a path of the flow, that a segment of the neural network is allocated to the network entity for execution.

19. The control entity (410) according to claim 18, wherein the network entity comprises a user device, an access network device, or an application function.

20. The control entity (410) according to claim 18 or 19, configured to: receive a third acknowldegement message from the network entity, wherein the third acknowldegement message indicates whether the segment of the neural network is successfully activated in the network entity.

21. A method (1200) performed by a user plane entity (400), comprising: receiving (1201) an activation message (401) from a control entity (410), wherein the activation message (401) is associated with a flow and comprises a configuration file (402) and flow-related information, wherein the configuration file (402) indicates a particular segment (403) of a neural network that is allocated to the user plane entity (400) for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers; deriving (1202) the configuration file (402) from the activation message (401); and composing (1203) the segment of the neural network for execution based on the configuration file (402).

22. A method (1300) performed by a control entity (410), comprising: sending (1301) an activation message (401) to a user plane entity (400), wherein the activation message (401) is associated with a flow and comprises a configuration file (402) and flow-related information, wherein the configuration file (402) indicates a particular segment of a neural network that is allocated to the user plane entity (400) for execution, wherein the neural network comprises a plurality of layers, and the neural network is partitioned into several segments, each segment comprising one or more consecutive layers of the plurality of layers.

23. A computer program product comprising a program code for carrying out, when implemented on a processor, the method according to claim 21 or claim 22.

Images