Method and device for training a neural network model

Abstract

A method and device for training a neural network model, the method comprising: a network-side device sending initial configuration information of a neural network model to a terminal, starting a model training process under a current hyperparameter configuration, training the neural network model, and saving model parameters of the neural network model after training is completed. Embodiments of the present application transmit forward propagation parameters and backward propagation parameters generated by hidden layers in a model training process between a network-side device and a terminal, thereby realizing collaborative training of the same neural network model between the network-side device and the terminal, so that the amount of data transmitted by the terminal to the network can be reduced, and the privacy and security of user data can be ensured. In addition, in the embodiments of the present application, the terminal side only needs to train a first model part of the neural network model, so that the requirement of model training on the computing power of the terminal is reduced.

Description

Technical Field

[0001] This invention relates to the field of mobile communication technology, and specifically to a method and apparatus for training a neural network model. Background Technology

[0002] Fifth-generation mobile communication (5G) and artificial intelligence (AI) are gradually being integrated into people's lives and bringing new economic growth trends. In the future sixth-generation mobile communication (6G) era, AI is likely to become part of the network architecture.

[0003] With the rapid development of AI technology, future network architectures will increasingly integrate with AI. AI technology demands significant computing power, but many existing terminals lack the necessary computing power to train models. Therefore, current AI models are trained on the network side. However, this training method requires transmitting user sample data to the network, making it difficult to guarantee user privacy. Summary of the Invention

[0004] At least one embodiment of the present invention provides a method, terminal and network device for training a neural network model, which can reduce the computing power requirements of the terminal for model training and ensure the privacy of user data.

[0005] According to one aspect of the present invention, at least one embodiment provides a method for training a neural network model, comprising:

[0006] The network-side device sends initial configuration information of the neural network model to the terminal. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer.

[0007] The model training process is initiated under the current hyperparameter configuration to train the neural network model, and the model parameters of the neural network model are saved after training. During the model training process:

[0008] The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process.

[0009] The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal.

[0010] Furthermore, according to at least one embodiment of the present invention, the initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model portion of the neural network model, and at least one set of hyperparameters of the neural network model.

[0011] Furthermore, according to at least one embodiment of the present invention, after initiating the model training process under the current hyperparameter configuration, training the neural network model, and saving the model parameters of the neural network model after training, the method further includes:

[0012] The model testing process is initiated, and the second forward inference parameters and the label of the test sample sent by the terminal are received. The second forward inference parameters are input into the second model part to obtain the prediction result of the test sample. The second forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model testing process.

[0013] Based on the prediction results and the labels of the test samples, the performance of the neural network model is evaluated to obtain the performance evaluation results of the neural network model under the current hyperparameter configuration.

[0014] Furthermore, according to at least one embodiment of the present invention, in the model training process, the method further includes:

[0015] The network-side device also receives round number indication information sent by the terminal, which indicates the first training round number corresponding to the first forward inference parameters and the label;

[0016] When the first training round number matches the network's expected training round number, the network-side device executes the step of inputting the first forward inference parameters into the (i+1)th hidden layer and training the second model part using the labels of the training samples.

[0017] Furthermore, according to at least one embodiment of the present invention, when the first forward inference parameter and the label are not received, the method further includes:

[0018] The network-side device sends a retransmission request to the terminal for the first forward inference parameter for the desired number of training epochs of the network; and after receiving the first forward inference parameter for the desired number of training epochs of the network, it continues to execute the model training process.

[0019] Furthermore, according to at least one embodiment of the present invention, it further includes:

[0020] The network-side device also sends round number indication information to the terminal, indicating the second training round number corresponding to the backpropagation parameters.

[0021] Furthermore, according to at least one embodiment of the present invention, the network-side device receives the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number sent by the terminal from the first RRC signaling, and sends the backpropagation parameters and the round number indication information for indicating the second training round number to the terminal through the second RRC signaling.

[0022] Furthermore, according to at least one embodiment of the present invention, after receiving the first forward inference parameters and the labels of the training samples sent by the terminal, the method further includes:

[0023] If the neural network model has converged or reached the maximum number of training rounds, the model parameters of the neural network model are saved and the model training process ends. Otherwise, the steps of inputting the first forward inference parameters into the (i+1)th hidden layer, training the second model part using the labels of the training samples, and sending the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer to the terminal are executed, and the next round of training process begins.

[0024] Furthermore, according to at least one embodiment of the present invention, the neural network model has multiple sets of hyperparameters, and the method further includes:

[0025] After obtaining the performance evaluation results of the neural network model under the current hyperparameter configuration, the model training process and model testing process are repeated under another set of hyperparameters in the multiple sets of hyperparameters to obtain the model parameters and performance evaluation results of the neural network model under the other set of hyperparameters.

[0026] After the model training and model testing processes have been executed for all hyperparameters, the hyperparameters and neural network model corresponding to the optimal performance evaluation result are taken as the final hyperparameters and final neural network model. The first model part of the final hyperparameters and final neural network model is saved on the network-side device, and the terminal is notified to save the second model part of the final hyperparameters and final neural network model.

[0027] Furthermore, according to at least one embodiment of the present invention, the network-side device sends an inference start message to the terminal, the inference start message being used to initiate a model inference process based on the neural network model;

[0028] Initiate the model inference process. In the model inference process, receive the third forward inference parameter sent by the terminal. The third forward inference parameter is the intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process.

[0029] The third forward inference parameter is input into the second model part of the neural network model to obtain the inference result output by the output layer of the neural network model.

[0030] According to another aspect of the present invention, at least one embodiment provides a method for training a neural network model, comprising:

[0031] The terminal receives initial configuration information of a neural network model sent by a network-side device. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer.

[0032] Under the current hyperparameter configuration, the model training process is initiated, the neural network model is trained using training samples and their labels, and the model parameters of the neural network model are saved after training. During the model training process:

[0033] Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process.

[0034] The system receives the backpropagation parameters from the (i+1)th hidden layer to the ith layer sent by the network-side device, and corrects the model parameters based on the backpropagation parameters.

[0035] Furthermore, according to at least one embodiment of the present invention, the initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model portion of the neural network model, and at least one set of hyperparameters of the neural network model.

[0036] Furthermore, according to at least one embodiment of the present invention, after initiating the model training process under the current hyperparameter configuration, training the neural network model, and saving the model parameters of the neural network model after training, the method further includes:

[0037] The model testing process is initiated, and the neural network model is tested using test samples and their labels. The second forward inference parameters and the labels of the test samples are sent to the network-side device. The second forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model testing process.

[0038] Furthermore, according to at least one embodiment of the present invention, in the model training process, the method further includes:

[0039] The terminal sends round number indication information to the network-side device, which indicates the first training round number corresponding to the first forward inference parameters and the label.

[0040] Furthermore, according to at least one embodiment of the present invention, the method further includes:

[0041] The terminal receives a retransmission request from the network-side device for a first forward inference parameter indicating the desired number of training epochs for the network; and, based on the retransmission request, sends the first forward inference parameter indicating the desired number of training epochs for the network to the network-side device.

[0042] Furthermore, according to at least one embodiment of the present invention, it further includes:

[0043] The terminal receives round number indication information sent by the network-side device, which indicates the second training round number corresponding to the backpropagation parameters;

[0044] When the second training round number matches the terminal's desired round number, the step of correcting the model parameters according to the backtransmission parameters is performed.

[0045] Furthermore, according to at least one embodiment of the present invention, the terminal sends the first forward inference parameters, a tag, and round number indication information for indicating the first training round number to the network-side device via a first RRC signaling, and receives the backpropagation parameters and round number indication information for indicating the second training round number sent by the network-side device via a second RRC signaling.

[0046] Furthermore, according to at least one embodiment of the present invention, it further includes:

[0047] After performing the step of correcting the model parameters based on the backpropagation parameters, the current training round number is updated.

[0048] Furthermore, according to at least one embodiment of the present invention, the neural network model has multiple sets of hyperparameters, and the method further includes:

[0049] After obtaining the performance evaluation result of the neural network model under the current hyperparameter configuration, the receiving network-side device repeats the model training process and the model testing process under another set of hyperparameters among the multiple sets of hyperparameters.

[0050] After the model training and testing processes have been executed for all hyperparameters, the terminal receives a notification message from the network-side device regarding the final hyperparameters and the final neural network model, and saves the second model portion of the final hyperparameters and the final neural network model locally.

[0051] Furthermore, according to at least one embodiment of the present invention, an inference start message sent by the network-side device is received, the inference start message being used to initiate a model inference process based on the neural network model;

[0052] The model inference process is initiated, and in the model inference process, the data to be inferred is input into the neural network model, and the third forward inference parameter is sent to the network-side device. The third forward inference parameter is the intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process.

[0053] According to another aspect of the present invention, at least one embodiment provides a network-side device, comprising:

[0054] An initial configuration module is used to send initial configuration information of a neural network model to the terminal. The neural network model includes a first model part located on the terminal side and a second model part located on the network side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer.

[0055] The model training module is used to initiate the model training process of the neural network model under the current hyperparameter configuration, train the neural network model, and save the model parameters of the neural network model after training. During the model training process:

[0056] The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process.

[0057] The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal.

[0058] According to another aspect of the present invention, at least one embodiment provides a network-side device including a transceiver and a processor, wherein,

[0059] The transceiver is used to send initial configuration information of a neural network model to a terminal. The neural network model includes a first model part located on the terminal side and a second model part located on the network side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer.

[0060] The processor is configured to, during the model training process of the neural network model, initiate the model training process under the current hyperparameter configuration, train the neural network model, and save the model parameters of the neural network model after training, wherein, during the model training process:

[0061] The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process.

[0062] The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal.

[0063] According to another aspect of the present invention, at least one embodiment provides a network-side device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method described above.

[0064] According to another aspect of the present invention, at least one embodiment provides a terminal comprising:

[0065] The first receiving module is used to receive the initial configuration information of the neural network model sent by the network-side device. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer.

[0066] The model training module is used to, during the model training process of the neural network model, initiate the model training process under the current hyperparameter configuration, train the neural network model using training samples and their labels, and save the model parameters of the neural network model after training. Specifically, during the model training process:

[0067] Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process.

[0068] The system receives the backpropagation parameters from the (i+1)th hidden layer to the ith layer sent by the network-side device, and corrects the model parameters based on the backpropagation parameters.

[0069] According to another aspect of the present invention, at least one embodiment provides a terminal including a transceiver and a processor, wherein,

[0070] The transceiver is used to receive initial configuration information of a neural network model sent by a network-side device. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer.

[0071] The processor is configured to, during the model training process of the neural network model, initiate the model training process under the current hyperparameter configuration, train the neural network model using training samples and their labels, and save the model parameters of the neural network model after training, wherein, during the model training process:

[0072] Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process.

[0073] The system receives the backpropagation parameters from the (i+1)th hidden layer to the ith layer sent by the network-side device, and corrects the model parameters based on the backpropagation parameters.

[0074] According to another aspect of the present invention, at least one embodiment provides a terminal comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method described above.

[0075] According to another aspect of the present invention, at least one embodiment provides a computer-readable storage medium on which a program is stored, which, when executed by a processor, implements the steps of the method described above.

[0076] Compared with the prior art, the neural network model training method and device provided in this embodiment of the invention...

[0077] This invention enables collaborative training of the same neural network model between the network-side device and the terminal by transmitting the forward propagation parameters and backward propagation parameters generated in the hidden layer during model training between the network-side device and the terminal. This reduces the amount of data sent from the terminal to the network and ensures the privacy and security of user data. Furthermore, in this invention, the terminal only needs to train the first part of the neural network model, thus reducing the computational requirements of the terminal for model training. Attached Figure Description

[0078] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0079] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention;

[0080] Figure 2This is a schematic diagram illustrating the deployment method of the neural network model according to an embodiment of the present invention;

[0081] Figure 3 This is a flowchart illustrating the application of the neural network model training method of this invention to a network-side device.

[0082] Figure 4 This is a schematic diagram illustrating the transmission of forward inference parameters according to an embodiment of the present invention;

[0083] Figure 5 This is a schematic diagram illustrating the transmission of backpropagation parameters according to an embodiment of the present invention;

[0084] Figure 6 A flowchart illustrating the application of the neural network model training method of this invention to a terminal;

[0085] Figure 7 This is an interactive flowchart of a neural network model training method according to an embodiment of the present invention;

[0086] Figure 8 A schematic diagram of a network device provided in an embodiment of the present invention;

[0087] Figure 9 This is another schematic diagram of the network device provided in an embodiment of the present invention;

[0088] Figure 10 A schematic diagram of a terminal provided in an embodiment of the present invention;

[0089] Figure 11 This is another structural schematic diagram of the terminal provided in an embodiment of the present invention. Detailed Implementation

[0090] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0091] The terms “first,” “second,” etc., used in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The terms “and / or” in the specification and claims indicate at least one of the connected objects.

[0092] The technologies described in this document are not limited to NR systems and Long Time Evolution (LTE) / LTE-Advanced (LTE-A) systems, but can also be used in various wireless communication systems such as Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Frequency Division Multiple Access (FDMA), Orthogonal Frequency Division Multiple Access (OFDMA), Single-carrier Frequency-Division Multiple Access (SC-FDMA), and other systems. The terms "system" and "network" are often used interchangeably. CDMA systems can implement radio technologies such as CDMA2000 and Universal Terrestrial Radio Access (UTRA). UTRA includes Wideband Code Division Multiple Access (WCDMA) and other CDMA variants. TDMA systems can implement radio technologies such as the Global System for Mobile Communication (GSM). OFDMA systems can implement radio technologies such as Ultra Mobile Broadband (UMB), Evolution-UTRA (E-UTRA), IEEE 802.21 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802.20, and Flash-OFDM. UTRA and E-UTRA are part of the Universal Mobile Telecommunications System (UMTS). LTE and more advanced LTE (such as LTE-A) are newer versions of UMTS that use E-UTRA. UTRA, E-UTRA, UMTS, LTE, LTE-A, and GSM are described in documents from an organization called the 3rd Generation Partnership Project (3GPP). CDMA2000 and UMB are described in documents from an organization called 3rd Generation Partnership Project 2 (3GPP2).The techniques described herein can be used with the systems and radio technologies mentioned above, as well as with other systems and radio technologies. However, the following description describes NR systems for illustrative purposes, and NR terminology is used in most of the following description, although these techniques can also be applied to applications beyond NR systems.

[0093] The following description provides examples and is not intended to limit the scope, applicability, or configuration set forth in the claims. Changes may be made to the function and arrangement of the elements discussed without departing from the spirit and scope of this disclosure. Various procedures or components may be appropriately omitted, substituted, or added to the examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described with reference to certain examples may be combined in other examples.

[0094] Please see Figure 1 , Figure 1 This diagram illustrates a block diagram of a wireless communication system applicable to an embodiment of the present invention. The wireless communication system includes a terminal 11 and a network device 12. The terminal 11 can also be referred to as a user terminal or user equipment (UE). The terminal 11 can be a mobile phone, tablet computer, laptop computer, personal digital assistant (PDA), mobile internet device (MID), wearable device, or in-vehicle device, etc. It should be noted that the specific type of terminal 11 is not limited in this embodiment of the present invention. Network device 12 can be a base station and / or a core network element. The base station can be a 5G or later version base station (e.g., gNB, 5G NR NB, etc.), or a base station in other communication systems (e.g., eNB, WLAN access point, or other access point, etc.). The base station can be referred to as a node B, evolved node B, access point, base transceiver station (BTS), radio base station, radio transceiver, basic service set (BSS), extended service set (ESS), B node, evolved B node (eNB), home B node, home evolved B node, WLAN access point, WiFi node, or any other suitable term in the field, as long as the same technical effect is achieved. The base station is not limited to specific technical terms. It should be noted that in the embodiments of the present invention, only the base station in the NR system is used as an example, but the specific type of base station is not limited.

[0095] The base station can communicate with terminal 11 under the control of a base station controller, which in various examples may be part of the core network or some base stations. Some base stations may communicate control information or user data with the core network via backhaul. In some examples, some of these base stations may communicate with each other directly or indirectly via backhaul links, which may be wired or wireless communication links. The wireless communication system may support operation on multiple carriers (waveform signals of different frequencies). A multi-carrier transmitter can transmit modulated signals on multiple carriers simultaneously. For example, each communication link may be a multi-carrier signal modulated according to various radio technologies. Each modulated signal may be transmitted on a different carrier and may carry control information (e.g., reference signals, control channels, etc.), overhead information, data, etc.

[0096] The base station can wirelessly communicate with terminal 11 via one or more access point antennas. Each base station can provide communication coverage for its respective coverage area. The coverage area of ​​an access point can be divided into sectors that constitute only a part of that coverage area. The wireless communication system may include different types of base stations (e.g., macro base stations, micro base stations, or pico base stations). Base stations may also utilize different radio technologies, such as cellular or WLAN radio access technologies. Base stations may be associated with the same or different access networks or operator deployments. The coverage areas of different base stations (including coverage areas of the same or different types of base stations, coverage areas using the same or different radio technologies, or coverage areas belonging to the same or different access networks) may overlap.

[0097] Communication links in a wireless communication system may include an uplink for carrying uplink (UL) transmissions (e.g., from terminal 11 to network device 12) or a downlink for carrying downlink (DL) transmissions (e.g., from network device 12 to terminal 11). UL transmissions may also be referred to as reverse link transmissions, and DL transmissions may also be referred to as forward link transmissions. Downlink transmissions may be carried out using licensed frequency bands, unlicensed frequency bands, or both. Similarly, uplink transmissions may be carried out using licensed frequency bands, unlicensed frequency bands, or both.

[0098] As described in the background section, when training a neural network model for AI, if training is performed on the network side, user data needs to be stored there, which is detrimental to ensuring user data privacy and security. The terminal side, however, receives more data across more dimensions than the network side. Therefore, training on the terminal side allows for the acquisition of more user data and better protection of user data privacy. However, AI technology has high computational power requirements, and existing terminals may lack sufficient computing power, sometimes making it difficult to complete model training.

[0099] The integration of existing network architecture with AI is still in its early stages of research. Many application scenarios have not yet been considered in the network architecture, and there is a lack of relevant AI training processes. For example, how to jointly train terminals and base stations has not been considered. Although terminals have certain computing power, in many cases, their computing resources are insufficient to independently train an AI model. Therefore, this invention utilizes the collaborative computing power of terminals and base stations to complete AI model training. This invention addresses this problem below and provides a method for training a model based on RRC signaling control. This method allows terminals and base stations to each provide a portion of computing power and complete model training by transmitting the required parameters over the air interface. This method does not require terminals to transmit training sample data to the network side, effectively reducing the amount of data that terminals need to transmit, reducing storage space, and lowering the computing power requirements on the terminal side.

[0100] To address at least one of the above problems, the neural network model training method provided in this embodiment of the invention implements a joint training mechanism between the terminal and the base station, and designs an RRC signaling flow to achieve the purpose of joint training between the terminal and the base station. The terminal utilizes the computing power of the base station to help it complete the model training. Since the terminal does not need to directly transmit user data to the base station for training, the privacy of user data is guaranteed.

[0101] Neural network models typically include an input layer, several hidden layers, and an output layer, such as... Figure 2 The image shows how the model is stored. Figure 2 Taking a fully connected network model as an example, this embodiment of the invention divides the neural network model into a first model part and a second model part. The first model part is located on the terminal side, and the second model part is located on the network side device (such as a base station). The first model part includes an input layer and a (1-i)th hidden layer, and the second model part includes (i+1-n)th hidden layers and an output layer. Here, it is assumed that there are a total of n hidden layers, where i is an integer greater than or equal to 1, and n is an integer greater than or equal to 2. Figure 2 Each small circle in the diagram represents a neuron. Figure 2 The number of neurons in each layer is only illustrative and does not represent the actual number of neurons. Furthermore, it should be noted that the symbol "-" in this article does not represent a minus sign, but rather a concise representation of a continuous relationship. For example, "hidden layer 1-i" represents hidden layers 1 to i, and "hidden layers (i+1-n)" represents hidden layers (i+1) to n. Steps 709-711 later refer to steps 709 to 711, and so on.

[0102] The first part of the network model is... Figure 2The left side of the model is stored on the terminal, while the second model part is stored on the network side (such as the base station). The base station sends RRC signaling to the UE side to configure the model and initiate training or inference. RRC signaling is used to carry the parameter transmission between the front and back hidden layers, enabling parallel training of the model. Since the input layer is located on the terminal side, user data does not need to be transmitted to the network side, protecting user privacy and allowing access to more dimensions of terminal data, which helps improve model performance. This model-splitting approach allows the terminal to perform only a portion of the computation in the training model, reducing the terminal's computational load and storage space. Furthermore, this embodiment of the invention also provides an overall process and parameter transmission method for distributed training via RRC signaling, tailored to this scenario. To address the possibility of parameter transmission failure due to poor signal transmission quality, RRC signaling is introduced to carry the transmission parameters, ensuring high data reliability.

[0103] Please refer to Figure 3 The neural network model training method provided in this embodiment of the invention, when applied to a network-side device, includes:

[0104] Step 31: The network-side device sends the initial configuration information of the neural network model to the terminal.

[0105] Here, the initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model part of the neural network model, and at least one set of hyperparameters of the neural network model. Preferably, the initial configuration information configures at least the structure of the first model part of the neural network model, such as the number of input layers and hidden layers included in the first model part.

[0106] The training / inference indicator can be a training indicator (indicating that the current process is a model training process) or an inference indicator (indicating that the current process is a model inference process). The model training process is the process of training a neural network model using training samples and labels. The model training process may also include a model testing process to test the performance of the trained model. The model inference process refers to the process of using the trained model to perform inference, that is, the model application process of using the trained model to solve problems.

[0107] The structure of the first model part of the neural network model can specifically indicate the number of hidden layers i included in the first model part.

[0108] The at least one set of hyperparameters can be one or more sets of hyperparameters. Specifically, each set of hyperparameters may include at least one of the following: learning rate, regularization parameter, number of layers in the neural network, number of neurons in the hidden layer, maximum number of training epochs (number of learning rounds), number of samples per training batch, encoding method of output neurons, cost function, weight initialization method, type of neuron activation function, and scale of training data. Hyperparameters can affect the learning speed of the neural network and the final classification result.

[0109] Step 32: Start the model training process under the current hyperparameter configuration, train the neural network model, and save the model parameters of the neural network model after training. During the model training process:

[0110] The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process.

[0111] The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal.

[0112] Here, when the initial configuration information includes a training / inference indicator bit, if the training / inference indicator bit indicates a model training process, then the model training process is started, specifically including model training processes at the terminal side and network side devices. If the initial configuration information does not carry a training / inference indicator bit, the network side device can start the model training process by sending a first configuration message carrying a training indicator bit to the terminal.

[0113] For example, if the initial configuration information includes a set of hyperparameters, the model is trained based on that set of hyperparameters when the model training process is started.

[0114] For example, when the initial configuration information includes multiple sets of hyperparameters, when starting the model training process, the first set of hyperparameters can be determined according to a preset order, and then model training can be performed based on the first set of hyperparameters. After the model training based on the first set of hyperparameters is completed, the network-side device can also send a second configuration message carrying a training indicator bit to the terminal to start model training based on the second set of hyperparameters. The second configuration message can also indicate the sequence number of the second set of hyperparameters.

[0115] For example, if the initial configuration information does not include any hyperparameters, the network-side device can send a third configuration message carrying hyperparameters to the terminal to indicate the hyperparameters used during training.

[0116] After starting the model training process under the current hyperparameter configuration, the neural network model will be trained using the training samples and labels stored locally on the terminal. At the end of the training, the model parameters of the neural network model will be saved on both the network-side device and the terminal. Specifically, after the model training is completed, the network-side device can save the model parameters of the second model part and notify the terminal to save the model parameters of the first model part.

[0117] Please refer to Figure 2 During model training, the terminal side inputs training data and labels into the input layer to train the first model part. During training, the i-th hidden layer of the first model part outputs intermediate parameters (first forward inference parameters, such as...) to the (i+1)-th hidden layer. Figure 2 a in i,1 a i,2 ...a i,m Here, m represents the number of first forward inference parameters and the labels of the training data. The terminal then sends the first forward inference parameters and the labels of the training data to the network-side device, specifically via RRC signaling. The network-side device can then receive the first forward inference parameters and the labels of the training data sent by the terminal, input the first forward inference parameters into the (i+1)th hidden layer, and use the labels of the training samples to train the second model part, thereby achieving the training of the second model part. Figure 4 An interactive diagram of the forward inference parameters is provided.

[0118] Similarly, during training, the (i+1)th layer of the second model part generates the backpropagation parameters (e.g., ...) for the output of the ith layer. Figure 2 a in i+1,1 a i+1,2 ...a i+1,k Here, k represents the number of backpropagation parameters. The network-side device then sends the backpropagation parameters to the terminal, specifically via RRC signaling. This allows the terminal to use the received backpropagation parameters to correct the model parameters. Figure 5 An interactive diagram of the backpropagation parameters is provided.

[0119] Through the above steps, this embodiment of the invention can transmit the forward propagation parameters and backward propagation parameters generated by the hidden layer during model training between the network-side device and the terminal, thereby enabling collaborative training of the same neural network model between the network-side device and the terminal. This eliminates the need to store training data and labels on the network-side device, reducing the amount of data sent from the terminal to the network and ensuring the privacy and security of user data. Furthermore, since the terminal only needs to train the first part of the neural network model, the computational power requirements on the terminal are reduced.

[0120] After completing model training based on a set of hyperparameters, the neural network model (i.e., the model parameters of the neural network model) can be obtained after training. At this point, the performance of the obtained model can be further tested. Specifically, the network-side device can initiate a model testing process, for example, by sending a model testing command to the terminal to start the model testing process, or by automatically entering the model testing process after model training is completed. In the model testing process, the terminal uses locally stored test samples and their labels to test the neural network model, and sends the second forward inference parameters and the labels of the test samples to the network-side device. The second forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model testing process. The network-side device receives the second forward inference parameters and the labels of the test samples sent by the terminal, inputs the second forward inference parameters into the second model part, and obtains the prediction result for the test samples. Then, the network-side device can perform a performance evaluation of the neural network model based on the prediction result and the labels of the test samples to obtain the performance evaluation result of the neural network model under the current hyperparameter configuration.

[0121] During the model training process described above, since multiple rounds of training may be performed, when the network-side device and the terminal exchange forward propagation parameters, labels, and backpropagation parameters, the sending device can carry the indication information of the training round number corresponding to these parameters. In this way, the receiving device can determine whether it matches the expected round number and then decide whether to perform subsequent training based on the received parameters.

[0122] For example, during the model training process, when the terminal sends the first forward inference parameters and labels, it also sends epoch indication information indicating the first training epoch corresponding to the first forward inference parameters and labels. Thus, the network-side device receives not only the first forward inference parameters and labels, but also epoch indication information indicating the first training epoch corresponding to the first forward inference parameters and labels.

[0123] Then, when the first training round number matches the network's expected training round number, the network-side device executes the step of inputting the first forward inference parameters into the (i+1)th hidden layer and training the second model part using the labels of the training samples.

[0124] If the first training epoch does not match the network's expected training epoch, or if the first forward inference parameters and labels are not received, the network-side device can send a retransmission request for the first forward inference parameters for the network's expected training epoch to the terminal. Upon receiving the retransmission request, the terminal will retransmit the first forward inference parameters, labels, and the epoch indication information for the network's expected training epoch. In this way, after receiving the first forward inference parameters for the network's expected training epoch, the network-side device continues to execute the model training process.

[0125] Similarly, when sending the backpropagation parameters, the network-side device can also send epoch indication information to the terminal, indicating the second training epoch number corresponding to the backpropagation parameters. Thus, when the second training epoch matches the terminal's desired epoch number, the terminal performs the step of correcting the model parameters based on the backpropagation parameters.

[0126] In this embodiment of the invention, the convergence condition or the maximum number of training rounds of the model can be pre-configured. During the model training process, after receiving the first forward inference parameters and the labels of the training samples sent by the terminal, the network-side device determines whether the neural network model has converged or reached the maximum number of training rounds. If so, the model parameters of the neural network model are saved and the model training process ends. Otherwise, the steps of inputting the first forward inference parameters into the (i+1)th hidden layer, training the second model part using the labels of the training samples, and sending the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer to the terminal are executed, and the next round of training process is entered. At this time, the number of rounds can be updated.

[0127] In this embodiment of the invention, the neural network model can have multiple sets of hyperparameters. After obtaining the performance evaluation result of the neural network model under the current hyperparameter configuration, the network-side device can also repeat the model training process and model testing process under another set of hyperparameters to obtain the model parameters and performance evaluation result of the neural network model under that other set of hyperparameters. If the terminal does not have the other set of hyperparameters, the network-side device can also send the other set of hyperparameters to the terminal before the model testing process. Through the above steps, this embodiment of the invention can execute the above-mentioned model training process and model testing process under different hyperparameters. After the model training process and model testing process have been executed for all hyperparameters, the hyperparameters and neural network model corresponding to the optimal performance evaluation result can be used as the final hyperparameters and final neural network model. The network-side device saves the first model part of the final hyperparameters and final neural network model, and notifies the terminal to save the second model part of the final hyperparameters and final neural network model, thereby saving the optimal neural network model on both the network-side device and the terminal.

[0128] Through the above steps, embodiments of the present invention can achieve model training based on different hyperparameters, thereby obtaining a better neural network model for subsequent inference processes. Of course, embodiments of the present invention can also train based on only one set of hyperparameters and then use the obtained model for inference.

[0129] Specifically, when performing inference using the trained neural network model, the network-side device can send an inference start message carrying an inference indicator bit to the terminal. This inference start message is used to initiate the model inference process based on the neural network model. Thus, the model inference process is initiated on both the network-side device and the terminal. In the model inference process, the terminal inputs the data to be inferred into the neural network model, specifically into the input layer of the neural network model, and sends a third forward inference parameter to the network-side device. This third forward inference parameter is the intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process. The network-side device receives the third forward inference parameter sent by the terminal and inputs it into the second model part of the neural network model, thereby obtaining the inference result output by the output layer of the neural network model, realizing the specific application of the model. After obtaining the inference result, the network-side device can also send the inference result back to the terminal.

[0130] It should also be noted that in the above model training, model testing, and model inference processes, the parameters exchanged between the network-side device and the terminal can be carried through RRC signaling. For example, the network-side device receives the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number sent by the terminal from the first RRC signaling, and sends the backpropagation parameters and the round number indication information for indicating the second training round number to the terminal through the second RRC signaling.

[0131] Please refer to Figure 6 The neural network model training method provided in this embodiment of the invention, when applied to a terminal, includes:

[0132] Step 61: The terminal receives the initial configuration information of the neural network model sent by the network-side device.

[0133] Here, the neural network model includes a first model part located on the terminal side and a second model part located on the network side device. The first model part includes an input layer and a (1-i)th hidden layer, and the second model part includes (i+1-n)th hidden layers and an output layer. The initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model part of the neural network model, and at least one set of hyperparameters of the neural network model. Preferably, the initial configuration information configures at least the structure of the first model part of the neural network model, such as the number of input and hidden layers included in the first model part.

[0134] Step 62: Under the current hyperparameter configuration, start the model training process, train the neural network model using training samples and their labels, and save the model parameters of the neural network model after training. During the model training process:

[0135] Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process.

[0136] The system receives the backpropagation parameters from the (i+1)th hidden layer to the ith layer sent by the network-side device, and corrects the model parameters based on the backpropagation parameters.

[0137] Through the above steps, this embodiment of the invention transmits the forward propagation parameters and backward propagation parameters generated by the hidden layer during model training between the network-side device and the terminal, thereby enabling collaborative training of the same neural network model between the network-side device and the terminal. This reduces the amount of data sent from the terminal to the network and ensures the privacy and security of user data. Furthermore, in this embodiment, the terminal only needs to train the first part of the neural network model, thus reducing the computational power requirements of the terminal for model training.

[0138] After training a model based on a set of hyperparameters, the neural network model (i.e., the model parameters of the neural network model) can be obtained. Next, the performance of the obtained model can be tested. Specifically, the network-side device can trigger the model testing process, for example, by sending a model testing command to the terminal, which then initiates the model testing process based on the command. Alternatively, the terminal can automatically enter the model testing process after model training is completed.

[0139] In the model testing process, the terminal uses locally stored test samples and their labels to test the neural network model, and sends the second forward inference parameters and the test sample labels to the network-side device. The second forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model testing process. Upon receiving the second forward inference parameters and the test sample labels, the network-side device inputs the second forward inference parameters into the second model part to obtain the prediction result for the test sample. Furthermore, the network-side device can also perform performance evaluation on the neural network model based on the prediction result and the test sample labels to obtain the performance evaluation result of the neural network model under the current hyperparameter configuration.

[0140] Since multiple training rounds may be performed during the training of a model based on a set of hyperparameters, when the network-side device and the terminal exchange forward propagation parameters, labels, and backpropagation parameters, the sending device can carry indication information of the training round number corresponding to these parameters. This allows the receiving device to determine whether it matches the expected round number and thus decide whether to proceed with subsequent training based on the received parameters. After the terminal corrects the model parameters according to the backpropagation parameters, it updates the current training round number.

[0141] For example, in the model training process, the terminal sends epoch indication information to the network-side device, indicating the first training epoch number corresponding to the first forward inference parameters and the label. This allows the network-side device to determine whether to continue model training based on the received parameters, and to request retransmission from the terminal if the epoch number corresponding to the received parameters does not match the network's desired epoch number. In this case, the terminal receives a retransmission request from the network-side device for the first forward inference parameters corresponding to the network's desired training epoch number; and, based on the retransmission request, sends the first forward inference parameters corresponding to the network's desired training epoch number to the network-side device.

[0142] For example, when the terminal receives the backpropagation parameters, it can also receive epoch indication information sent by the network-side device, indicating the second training epoch number corresponding to the backpropagation parameters. Thus, if the second training epoch number matches the terminal's desired epoch number, the step of correcting the model parameters based on the backpropagation parameters is executed. If the second training epoch number does not match the terminal's desired epoch number, training can be paused or a retransmission request can be made to the network side.

[0143] The parameters exchanged between the network side and the terminal can be carried through RRC signaling. For example, the terminal sends the first forward inference parameters, the tag, and round number indication information for indicating the first training round number to the network-side device via the first RRC signaling, and receives the backpropagation parameters and round number indication information for indicating the second training round number sent by the network-side device via the second RRC signaling.

[0144] When the neural network model has multiple sets of hyperparameters, the terminal can also perform the above model training and model testing processes separately for each set of hyperparameters. For example, after receiving the performance evaluation results of the neural network model under the current hyperparameter configuration from the network-side device, the terminal repeats the model training and model testing processes under another set of hyperparameters. After the model training and model testing processes have been performed for all hyperparameters, the terminal receives a notification message from the network-side device regarding the final hyperparameters and the final neural network model, and saves the second model portion of the final hyperparameters and the final neural network model locally.

[0145] Through the above steps, embodiments of the present invention can achieve model training based on different hyperparameters, thereby obtaining a better neural network model for subsequent inference processes. Of course, embodiments of the present invention can also train based on only one set of hyperparameters and then use the obtained model for inference.

[0146] Specifically, the terminal can receive an inference start message sent by the network-side device. This inference start message initiates the model inference process based on the neural network model. The terminal initiates the model inference process according to this message, inputting the data to be inferred into the neural network model and sending a third forward inference parameter to the network-side device. This third forward inference parameter is an intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process. Subsequently, the terminal can also receive inference results sent by the network-side device.

[0147] Alternatively, the above model inference process can also be initiated by the terminal requesting the network side. For example, when the terminal needs to perform model inference, it sends a model inference request message to the network device. Based on this request message, the network device sends the aforementioned inference start message when model inference is permitted. The subsequent process can be referred to above.

[0148] As can be seen from the above, the embodiments of the present invention decompose the network structure of the model to reduce the processing power of the terminal and the privacy risks of the data. The neural network (AI) model is stored on the terminal side and the base station side respectively. Since the neural network is a multi-layer network model, which includes an input layer, a hidden layer and an output layer, the input layer and part of the hidden layer are stored on the terminal side, and the latter half of the hidden layer and the output layer are stored on the network side. The parameters of the first layer of the model are trained on the terminal side using user data, and the trained parameters are transmitted to the network side through the air interface for subsequent training. During backpropagation, the parameters are transmitted back to the terminal through the air interface.

[0149] Figure 7 An example of a process for training and determining an optimal model based on multiple sets of hyperparameters is given. In this example:

[0150] Figure 7 An example of a process for training and determining an optimal model based on multiple sets of hyperparameters is given. In this example:

[0151] 701. The network management side OAM first configures the network side (gNB) with the number of layers n of the neural network structure, the type of each network structure, its configuration parameters, and the calculation rules of each network.

[0152] 702-704, signaling from the gNB (base station) side to the terminal side via RRC connection establishment or RRC connection reconfiguration configures initial configuration parameters. Specifically, the aforementioned RRC signaling may include one or more of the following information:

[0153] 1. Training / Inference Indicator: This indicates whether the current process is model training or inference. If it's an inference process, only forward inference transmission is needed. If it's a training process, both forward and backward propagation transmissions are required. For example, if the current process is inference, it's marked as 0, meaning only forward inference calculations and transmissions are needed. If it's a training process, it's marked as 1, meaning both forward and backward propagation parameters need to be calculated.

[0154] 2. Maximum number of training rounds N: The process of training all training samples once is called one round of training (single training). When the number of training rounds reaches N, training can be stopped.

[0155] 3. The batch size (number of samples per training batch): When selecting training samples, if all samples are selected, the calculated descent direction will accurately point to the direction of the extremum. However, due to limited device memory, it may not be possible to read all the data for training at once. Therefore, it is necessary to set the batch size so that the model trains only a portion of the samples each time. That is, each training session will select batch size samples for training, and the network weights will be updated once every batch size of samples.

[0156] 4. Gradient Threshold: When calculating backpropagation in deep learning, it's necessary to monitor whether the gradient value is too large. If the gradient is too large, gradient explosion will occur during backpropagation, ultimately causing the model to fail to converge. Therefore, a gradient threshold needs to be set. If the gradient exceeds the threshold, the gradient can be set to that threshold value before parameter update calculation.

[0157] 5. Learning rate: This is a parameter that controls how fast the model learns. During model training, when the value of the loss function on the base station side tends to remain constant, the learning rate can be decayed by issuing a command through RRC signaling, that is, reducing the size of the learning rate, thereby slowing down the gradient descent speed and reducing the intensity of model learning, making it easier to find the minimum point of the current region during training.

[0158] 6. Define the neural network computation rules and their configuration parameters, and configure the neural network model information by introducing the neural network model domain. This information can be customized according to predefined fields defined by both parties, such as: the number of layers n in the current neural network, the network structure type of each layer, its configuration parameters, and the computation rules for each layer. Take a common centralized network structure as an example:

[0159] (A) Convolutional Neural Network (CNN): The parameters that need to be configured are the number of convolutional kernels (filters), the kernel size (size), the stride, the padding (pad), the batch normalization (bn), the activation function (such as sigmoid, tanh, relu, leaky relu, etc.), and the initial weights.

[0160] (B) Fully Connected Network (FC): The parameters that need to be configured are the number of convolutional kernels, filters, activation functions, and initial weights.

[0161] (C) Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM): The parameters that need to be configured are the number of hidden nodes (hiddensize), the number of network layers (numlayer), the activation function, and the initial weights.

[0162] (D) Pooling layer: The parameters that need to be configured are the pooling stride and the pooling kernel size.

[0163] 705-706. Based on the currently configured hyperparameters, the terminal begins the model training process and transmits the forward inference parameters and labels obtained during the training process to the network side via RRC signaling. The network side first determines whether the maximum number of training epochs or the convergence threshold has been reached. If not, the network side performs backpropagation and sends the corresponding backpropagation parameters to the terminal side via RRC signaling. The terminal corrects the model parameters based on the backpropagation transmission and continues forward propagation, repeating the above process until the convergence threshold is met or the maximum number of training epochs is reached. If the threshold is met, the training process ends, and the model testing process begins.

[0164] 707-708, during the model testing process, the performance of the model is tested.

[0165] 709-711, reconfiguration model, the base station configures another set of hyperparameters for the UE.

[0166] 712-717, the terminal and network repeatedly execute the model training process, testing process and model reconfiguration process until all hyperparameters are trained and tested, and the optimal hyperparameters are retained.

[0167] In this embodiment of the invention, the terminal side pre-stores the training dataset and the test dataset, and an RRC connection is established between the terminal and the network-side device.

[0168] During the model configuration phase, the base station sends the model configuration information to the terminal via RRC signaling, indicating that the terminal is currently in the training phase. The terminal builds the model according to the base station's requirements, including the model type and model parameters. The terminal saves parameters such as the learning rate, gradient threshold, number of samples per training iteration, and maximum number of training epochs.

[0169] During model training, the model parameter optimization phase involves forward inference and backpropagation using the local training dataset, meaning both samples and their labels reside on the terminal side. The forward inference parameters and sample labels calculated by the terminal are transmitted to the base station via uplink data over the air interface. The backpropagation parameters calculated by the base station are transmitted to the terminal via downlink data over the air interface. For example, ... Figure 4 As shown, the forward inference parameters corresponding to the xth training round are labeled as a. i,1 a i,2 ... a i,m The base station side correctly receives a i,1 a i,2 ... a i,m Then, the parameter b corresponding to backpropagation is calculated. i+1,1 b i+1,2 ... b i+1,k The base station sends backpropagation data to the terminal via RRC signaling. This allows model parameters to be updated based on the learning rate, gradient threshold, number of samples per training iteration, maximum number of training epochs, and sample labels. When the model performs the x-th training epoch, the terminal transmits parameters and labels over the air interface for the x-th forward propagation. The base station calculates the gradient descent direction of the loss function and uses this to calculate the new weights for each layer. The base station then transmits the backpropagation parameters to the terminal over the air interface. After backpropagation, the forward inference parameters for the (x+1)-th epoch are updated. This process is repeated until the configured maximum number of training epochs or the convergence threshold is reached.

[0170] During the model testing phase, the terminal sends intermediate parameters and labels to the base station via uplink through forward propagation. The trained model is then used to predict test samples at the base station, and the model's performance is evaluated at the base station.

[0171] During the model inference phase, the base station sends out configurations, specifying the current state as model inference and batch size. Upon receiving this message, the terminal performs forward inference and sends the intermediate parameters to the base station via the uplink. The trained model then predicts the samples on the base station side.

[0172] During model training, if the signal transmission quality is poor, the parameters may fail to reach the base station or terminal side when transmitted over the air interface. In this case, the model training will fail to produce the correct weights due to the incorrect parameter transmission. To address this issue, this example proposes a method using RRC signaling to transmit training parameters.

[0173] 1. When sending forward inference data, the terminal triggers the generation of forward inference RRC signaling, indicating the number of training rounds for transmitting the current parameters and the corresponding forward inference parameters in the RRC signaling. After receiving the forward inference RRC signaling, the base station verifies whether the number of training rounds is the expected number of training rounds. If it is, it updates the number of training rounds. If not, it discards the signal.

[0174] 2. The base station can only perform backpropagation calculations when the training round number corresponding to the forward inference is the expected round number. If no forward inference signaling is received, the base station triggers a forward inference retransmission request for the corresponding round number until the forward inference for the corresponding round number is received or the forward inference timer times out.

[0175] 3. After receiving the backpropagation parameters generated after the xth training iteration, the terminal updates the forward inference parameters for x+1. If the terminal does not receive the backpropagation parameters, it sets the backpropagation parameters to 0 after the backpropagation timer expires, and then continues to update the forward inference parameters for x+1.

[0176] 4. Repeat the above process until the maximum number of training rounds configured or the convergence threshold is reached.

[0177] The following is given Figure 7 Examples of the above types of RRC signaling:

[0178] One possible form for forward and backward propagation RRC signaling is as follows:

[0179] AIParaInformation-IEs::=SEQUENCE{

[0180] CycleID----Training Round Number

[0181] y----tag

[0182] ForwardPara sequence(SIZE(1..maxParaNum))Para, --- forward propagation parameters, backward propagation parameters

[0183] },

[0184] One possible form of forward inference retransmission request (RRC) signaling is as follows:

[0185] AIParaRequestInformation-IEs::=SEQUENCE{

[0186] Batchsize----Number of test images

[0187] },

[0188] As can be seen from the above examples, the embodiments of the present invention can reduce the computational load and model storage load on the terminal side while ensuring user data privacy. By using RRC signaling, the embodiments of the present invention can effectively and reliably ensure joint training of the model on the terminal side and the base station side.

[0189] The various methods of the embodiments of the present invention have been described above. Apparatus for implementing the above methods will now be provided.

[0190] The embodiments of the present invention provide Figure 8 A network-side device shown includes:

[0191] The initial configuration module 81 is used to send initial configuration information of the neural network model to the terminal. The neural network model includes a first model part located on the terminal side and a second model part located on the network side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer.

[0192] Model training module 82 is used to initiate the model training process under the current hyperparameter configuration during the model training process of the neural network model, train the neural network model, and save the model parameters of the neural network model after training. During the model training process:

[0193] The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process.

[0194] The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal.

[0195] Optionally, the initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model portion of the neural network model, and at least one set of hyperparameters of the neural network model.

[0196] Optionally, the network-side device further includes:

[0197] The model testing module is used to initiate the model training process under the current hyperparameter configuration, train the neural network model, and after saving the model parameters of the neural network model after training, initiate the model testing process. It receives the second forward inference parameters and the labels of the test samples sent by the terminal, inputs the second forward inference parameters into the second model part, and obtains the prediction results for the test samples. The second forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model testing process. Based on the prediction results and the labels of the test samples, it performs a performance evaluation of the neural network model to obtain the performance evaluation results of the neural network model under the current hyperparameter configuration.

[0198] Optionally, the model training module is further configured to receive, during the model training process, round number indication information sent by the terminal indicating the first training round number corresponding to the first forward inference parameters and labels; and when the first training round number matches the network's expected training round number, execute the step of inputting the first forward inference parameters into the (i+1)th hidden layer and training the second model part using the labels of the training samples.

[0199] Optionally, the model training module is further configured to send a retransmission request for the first forward inference parameter for the desired number of training rounds of the network to the terminal when the first forward inference parameter and label are not received; and to continue executing the model training process after receiving the first forward inference parameter for the desired number of training rounds of the network.

[0200] Optionally, the model training module is further configured to send epoch indication information to the terminal, indicating the second training epoch number corresponding to the backpropagation parameters.

[0201] Optionally, the model training module is further configured to receive the first forward inference parameters, the label, and the round number indication information for indicating the first training round number sent by the terminal from the first RRC signaling, and to send the backpropagation parameters and the round number indication information for indicating the second training round number to the terminal via the second RRC signaling.

[0202] Optionally, the model training module is further configured to, after receiving the first forward inference parameters and the labels of the training samples sent by the terminal, determine whether the neural network model has converged or reached the maximum number of training rounds. If so, save the model parameters of the neural network model and end the model training process; otherwise, execute the steps of inputting the first forward inference parameters into the (i+1)th hidden layer, training the second model part using the labels of the training samples, and sending the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer to the terminal, and enter the next round of training process.

[0203] Optionally, the neural network model has multiple sets of hyperparameters. The model training module is further configured to, after obtaining the performance evaluation result of the neural network model under the current hyperparameter configuration, repeatedly execute the model training process and model testing process under another set of hyperparameters in the multiple sets of hyperparameters to obtain the model parameters and performance evaluation result of the neural network model under the other set of hyperparameters; after the model training process and model testing process have been executed for all hyperparameters, the hyperparameters and neural network model corresponding to the optimal performance evaluation result are taken as the final hyperparameters and final neural network model, and the first model part of the final hyperparameters and final neural network model is saved on the network-side device, and the terminal is notified to save the second model part of the final hyperparameters and final neural network model.

[0204] Optionally, the network-side device further includes:

[0205] The model inference module is used to send an inference start message to the terminal, which is used to start the model inference process based on the neural network model; in the model inference process, the module receives a third forward inference parameter sent by the terminal, which is an intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process; the module inputs the third forward inference parameter to the second model part of the neural network model to obtain the inference result output by the output layer of the neural network model.

[0206] It should be noted that the device in this embodiment is the same as the one described above. Figure 3 The devices corresponding to the methods shown are all applicable to the embodiments of the above-described devices, and can achieve the same technical effects. It should be noted that the devices provided in this embodiment can implement all the method steps implemented in the above-described method embodiments and achieve the same technical effects. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0207] Please refer to Figure 9 This invention provides a schematic diagram of a network-side device, including: a processor 901, a transceiver 902, a memory 903, and a bus interface, wherein:

[0208] In this embodiment of the invention, the network-side device further includes: a program stored on a memory 903 and executable on a processor 901, wherein the program, when executed by the processor 901, performs the following steps:

[0209] Send initial configuration information of a neural network model to the terminal, wherein the neural network model includes a first model part located on the terminal side and a second model part located on the network side device, the first model part includes an input layer and a hidden layer of layer 1-i, and the second model part includes a hidden layer of layer i+1-n and an output layer;

[0210] The model training process is initiated under the current hyperparameter configuration to train the neural network model, and the model parameters of the neural network model are saved after training. During the model training process:

[0211] The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process.

[0212] The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal.

[0213] Optionally, the initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model portion of the neural network model, and at least one set of hyperparameters of the neural network model.

[0214] Optionally, when the processor executes the program, it further implements the following steps:

[0215] The model training process is started under the current hyperparameter configuration. The neural network model is trained, and after the model parameters of the neural network model are saved after training, the model testing process is started. The second forward inference parameters and the label of the test sample sent by the terminal are received. The second forward inference parameters are input into the second model part to obtain the prediction result of the test sample. The second forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model testing process.

[0216] Based on the prediction results and the labels of the test samples, the performance of the neural network model is evaluated to obtain the performance evaluation results of the neural network model under the current hyperparameter configuration.

[0217] Optionally, when the processor executes the program, it further implements the following steps:

[0218] In the model training process, the receiving terminal sends round number indication information to indicate the first training round number corresponding to the first forward inference parameters and the label;

[0219] When the first training round number matches the network's expected training round number, the step of inputting the first forward inference parameters into the (i+1)th hidden layer and training the second model part using the labels of the training samples is executed.

[0220] Optionally, when the processor executes the program, it further implements the following steps:

[0221] If the first forward inference parameters and the label are not received, a retransmission request for the first forward inference parameters for the desired number of training epochs of the network is sent to the terminal; and after receiving the first forward inference parameters for the desired number of training epochs of the network, the model training process continues.

[0222] Optionally, when the processor executes the program, it further implements the following steps:

[0223] Send round number indication information to the terminal to indicate the second training round number corresponding to the backpropagation parameters.

[0224] Optionally, when the processor executes the program, it further implements the following steps:

[0225] The system receives the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number from the terminal in the first RRC signaling, and sends the backpropagation parameters and the round number indication information for indicating the second training round number to the terminal via the second RRC signaling.

[0226] Optionally, when the processor executes the program, it further implements the following steps:

[0227] After receiving the first forward inference parameters and the labels of the training samples sent by the terminal, it is determined whether the neural network model has converged or reached the maximum number of training rounds. If so, the model parameters of the neural network model are saved and the model training process ends. Otherwise, the steps of inputting the first forward inference parameters into the (i+1)th hidden layer, training the second model part using the labels of the training samples, and sending the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer to the terminal are executed, and the next round of training process begins.

[0228] Optionally, the neural network model has multiple sets of hyperparameters; optionally, the processor, when executing the program, further implements the following steps:

[0229] After obtaining the performance evaluation results of the neural network model under the current hyperparameter configuration, the model training process and model testing process are repeated under another set of hyperparameters in the multiple sets of hyperparameters to obtain the model parameters and performance evaluation results of the neural network model under the other set of hyperparameters.

[0230] After the model training and model testing processes have been executed for all hyperparameters, the hyperparameters and neural network model corresponding to the optimal performance evaluation result are taken as the final hyperparameters and final neural network model. The first model part of the final hyperparameters and final neural network model is saved on the network-side device, and the terminal is notified to save the second model part of the final hyperparameters and final neural network model.

[0231] Optionally, when the processor executes the program, it further implements the following steps:

[0232] Send an inference start message to the terminal, the inference start message being used to initiate the model inference process based on the neural network model;

[0233] Initiate the model inference process. In the model inference process, receive the third forward inference parameter sent by the terminal. The third forward inference parameter is the intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process.

[0234] The third forward inference parameter is input into the second model part of the neural network model to obtain the inference result output by the output layer of the neural network model.

[0235] Understandably, in this embodiment of the invention, the computer program executed by the processor 901 can achieve the above-mentioned functions. Figure 3 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0236] exist Figure 9 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 901) and memory (memory 903). The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 902 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium.

[0237] The processor 901 is responsible for managing the bus architecture and general processing, while the memory 903 can store the data used by the processor 901 when performing operations.

[0238] It should be noted that the terminal in this embodiment is the same as the one described above. Figure 3The device corresponding to the method shown above, and the implementation methods in each embodiment, are all applicable to the embodiments of this terminal, and can achieve the same technical effect. In this device, the transceiver 902 and the memory 903, as well as the transceiver 902 and the processor 901, can be connected for communication via a bus interface. The function of the processor 901 can also be implemented by the transceiver 902, and the function of the transceiver 902 can also be implemented by the processor 901. It should be noted that the device provided by the embodiments of the present invention can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Therefore, the parts that are the same as those in the method embodiments and the beneficial effects will not be described in detail here.

[0239] In some embodiments of the present invention, a computer-readable storage medium is also provided, on which a program is stored, which, when executed by a processor, performs the following steps:

[0240] Send initial configuration information of a neural network model to the terminal, wherein the neural network model includes a first model part located on the terminal side and a second model part located on the network side device, the first model part includes an input layer and a hidden layer of layer 1-i, and the second model part includes a hidden layer of layer i+1-n and an output layer;

[0241] The model training process is initiated under the current hyperparameter configuration to train the neural network model, and the model parameters of the neural network model are saved after training. During the model training process:

[0242] The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process.

[0243] The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal.

[0244] When executed by the processor, this program can implement all the above-mentioned methods for training neural network models applied to network-side devices, and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0245] Please refer to Figure 10 This invention provides a terminal, comprising:

[0246] The first receiving module 101 is used to receive the initial configuration information of the neural network model sent by the network-side device. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer.

[0247] Model training module 102 is used to, during the model training process of the neural network model, initiate the model training process under the current hyperparameter configuration, train the neural network model using training samples and their labels, and save the model parameters of the neural network model after training is completed. In the model training process:

[0248] Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process.

[0249] The system receives the backpropagation parameters from the (i+1)th hidden layer to the ith layer sent by the network-side device, and corrects the model parameters based on the backpropagation parameters.

[0250] Optionally, the initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model portion of the neural network model, and at least one set of hyperparameters of the neural network model.

[0251] Optionally, the terminal further includes:

[0252] The model testing module is used to start the model training process under the current hyperparameter configuration, train the neural network model, and after saving the model parameters of the neural network model after training, start the model testing process, test the neural network model using test samples and their labels, and send the second forward inference parameters and the labels of the test samples to the network-side device. The second forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model testing process.

[0253] Optionally, the model training module is further configured to send round number indication information to the network-side device during the model training process, indicating the first training round number corresponding to the first forward inference parameters and the label.

[0254] Optionally, the model training module is further configured to receive a retransmission request for a first forward inference parameter for the desired number of training epochs of the network sent by the network-side device; and, according to the retransmission request, send the first forward inference parameter for the desired number of training epochs of the network to the network-side device.

[0255] Optionally, the model training module is further configured to receive round number indication information sent by the network-side device, which indicates the second training round number corresponding to the backpropagation parameters; and when the second training round number matches the terminal's desired round number, to perform the step of correcting the model parameters according to the backpropagation parameters.

[0256] Optionally, the model training module is further configured to send the first forward inference parameters, the label, and the round number indication information for indicating the first training round number to the network-side device via the first RRC signaling, and to receive the backpropagation parameters and the round number indication information for indicating the second training round number sent by the network-side device via the second RRC signaling.

[0257] Optionally, the model training module is further configured to update the current training round number after performing the step of correcting the model parameters according to the backtransmission parameters.

[0258] Optionally, the neural network model has multiple sets of hyperparameters. Optionally, the model training module is further configured to receive, after the network-side device obtains the performance evaluation result of the neural network model under the current hyperparameter configuration, repeatedly execute the model training process and model testing process under another set of hyperparameters among the multiple sets of hyperparameters; after the model training process and model testing process have been executed for all hyperparameters, receive the notification message of the final hyperparameters and the final neural network model sent by the network-side device, and save the second model part of the final hyperparameters and the final neural network model locally on the terminal.

[0259] Optionally, the terminal further includes:

[0260] The model inference module is used to receive an inference start message sent by the network-side device. The inference start message is used to start the model inference process based on the neural network model. The model inference process is started, and in the model inference process, the data to be inferred is input into the neural network model, and a third forward inference parameter is sent to the network-side device. The third forward inference parameter is an intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process.

[0261] It should be noted that the device in this embodiment is the same as the one described above. Figure 6 The devices corresponding to the methods shown are all applicable to the embodiments of the above-described devices, and can achieve the same technical effects. The devices provided by the embodiments of the present invention can implement all the method steps implemented in the above-described method embodiments and can achieve the same technical effects. Therefore, the parts that are the same as those in the method embodiments and the beneficial effects will not be described in detail here.

[0262] Please refer to Figure 11A schematic diagram of a terminal provided in an embodiment of the present invention includes: a processor 1101, a transceiver 1102, a memory 1103, a user interface 1104, and a bus interface.

[0263] In this embodiment of the invention, the terminal further includes a program stored on memory 1103 and executable on processor 1101.

[0264] When the processor 1101 executes the program, it performs the following steps:

[0265] Receive initial configuration information of a neural network model sent by a network-side device, wherein the neural network model includes a first model part located on the terminal side and a second model part located on the network-side device, the first model part includes an input layer and a hidden layer of layer 1-i, and the second model part includes a hidden layer of layer i+1-n and an output layer;

[0266] Under the current hyperparameter configuration, the model training process is initiated, the neural network model is trained using training samples and their labels, and the model parameters of the neural network model are saved after training. During the model training process:

[0267] Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process.

[0268] The system receives the backpropagation parameters from the (i+1)th hidden layer to the ith layer sent by the network-side device, and corrects the model parameters based on the backpropagation parameters.

[0269] Optionally, the initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model portion of the neural network model, and at least one set of hyperparameters of the neural network model.

[0270] Optionally, when the processor executes the program, it further implements the following steps:

[0271] The model training process is initiated under the current hyperparameter configuration to train the neural network model. After the training is completed, the model parameters of the neural network model are saved. Then, the model testing process is initiated to test the neural network model using test samples and their labels. The second forward inference parameters and the labels of the test samples are sent to the network-side device. The second forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model testing process.

[0272] Optionally, when the processor executes the program, it further implements the following steps:

[0273] During the model training process, round number indication information is sent to the network-side device to indicate the first training round number corresponding to the first forward inference parameters and the label.

[0274] Optionally, when the processor executes the program, it further implements the following steps:

[0275] The system receives a retransmission request from a network-side device for a first forward inference parameter indicating the desired number of training epochs for the network; and, based on the retransmission request, sends the first forward inference parameter indicating the desired number of training epochs for the network to the network-side device.

[0276] Optionally, when the processor executes the program, it further implements the following steps:

[0277] Receive round number indication information sent by the network-side device, which indicates the second training round number corresponding to the backpropagation parameters;

[0278] When the second training round number matches the terminal's desired round number, the step of correcting the model parameters according to the backtransmission parameters is performed.

[0279] Optionally, when the processor executes the program, it further implements the following steps:

[0280] The system sends the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number to the network-side device via the first RRC signaling, and receives the backpropagation parameters and the round number indication information for indicating the second training round number sent by the network-side device via the second RRC signaling.

[0281] Optionally, when the processor executes the program, it further implements the following steps:

[0282] After performing the step of correcting the model parameters based on the backpropagation parameters, the current training round number is updated.

[0283] Optionally, the neural network model has multiple sets of hyperparameters, and the processor, when executing the program, further implements the following steps:

[0284] After obtaining the performance evaluation result of the neural network model under the current hyperparameter configuration, the receiving network-side device repeats the model training process and the model testing process under another set of hyperparameters among the multiple sets of hyperparameters.

[0285] After the model training and testing processes have been executed for all hyperparameters, the terminal receives a notification message from the network-side device regarding the final hyperparameters and the final neural network model, and saves the second model portion of the final hyperparameters and the final neural network model locally.

[0286] Optionally, when the processor executes the program, it further implements the following steps:

[0287] Receive an inference start message sent by the network-side device, the inference start message being used to initiate the model inference process based on the neural network model;

[0288] The model inference process is initiated, and in the model inference process, the data to be inferred is input into the neural network model, and the third forward inference parameter is sent to the network-side device. The third forward inference parameter is the intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process.

[0289] Understandably, in this embodiment of the invention, the computer program executed by the processor 1101 can achieve the above-mentioned functions. Figure 6 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0290] exist Figure 11 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 1101) and memory (memory 1103). The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 1102 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. For different user equipment, the user interface 1104 can also be an interface capable of connecting external or internal devices, including but not limited to keypads, displays, speakers, microphones, joysticks, etc.

[0291] The processor 1101 is responsible for managing the bus architecture and general processing, and the memory 1103 can store the data used by the processor 1101 when performing operations.

[0292] It should be noted that the device in this embodiment is the same as the one described above. Figure 6The device corresponding to the method shown above is applicable to the embodiments of this device and can achieve the same technical effect. In this device, the transceiver 1102 and the memory 1103, as well as the transceiver 1102 and the processor 1101, can be connected via a bus interface. The functions of the processor 1101 can also be implemented by the transceiver 1102, and vice versa. It should be noted that the device provided in this embodiment can implement all the method steps implemented in the above method embodiments and achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0293] In some embodiments of the present invention, a computer-readable storage medium is also provided, on which a program is stored, which, when executed by a processor, performs the following steps:

[0294] Receive initial configuration information of a neural network model sent by a network-side device, wherein the neural network model includes a first model part located on the terminal side and a second model part located on the network-side device, the first model part includes an input layer and a hidden layer of layer 1-i, and the second model part includes a hidden layer of layer i+1-n and an output layer;

[0295] Under the current hyperparameter configuration, the model training process is initiated, the neural network model is trained using training samples and their labels, and the model parameters of the neural network model are saved after training. During the model training process:

[0296] Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process.

[0297] The system receives the backpropagation parameters from the (i+1)th hidden layer to the ith layer sent by the network-side device, and corrects the model parameters based on the backpropagation parameters.

[0298] When executed by the processor, this program can implement all the above-mentioned methods for training neural network models applied to the terminal side, and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0299] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0300] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0301] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0302] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0303] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0304] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0305] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for training a neural network model, characterized in that, include: The network-side device sends initial configuration information of the neural network model to the terminal. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer. The model training process is initiated under the current hyperparameter configuration to train the neural network model, and the model parameters of the neural network model are saved after training. During the model training process: The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process. The network-side device also receives round number indication information sent by the terminal, which indicates the first training round number corresponding to the first forward inference parameters and the label; The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal. The network-side device also sends round number indication information to the terminal to indicate the second training round number corresponding to the backpropagation parameters; The network-side device receives the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number sent by the terminal from the first RRC signaling, and sends the back propagation parameters and the round number indication information for indicating the second training round number to the terminal through the second RRC signaling; When the first forward inference parameter and label are not received, the method further includes: The network-side device sends a retransmission request to the terminal for the first forward inference parameter for the desired number of training epochs of the network; and after receiving the first forward inference parameter for the desired number of training epochs of the network, it continues to execute the model training process.

2. The method as described in claim 1, characterized in that, The initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model part of the neural network model, and at least one set of hyperparameters of the neural network model.

3. The method as described in claim 1, characterized in that, After initiating the model training process under the current hyperparameter configuration, training the neural network model, and saving the model parameters of the neural network model after training, the process also includes: The model testing process is initiated, and the second forward inference parameters and the label of the test sample sent by the terminal are received. The second forward inference parameters are input into the second model part to obtain the prediction result of the test sample. The second forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model testing process. Based on the prediction results and the labels of the test samples, the performance of the neural network model is evaluated to obtain the performance evaluation results of the neural network model under the current hyperparameter configuration.

4. The method as described in claim 1, characterized in that, In the model training process, the method further includes: When the first training round number matches the network's expected training round number, the network-side device executes the step of inputting the first forward inference parameters into the (i+1)th hidden layer and training the second model part using the labels of the training samples.

5. The method as described in claim 1, characterized in that, After receiving the first forward inference parameters and the labels of the training samples sent by the terminal, the method further includes: If the neural network model has converged or reached the maximum number of training rounds, the model parameters of the neural network model are saved and the model training process ends. Otherwise, the steps of inputting the first forward inference parameters into the (i+1)th hidden layer, training the second model part using the labels of the training samples, and sending the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer to the terminal are executed, and the next round of training process begins.

6. The method as described in claim 1, characterized in that, The neural network model has multiple sets of hyperparameters, and the method further includes: After obtaining the performance evaluation results of the neural network model under the current hyperparameter configuration, the model training process and model testing process are repeated under another set of hyperparameters in the multiple sets of hyperparameters to obtain the model parameters and performance evaluation results of the neural network model under the other set of hyperparameters. After the model training and model testing processes have been executed for all hyperparameters, the hyperparameters and neural network model corresponding to the optimal performance evaluation result are taken as the final hyperparameters and final neural network model. The first model part of the final hyperparameters and final neural network model is saved on the network-side device, and the terminal is notified to save the second model part of the final hyperparameters and final neural network model.

7. The method as described in claim 1, characterized in that, The network-side device sends an inference start message to the terminal, which is used to start the model inference process based on the neural network model; Initiate the model inference process. In the model inference process, receive the third forward inference parameter sent by the terminal. The third forward inference parameter is the intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process. The third forward inference parameter is input into the second model part of the neural network model to obtain the inference result output by the output layer of the neural network model.

8. A method for training a neural network model, characterized in that, include: The terminal receives initial configuration information of a neural network model sent by a network-side device. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer. Under the current hyperparameter configuration, the model training process is initiated, the neural network model is trained using training samples and their labels, and the model parameters of the neural network model are saved after training. During the model training process: Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process. The terminal sends round number indication information to the network-side device to indicate the first training round number corresponding to the first forward inference parameters and the label; Receive the backpropagation parameters output from the (i+1)th hidden layer to the ith layer from the network-side device, and correct the model parameters according to the backpropagation parameters; The terminal receives round number indication information sent by the network-side device, which indicates the second training round number corresponding to the backpropagation parameters; When the second training round number matches the terminal's desired round number, the step of correcting the model parameters according to the reverse transmission parameters is executed; The terminal sends the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number to the network-side device via the first RRC signaling, and receives the back propagation parameters and the round number indication information for indicating the second training round number sent by the network-side device via the second RRC signaling. The method further includes: The terminal receives a retransmission request from the network-side device for a first forward inference parameter indicating the desired number of training epochs for the network; and, based on the retransmission request, sends the first forward inference parameter indicating the desired number of training epochs for the network to the network-side device.

9. The method as described in claim 8, characterized in that, The initial configuration information includes at least one of the following: training / inference indicator bits, the structure of the first model part of the neural network model, and at least one set of hyperparameters of the neural network model.

10. The method as described in claim 8, characterized in that, After initiating the model training process under the current hyperparameter configuration, training the neural network model, and saving the model parameters of the neural network model after training, the process also includes: The model testing process is initiated, and the neural network model is tested using test samples and their labels. The second forward inference parameters and the labels of the test samples are sent to the network-side device. The second forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model testing process.

11. The method as described in claim 8, characterized in that, Also includes: After performing the step of correcting the model parameters based on the backpropagation parameters, the current training round number is updated.

12. The method as described in claim 8, characterized in that, The neural network model has multiple sets of hyperparameters, and the method further includes: After obtaining the performance evaluation result of the neural network model under the current hyperparameter configuration, the receiving network-side device repeats the model training process and the model testing process under another set of hyperparameters among the multiple sets of hyperparameters. After the model training and testing processes have been executed for all hyperparameters, the terminal receives a notification message from the network-side device regarding the final hyperparameters and the final neural network model, and saves the second model portion of the final hyperparameters and the final neural network model locally.

13. The method as described in claim 8, characterized in that, Receive an inference start message sent by the network-side device, the inference start message being used to initiate the model inference process based on the neural network model; The model inference process is initiated, and in the model inference process, the data to be inferred is input into the neural network model, and the third forward inference parameter is sent to the network-side device. The third forward inference parameter is the intermediate parameter output from the i-th hidden layer to the (i+1)-th layer in the model inference process.

14. A network-side device, characterized in that, include: An initial configuration module is used to send initial configuration information of a neural network model to the terminal. The neural network model includes a first model part located on the terminal side and a second model part located on the network side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer. The model training module is used to initiate the model training process of the neural network model under the current hyperparameter configuration, train the neural network model, and save the model parameters of the neural network model after training. During the model training process: The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process. The network-side device also receives round number indication information sent by the terminal, which indicates the first training round number corresponding to the first forward inference parameters and the label; The first forward inference parameters are input into the (i+1)th hidden layer, the second model part is trained using the labels of the training samples, and the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer are sent to the terminal. The network-side device also sends round number indication information to the terminal to indicate the second training round number corresponding to the backpropagation parameters; The network-side device receives the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number sent by the terminal from the first RRC signaling, and sends the back propagation parameters and the round number indication information for indicating the second training round number to the terminal through the second RRC signaling; If the first forward inference parameters and labels are not received, the model training process further includes: The network-side device sends a retransmission request to the terminal for the first forward inference parameter for the desired number of training epochs of the network; and after receiving the first forward inference parameter for the desired number of training epochs of the network, it continues to execute the model training process.

15. A network-side device, characterized in that, Includes transceivers and processors, among which, The transceiver is used to send initial configuration information of a neural network model to a terminal. The neural network model includes a first model part located on the terminal side and a second model part located on the network side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer. The processor is configured to, during the model training process of the neural network model, initiate the model training process under the current hyperparameter configuration, train the neural network model, and save the model parameters of the neural network model after training, wherein, during the model training process: The system receives the first forward inference parameters and the labels of the training samples sent by the terminal. The first forward inference parameters are intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer during the model training process. The network-side device also receives round number indication information sent by the terminal, which indicates the first training round number corresponding to the first forward inference parameters and the label; inputs the first forward inference parameters into the (i+1)th hidden layer, trains the second model part using the labels of the training samples, and sends the backpropagation parameters output from the (i+1)th hidden layer to the i-th layer to the terminal. The network-side device also sends round number indication information to the terminal to indicate the second training round number corresponding to the backpropagation parameters; The network-side device receives the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number sent by the terminal from the first RRC signaling, and sends the back propagation parameters and the round number indication information for indicating the second training round number to the terminal through the second RRC signaling; If the first forward inference parameters and labels are not received, the model training process further includes: The network-side device sends a retransmission request to the terminal for the first forward inference parameter for the desired number of training epochs of the network; and after receiving the first forward inference parameter for the desired number of training epochs of the network, it continues to execute the model training process.

16. A network-side device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 7.

17. A terminal, characterized in that, include: The first receiving module is used to receive the initial configuration information of the neural network model sent by the network-side device. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer. The model training module is used to, during the model training process of the neural network model, initiate the model training process under the current hyperparameter configuration, train the neural network model using training samples and their labels, and save the model parameters of the neural network model after training. Specifically, during the model training process: Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process. The terminal sends round number indication information to the network-side device to indicate the first training round number corresponding to the first forward inference parameters and the label; Receive the backpropagation parameters output from the (i+1)th hidden layer to the ith layer from the network-side device, and correct the model parameters according to the backpropagation parameters; The terminal receives round number indication information sent by the network-side device, which indicates the second training round number corresponding to the backpropagation parameters; When the second training round number matches the terminal's desired round number, the step of correcting the model parameters according to the reverse transmission parameters is executed; The terminal sends the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number to the network-side device via the first RRC signaling, and receives the back propagation parameters and the round number indication information for indicating the second training round number sent by the network-side device via the second RRC signaling. The model training process also includes: The terminal receives a retransmission request from the network-side device for a first forward inference parameter indicating the desired number of training epochs for the network; and, based on the retransmission request, sends the first forward inference parameter indicating the desired number of training epochs for the network to the network-side device.

18. A terminal, characterized in that, Includes transceivers and processors, among which, The transceiver is used to receive initial configuration information of a neural network model sent by a network-side device. The neural network model includes a first model part located on the terminal side and a second model part located on the network-side device. The first model part includes an input layer and a hidden layer of layer 1-i. The second model part includes a hidden layer of layer i+1-n and an output layer. The processor is configured to, during the model training process of the neural network model, initiate the model training process under the current hyperparameter configuration, train the neural network model using training samples and their labels, and save the model parameters of the neural network model after training, wherein, during the model training process: Send the first forward inference parameters and the labels of the training samples to the network-side device. The first forward inference parameters are the intermediate parameters output from the i-th hidden layer to the (i+1)-th hidden layer in the model training process. The terminal sends round number indication information to the network-side device to indicate the first training round number corresponding to the first forward inference parameters and the label; Receive the backpropagation parameters output from the (i+1)th hidden layer to the ith layer from the network-side device, and correct the model parameters according to the backpropagation parameters; The terminal receives round number indication information sent by the network-side device, which indicates the second training round number corresponding to the backpropagation parameters; When the second training round number matches the terminal's desired round number, the step of correcting the model parameters according to the reverse transmission parameters is executed; The terminal sends the first forward inference parameters, the tag, and the round number indication information for indicating the first training round number to the network-side device via the first RRC signaling, and receives the back propagation parameters and the round number indication information for indicating the second training round number sent by the network-side device via the second RRC signaling. The model training process also includes: The terminal receives a retransmission request from the network-side device for a first forward inference parameter indicating the desired number of training epochs for the network; and, based on the retransmission request, sends the first forward inference parameter indicating the desired number of training epochs for the network to the network-side device.

19. A terminal, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method as described in any one of claims 8 to 13.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 13.

Images