Model training method, communication device, and medium

Abstract

The present disclosure relates to a model training method, a communication device, and a medium. The method executed by a first node comprises: when a first model is obtained by means of training by a first node, training the first model on the basis of first training data; and when the first model is obtained by means of training by a plurality of nodes including the first node, during the process of training the first model on the basis of the first training data, exchanging model iteration information of the first model with a second node. The present disclosure optimizes the model training process and improves the efficiency and accuracy of model training.

Description

Model training method, communication device and medium TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of communication, and particularly relates to a model training method, a communication device and a medium. BACKGROUND

[0002] Machine Learning (ML) is one of the most important implementation methods of Artificial Intelligence (AI) technology. Machine Learning can train a model according to a large amount of training data, and can use the trained model to predict events. In many fields, the model trained by using Machine Learning can obtain very accurate prediction results. In the field of communication technology, a model can also be applied to predict events.

[0003] SUMMARY

[0004] Embodiments of the present disclosure provide a model training method, a communication device and a medium.

[0005] According to a first aspect of embodiments of the present disclosure, a model training method is provided, which is performed by a first node, and the method comprises: in a case where a first model is trained by the first node, training the first model according to first training data; in a case where the first model is trained by a plurality of nodes including the first node, during training of the first model according to the first training data, interacting model iteration information of the first model with a second node.

[0006] According to a second aspect of embodiments of the present disclosure, a communication device is provided, comprising: a processing module configured to, in a case where a first model is trained by the first node, train the first model according to first training data; in a case where the first model is trained by a plurality of nodes including the first node, during training of the first model according to the first training data, interact model iteration information of the first model with a second node.

[0007] According to a third aspect of embodiments of the present disclosure, a communication device is provided, comprising: one or more processors; a memory coupled to the processors, the memory having stored thereon executable instructions that, as a result of execution by the processors, cause the model training method of the first aspect to be performed.

[0008] According to a fourth aspect of embodiments of the present disclosure, a communication system is provided, comprising a terminal and a network device, wherein the terminal is configured to implement the model training method of the first aspect, and / or the network device is configured to implement the model training method of the first aspect.

[0009] According to a fifth aspect of the embodiments of the present disclosure, a storage medium is provided, which stores instructions, when the instructions are executed on a communication device, causing the communication device to perform the model training method of the first aspect.

[0010] According to a sixth aspect of the embodiments of the present disclosure, a computer program product is provided, which includes a computer program and / or instructions, when the computer program and / or instructions are executed by a communication device, implementing the model training method of the first aspect.

[0011] The above technical solutions can achieve at least the following beneficial technical effects:

[0012] By intelligently controlling the interaction of model iteration information, the model training process is optimized, and the model training efficiency and accuracy are improved. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the following describes the drawings required for the embodiments, and the following drawings are only some embodiments of the present disclosure, and do not specifically limit the protection scope of the present disclosure.

[0014] FIG. 1 is an exemplary schematic diagram of an architecture of a communication system according to an embodiment of the present disclosure.

[0015] FIG. 2A is an interaction schematic diagram of a model training method according to an embodiment of the present disclosure.

[0016] FIG. 2B is an interaction schematic diagram of a model training method according to an embodiment of the present disclosure.

[0017] FIG. 2C is an interaction schematic diagram of a model training method according to an embodiment of the present disclosure.

[0018] FIG. 2D is an interaction schematic diagram of a model training method according to an embodiment of the present disclosure.

[0019] FIG. 2E is an interaction schematic diagram of a model training method according to an embodiment of the present disclosure.

[0020] FIG. 2F is an interaction schematic diagram of a model training method according to an embodiment of the present disclosure.

[0021] FIG. 3A is a flow schematic diagram of a model training method according to an embodiment of the present disclosure.

[0022] FIG. 3B is a flow schematic diagram of a model training method according to an embodiment of the present disclosure.

[0023] FIG. 3C is a flow schematic diagram of a model training method according to an embodiment of the present disclosure.

[0024] FIG. 3D is a flow diagram of a model training method according to an embodiment of the present disclosure.

[0025] FIG. 3E is a flow diagram of a model training method according to an embodiment of the present disclosure.

[0026] FIG. 3F is a flow diagram of a model training method according to an embodiment of the present disclosure.

[0027] FIG. 4 is a schematic diagram of a bilateral model according to an embodiment of the present disclosure.

[0028] FIG. 5 is a schematic diagram of a structure of a first node according to an embodiment of the present disclosure.

[0029] FIG. 6A is a schematic diagram of a structure of a communication device according to an embodiment of the present disclosure.

[0030] FIG. 6B is a schematic diagram of a structure of a chip according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0031] The embodiments of the present disclosure provide a model training method, a communication device and a medium.

[0032] In a first aspect, the embodiments of the present disclosure provide a model training method, executed by a first node, comprising: training a first model according to first training data, in a case that the first model is trained by the first node; and in a case that the first model is trained by a plurality of nodes including the first node, interacting, with a second node, model iteration information of the first model in a process of training the first model according to the first training data.

[0033] Optionally, the model iteration information is process quantity information generated in a process of training the first model.

[0034] In the process of training the first model according to the first training data, it is determined whether to interact, with the second node, model iteration information of the first model according to a model training manner, the model iteration information being information of a process quantity generated in the process of training the first model, and the model iteration information including at least one of the following: gradient information; model parameter iteration information.

[0035] In the above embodiments, the first node can determine whether to interact with the second node to share model iteration information according to a model training manner during training of the first model, the model iteration information including gradient information and / or model parameter iteration information, the model iteration information being an important process quantity generated in the model training process and being crucial to optimization, adjustment and convergence of the model. In this way, the disclosure can effectively manage and utilize the process quantity generated in the model training process, thereby improving the efficiency and accuracy of model training, accelerating the model convergence speed, and reducing resource consumption and time cost in the training process.

[0036] In some embodiments of the first aspect, the method further includes collecting the first training data, wherein the first training data includes at least one of the following: the first data sent by the second node; and the second data obtained by measuring the first signal sent by the second node.

[0037] In the above embodiments, the second node can assist the first node to quickly obtain high-quality first training data. For example, the first node receives the first data sent by the second node, and / or the first node measures the first signal sent by the second node to obtain the second data, and uses the first data and / or the second data as the first training data. Training the first model using the first training data obtained in this way can enhance the ability of the first model to adapt to actual application scenarios.

[0038] In some embodiments of the first aspect, before collecting the first training data, the method includes sending a first message to the second node, the first message including at least one of the following:

[0039] a first request for requesting the second node to allow the first node to train the first model online;

[0040] a second request for requesting the second node to send the first data and / or the first signal to the first node;

[0041] performance information of the first model;

[0042] model training capability of the first node.

[0043] In the above embodiments, before collecting the first training data, the first node can send the first message to the second node to achieve the purpose of sending at least one of the first request, the second request, the performance information of the first model, and the model training capability of the first node to the second node.

[0044] In some embodiments combined with the first aspect, in some embodiments, the method further comprises: receiving a second message sent by the second node, the second message comprising at least one of:

[0045] a first indication, the first indication being used to indicate that the first node is online training the first model and / or initial model parameters of the first model;

[0046] a second indication, the second indication being used to indicate that the first node is collecting the first training data.

[0047] In the above embodiments, the first node can receive the second message fed back by the second node for the first message, and the second message comprises the first indication and / or the second indication.

[0048] In the above embodiments, by interacting the first message and / or the second message between the first node and the second node, the way that the first node actively requests online training of the first model and the way that the first node passively online trains the first model according to the indication of the second node can be realized. For example, the first indication can be a response to the first request, which can realize the way that the first node actively requests the second node to online train the first model on the first node. For example, the first indication can be decided according to the performance information of the first model and / or the model training capability of the first node, which can realize the way that the first node passively online trains the first model on the first node according to the indication of the second node.

[0049] In some embodiments combined with the first aspect, in some embodiments, the collecting the first training data comprises: collecting the first training data at multiple time instants.

[0050] In the above embodiments, training the first model according to the first training data collected at multiple time instants can increase the relevance of the first model in the time domain.

[0051] In some embodiments combined with the first aspect, in some embodiments, the method further comprises: detecting that a first condition is met, sending a third message to the second node, the third message comprising at least one of:

[0052] a third indication, the third indication being used to indicate that the trained first model meets a performance requirement;

[0053] a fourth indication, the fourth indication being used to indicate that the second node stops sending the first data and / or the first signal;

[0054] a fifth indication, the fifth indication being used to indicate that the trained first model does not meet the performance requirement and / or the reason why the trained first model does not meet the performance requirement.

[0055] In the above embodiment, in the training process of the first model, the third message can be sent to the second node when it is detected that the first condition is met, and the third message can include at least one of the third indication, the fourth indication, and the fifth indication, which achieves the purpose of timely reporting the first model training situation of the first node to the second node.

[0056] In some embodiments of the first aspect, the method further includes: receiving a fourth message sent by the second node, the fourth message being sent by the second node according to the fifth indication; and performing a first operation based on the fourth message, the first operation including at least one of:

[0057] switching the first model to a second model, the second model having the same function as the first model;

[0058] replacing the first model with a first algorithm, the first algorithm having the same function as the first model;

[0059] increasing the number of iterations of model training and continuing to train the first model;

[0060] increasing the length of time of model training and continuing to train the first model;

[0061] continuing to obtain the first training data and training the first model;

[0062] giving up training the first model.

[0063] In the above embodiment, the first node can receive the fourth message sent by the second node for the fifth indication, so as to cope with the situation that the trained first model does not meet the performance requirement, and improve the coping ability of the first node.

[0064] In some embodiments of the first aspect, before collecting the first training data, the method includes: receiving a fifth message sent by the second node, the fifth message including at least one of:

[0065] a third request for requesting the first node to train the first model online;

[0066] performance information of the first model;

[0067] initial model parameters of the first model.

[0068] In the above embodiment, the first node can receive the fifth message sent by the second node before collecting the first training data, and the fifth message can include at least one of the third request, the performance information of the first model, and the initial model parameters of the first model.

[0069] In some embodiments combined with the first aspect, in some embodiments, the method further comprises: sending, to the second node, a sixth message, the sixth message comprising at least one of:

[0070] a sixth indication for instructing the first node to online train the first model;

[0071] a seventh indication for instructing the second node to send the first data and / or the first signal to the first node.

[0072] In the above embodiments, the first node can send, to the second node, the sixth message, the sixth message comprising the sixth indication and / or the seventh indication.

[0073] In some embodiments, by interacting the fifth message and / or the sixth message between the first node and the second node, the manner that the first node actively online trains the first model and the manner that the first node passively online trains the first model according to the request of the second node can be realized. For example, the sixth indication can be a response to the third request, thereby realizing the manner that the first node passively online trains the first model according to the request of the second node. For example, the sixth indication can be an active decision, thereby the manner that the first node actively online trains the first model can be realized.

[0074] In some embodiments combined with the first aspect, in some embodiments, the method further comprises: detecting that a first condition is satisfied, sending, to the second node, a seventh message, the seventh message comprising at least one of:

[0075] an eighth indication for instructing the second node to stop sending the first data and / or the first signal;

[0076] a ninth indication for informing the second node that the first model training fails and / or the first node will perform a second operation;

[0077] a tenth indication for instructing the second node to perform the second operation;

[0078] wherein the second operation comprises at least one of:

[0079] switching the first model to a second model, a function of the second model being same as a function of the first model;

[0080] replacing the first model with a first algorithm, a function of the first algorithm being same as the function of the first model.

[0081] In the above embodiments, in the process of training the first model, if the first condition is detected, the first node can send a seventh message to the second node to inform the training of the first model and the operation to be performed by the first node and / or the second node.

[0082] With reference to some embodiments of the first aspect, in some embodiments, the method further includes: the application node of the first model sending, to the second node, at least one of the trained first model, parameter information of the trained first model, and update information of the trained first model, if the trained first model meets the performance requirement.

[0083] In the above embodiments, if the application node of the first model trained online by the first node is the second node, the second node can be sent the first model trained to meet the performance requirement.

[0084] With reference to some embodiments of the first aspect, in some embodiments, the first model includes a first partial model, and the application node of the first partial model is the second node, and the method further includes: sending, to the second node, at least one of the first partial model in the first model that meets the performance requirement, update information of the first partial model, and update information of the first partial model, if the trained first model meets the performance requirement.

[0085] In the above embodiments, if the first model trained online by the first node includes a first partial model whose application node is the second node, the second node can be sent the first partial model in the first model trained to meet the performance requirement.

[0086] With reference to some embodiments of the first aspect, in some embodiments, the first node is a terminal, and the second node is a network device.

[0087] In the above embodiments, it is specified that the first node is a terminal, and the second node is a network device.

[0088] With reference to some embodiments of the first aspect, in some embodiments, the first node is a network device, and the second node is a terminal.

[0089] In the above embodiments, it is specified that the first node is a network device, and the second node is a terminal.

[0090] With reference to some embodiments of the first aspect, in some embodiments, if the first model is trained by the first node, the first model includes at least one of:

[0091] a one-sided model;

[0092] The multi-edge model includes a plurality of partial models, and one partial model corresponds to one application node.

[0093] The at least one partial model allows independent training.

[0094] In the above embodiment, the model capable of completing the entire training process on a single entity includes at least one of a single-edge model, a double-edge model, and at least one partial model. For these models, a single-entity training method can be used, and in the training process, model iteration information can not be exchanged, thereby improving the efficiency and accuracy of model training, while reducing the interaction resource consumption and interaction time cost in the training process and accelerating the model convergence speed.

[0095] In combination with some embodiments of the first aspect, in some embodiments, when the first model is obtained by training a plurality of nodes including the first node, the first model is at least one partial model in a multi-edge model, the multi-edge model includes a plurality of partial models, and one partial model corresponds to one application node; the third model is included on the second node, and the third model is at least one partial model in the multi-edge model except the first model.

[0096] In the above embodiment, when the first model on the first node is at least one partial model in a multi-edge model, the third model is included on the second node, and the first model and the third model are components of a double-edge model, a multi-entity cooperative training method can be used to train the first model, thereby accelerating the model training speed and improving the model robustness.

[0097] In combination with some embodiments of the first aspect, in some embodiments, the model iteration information includes gradient information, the gradient information includes forward gradient information and / or backward gradient information, and the interaction of the first model with the second node includes at least one of the following:

[0098] The forward gradient information output by the first model is sent to the second node.

[0099] The backward gradient information output by the third model is received from the second node.

[0100] The backward gradient information output by the first model is sent to the second node.

[0101] The forward gradient information output by the third model is received from the second node.

[0102] In the above embodiment, the specific implementation of the model iteration information of the first model exchanged between the first node and the second node is specified.

[0103] In some embodiments of the first aspect, in the process of training the first model according to the first training data, the model parameters of the first model and the model parameters of the third model are both updated, or the model parameters of the first model are updated and the model parameters of the third model are not updated.

[0104] In the above embodiment, the model parameters of the third model can be updated as the model parameters of the first model are updated, or the model parameters of the third model are fixed in the process of training the first model by multiple entities. This can improve the training effect of the first model and accelerate the convergence speed of the first model.

[0105] In some embodiments of the first aspect, in the case where the first model is trained by multiple nodes including the first node, the first model is at least one of:

[0106] a single-sided model;

[0107] a multi-sided model including multiple partial models, one partial model corresponding to one application node;

[0108] at least one partial model allowing independent training.

[0109] In the above embodiment, the first model capable of being trained by multiple entities includes at least one of a single-sided model, a double-sided model, and at least one partial model. For these models, model iteration information can be exchanged during training, thereby accelerating model convergence and improving the generalization of the first model on different nodes and different data sets.

[0110] In some embodiments of the first aspect, the first node is a training node, and the second node is a center node. The first node exchanges the model iteration information of the first model with the second node, including: sending first model iteration information to the second node under the condition that a second condition is met, the first model iteration information including first gradient information and / or first model parameter iteration information, the first gradient information being output by the first model, and the first model parameter information being obtained after the first node trains the first model; and receiving second model iteration information sent by the second node, the second model iteration information being determined by the second node according to the first gradient information and / or the first model parameter iteration information uploaded by all training nodes.

[0111] In some embodiments combined with the first aspect, in some embodiments, the method further comprises: updating the first model according to the second model parameter iteration information; and training the updated first model according to the first training data.

[0112] In the above embodiments, the first node as the training node sends the first gradient information and / or the first model parameter iteration information to the second node when the second condition is met, receives the second model iteration information sent by the second node, updates the first model according to the second model iteration information, and continues to train the updated first model according to the first training data. In this way, the distributed training of the first model can be implemented, the training intensity of each training node can be reduced, the model training speed can be accelerated, the training effect of the first model can still be guaranteed, and the generalization of the first model on different nodes and different data sets can be improved.

[0113] In some embodiments combined with the first aspect, in some embodiments, the first node is a node selected by the second node and having the capability of multi-entity joint training.

[0114] In the above embodiments, the node selected by the first node for the second node and having the capability of multi-entity joint training is specified.

[0115] In some embodiments combined with the first aspect, in some embodiments, the first condition comprises at least one of the following:

[0116] The first model meets a performance requirement;

[0117] The data amount of the first training data reaches a first threshold;

[0118] The number of iterations of the training model reaches a second threshold;

[0119] The time length of the training model reaches a third threshold.

[0120] In the above embodiments, the types of the first condition are specified to adapt to different scene requirements.

[0121] In some embodiments combined with the first aspect, in some embodiments, the second condition comprises at least one of the following:

[0122] The first model iteration information including the first gradient information and / or the first model parameter iteration information is reported at a time point;

[0123] It is detected that the error of the first model is in a first range;

[0124] The number of local iterations of the first model reaches a fourth threshold;

[0125] The first node is unable to continue training the first model.

[0126] In the above embodiments, multiple categories of the second condition are specified to adapt to different scene requirements.

[0127] In a second aspect, the embodiments of the present disclosure provide a communication device, which comprises at least one of a transceiver module and a processing module; wherein the communication device is configured to perform the optional implementation manners of the first aspect.

[0128] In a third aspect, the embodiments of the present disclosure provide a communication device, which comprises one or more processors; wherein the communication device is configured to perform the optional implementation manners of the first aspect.

[0129] In a fourth aspect, the embodiments of the present disclosure provide a communication system, which comprises a terminal and a network device; wherein the terminal is configured to perform the method described in the optional implementation manners of the first aspect, and the network device is configured to perform the method described in the optional implementation manners of the first aspect.

[0130] In a fifth aspect, the embodiments of the present disclosure provide a storage medium, which stores instructions, when the instructions run on a communication device, cause the communication device to perform the method described in the optional implementation manners of the first aspect.

[0131] In a sixth aspect, the embodiments of the present disclosure provide a program product, which, when executed by a communication device, causes the communication device to perform the method described in the optional implementation manners of the first aspect.

[0132] In a seventh aspect, the embodiments of the present disclosure provide a computer program, which, when running on a computer, causes the computer to perform the method described in the optional implementation manners of the first aspect.

[0133] In an eighth aspect, the embodiments of the present disclosure provide a chip or chip system. The chip or chip system comprises processing circuitry configured to perform the method described in the optional implementation manners of the first aspect.

[0134] It can be understood that the first node, the communication device, the communication system, the storage medium, the program product, the computer program, the chip or the chip system are all used to perform the method proposed in the embodiments of the present disclosure. Therefore, the beneficial effects they can achieve can refer to the beneficial effects in the corresponding method, which will not be repeated here.

[0135] The model training method, the information processing method, the communication method, and the like can be replaced with each other, the model training apparatus, the information processing apparatus, the communication apparatus, and the like can be replaced with each other, and the communication system, the information processing system, the model training system, and the like can be replaced with each other.

[0136] The embodiments of the present disclosure are not exhaustive, but only illustrate some embodiments, and are not specific limitations on the protection scope of the present disclosure. In the case of no contradiction, each step in an embodiment can be implemented as an independent embodiment, and the steps can be combined arbitrarily, for example, the scheme after removing some steps in an embodiment can also be implemented as an independent embodiment, and the order of the steps in an embodiment can be exchanged arbitrarily, in addition, the optional implementation manners in an embodiment can be combined arbitrarily; in addition, the embodiments can be combined arbitrarily, for example, some or all steps of different embodiments can be combined arbitrarily, an embodiment can be combined with optional implementation manners of other embodiments.

[0137] In the embodiments of the present disclosure, the terms and / or descriptions between the embodiments are consistent and can be referred to each other if there is no special description and logical conflict, and the technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0138] The terms used in the embodiments of the present disclosure are only for the purpose of describing the specific embodiments, and not as a limitation on the present disclosure.

[0139] In the embodiments of the present disclosure, unless otherwise specified, the elements expressed in singular form, such as “one”, “a”, “the”, “above”, “said”, “preceding”, “this”, and the like, can represent “one and only one”, or “one or more”, “at least one”, and the like. For example, in the case of using articles such as “a”, “an”, “the” in English, the noun after the article can be understood as singular expression, or as plural expression.

[0140] In the embodiments of the present disclosure, “a plurality of” means two or more.

[0141] In some embodiments, the terms “at least one of”, “one or more”, “a plurality of”, “multiple”, and the like can be replaced with each other.

[0142] In some embodiments, "at least one of A, B", "A and / or B", "in one case A, in another case B", "responsive to case A, responsive to case B" and the like, can be interpreted to include both cases, A and B, in some embodiments, A (A is performed regardless of B), in some embodiments, B (B is performed regardless of A), in some embodiments, selected from the group consisting of A and B (the selection between A and B is an option), in some embodiments, A and B (both A and B are performed).

[0143] In some embodiments, "A or B" and the like, can be interpreted to include both cases, A and B, in some embodiments, A (A is performed regardless of B), in some embodiments, B (B is performed regardless of A), in some embodiments, selected from the group consisting of A and B (the selection between A and B is an option).

[0144] In some embodiments, the prefix words "first", "second", and the like, are used only to distinguish different descriptive objects, and do not limit the position, order, priority, quantity, or content of the descriptive objects, and the description of the descriptive objects is referred to the description in the claims or embodiments, and should not be construed as redundant limitations. For example, the descriptive object is "field", and the ordinal words before "field" in "first field" and "second field" do not limit the position or order between "fields", and "first" and "second" do not limit whether the "fields" modified by them are in the same message or not, nor limit the order of "first field" and "second field". For another example, the descriptive object is "level", and the ordinal words before "level" in "first level" and "second level" do not limit the priority between "levels". For another example, the quantity of the descriptive object is not limited by the ordinal words, and can be one or more, for example, "first device", where the quantity of "device" can be one or more. In addition, the objects modified by different prefix words can be the same or different, for example, the descriptive object is "device", and "first device" and "second device" can be the same device or different devices, and their types can be the same or different; for another example, the descriptive object is "information", and "first information" and "second information" can be the same information or different information, and their contents can be the same or different.

[0145] In some embodiments, "including A", "containing A", "for indicating A", "carrying A", can be interpreted as directly carrying A, or indirectly indicating A.

[0146] In some embodiments, the terms "in response to", "in response to determining", "in the case of", "when", "when", "if", "if" and the like can be replaced with each other.

[0147] In some embodiments, the terms "greater than", "greater than or equal to", "not less than", "more than", "more than or equal to", "not less than", "higher than", "higher than or equal to", "not lower than", "above", and the like can be replaced with each other, and the terms "less than", "less than or equal to", "not greater than", "fewer than", "fewer than or equal to", "not more than", "lower than", "lower than or equal to", "not higher than", "below", and the like can be replaced with each other.

[0148] In some embodiments, an apparatus and the like can be interpreted as an entity, and can also be interpreted as virtual, and the name thereof is not limited to the name recited in the embodiments, and the terms "apparatus", "equipment", "device", "circuit", "network element", "node", "function", "unit", "section", "system", "network", "chip", "chip system", "entity", "subject", and the like can be replaced with each other.

[0149] In some embodiments, "network" can be interpreted as an apparatus (for example, an access network device, a core network device, and the like) included in the network.

[0150] In some embodiments, the terms “access network device (AN device),” “radio access network device (RAN device),” “base station (BS),” “radio base station,” “fixed station,” “node,” “access point,” “transmission point (TP),” “reception point (RP),” “transmission / reception point (TRP),” “panel,” “antenna panel,” “antenna array,” “cell,” “macro cell,” “small cell,” “femto cell,” “pico cell,” “sector,” “cell group,” “serving cell,” “carrier,” “component carrier,” “bandwidth part (BWP),” and the like can be used interchangeably.

[0151] In some embodiments, the terms "terminal," "terminal device," "user equipment (UE)," "user terminal," "mobile station (MS)," "mobile terminal (MT)," "subscriber station," "mobile unit," "subscriber unit," "wireless unit," "remote unit," "mobile device," "wireless device," "wireless communication device," "remote device," "mobile subscriber station," "access terminal," "mobile terminal," "wireless terminal," "remote terminal," "handset," "user agent," "mobile client," "client," and so on can be replaced with each other.

[0152] In some embodiments, the access network device, the core network device, or the network device can be replaced with a terminal. For example, the embodiments of the present disclosure can also be applied to a structure in which communication between the access network device, the core network device, or the network device and the terminal is replaced with communication between a plurality of terminals (e.g., device-to-device (D2D), vehicle-to-everything (V2X), etc.). In this case, the terminal can also be configured to have all or part of the functions of the access network device. In addition, the terms "uplink," "downlink," and the like can also be replaced with terms corresponding to the inter-terminal communication (e.g., "side"). For example, the uplink channel, the downlink channel, and the like can be replaced with the side channel, and the uplink, the downlink, and the like can be replaced with the sidelink.

[0153] In some embodiments, the terminal can be replaced with the access network device, the core network device, or the network device. In this case, the access network device, the core network device, or the network device can also be configured to have all or part of the functions of the terminal.

[0154] In some embodiments, obtaining data, information, and the like can comply with laws and regulations of the country where the location is situated.

[0155] In some embodiments, data, information, and the like can be obtained after obtaining consent of a user.

[0156] In addition, each element, each row, or each column in the table of the embodiments of the present disclosure can be implemented as an independent embodiment, and any combination of any element, any row, or any column can also be implemented as an independent embodiment.

[0157] FIG. 1 is a schematic diagram of an architecture of a communication system according to an embodiment of the present disclosure. As shown in FIG. 1, the communication system 100 can include a terminal 101 and a network device 102.

[0158] Optionally, the terminal 101 is a first node, and the network device 102 is a second node.

[0159] Optionally, the terminal 101 is a second node, and the network device 102 is a first node.

[0160] In some embodiments, the terminal 101 includes at least one of a mobile phone, a wearable device, an Internet of Things device, a communication-capable automobile, a smart automobile, a tablet computer (Pad), a wireless transceiver-enabled computer, a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, a wireless terminal device in industrial control, a wireless terminal device in self-driving, a wireless terminal device in remote medical surgery, a wireless terminal device in a smart grid, a wireless terminal device in transportation safety, a wireless terminal device in a smart city, a wireless terminal device in a smart home, and the like, but is not limited thereto.

[0161] In some embodiments, the network device 102 can include at least one of an access network device and a core network device.

[0162] Optionally, the network device 102 is an access network device. Optionally, the access network device is at least one of a node or device that accesses a terminal to a wireless network, and can include at least one of an evolved NodeB (eNB) in a 5G communication system, a next generation eNB (ng-eNB), a next generation NodeB (gNB), a node B (NB), a home node B (HNB), a home evolved node B (HeNB), a wireless backhaul device, a radio network controller (RNC), a base station controller (BSC), a base transceiver station (BTS), a base band unit (BBU), a mobile switching center, a base station in a 6G communication system, an Open RAN, a Cloud RAN, a base station in other communication systems, an access node in a Wi-Fi system, but is not limited thereto.

[0163] In some embodiments, the network device 102 is a base station. Optionally, the base station is at least one of a macro base station, a micro base station (also referred to as a small station), a relay station, an access point, a 5G base station or a future base station, a satellite, a Transmitting and Receiving Point (TRP), a Transmitting Point (TP), a mobile switching center, or other devices that perform a base station function in a communication system, etc., and the embodiments of the present disclosure are not limited thereto. For convenience of description, in all embodiments of the present disclosure, devices that provide a wireless communication function for a terminal device are collectively referred to as network devices or base stations.

[0164] In some embodiments, the network device 102 is a core network device. Optionally, the core network device can be one device including all or part of a first network element, a second network element, etc., or can be a plurality of devices or device groups including all or part of the first network element, the second network element, etc. The network element can be virtual or physical. The core network includes at least one of an Evolved Packet Core (EPC), a 5G Core Network (5GCN), a Next Generation Core (NGC), etc.

[0165] In some embodiments, the technical solutions of the present disclosure can be applied to an Open RAN architecture, at this time, the interfaces between or within the access network devices involved in the embodiments of the present disclosure can become internal interfaces of the Open RAN, and the processes and information interactions between these internal interfaces can be implemented through software or programs.

[0166] In some embodiments, the access network device can be composed of a central unit (CU) and a distributed unit (DU), where the CU can also be referred to as a control unit. The CU-DU structure can split the protocol layers of the access network device, and some of the protocol layers are controlled by the CU, and the rest or all of the protocol layers are distributed in the DU and controlled by the CU. However, the present disclosure is not limited thereto.

[0167] It can be understood that the communication system described in the embodiments of the present disclosure is for more clearly illustrating the technical solutions of the embodiments of the present disclosure, and does not constitute a limitation on the technical solutions proposed by the embodiments of the present disclosure. Those skilled in the art can know that, as the system architecture evolves and new business scenarios appear, the technical solutions proposed by the embodiments of the present disclosure are also applicable to similar technical problems.

[0168] The following embodiments of the present disclosure can be applied to the communication system 100 shown in FIG. 1 or part of the subject, but are not limited thereto. The subjects shown in FIG. 1 are exemplary, and the communication system can include all or part of the subjects in FIG. 1, or other subjects other than FIG. 1. The number and form of each subject is arbitrary, each subject can be physical or virtual, the connection relationship between each subject is exemplary, each subject can not be connected or can be connected, the connection can be in any way, can be direct connection or indirect connection, can be wired connection or wireless connection.

[0169] Embodiments of the present disclosure can be applied to Long Term Evolution (LTE), LTE-Advanced (LTE-A), LTE-Beyond (LTE-B), SUPER 3G, IMT-Advanced, 4th generation mobile communication system (4G), 5th generation mobile communication system (5G), 5G new radio (NR), Future Radio Access (FRA), New-Radio Access Technology (RAT), New Radio (NR), New radio access (NX), Future generation radio access (FX), Global System for Mobile communications (GSM (registered trademark)), CDMA2000, Ultra Mobile Broadband (UMB), IEEE 802.11 (Wi-Fi (registered trademark)), IEEE 802.16 (WiMAX (registered trademark)), IEEE 802.20, Ultra-WideBand (UWB), Bluetooth (Bluetooth (registered trademark)), Public Land Mobile Network (PLMN) network, Device-to-Device (D2D) system, Machine to Machine (M2M) system, Internet of Things (IoT) system, Vehicle-to-Everything (V2X), system using other communication methods, next-generation system expanded based on them, and the like. In addition, a plurality of systems can be combined (for example, combination of LTE or LTE-A and 5G, and the like).

[0170] In some embodiments, as artificial intelligence (AI / ML) technology rapidly develops, there is an increasing demand for using AI / ML models for various tasks, such as using AI / ML models for image recognition, speech recognition, natural language processing, replacing traditional wireless communication algorithms, and the like. The AI / ML model is obtained through training of the AI / ML model.

[0171] In some embodiments, AI / ML model training refers to learning the relationship between input and output through data-driven learning to obtain an AI / ML model for inference. AI / ML model training is further divided into offline training and online training.

[0172] In some embodiments, offline training and online training are defined as follows:

[0173] AI / ML model offline training: it is an AI / ML training process, in which the model is trained based on a collected dataset, and after the training is completed, the model is used for inference or delivered for use.

[0174] AI / ML model online training: it is a process of AI / ML model training, in which the model for inference is usually trained (nearly) in real time as new training samples arrive (usually continuously).

[0175] In some embodiments, the above definitions of online training and offline training are for reference only. There may be cases that do not fully meet this definition, but can still be classified as online training according to generally accepted conventions. If a consensus is reached, even if some cases do not meet the definition, they can still be considered offline training.

[0176] In some embodiments, according to the node (i.e. entity) division required for AI / ML model inference, AI / ML model includes terminal side (UE) model, network side (NW) side model and bilateral model, and the definitions of these models are as follows:

[0177] Single-sided model, the function of single-sided model is implemented on one node or entity. Single-sided model includes but is not limited to UE side model and NW side model.

[0178] UE side model: AI / ML model applied on UE device to perform prediction or inference.

[0179] NW side model: AI / ML model applied on NW device to perform prediction or inference.

[0180] Bilateral model: a pair of models for joint inference, in which joint inference is AI / ML model inference jointly performed on terminal user device and network device. That is, UE performs inference or prediction of the first part of the model, and network device performs inference or prediction of the second part of the model, and vice versa. In short, one part of the bilateral model is applied on entity A, and the other part is applied on entity B, and the function of the bilateral model needs to be realized by entity A and B together.

[0181] Multi-sided model, that is, the model includes multiple parts, one part is applied on one entity, and the function of the multi-sided model needs to be completed by multiple entities together.

[0182] In some embodiments, for the above-mentioned several models, the node that trains the model can not be the node that performs inference. For example, for the UE-side model, the model can be trained by the NW-side device in an offline or online manner, and then the trained model is passed to the UE, and the UE uses the trained model to perform inference or prediction. The UE-side model or the NW-side model described above can complete the entire model training on one node. For the bilateral model, it can be trained on one node or jointly trained on two nodes. In order to improve the efficiency of the learning process and shorten the model training time, or to solve the problem that some complex models cannot be trained on one node, distributed training can also be used. It is a technology and method of dispersing computing tasks to multiple computing nodes for learning. The core idea is to distribute data and computing load to different machines or computing units. Distributed learning is usually used to handle large-scale data sets and complex models. Distributed training can be divided into the following types:

[0183] The first type is data parallelism. In this method, data is divided into multiple subsets, and each subset is processed on a different computing node. Each node has a copy of the same model and only processes different data subsets. The calculation results (such as gradients, parameter updates) are aggregated and averaged among nodes to update the model parameters.

[0184] The first type is model parallelism. In model parallelism, the model is divided into different parts, and each part is run on a different computing node. This method is suitable for very large models, as a single computing node can not be able to accommodate the entire model. Frequent communication between nodes is required to ensure coordination and synchronization of the model parts.

[0185] The first type is hybrid parallelism. This method combines the advantages of data parallelism and model parallelism. It is used in scenarios that require handling large-scale data and complex models simultaneously. Data and models are distributed and parallel computed on different nodes.

[0186] In some embodiments, in order to protect the data privacy of users or reduce data transmission between nodes, the present disclosure also proposes a model training of federated learning. It is a distributed training method. Unlike traditional centralized learning methods, federated learning allows training models on local devices where data resides, rather than centralizing data on servers for processing. The basic principles of federated learning include:

[0187] Local training: data is kept on local devices (e.g., smartphones, sensors, etc.), and each device independently trains a local copy of the model. The goal of local training is to update and optimize the model using data on the device.

[0188] Model update: each device sends updates (e.g., gradients or weights) of the local model to the central server instead of sending raw data. The central server aggregates the model updates from individual devices and computes the average or weighted average of the global model.

[0189] Global model: the central server sends the updated global model back to individual devices. The devices further train with the global model for local training, and this process is repeated until the model converges.

[0190] In some embodiments, unlike offline training, online training needs to consider the impact of online training data collection, training duration, etc. In addition, for different AI / ML models, the workflow of online training may also be different. In particular, for the case of applying a model to one or more nodes, how to implement online training of the model through one or more nodes to obtain a trained model is a problem to be solved. Currently, there is no workflow or method for online training of different models. Therefore, the embodiments of the present disclosure propose a model training method, a communication device and a medium, which intelligently control the interaction of model iteration information according to different training methods corresponding to different models, thereby optimizing the model training process and improving the model training efficiency and accuracy.

[0191] FIG. 2A is an interaction diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 2A, the embodiments of the present disclosure relate to a model training method, and in particular to a model training method for training a first model through a single entity training method, which is performed by a communication system 100. In the embodiment shown in FIG. 2A, a terminal 101 is a first node, and a network device 102 is a second node. The first node can be referred to as a model training node, i.e., a node that performs a model training process, which includes steps of inputting training data into a model, determining model parameter update information, and updating model parameters. The model training method shown in FIG. 2A can include the following steps:

[0192] In step S2101, the terminal 101 sends a first message to the network device 102.

[0193] In some embodiments, the network device receives the first message. For example, the network device receives the first message sent by the terminal.

[0194] In some embodiments, the terminal sends the first message to the network device, and the first message includes at least one of the following contents:

[0195] The first request is used to request the network device to allow the terminal to train the first model online.

[0196] The second request is used to request the network device to send the terminal the first data and / or the first signal, wherein the first data and / or the first signal are used for the terminal to collect the first training data, and the first training data are used for training the first model.

[0197] The performance information of the first model.

[0198] The model training capability of the terminal.

[0199] In some embodiments, one or more of the first request, the second request, the performance information of the first model, and the model training capability of the terminal can be sent simultaneously or separately, which is not limited in the disclosure.

[0200] In step S2102, the network device 102 sends a second message to the terminal 101.

[0201] In some embodiments, the terminal receives the second message. For example, the terminal receives the second message sent by the network device.

[0202] In some embodiments, the network device sends the terminal the second message, and the second message includes at least one of the following:

[0203] The first indication is used to indicate the terminal to train the first model online and / or to indicate the terminal the initial model parameters of the first model.

[0204] The second indication is used to indicate the terminal to collect the first training data.

[0205] In some embodiments, one or more of the first indication and the second indication can be sent simultaneously or separately, which is not limited in the disclosure.

[0206] For example, assuming that in step S2101, the first message sent by the terminal to the network device includes the first request, then in step S2102, the second message sent by the network device to the terminal can include the first indication, which can be the response of the network device to the first request. This is a way for the terminal to actively request the network device to train the first model online.

[0207] For example, assuming that the first message sent by the terminal to the network device in step S2101 does not include the first request but includes the performance information of the first model, the second message sent by the network device to the terminal in step S2102 can include the first indication, which can be sent by the network device according to the performance information of the first model reported by the terminal. This is a way in which the terminal passively trains the first model online according to the indication of the network device.

[0208] For example, assuming that the first message sent by the terminal to the network device in step S2101 does not include the first request but includes the model training capability of the terminal, the second message sent by the network device to the terminal in step S2102 can include the first indication, which can be sent by the network device according to the model training capability of the terminal reported by the terminal. This is a way in which the terminal passively trains the first model online according to the indication of the network device.

[0209] For example, assuming that the first message sent by the terminal to the network device in step S2101 does not include the first request and does not include the performance information of the first model, the second message sent by the network device to the terminal in step S2102 can include the first indication, which can be sent by the network device to the terminal actively, for example, the network device actively sends the second message including the first indication to the terminal when detecting that the first model training time reaches a preset value. This is a way in which the terminal passively trains the first model online according to the indication of the network device.

[0210] For example, step S2101 can be omitted, and the second message sent by the network device to the terminal in step S2102 can include the first indication, which can be sent by the network device to the terminal actively, which is a way in which the terminal passively trains the first model online according to the indication of the network device.

[0211] It should be noted that in the above examples, the premise for the network device to send the first indication to the terminal is to confirm that the model training capability of the terminal can meet the requirement of training the first model. In the above embodiment, the first indication sent by the network device to the terminal indicates that the network device allows the terminal to train the first model online.

[0212] In some embodiments, assuming that the first message sent by the terminal to the network device in step S2101 includes the second request, the second message sent by the network device to the terminal in step S2102 can include the second indication, which can be a response of the network device to the second request.

[0213] In step S2103, the terminal 101 collects the first training data.

[0214] In some embodiments, the terminal starts collecting the first training data in response to receiving the second message sent by the network device.

[0215] In some embodiments, the terminal collects the first training data, wherein the first training data comprises at least one of:

[0216] the first data sent by the network device;

[0217] the second data obtained by measuring the first signal sent by the network device.

[0218] The first signal can be a pilot signal or other signal used by the terminal to obtain the second data.

[0219] In some embodiments, the network device can send the first data to the terminal. In some embodiments, the network device can send the first signal, and the terminal can measure the first signal to obtain the second data.

[0220] In some embodiments, in step S2102, the second indication in the second message can also be used to indicate a resource for sending the first signal, the resource comprising at least one of a time domain resource, a frequency domain resource, and a spatial domain resource. The implementation of the terminal collecting the first training data comprises measuring the first signal on the corresponding time domain, frequency domain, and spatial domain to obtain the second data.

[0221] In some embodiments, in step S2102, the second indication in the second message can also be used to indicate a data format of the first data. The implementation of the terminal collecting the first training data comprises receiving the first data in the data format.

[0222] In some embodiments, the terminal can collect the first training data at multiple time instants, and the first training data comprises the first data and / or the second data.

[0223] In some embodiments, the first training data comprises input samples and / or corresponding output samples of the first model.

[0224] In step S2104, the terminal 101 trains the first model according to the first training data.

[0225] In some embodiments, the first model is trained by the first node (terminal 101), and the first model is trained according to the first training data. In the case where the first model is trained by multiple nodes including the first node (terminal 101), the model iteration information of the first model is interacted with the second node (network device 102) in the process of training the first model according to the first training data.

[0226] In some embodiments, if the first model is trained by one first node, the first node can not interact the model iteration information of the first model with the second node. For example, assuming that the first node is the terminal 101 and the second node is the network device 102, if the first model is trained by the terminal 101, the terminal 101 can not interact the model iteration information of the first model with the network device 102.

[0227] In some embodiments, if the first model is trained by multiple nodes including the first node, the first node can interact the model iteration information of the first model with the second node. For example, assuming that the first node is the terminal 101 and the second node is the network device 102, if the first model is trained by the terminal 101 and the network device 102 together, the terminal 101 can interact the model iteration information of the first model with the network device 102. For example, assuming that the first node is the terminal 101 and the second node is the network device 102, if the first model is trained by the terminal 101 and one or more other terminals except the terminal 101 together, the terminal 101 can interact the model iteration information of the first model with the network device 102. For example, assuming that the first node is the terminal 101 and the second node is the network device 102, if the first model is trained by the terminal 101, the network device 102 and one or more other network devices except the network device 102 together, the terminal 101 can interact the model iteration information of the first model with the network device 102.

[0228] In some embodiments, in the process of training the first model according to the first training data, it is known by one or more nodes training the first model whether to interact the model iteration information of the first model with the network device by default.

[0229] In some embodiments, in the process of training the first model according to the first training data, the model training manner can be distinguished according to the number of nodes training the first model, and then it is known by default whether to interact the model iteration information of the first model with the network device according to the manner of training the first model.

[0230] In some embodiments, the model iteration information is the information of the process quantity generated in the process of training the first model, and the model iteration information can optionally include at least one of the following:

[0231] Gradient information;

[0232] Model parameter iteration information.

[0233] In some embodiments, if the terminal trains the first model in a single entity training manner, the terminal can not interact with the network device model iteration information of the first model. The single entity training manner refers to that the model training process is completed by one node, and the model training process includes the steps of inputting training data into the model, determining model parameter update information, and updating model parameters.

[0234] In some embodiments, the first model applicable to the single entity training manner includes at least one of:

[0235] a single edge model;

[0236] a multi-edge model including a plurality of partial models, one partial model corresponding to one application node;

[0237] at least one partial model that can be independently trained.

[0238] The single edge model can be an edge model applied to the terminal, or the single edge model can be an edge model applied to the network device.

[0239] In step S2105, the terminal 101 detects that the first condition is met, and sends a third message to the network device 102.

[0240] In some embodiments, the network device receives the third message. For example, the network device receives the third message sent by the terminal.

[0241] In some embodiments, the terminal determines whether the first condition is met during the process of training the first model according to the first training data. Optionally, the first condition includes at least one of:

[0242] the first model meets a performance requirement;

[0243] the data amount of the first training data reaches a first threshold value;

[0244] the iteration number of the training model reaches a second threshold value;

[0245] the time length of the training model reaches a third threshold value.

[0246] At least one of the performance requirement, the first threshold value, the second threshold value, and the third threshold value is network device indicated or protocol specified preset information.

[0247] In some embodiments, the terminal can send the third message to the network device when it is determined that the first condition is met. The third message is used for the terminal to report the training situation of the first model to the network device.

[0248] In some embodiments, the third message includes at least one of:

[0249] a third indication indicating that the trained first model meets the performance requirement;

[0250] a fourth indication indicating that the network device stops sending the first data and / or the first signal;

[0251] a fifth indication indicating that the trained first model does not meet the performance requirement and / or a reason why the trained first model does not meet the performance requirement.

[0252] In some embodiments, if the third message is used to report that the trained first model meets the performance requirement, the third message can include the third indication and / or the fourth indication.

[0253] In some embodiments, if the third message is used to report that the trained first model does not meet the performance requirement, the third message can include the fifth indication.

[0254] At step S2106, the network device 102 sends a fourth message to the terminal 101.

[0255] In some embodiments, the terminal receives the fourth message. For example, the terminal receives the fourth message sent by the network device.

[0256] In some embodiments, the fourth message is a response to the third message including the fifth indication.

[0257] In some embodiments, the terminal receives the fourth message sent by the network device, the fourth message being sent by the network device according to the received fifth indication, and the fourth message being used to instruct the terminal to perform a first operation.

[0258] In some embodiments, the terminal performs the first operation in response to receiving the fourth message. Optionally, the first operation includes at least one of the following:

[0259] switching the first model to a second model, the second model having the same function as the first model;

[0260] replacing the first model with a first algorithm, the first algorithm having the same function as the first model;

[0261] increasing the number of iterations of model training and continuing to train the first model;

[0262] increasing the time length of model training and continuing to train the first model;

[0263] continuing to obtain the first training data and training the first model;

[0264] giving up training the first model.

[0265] In some embodiments, when the first model is a unilateral model on the terminal side, the first operation can include at least one of the following:

[0266] switching the first model to a second model, the second model having the same function as the first model;

[0267] replacing the first model with a first algorithm, the first algorithm having the same function as the first model;

[0268] increasing the number of iterations of model training and continuing to train the first model;

[0269] increasing the length of model training and continuing to train the first model;

[0270] continuing to obtain the first training data and training the first model;

[0271] giving up training the first model.

[0272] In some embodiments, when the first model is a one-sided model on the network device side, the first operation can include at least one of:

[0273] increasing the number of iterations of model training and continuing to train the first model;

[0274] increasing the length of model training and continuing to train the first model;

[0275] continuing to obtain the first training data and training the first model;

[0276] giving up training the first model.

[0277] In some embodiments, when the first model is a multi-sided model, the first operation can include at least one of:

[0278] increasing the number of iterations of model training and continuing to train the first model;

[0279] increasing the length of model training and continuing to train the first model;

[0280] continuing to obtain the first training data and training the first model;

[0281] giving up training the first model.

[0282] In some embodiments, when the first model is at least one partial model in the multi-sided model, the first operation can include at least one of:

[0283] increasing the number of iterations of model training and continuing to train the first model;

[0284] increasing the length of model training and continuing to train the first model;

[0285] continuing to obtain the first training data and training the first model;

[0286] giving up training the first model.

[0287] In some embodiments, step S2106 can be omitted in the case where the third message in step S2105 does not include the fifth indication. For example, step S2106 can be omitted in the case where the third message includes the third indication and / or the fourth indication.

[0288] In some embodiments, if the application node of the first model is the network device, the first model that meets the performance requirement or the update information of the first model can be sent to the network device when the trained first model meets the performance requirement.

[0289] In some embodiments, if the first model includes a first partial model and the application node of the first partial model is the network device, the first partial model in the first model that meets the performance requirement or the update information of the first partial model can be sent to the network device when the trained first model meets the performance requirement.

[0290] In some embodiments, the names of information, etc. are not limited to the names described in the embodiments, and the terms such as "information", "message", "signal", "signaling", "report", "configuration", "indication", "instruction", "command", "channel", "parameter", "domain", "field", "symbol", "symbol", "codebook", "codeword", "code point", "bit", "data", "program", "chip", and the like can be replaced with each other.

[0291] In some embodiments, the terms such as "synchronization signal (SS)", "synchronization signal block (SSB)", "reference signal (RS)", "pilot", "pilot signal", and the like can be replaced with each other.

[0292] In some embodiments, the terms such as "time", "time point", "time", "time position", and the like can be replaced with each other, and the terms such as "duration", "period", "time window", "window", "time", and the like can be replaced with each other.

[0293] In some embodiments, "acquire", "obtain", "get", "receive", "transmit", "bidirectional transmission", "send and / or receive" can be replaced with each other, which can be interpreted as receiving from other subjects, acquiring from protocols, acquiring from higher layers, processing to obtain by itself, autonomously implementing, and the like.

[0294] In some embodiments, the terms "send", "transmit", "report", "issue", "transmit", "bidirectional transmission", "send and / or receive", and the like can be replaced with each other.

[0295] In some embodiments, the terms "certain", "preset", "preset", "set", "indicated", "certain", "arbitrary", "first", and the like can be replaced with each other. "Certain A", "preset A", "preset A", "set A", "indicated A", "certain A", "arbitrary A", "first A" can be interpreted as A specified in advance in protocols and the like, can be interpreted as A obtained by setting, configuring, or indicating, and the like, and can be interpreted as certain A, certain A, arbitrary A, or first A, but not limited thereto.

[0296] In some embodiments, determination or judgment can be made by a value represented by 1 bit (0 or 1), or by a true or false value (Boolean value) represented by true or false, or by comparison of numerical values (for example, comparison with a predetermined value), but not limited thereto.

[0297] The model training method related to the embodiments of the present disclosure can also be referred to as a single entity training method. The model training method shown in FIG. 2A is applicable to terminal training of at least one of the following models:

[0298] Terminal side model;

[0299] Network side model;

[0300] Bilateral model;

[0301] At least part of the bilateral model, which allows independent entities to be trained.

[0302] The model training method related to the embodiments of the present disclosure can include at least one of steps S2101-S2106. For example, step S2104 can be implemented as an independent embodiment, steps S2101 and S2102 can be implemented as independent embodiments, and steps S2103 and S2104 can be implemented as independent embodiments, but not limited thereto.

[0303] In some embodiments, the order between any two of steps S2101-S2106 can be exchanged or performed simultaneously.

[0304] In some embodiments, one or more of steps S2101-S2106 are optional, and one or more of these steps can be omitted or replaced in different embodiments.

[0305] In some embodiments, other optional implementations described before or after the description corresponding to FIG. 2A can be referred to.

[0306] In some embodiments, the model training method shown in FIG. 2A can be modified as long as it is allowed, for example, the network device performs the operation of the first node, and the terminal performs the operation of the second node, that is, the operations performed by the terminal and the network device in FIG. 2A are exchanged. For example, the terminal in FIG. 2A is replaced by another network device. For example, the network device in FIG. 2A is replaced by another terminal.

[0307] FIG. 2B is an interaction diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 2B, the embodiment of the present disclosure relates to a model training method, and in particular to a model training method for training a first model by a single entity training method, which is performed by a communication system 100. In the embodiment shown in FIG. 2B, the terminal 101 is a second node, and the network device 102 is a first node. The first node can be referred to as a model training node, that is, a node performing a model training process, which includes steps of inputting training data into a model, determining model parameter update information, and updating model parameters. The model training method shown in FIG. 2B can include the following steps:

[0308] Step S2201, the terminal 101 sends a fifth message to the network device 102.

[0309] In some embodiments, the network device receives the fifth message. For example, the network device receives the fifth message sent by the terminal.

[0310] In some embodiments, the terminal sends the fifth message to the network device, and the fifth message includes at least one of the following:

[0311] A third request for requesting the network device to train the first model online;

[0312] Performance information of the first model;

[0313] Initial model parameters of the first model.

[0314] In some embodiments, one or more of the third request, the performance information of the first model, and the initial model parameters of the first model can be sent simultaneously or separately, and the present disclosure does not make a specific limitation thereon.

[0315] In step S2202, the network device 102 sends a sixth message to the terminal 101.

[0316] In some embodiments, the terminal receives the sixth message. For example, the terminal receives the sixth message sent by the network device.

[0317] In some embodiments, the network device sends the sixth message to the terminal, and the sixth message comprises at least one of:

[0318] A sixth indication for instructing the network device to online train the first model;

[0319] A seventh indication for instructing the terminal to send the first data and / or the first signal to the network device, wherein the first data and / or the first signal are used for the network device to collect the first training data, and the first training data are used for training the first model.

[0320] In some embodiments, one or more of the sixth indication and the seventh indication can be sent simultaneously or separately, and the present disclosure does not make a specific limitation thereon.

[0321] For example, assuming that in step S2201, the fifth message sent by the terminal to the network device comprises the third request, then in step S2202, the sixth message sent by the network device to the terminal can comprise the sixth indication, which can be a response of the network device to the third request. This is a passive way of online training the first model by the network device according to the request of the terminal.

[0322] For example, assuming that in step S2201, the fifth message sent by the terminal to the network device does not comprise the third request but comprises the performance information of the first model, then in step S2202, the sixth message sent by the network device to the terminal can comprise the sixth indication, which can be sent by the network device according to the performance information of the first model reported by the terminal. This is an active way of online training the first model by the network device.

[0323] For example, step S2201 can be omitted, and in step S2202, the sixth message sent by the network device to the terminal can comprise the sixth indication, which can be sent by the network device to the terminal actively, for example, sent by the network device to the terminal actively when it is detected that the first model training time reaches a preset value, or sent by the network device to the terminal actively when it is detected that the performance information of the first model does not meet the requirements. This is an active way of online training the first model by the network device.

[0324] It should be noted that in the above example, the seventh indication sent by the network device to the terminal can default that the network device will train the first model online.

[0325] In step S2203, the network device 102 collects first training data.

[0326] In some embodiments, the network device starts to collect the first training data when the sixth message is sent to the terminal.

[0327] In some embodiments, the network device collects the first training data, wherein the first training data comprises at least one of the following:

[0328] The first data sent by the terminal;

[0329] The second data obtained by measuring the first signal sent by the terminal.

[0330] The first signal can be a pilot signal or other signal used by the network device to obtain the second data.

[0331] In some embodiments, the terminal can send the first data to the network device. In some embodiments, the terminal can send the first signal, and the network device can measure the first signal to obtain the second data.

[0332] In some embodiments, the network device can collect the first training data at multiple time points, and the first training data comprises the first data and / or the second data.

[0333] In some embodiments, the first training data comprises input samples and / or corresponding output samples of the first model.

[0334] In step S2204, the network device 102 trains the first model according to the first training data.

[0335] In some embodiments, the first model is trained according to the first training data when the first model is trained by the first node (the network device 102). When the first model is trained by multiple nodes including the first node (the network device 102), the model iteration information of the first model is interacted with the second node (the terminal 101) during the training of the first model according to the first training data.

[0336] In some embodiments, if the first model is trained by one first node, the first node can not interact the model iteration information of the first model with the second node. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102, the network device 102 can not interact the model iteration information of the first model with the terminal 101.

[0337] In some embodiments, if the first model is trained by multiple nodes including the first node, the first node can interact the model iteration information of the first model with the second node. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102 and the terminal 101, the network device 102 can interact the model iteration information of the first model with the terminal 101. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102 and one or more other network devices except the network device 102, the network device 102 can interact the model iteration information of the first model with the terminal 101. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102, the terminal 101 and one or more other terminals except the terminal 101, the network device 102 can interact the model iteration information of the first model with the terminal 101.

[0338] In some embodiments, during the training of the first model according to the first training data, the number of nodes training the first model is one or more default to know whether to interact the model iteration information of the first model with the terminal.

[0339] In some embodiments, during the training of the first model according to the first training data, the number of nodes training the first model can be distinguished according to the model training manner, and then the number of nodes training the first model is one or more default to know whether to interact the model iteration information of the first model with the terminal.

[0340] In some embodiments, the model iteration information is the information of the process quantity generated in the process of training the first model, and optionally, the model iteration information includes at least one of the following:

[0341] Gradient information;

[0342] Model parameter iteration information.

[0343] In some embodiments, if the network device trains the first model in a single entity training manner, the network device can not interact the model iteration information of the first model with the terminal. The single entity training manner refers to that the model training process is completed by one node, and the model training process includes the steps of inputting the training data into the model, determining the model parameter update information, and updating the model parameters.

[0344] In some embodiments, the first model suitable for the single entity training manner includes at least one of the following:

[0345] One-sided model;

[0346] The multi-edge model comprises a plurality of sub-models, and one sub-model corresponds to one application node.

[0347] The at least one sub-model is allowed to be independently trained.

[0348] The single-edge model can be a single-edge model applied to a terminal, or the single-edge model can be a single-edge model applied to a network device.

[0349] In step S2205, the network device 102 detects that the first condition is met, and sends a seventh message to the terminal 101.

[0350] In some embodiments, the terminal receives the seventh message. For example, the terminal receives the seventh message sent by the network device.

[0351] In some embodiments, the network device determines whether the first condition is met in the process of training the first model according to the first training data. Optionally, the first condition comprises at least one of the following:

[0352] The first model meets a performance requirement;

[0353] The amount of data of the first training data reaches a first threshold value;

[0354] The number of iterations of training the model reaches a second threshold value;

[0355] The duration of training the model reaches a third threshold value.

[0356] At least one of the performance requirement, the first threshold value, the second threshold value, and the third threshold value is preset information specified by a protocol.

[0357] In some embodiments, the network device can send the seventh message to the terminal when it is determined that the first condition is met. The seventh message is used for the network device to inform the terminal of the training situation of the first model.

[0358] In some embodiments, the seventh message comprises at least one of the following:

[0359] An eighth indication for instructing the terminal to stop sending the first data and / or the first signal;

[0360] A ninth indication for informing the terminal that the first model training fails and / or the network device will perform a second operation;

[0361] A tenth indication for instructing the terminal to perform the second operation;

[0362] The second operation comprises at least one of the following:

[0363] Switching the first model to a second model, and the function of the second model is the same as that of the first model.

[0364] using a first algorithm to replace the first model, the first algorithm having the same function as the first model.

[0365] In some embodiments, if the first model is a unilateral model on the network device side, the seventh message comprises at least one of:

[0366] an eighth indication for instructing the terminal to stop sending the first data and / or the first signal;

[0367] a ninth indication for informing the terminal that the first model training fails and / or the network device will perform a second operation.

[0368] In some embodiments, if the first model is a unilateral model on the terminal side, the seventh message comprises at least one of:

[0369] an eighth indication for instructing the terminal to stop sending the first data and / or the first signal;

[0370] a ninth indication for informing the terminal that the first model training fails;

[0371] a tenth indication for instructing the terminal to perform a second operation.

[0372] In some embodiments, if the application node of the first model is the terminal, when the trained first model meets the performance requirement, the terminal can be sent the first model meeting the performance requirement or update information of the first model.

[0373] In some embodiments, if the first model comprises a first partial model, the application node of the first partial model is the terminal, and when the trained first model meets the performance requirement, the terminal is sent the first partial model in the first model meeting the performance requirement or update information of the first partial model.

[0374] In some embodiments, the names of information and the like are not limited to the names described in the embodiments, and terms such as "information", "message", "signal", "signaling", "report", "configuration", "indication", "instruction", "command", "channel", "parameter", "domain", "field", "symbol", "codebook", "codeword", "codepoint", "bit", "data", "program", "chip", and the like can be replaced with each other.

[0375] In some embodiments, terms such as "synchronization signal (SS)", "synchronization signal block (SSB)", "reference signal (RS)", "pilot", "pilot signal", and the like can be replaced with each other.

[0376] In some embodiments, terms such as "time", "time point", "time position", and the like can be replaced with each other, and terms such as "duration", "period", "time window", "window", and "time" can be replaced with each other.

[0377] In some embodiments, "acquire", "obtain", "get", "receive", "transmit", "bidirectional transmission", "send and / or receive", and the like can be replaced with each other, and can be interpreted as receiving from other subjects, acquiring from protocols, acquiring from higher layers, obtaining by self-processing, autonomously implementing, and the like.

[0378] In some embodiments, terms such as "send", "transmit", "report", "issue", "transmit", "bidirectional transmission", "send and / or receive", and the like can be replaced with each other.

[0379] In some embodiments, the terms "certain", "preset", "pre-set", "set", "indicated", "any", "first", and the like can be replaced with each other, and "certain A", "preset A", "pre-set A", "set A", "indicated A", "any A", "first A" can be interpreted as A predetermined in a protocol or the like, or A obtained by setting, configuring, or indicating, or a specific A, any A, or first A, but are not limited thereto.

[0380] In some embodiments, the determination or judgment can be made by a value represented by 1 bit (0 or 1), or by a true or false value (Boolean value) represented by true or false, or by comparison of numerical values (for example, comparison with a predetermined value), but is not limited thereto.

[0381] The model training method related to the embodiments of the present disclosure can also be referred to as a single entity training method, and the model training method shown in FIG. 2B is applicable to training at least one of the following models by the network device:

[0382] a terminal side model;

[0383] a network side model;

[0384] a bilateral model;

[0385] at least part of the bilateral model, which allows to be trained by independent entities.

[0386] The model training method related to the embodiments of the present disclosure can include at least one of steps S2201-S2205. For example, step S2204 can be implemented as an independent embodiment, steps S2201 and S2202 can be implemented as independent embodiments, and steps S2203 and S2204 can be implemented as independent embodiments, but are not limited thereto.

[0387] In some embodiments, the order of any two of steps S2201-S2205 can be exchanged or executed simultaneously.

[0388] In some embodiments, one or more of steps S2201-S2205 are optional, and one or more of these steps can be omitted or replaced in different embodiments.

[0389] In some embodiments, reference can be made to other optional implementations described before or after the corresponding description of FIG. 2B.

[0390] In some embodiments, the model training method shown in FIG. 2B can be modified as long as it is allowed, for example, the network device performs the operation of the second node, the terminal performs the operation of the first node, that is, the operations performed by the terminal and the network device in FIG. 2B are exchanged. For example, replace the terminal in FIG. 2B with another network device. For example, replace the network device in FIG. 2B with another terminal.

[0391] FIG. 2C is an interaction diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 2C, the embodiments of the present disclosure relate to a model training method, and in particular to a model training method for training at least part of a multi-edge model by a multi-entity collaborative training manner, which is performed by the communication system 100. In the embodiment shown in FIG. 2C, the terminal 101 is the first node, and the network device 102 is the second node. The first model on the first node and the third model on the second node are components of the multi-edge model. The first node (terminal) and the second node (network device) can achieve the purpose of training and updating the first model and / or the third model by collaborative training. In this embodiment, it is assumed that neither the first model nor the third model can be trained independently. The model training method shown in FIG. 2C can include the following steps:

[0392] In step S2301, the terminal 101 sends a first message to the network device 102.

[0393] In some embodiments, the network device receives the first message. For example, the network device receives the first message sent by the terminal.

[0394] In some embodiments, the terminal sends the first message to the network device, and the first message includes at least one of the following contents:

[0395] A first request for requesting the network device to allow the terminal to train the first model online;

[0396] A second request for requesting the network device to send the terminal with first data and / or a first signal, wherein the first data and / or the first signal are used by the terminal to collect first training data, and the first training data are used to train the multi-edge model, and the multi-edge model includes the first model and / or the third model, that is, the first training data are used to train the first model and / or the third model;

[0397] Performance information of the first model;

[0398] Model training capability of the terminal.

[0399] In some embodiments, one or more of the first request, the second request, the performance information of the first model, and the model training capability of the terminal can be sent simultaneously or separately, and the present disclosure does not make a specific limitation thereon.

[0400] In some embodiments, since the first model cannot be trained independently, i.e., the first model needs to be trained by means of cooperative training, or the first model needs to participate in cooperative training to train the third model, the first request can be understood as a request of the terminal to the network device for online cooperative training of the model, and the purpose of online cooperative training includes training and optimizing / updating the first model and / or the third model. Alternatively, the online cooperative training includes training and optimizing / updating the first model and / or the third model.

[0401] In step S2302, the network device 102 sends a second message to the terminal 101.

[0402] In some embodiments, the terminal receives the second message. For example, the terminal receives the second message sent by the network device.

[0403] In some embodiments, the network device sends the second message to the terminal, and the second message includes at least one of the following contents:

[0404] The first indication is used to instruct the terminal to train the first model online and / or to indicate the initial model parameters of the first model to the terminal.

[0405] The second indication is used to instruct the terminal to collect the first training data.

[0406] In some embodiments, one or more of the first indication and the second indication can be sent simultaneously or separately. The present disclosure does not limit this.

[0407] In some embodiments, since the first model cannot be trained independently, i.e., the first model needs to be trained by means of cooperative training, the first indication can be understood as instructing the terminal to train the model online, and the purpose of online cooperative training includes training and optimizing / updating the first model. Alternatively, the online cooperative training includes training and optimizing / updating the first model and the third model.

[0408] For example, assuming that in step S2301, the first message sent by the terminal to the network device includes the first request, in step S2302, the second message sent by the network device to the terminal can include the first indication, which can be a response of the network device to the first request. This is a way for the terminal to actively request the network device to train the model online.

[0409] For example, assuming that in step S2301, the first message sent by the terminal to the network device does not include the first request but includes the performance information of the first model, then in step S2302, the second message sent by the network device to the terminal can include the first indication, which can be sent by the network device according to the performance information of the first model reported by the terminal. This is a way for the terminal to passively train the model online according to the indication of the network device.

[0410] For example, assuming that in step S2301, the first message sent by the terminal to the network device does not include the first request but includes the model training capability of the terminal, then in step S2302, the second message sent by the network device to the terminal can include the first indication, which can be sent by the network device according to the model training capability of the terminal reported by the terminal. This is a way for the terminal to passively train the model online according to the indication of the network device.

[0411] For example, assuming that in step S2301, the first message sent by the terminal to the network device does not include the first request and does not include the performance information of the first model, then in step S2302, the second message sent by the network device to the terminal can include the first indication, which can be sent by the network device to the terminal, for example, the network device sends the second message including the first indication to the terminal when detecting that the first model training time reaches a preset value. This is a way for the terminal to passively train the model online according to the indication of the network device.

[0412] For example, step S2301 can be omitted, and in step S2302, the second message sent by the network device to the terminal can include the first indication, which can be sent by the network device to the terminal, for example, the network device sends the first indication to the terminal when determining that the first model and / or the third model need to be trained online. This is a way for the terminal to passively train the model online according to the indication of the network device.

[0413] It should be noted that in the above examples, the premise of the network device sending the first indication to the terminal is that the model training capability of the terminal is confirmed to be able to meet the requirement of training the first model online. In the above embodiment, the first indication sent by the network device to the terminal indicates that the network device allows the model to be trained online.

[0414] In some embodiments, assuming that in step S2301, the first message sent by the terminal to the network device includes the second request, then in step S2302, the second message sent by the network device to the terminal can include the second indication, which can be a response of the network device to the second request.

[0415] Step S2303, the terminal 101 collects the first training data.

[0416] In some embodiments, the terminal starts to collect the first training data in response to receiving the second message sent by the network device.

[0417] In some embodiments, the terminal collects the first training data, wherein the first training data comprises at least one of:

[0418] the first data sent by the network device;

[0419] the second data obtained by measuring the first signal sent by the network device.

[0420] The first signal can be a pilot signal or other signal used by the terminal to obtain the second data.

[0421] In some embodiments, the network device can send the first data to the terminal. In some embodiments, the network device can send the first signal, and the terminal can measure the first signal to obtain the second data.

[0422] In some embodiments, in step S2302, the second indication in the second message can also be used to indicate a resource for sending the first signal, the resource comprising at least one of a time domain resource, a frequency domain resource, and a spatial domain resource. The implementation of the terminal collecting the first training data comprises measuring the first signal on the corresponding time domain, frequency domain, and spatial domain to obtain the second data.

[0423] In some embodiments, in step S2302, the second indication in the second message can also be used to indicate a data format of the first data. The implementation of the terminal collecting the first training data comprises receiving the first data in the data format.

[0424] In some embodiments, the terminal can collect the first training data at multiple time instants, and the first training data comprises the first data and / or the second data.

[0425] In some embodiments, the first training data comprises input samples and / or corresponding output samples of the multi-edge model.

[0426] In some embodiments, the first training data comprises input samples of the first model and / or output samples of the third model.

[0427] In some embodiments, the first training data comprises input samples of the third model and / or output samples of the first model.

[0428] In some embodiments, the first model is trained according to the first training data if the first model is trained by the first node (terminal 101). If the first model is trained by a plurality of nodes including the first node (terminal 101), the model iteration information of the first model is interacted with the second node (network device 102) in the process of training the first model according to the first training data.

[0429] In some embodiments, if the first model is trained by one first node, the first node can not interact the model iteration information of the first model with the second node. For example, assuming that the first node is terminal 101 and the second node is network device 102, if the first model is trained by terminal 101, terminal 101 can not interact the model iteration information of the first model with network device 102.

[0430] In some embodiments, if the first model is trained by a plurality of nodes including the first node, the first node can interact the model iteration information of the first model with the second node. For example, assuming that the first node is terminal 101 and the second node is network device 102, if the first model is trained by terminal 101 and network device 102 together, terminal 101 can interact the model iteration information of the first model with network device 102. For example, assuming that the first node is terminal 101 and the second node is network device 102, if the first model is trained by terminal 101 and one or more other terminals except terminal 101 together, terminal 101 can interact the model iteration information of the first model with network device 102. For example, assuming that the first node is terminal 101 and the second node is network device 102, if the first model is trained by terminal 101, network device 102, and one or more other network devices except network device 102, terminal 101 can interact the model iteration information of the first model with network device 102.

[0431] In some embodiments, in the process of training the first model according to the first training data, the step of interacting the model iteration information of the first model with the network device is performed or not performed according to the number of nodes training the first model.

[0432] In some embodiments, in the process of training the first model according to the first training data, the model training manner can be distinguished according to the number of nodes training the first model, and then the step of interacting the model iteration information of the first model with the network device is performed or not performed according to the training manner of the first model.

[0433] In some embodiments, if the manner in which the terminal trains the first model is a multi-entity collaborative training manner, the terminal can interact model iteration information of the first model with the network device. Optionally, the first model is at least one partial model in a multi-edge model, the multi-edge model including a plurality of partial models, one partial model corresponding to one application node. Optionally, the network device includes a third model, the third model being at least one partial model in the multi-edge model other than the first model.

[0434] In some embodiments, the gradient information includes forward gradient information and / or backward gradient information. It should be explained that the forward gradient information of the model refers to the process of calculating the output from the input layer to the output layer through the neural network, i.e., forward propagation. While the backward gradient information (i.e., reverse gradient information) refers to the propagation of errors from the output layer to the input layer for calculating the gradient of each parameter on the loss, i.e., back propagation.

[0435] In some embodiments, the implementation in which the terminal interacts the model iteration information of the first model with the network device includes at least one of the following:

[0436] The terminal sends the forward gradient information to the network device, the forward gradient information being output or generated by the first model;

[0437] The terminal receives the backward gradient information sent by the network device, the backward gradient information being output or generated by the third model;

[0438] The terminal receives the forward gradient information sent by the network device, the forward gradient information being output or generated by the third model;

[0439] The terminal sends the backward gradient information to the network device, the backward gradient information being output or generated by the first model.

[0440] After step S2303, at least one of steps S2304-S2308 is selected to be executed. Steps S2304-S2308 are exemplarily described taking the output of the first model as the input of the third model. Steps S2304-S2308 can be executed one or more times. The purpose of executing steps S2304-S2308 one or more times is to reduce the prediction error of the first model by adjusting the model parameters of the first model. After the Nth execution of steps S2304-S2308, if the prediction error of the first model is less than a preset value, the training of the first model is stopped.

[0441] In step S2304, the terminal 101 inputs the input sample of the first model into the first model.

[0442] Step S2305, the terminal 101 sends the prediction data of the first model, the forward gradient information, and the output sample of the third model to the network device 102.

[0443] In some embodiments, the prediction data of the first model, the forward gradient information, and the output sample of the third model can be sent simultaneously or separately. The present disclosure does not make any limitation in this regard.

[0444] Step S2306, the network device 102 inputs the prediction data of the first model and the forward gradient information into the third model, determines the error according to the prediction data of the third model and the output sample of the third model, and determines the backward gradient information.

[0445] Optionally, the network device adjusts the model parameters of the third model according to the backward gradient information.

[0446] Step S2307, the network device 102 sends the backward gradient information to the terminal 101.

[0447] Step S2308, the terminal 101 adjusts the model parameters of the first model according to the backward gradient information.

[0448] In some embodiments, steps S2304-S2308 can be replaced by the following steps 1-6, which are exemplarily described by taking the output of the third model as the input of the first model.

[0449] Step 1, the terminal sends the input sample of the third model to the network device.

[0450] Step 2, the network device inputs the input sample of the third model into the third model.

[0451] Step 3, the network device sends the prediction data of the third model and the forward gradient information to the terminal.

[0452] Step 4, the terminal inputs the prediction data of the third model and the forward gradient information into the first model, determines the error according to the prediction data of the first model and the output sample of the first model, determines the backward gradient information, and adjusts the model parameters of the first model according to the backward gradient information.

[0453] Step 5, the terminal sends the backward gradient information to the network device.

[0454] Step 6, the network device adjusts the model parameters of the third model according to the backward gradient information.

[0455] Step S2309, the terminal 101 detects that the first condition is met, and sends a third message to the network device 102.

[0456] In some embodiments, the network device receives a third message. For example, the network device receives the third message sent by the terminal.

[0457] In some embodiments, the terminal determines whether a first condition is met during the cooperative training of the first model. For example, the determination is made in real time, for example, once when the first model is iterated n (n is an integer greater than 0) times. Optionally, the first condition includes at least one of the following:

[0458] The first model meets a performance requirement;

[0459] The amount of data of the first training data reaches a first threshold;

[0460] The number of iterations of the trained model reaches a second threshold;

[0461] The duration of the trained model reaches a third threshold.

[0462] At least one of the performance requirement, the first threshold, the second threshold, and the third threshold is indicated by the network device or is preset information specified by a protocol.

[0463] In some embodiments, the terminal can send a third message to the network device if it is determined that the first condition is met. The third message is used by the terminal to report the training situation of the first model to the network device.

[0464] In some embodiments, the third message includes at least one of the following:

[0465] A third indication indicating that the trained first model meets the performance requirement;

[0466] A fourth indication indicating that the network device stops sending the first data and / or the first signal;

[0467] A fifth indication indicating that the trained first model does not meet the performance requirement and / or the reason why the trained first model does not meet the performance requirement.

[0468] In some embodiments, if the third message is used to report that the trained first model meets the performance requirement, the third message can include the third indication and / or the fourth indication.

[0469] In some embodiments, if the third message is used to report that the trained first model does not meet the performance requirement, the third message can include the fifth indication.

[0470] Step S2310: The network device 102 sends a fourth message to the terminal 101.

[0471] In some embodiments, the terminal receives the fourth message. For example, the terminal receives the fourth message sent by the network device.

[0472] In some embodiments, the fourth message is a response to the third message including the fifth indication.

[0473] In some embodiments, the terminal receives a fourth message sent by the network device, the fourth message being sent by the network device according to the received fifth indication, and the fourth message being used to instruct the terminal to perform the first operation.

[0474] In some embodiments, the terminal performs the first operation in response to receiving the fourth message. Optionally, the first operation includes at least one of the following:

[0475] switching the first model to a second model, the second model having the same function as the first model;

[0476] replacing the first model with a first algorithm, the first algorithm having the same function as the first model;

[0477] increasing the number of iterations of model training and continuing to train the first model;

[0478] increasing the length of model training and continuing to train the first model;

[0479] continuing to obtain the first training data and training the first model;

[0480] giving up training the first model.

[0481] In some embodiments, the step S2310 can be omitted in the case where the third message in the step S2309 does not include the fifth indication. For example, the step S2310 can be omitted in the case where the third message includes the third indication and / or the fourth indication.

[0482] In some embodiments, if the application node of the first model is the network device, the first model or the update information of the first model reaching the performance requirement can be sent to the network device when the trained first model reaches the performance requirement.

[0483] In some embodiments, if the first model includes a first partial model and the application node of the first partial model is the network device, the first partial model or the update information of the first partial model in the first model reaching the performance requirement can be sent to the network device when the trained first model reaches the performance requirement.

[0484] In some embodiments, the names of information and the like are not limited to the names described in the embodiments, and terms such as "information", "message", "signal", "signaling", "report", "configuration", "indication", "instruction", "command", "channel", "parameter", "domain", "field", "symbol", "codebook", "codeword", "codepoint", "bit", "data", "program", "chip", and the like can be replaced with each other.

[0485] In some embodiments, terms such as "synchronization signal (SS)", "synchronization signal block (SSB)", "reference signal (RS)", "pilot", "pilot signal", and the like can be replaced with each other.

[0486] In some embodiments, terms such as "time", "time point", "time", "time position", and the like can be replaced with each other, and terms such as "duration", "period", "time window", "window", "time", and the like can be replaced with each other.

[0487] In some embodiments, "acquire", "obtain", "get", "receive", "transmit", "bidirectional transmission", "send and / or receive", and the like can be replaced with each other, and can be interpreted as receiving from other subjects, acquiring from protocols, acquiring from higher layers, obtaining by self-processing, autonomously implementing, and the like.

[0488] In some embodiments, terms such as "send", "transmit", "report", "issue", "transmit", "bidirectional transmission", "send and / or receive", and the like can be replaced with each other.

[0489] In some embodiments, the terms "certain", "preset", "pre-set", "set", "indicated", "any", "first", and the like can be replaced with each other, and "certain A", "preset A", "pre-set A", "set A", "indicated A", "any A", "first A" can be interpreted as A predetermined in a protocol or the like, or A obtained by setting, configuring, or indicating, or a specific A, any A, or first A, but are not limited thereto.

[0490] In some embodiments, the determination or judgment can be performed by a value represented by 1 bit (0 or 1), or by a true or false value (Boolean value) represented by true or false, or by comparison of numerical values (for example, comparison with a predetermined value), but is not limited thereto.

[0491] The model training method related to the embodiments of the present disclosure can also be referred to as a multi-entity collaborative training method, and the model training method shown in FIG. 2C is applicable to training at least part of the multi-edge model by the terminal and the network device in collaboration, and the at least part of the model is not allowed to be trained by an independent entity.

[0492] The model training method related to the embodiments of the present disclosure can include at least one of steps S2301-S2310. For example, step S2305 can be implemented as an independent embodiment, step S2307 can be implemented as an independent embodiment, steps S2301 and S2302 can be implemented as independent embodiments, steps S2305 and S2307 can be implemented as independent embodiments, and steps S2304-S2308 can be implemented as independent embodiments, but are not limited thereto.

[0493] In some embodiments, the order of any two steps in steps S2301-S2310 can be exchanged or executed simultaneously.

[0494] In some embodiments, one or more steps in steps S2301-S2310 are optional, and one or more steps in these steps can be omitted or replaced in different embodiments.

[0495] In some embodiments, other optional implementations described before or after the corresponding description of FIG. 2C can be referred to.

[0496] In some embodiments, the model training method shown in FIG. 2C can be modified as long as it is allowed, for example, the network device performs the operation of the first node, the terminal performs the operation of the second node, that is, the operations performed by the terminal and the network device in FIG. 2C are exchanged. For example, the terminal in FIG. 2C is replaced by another network device. For example, the network device in FIG. 2C is replaced by another terminal.

[0497] FIG. 2D is an interaction diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 2D, the embodiments of the present disclosure relate to a model training method, and in particular to a model training method for training at least part of a multi-edge model by a multi-entity collaborative training manner, which is performed by the communication system 100. In the embodiment shown in FIG. 2D, the terminal 101 is the second node, and the network device 102 is the first node. The first model on the first node and the third model on the second node are components of the multi-edge model. The first node (network device) and the second node (terminal) can achieve the purpose of training and updating the first model and / or the third model by collaborative training. In this embodiment, it is assumed that neither the first model nor the third model can be trained independently.

[0498] The model training method shown in FIG. 2D can include the following steps:

[0499] In step S2401, the network device 102 sends a sixth message to the terminal 101.

[0500] In some embodiments, the terminal receives the sixth message. For example, the terminal receives the sixth message sent by the network device.

[0501] In some embodiments, the network device sends the sixth message to the terminal, and the sixth message includes at least one of the following:

[0502] The sixth indication is used to indicate that the network device will train the first model online;

[0503] The seventh indication is used to instruct the terminal to send the first data and / or the first signal to the network device, wherein the first data and / or the first signal are used for the network device to collect the first training data, the first training data are used for training the multi-edge model, the multi-edge model includes the first model and / or the third model, and the first training data are used for training the first model and / or the third model.

[0504] In some embodiments, one or more of the sixth indication and the seventh indication can be sent simultaneously or separately. The present disclosure does not limit this.

[0505] In some embodiments, the first model cannot be trained independently, i.e., the first model needs to be trained by means of cooperative training or the first model needs to participate in cooperative training to train the third model, and thus the sixth indication can indicate to the terminal that the network device will train the model online, and the purpose of online cooperative training includes training and / or updating the first model and / or the third model. Alternatively, the online cooperative training includes training and / or updating the first model and / or the third model.

[0506] For example, in step S2401, the sixth message sent by the network device to the terminal can include the sixth indication, which can be sent by the network device to the terminal actively, e.g., sent by the network device to the terminal actively when the network device detects that the preset first model training time is reached, e.g., sent by the network device to the terminal actively when the network device detects that the performance information of the first model does not meet the requirements. This is a way for the network device to actively indicate the terminal to train the model online.

[0507] For example, in step S2401, the sixth message sent by the network device to the terminal can include the sixth indication, which can be sent by the network device to the terminal actively, e.g., sent by the network device to the terminal actively when the network device detects that the preset third model training time is reached, e.g., sent by the network device to the terminal actively when the network device detects that the performance information of the third model does not meet the requirements. This is a way for the network device to actively indicate the terminal to train the model online.

[0508] It should be noted that in the above example, when the network device sends the seventh indication to the terminal, the network device and the terminal can default to train the first model and / or the third model online.

[0509] In step S2402, the network device 102 collects the first training data.

[0510] In some embodiments, the network device starts to collect the first training data when the network device sends the sixth message to the terminal.

[0511] In some embodiments, the network device collects the first training data, wherein the first training data includes at least one of the following:

[0512] The first data sent by the terminal;

[0513] The second data obtained by measuring the first signal sent by the terminal.

[0514] The first signal can be a pilot signal or other signal used by the network device to obtain the second data.

[0515] In some embodiments, the terminal can send first data to the network device. In some embodiments, the terminal can send a first signal, and the network device can measure the first signal to obtain second data.

[0516] In some embodiments, the network device can collect first training data at multiple time instants, the first training data comprising the first data and / or the second data.

[0517] In some embodiments, the first training data comprises input samples and / or corresponding output samples of the multi-edge model.

[0518] In some embodiments, the first training data comprises input samples of the first model and / or output samples of the third model.

[0519] In some embodiments, the first training data comprises input samples of the third model and / or output samples of the first model.

[0520] In some embodiments, if the first model is trained by a first node (the network device 102), the first model is trained according to the first training data. If the first model is trained by multiple nodes including the first node (the network device 102), the model iteration information of the first model is interacted with the second node (the terminal 101) in the process of training the first model according to the first training data.

[0521] In some embodiments, if the first model is trained by one first node, the first node can not interact the model iteration information of the first model with the second node. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102, the network device 102 can not interact the model iteration information of the first model with the terminal 101.

[0522] In some embodiments, if the first model is trained by multiple nodes including the first node, the first node can interact the model iteration information of the first model with the second node. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102 and the terminal 101, the network device 102 can interact the model iteration information of the first model with the terminal 101. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102 and one or more other network devices except the network device 102, the network device 102 can interact the model iteration information of the first model with the terminal 101. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102, the terminal 101 and one or more other terminals except the terminal 101, the network device 102 can interact the model iteration information of the first model with the terminal 101.

[0523] In some embodiments, during the training of the first model according to the first training data, the number of nodes training the first model is one or more default to know whether to interact the model iteration information of the first model with the terminal.

[0524] In some embodiments, during the training of the first model according to the first training data, the number of nodes training the first model can be used to distinguish the model training manner, and then according to the manner of training the first model, it is known by default whether to interact the model iteration information of the first model with the terminal.

[0525] In some embodiments, if the manner of training the first model by the network device is the multi-entity collaborative training manner, the model iteration information of the first model can be interacted with the terminal. Optionally, the first model is at least one partial model in a multi-edge model, the multi-edge model includes a plurality of partial models, and one partial model corresponds to one application node. Optionally, the terminal includes a third model, and the third model is at least one partial model in the multi-edge model except the first model.

[0526] In some embodiments, the gradient information includes forward gradient information and / or backward gradient information. It should be explained that the forward gradient information of the model refers to the process of calculating the output from the input layer to the output layer through the neural network, i.e., forward propagation. And the backward gradient information (i.e., the reverse gradient information) refers to the propagation of errors from the output layer to the input layer, which is used to calculate the gradient of each parameter on the loss, i.e., back propagation.

[0527] In some embodiments, the network device interacts with the terminal to implement the model iteration information of the first model, including at least one of the following:

[0528] The network device sends the forward gradient information to the terminal, and the forward gradient information is output or generated by the first model;

[0529] The network device receives the backward gradient information sent by the terminal, and the backward gradient information is output or generated by the third model;

[0530] The network device receives the forward gradient information sent by the terminal, and the forward gradient information is output or generated by the third model;

[0531] The network device sends the backward gradient information to the terminal, and the backward gradient information is output or generated by the first model.

[0532] After step S2402, at least one of steps S2403-S2407 is selected to be executed. Steps S2403-S2407 are exemplarily described taking the output of the first model as the input of the third model. Steps S2403-S2407 can be executed one or more times.

[0533] In step S2403, the network device 102 inputs the input sample of the first model into the first model.

[0534] In step S2404, the network device 102 sends the prediction data of the first model, the forward gradient information, and the output sample of the third model to the terminal 101.

[0535] In some embodiments, the prediction data of the first model, the forward gradient information, and the output sample of the third model can be sent simultaneously or separately. The present disclosure does not make any limitation in this regard.

[0536] In step S2405, the terminal 101 inputs the prediction data of the first model and the forward gradient information into the third model, determines the error according to the prediction data of the third model and the output sample of the third model, and determines the backward gradient information.

[0537] Optionally, the terminal adjusts the model parameters of the third model according to the backward gradient information.

[0538] In step S2406, the terminal 101 sends the backward gradient information to the network device 102.

[0539] In step S2407, the network device 102 adjusts the model parameters of the first model according to the backward gradient information.

[0540] In some embodiments, steps S2403-S2407 can be replaced by the following steps 1-6, which are exemplarily described taking the output of the third model as the input of the first model as an example.

[0541] Step 1, the network device sends the input sample of the third model to the terminal.

[0542] Step 2, the terminal inputs the input sample of the third model into the third model.

[0543] Step 3, the terminal sends the predicted data of the third model and the forward gradient information to the network device.

[0544] Step 4, the network device inputs the predicted data of the third model and the forward gradient information into the first model, determines the error according to the predicted data of the first model and the output sample of the first model, determines the backward gradient information, and adjusts the model parameters of the first model according to the backward gradient information.

[0545] Step 5, the network device sends the backward gradient information to the terminal.

[0546] Step 6, the terminal adjusts the model parameters of the third model according to the backward gradient information.

[0547] Step S2408, the network device 102 detects that the first condition is met, and sends a seventh message to the terminal 101.

[0548] In some embodiments, the terminal receives the seventh message. For example, the terminal receives the seventh message sent by the network device.

[0549] In some embodiments, the network device determines whether the first condition is met during the process of cooperatively training the first model. For example, real-time determination, for example, determining once when the first model is iterated n (n is an integer greater than 0) times. Optionally, the first condition includes at least one of the following:

[0550] The first model meets the performance requirement;

[0551] The data amount of the first training data reaches a first threshold;

[0552] The number of iterations of the training model reaches a second threshold;

[0553] The time length of the training model reaches a third threshold.

[0554] Among them, at least one of the performance requirement, the first threshold, the second threshold, and the third threshold is protocol-specified preset information.

[0555] In some embodiments, the network device can send a seventh message to the terminal in a case where it is determined that the first condition is met. The seventh message is used for the network device to inform the terminal of the training of the first model.

[0556] In some embodiments, the seventh message comprises at least one of the following:

[0557] an eighth indication for instructing the terminal to stop sending the first data and / or the first signal;

[0558] a ninth indication for informing the terminal that the training of the first model fails and / or the network device will perform a second operation;

[0559] wherein the second operation comprises at least one of the following:

[0560] switching the first model to a second model, the second model having the same function as the first model;

[0561] replacing the first model with a first algorithm, the first algorithm having the same function as the first model.

[0562] In some embodiments, if the application node of the first model is the terminal, the terminal can be sent the first model meeting the performance requirement or the update information of the first model when the trained first model meets the performance requirement.

[0563] In some embodiments, if the first model comprises a first partial model and the application node of the first partial model is the terminal, the terminal can be sent the first partial model in the first model meeting the performance requirement or the update information of the first partial model when the trained first model meets the performance requirement.

[0564] In some embodiments, the names of information, etc. are not limited to the names described in the embodiments, and the terms such as “information”, “message”, “signal”, “signaling”, “report”, “configuration”, “indication”, “instruction”, “command”, “channel”, “parameter”, “domain”, “field”, “symbol”, “symbol”, “codebook”, “codeword”, “code point”, “bit”, “data”, “program”, “chip”, etc. can be replaced with each other.

[0565] In some embodiments, the terms “synchronization signal (SS)”, “synchronization signal block (SSB)”, “reference signal (RS)”, “pilot”, “pilot signal” and the like can be replaced with each other.

[0566] In some embodiments, the terms “moment”, “time point”, “time”, “time position” and the like can be replaced with each other, and the terms “time length”, “time period”, “time window”, “window”, “time” and the like can be replaced with each other.

[0567] In some embodiments, “acquire”, “obtain”, “get”, “receive”, “transmit”, “bidirectional transmission”, “send and / or receive” and the like can be replaced with each other, which can be interpreted as receiving from other subjects, acquiring from protocols, acquiring from higher layers, obtaining by self-processing, implementing autonomously and the like.

[0568] In some embodiments, the terms “send”, “transmit”, “report”, “issue”, “transmit”, “bidirectional transmission”, “send and / or receive” and the like can be replaced with each other.

[0569] In some embodiments, the terms “certain”, “preseted”, “preset”, “set”, “indicated”, “certain”, “arbitrary”, “first” and the like can be replaced with each other, and “certain A”, “preset A”, “preset A”, “set A”, “indicated A”, “certain A”, “arbitrary A”, “first A” can be interpreted as A specified in advance in protocols and the like, can be interpreted as A obtained by setting, configuring or indicating and the like, and can be interpreted as certain A, certain A, arbitrary A or first A, but are not limited thereto.

[0570] In some embodiments, determination or judgment can be performed by a value represented by 1 bit (0 or 1), by a true or false value (Boolean value) represented by true or false, or by comparison of numerical values (for example, comparison with a predetermined value), but is not limited thereto.

[0571] The model training method related to the embodiments of the present disclosure can also be referred to as a multi-entity cooperative training method. The model training method shown in FIG. 2D is applicable to training at least part of the multi-edge model by the terminal and the network device in cooperation, and the at least part of the model is not allowed to be trained by an independent entity.

[0572] The model training method related by the embodiments of the present disclosure can include at least one of steps S2401-S2408. For example, step S2404 can be implemented as an independent embodiment, step S2406 can be implemented as an independent embodiment, step S2404 and step S2406 can be implemented as independent embodiments, but are not limited thereto.

[0573] In some embodiments, the order between any two of steps S2401-S2408 can be exchanged or executed simultaneously.

[0574] In some embodiments, one or more of steps S2401-S2408 are optional, and one or more of these steps can be omitted or replaced in different embodiments.

[0575] In some embodiments, other optional implementations described before or after the description corresponding to FIG. 2D can be referred to.

[0576] In some embodiments, the model training method shown in FIG. 2D can be deformed as long as it is allowed, for example, the network device performs the operation of the second node, the terminal performs the operation of the first node, that is, the operations performed by the terminal and the network device in FIG. 2D are exchanged. For example, the terminal in FIG. 2D is replaced by another network device. For example, the network device in FIG. 2D is replaced by another terminal.

[0577] FIG. 2E is an interaction diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 2D, the embodiments of the present disclosure relate to a model training method, in particular to a model training method for training a first model by a multi-entity joint training method, which is performed by the communication system 100. In the embodiment shown in FIG. 2E, the terminal 101 is the first node, the first node is the training node, the network device 102 is the second node, and the second node is the center node. The model training method shown in FIG. 2E can include the following steps:

[0578] Step S2501, the terminal 101 sends a first message to the network device 102.

[0579] In some embodiments, the network device receives the first message. For example, the network device receives the first message sent by the terminal.

[0580] In some embodiments, the terminal sends the first message to the network device, and the first message includes at least one of the following contents:

[0581] A first request for requesting the network device to allow the terminal to train the first model online;

[0582] a second request for requesting the network device to send first data and / or a first signal to the terminal, wherein the first data and / or the first signal are used for the terminal to collect first training data, and the first training data are used for training the first model;

[0583] performance information of the first model;

[0584] model training capability of the terminal.

[0585] In some embodiments, one or more of the first request, the second request, the performance information of the first model, and the model training capability of the terminal can be sent simultaneously or separately, which is not limited in the present disclosure.

[0586] In some embodiments, since the training manner of the first model is a multi-entity joint training manner, the first request can be understood as a request for the network device to allow the terminal to participate in training the first model, i.e., a request for the network device to allow the terminal to train the first model as a training node.

[0587] It should be explained that the multi-entity joint training manner refers to that multiple entities / nodes jointly train a model, each entity / node independently trains the model according to local training data, each entity / node sends local model updates (such as gradients or weights) to a central server (e.g., the network device), the central server aggregates the model updates from various entities / nodes to calculate the average or weighted average of the global model. The central server sends the updated global model information back to each entity / node. Each entity / node updates the local model according to the global model information and continues the model training, and so on until the model training is completed.

[0588] In some embodiments, the implementation manner of the terminal sending the model training capability of the terminal to the network device can be that the terminal actively reports the model training capability to the network device.

[0589] In some embodiments, the implementation manner of the terminal sending the model training capability of the terminal to the network device can be that the terminal sends the model training capability to the network device in response to receiving a model training capability reporting indication sent by the network device through one or more of a broadcast message, a multicast message, RRC, MAC-CE, and DCI.

[0590] In step S2502, the network device 102 sends a second message to the terminal 101.

[0591] In some embodiments, the terminal receives the second message. For example, the terminal receives the second message sent by the network device.

[0592] In some embodiments, the network device sends the second message to the terminal, and the second message includes at least one of the following:

[0593] a first indication for instructing the terminal to train the first model online and / or for instructing the terminal with initial model parameters of the first model;

[0594] a second indication for instructing the terminal to collect the first training data.

[0595] In some embodiments, one or more of the first indication and the second indication can be sent simultaneously or separately. The present disclosure does not limit this.

[0596] For example, assuming that in step S2501, the first message sent by the terminal to the network device includes the first request, then in step S2502, the second message sent by the network device to the terminal can include the first indication, which can be a response of the network device to the first request. This is a way in which the terminal actively requests the network device to participate in the training of the first model in the multi-entity co-training manner.

[0597] For example, assuming that in step S2501, the first message sent by the terminal to the network device does not include the first request but includes the performance information of the first model, then in step S2502, the second message sent by the network device to the terminal can include the first indication, which can be sent by the network device according to the performance information of the first model reported by the terminal. This is a way in which the terminal passively participates in the training of the first model in the multi-entity co-training manner according to the indication of the network device.

[0598] For example, assuming that in step S2501, the first message sent by the terminal to the network device does not include the first request but includes the model training capability of the terminal, then in step S2502, the second message sent by the network device to the terminal can include the first indication, which can be sent by the network device according to the model training capability of the terminal reported by the terminal. This is a way in which the terminal passively participates in the training of the first model in the multi-entity co-training manner according to the indication of the network device.

[0599] For example, assuming that in step S2501, the first message sent by the terminal to the network device does not include the first request and does not include the performance information of the first model, then in step S2502, the second message sent by the network device to the terminal can include the first indication, which can be sent by the network device to the terminal, for example, the network device sends the second message including the first indication to the terminal when detecting that the preset first model training time is reached. This is a way in which the terminal passively participates in the training of the first model in the multi-entity co-training manner according to the indication of the network device.

[0600] For example, step S2501 can be omitted, and the first indication can be included in the second message sent by the network device to the terminal in step S2502. The first indication can be sent by the network device to the terminal actively, which is a way of the terminal participating in the training of the first model in the multi-entity co-training manner according to the indication of the network device passively.

[0601] It should be noted that in the above examples, the premise of the network device sending the first indication to the terminal is to confirm that the model training capability of the terminal can meet the requirement of training the first model. That is, the terminal is a training node with the capability of multi-entity co-training selected by the network device as the center node.

[0602] In the above embodiment, the network device sending the first indication to the terminal means that the network device as the center node allows the terminal to participate in the training of the first model in the multi-entity co-training manner.

[0603] In some embodiments, assuming that in step S2501, the first message sent by the terminal to the network device includes the second request, then in step S2502, the second message sent by the network device to the terminal can include the second indication, which can be a response of the network device to the second request.

[0604] In some embodiments, in step S2501, the first message sent by the terminal to the network device can not include the second request, and the terminal can obtain the first training data through other ways.

[0605] In step S2503, the terminal 101 collects the first training data.

[0606] In some embodiments, the terminal starts to collect the first training data in response to receiving the second message sent by the network device.

[0607] In some embodiments, the terminal collects the first training data, wherein the first training data includes at least one of the following:

[0608] The first data sent by the network device;

[0609] The second data obtained by measuring the first signal sent by the network device.

[0610] The first signal can be a pilot signal or other signal used by the terminal to obtain the second data.

[0611] In some embodiments, the network device can send the first data to the terminal. In some embodiments, the network device can send the first signal, and the terminal can measure the first signal to obtain the second data.

[0612] In some embodiments, in step S2502, the second indication in the second message can also be used to indicate a resource for sending the first signal, the resource including at least one of a time domain resource, a frequency domain resource, and a spatial domain resource. The implementation of the terminal collecting the first training data includes measuring the first signal on the corresponding time domain, frequency domain, and spatial domain to obtain the second data.

[0613] In some embodiments, in step S2502, the second indication in the second message can also be used to indicate a data format of the first data. The implementation of the terminal collecting the first training data includes receiving the first data in the data format.

[0614] In some embodiments, the terminal can collect the first training data at multiple time instants, the first training data including the first data and / or the second data.

[0615] In some embodiments, the first training data includes input samples and / or corresponding output samples of the first model.

[0616] In step S2504, the terminal 101 trains the first model according to the first training data.

[0617] In some embodiments, in the case where the first model is trained by the first node (the terminal 101), the first model is trained according to the first training data. In the case where the first model is trained by multiple nodes including the first node (the terminal 101), the model iteration information of the first model is interacted with the second node (the network device 102) in the process of training the first model according to the first training data.

[0618] In some embodiments, if the first model is trained by one first node, the first node can not interact the model iteration information of the first model with the second node. For example, assuming that the first node is the terminal 101 and the second node is the network device 102, if the first model is trained by the terminal 101, the terminal 101 can not interact the model iteration information of the first model with the network device 102.

[0619] In some embodiments, if the first model is trained by multiple nodes including the first node, the first node can interact the model iteration information of the first model with the second node. For example, assuming that the first node is the terminal 101 and the second node is the network device 102, if the first model is trained by the terminal 101 and the network device 102, the terminal 101 can interact the model iteration information of the first model with the network device 102. For example, assuming that the first node is the terminal 101 and the second node is the network device 102, if the first model is trained by the terminal 101 and one or more other terminals except the terminal 101, the terminal 101 can interact the model iteration information of the first model with the network device 102. For example, assuming that the first node is the terminal 101 and the second node is the network device 102, if the first model is trained by the terminal 101, the network device 102, and one or more other network devices except the network device 102, the terminal 101 can interact the model iteration information of the first model with the network device 102.

[0620] In some embodiments, in the process of training the first model according to the first training data, it is known by one or more nodes training the first model whether to interact the model iteration information of the first model with the network device according to the number of nodes training the first model.

[0621] In some embodiments, in the process of training the first model according to the first training data, the model training manner can be distinguished according to the number of nodes training the first model, and then it is known by the nodes training the first model whether to interact the model iteration information of the first model with the network device according to the training manner of the first model.

[0622] In some embodiments, if the terminal trains the first model in the multi-entity joint training manner, the terminal can interact the model iteration information of the first model with the network device.

[0623] In some embodiments, the first model is at least one of the following:

[0624] a single-sided model;

[0625] a multi-sided model including multiple partial models, one partial model corresponding to one application node;

[0626] at least one partial model allowing to be independently trained.

[0627] The single-sided model can be a terminal-side model applied to a terminal or a network device-side model applied to a network device.

[0628] In some embodiments, the first node, i.e., the terminal, is a training node, and the second node, i.e., the network device, is a center node.

[0629] In some embodiments, the terminal interacts with the network device to implement the model iteration information of the first model includes steps S2505 and / or S2506.

[0630] In step S2505, the terminal 101 sends the first model iteration information to the network device 102 when the second condition is met.

[0631] In some embodiments, the first model iteration information includes first gradient information and / or first model parameter iteration information. Wherein, the first gradient information is output by the first model, and the first model parameter information is obtained by the terminal after training the first model locally.

[0632] In some embodiments, the gradient information includes forward gradient information and / or backward gradient information. It should be explained that the forward gradient information of the model refers to the process of calculating the output from the input layer to the output layer through the neural network, that is, forward propagation. While the backward gradient information (i.e. backward gradient information) refers to the propagation of errors from the output layer to the input layer for calculating the gradient of each parameter on the loss, that is, back propagation.

[0633] In some embodiments, the second condition includes at least one of the following:

[0634] The time when the first model iteration information is reported;

[0635] Detecting that the error of the first model is in a first range;

[0636] The number of local iterations of the first model reaches a fourth threshold;

[0637] The first node (i.e. the terminal) cannot continue to train the first model.

[0638] In some embodiments, one or more of the above time, the first range, and the fourth threshold are preset values indicated by the network device or specified by the protocol.

[0639] In step S2506, the network device 102 sends the second model iteration information to the terminal 101.

[0640] In some embodiments, the second model iteration information includes second model parameter iteration information.

[0641] In some embodiments, the second model parameter information is determined by the network device according to the first gradient information and / or the first model parameter iteration information uploaded by all training nodes. For example, the network device averages or weighted averages the first gradient information and / or the first model parameter iteration information uploaded by all training nodes to obtain the second model parameter iteration information.

[0642] In some embodiments, the training nodes participating in training the first model can be multiple, and the training nodes can be terminals or network devices. Each training node performs operations similar to those performed by the terminal 101 in this embodiment, and thus the details are not repeated in this embodiment.

[0643] At step S2507, the terminal 101 updates the first model according to the second model parameter iteration information.

[0644] At step S2508, the terminal 101 trains the updated first model according to the first training data.

[0645] In some embodiments, steps S2504 to S2508 can be performed one or more times in a loop.

[0646] In some embodiments, the terminal stops training the first model upon receiving information indicating that the training of the first model is completed, which is sent by the network device. The network device can indicate that the training is completed to the terminal when the first model reaches the performance requirement.

[0647] In some embodiments, if the application node of the first model is the network device, the first model or the update information of the first model that reaches the performance requirement can be sent to the network device when the trained first model reaches the performance requirement.

[0648] In some embodiments, if the first model includes a first partial model and the application node of the first partial model is the network device, the first partial model or the update information of the first partial model that reaches the performance requirement can be sent to the network device when the trained first model reaches the performance requirement.

[0649] In some embodiments, the names of information, etc. are not limited to the names described in the embodiments, and the terms such as “information”, “message”, “signal”, “signaling”, “report”, “configuration”, “indication”, “instruction”, “command”, “channel”, “parameter”, “domain”, “field”, “symbol”, “symbol”, “codebook”, “codeword”, “codepoint”, “bit”, “data”, “program”, “chip”, and the like can be replaced with each other.

[0650] In some embodiments, the terms "synchronization signal (SS)", "synchronization signal block (SSB)", "reference signal (RS)", "pilot", "pilot signal" and the like can be replaced with each other.

[0651] In some embodiments, the terms "moment", "time point", "time", "time position" and the like can be replaced with each other, and the terms "time length", "time period", "time window", "window", "time" and the like can be replaced with each other.

[0652] In some embodiments, the terms "acquire", "obtain", "get", "receive", "transmit", "bidirectional transmission", "send and / or receive" and the like can be replaced with each other, and can be interpreted as receiving from other subjects, acquiring from protocols, acquiring from higher layers, obtaining by self-processing, autonomously implementing and the like.

[0653] In some embodiments, the terms "send", "transmit", "report", "issue", "transmit", "bidirectional transmission", "send and / or receive" and the like can be replaced with each other.

[0654] In some embodiments, the terms "certain", "preset", "preset", "set", "indicated", "certain", "arbitrary", "first" and the like can be replaced with each other, and "certain A", "preset A", "preset A", "set A", "indicated A", "certain A", "arbitrary A", "first A" can be interpreted as A specified in advance in protocols and the like, can be interpreted as A obtained by setting, configuring or indicating and the like, and can be interpreted as certain A, certain A, arbitrary A or first A, but are not limited thereto.

[0655] In some embodiments, determination or judgment can be performed by a value represented by 1 bit (0 or 1), can be performed by a true or false value (Boolean value) represented by true or false, or can be performed by comparison of numerical values (for example, comparison with a predetermined value), but is not limited thereto.

[0656] The model training method related to the embodiments of the present disclosure can also be referred to as a multi-entity joint training method, and the model training method shown in FIG. 2E is applicable to training at least one of the following models by the terminal:

[0657] Unilateral model;

[0658] The multi-edge model includes a plurality of sub-models, and one sub-model corresponds to one application node.

[0659] The at least one sub-model allows independent training by one training node.

[0660] The model training method related to the embodiments of the present disclosure can include at least one of steps S2501-S2508. For example, step S2505 can be implemented as an independent embodiment, step S2506 can be implemented as an independent embodiment, step S2505 and step S2506 can be implemented as independent embodiments, but are not limited thereto.

[0661] In some embodiments, the order between any two of steps S2501-S2508 can be exchanged or executed simultaneously.

[0662] In some embodiments, one or more of steps S2501-S2508 are optional, and one or more of these steps can be omitted or replaced in different embodiments.

[0663] In some embodiments, other optional implementations described before or after the corresponding description of FIG. 2E can be referred to.

[0664] In some embodiments, the model training method shown in FIG. 2E can be modified as long as it is allowed, for example, the network device performs the operation of the first node, the terminal performs the operation of the second node, that is, the operations performed by the terminal and the network device in FIG. 2E are exchanged. For example, the terminal in FIG. 2E is replaced by another network device. For example, the network device in FIG. 2E is replaced by another terminal.

[0665] FIG. 2F is an interaction diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 2F, the embodiments of the present disclosure relate to a model training method, in particular, to a model training method for training a first model by a multi-entity joint training method, which is performed by the communication system 100. In the embodiment shown in FIG. 2F, the terminal 101 is a second node, the second node is a center node, the network device 102 is a first node, and the first node is a training node. The model training method shown in FIG. 2F can include the following steps:

[0666] In step S2601, the terminal 101 sends a fifth message to the network device 102.

[0667] In some embodiments, the network device receives the fifth message. For example, the network device receives the fifth message sent by the terminal.

[0668] In some embodiments, the terminal sends a fifth message to the network device, and the fifth message comprises at least one of the following:

[0669] a third request for requesting the network device to train the first model online;

[0670] performance information of the first model;

[0671] initial model parameters of the first model.

[0672] In some embodiments, one or more of the third request, the performance information of the first model, and the initial model parameters of the first model can be sent simultaneously or separately, which is not limited in the present disclosure.

[0673] In some embodiments, since the training manner of the first model is the multi-entity joint training manner, the third request can be understood as a request for the network device to participate in training the first model, i.e., a request for the network device to train the first model as a training node.

[0674] It should be explained that the multi-entity joint training manner refers to that multiple entities / nodes jointly train a model, each entity / node independently trains the model according to local training data, each entity / node sends local model updates (such as gradients or weights) to a central server (e.g., the terminal), the central server aggregates the model updates from each entity / node to calculate the average or weighted average of the global model. The central server sends the updated global model information back to each entity / node. Each entity / node updates the local model according to the global model information and continues to train the model, and so on until the model training is completed.

[0675] In some embodiments, the terminal sends the third request to the network device on the premise that the model training capability of the network device can meet the requirement of training the first model. That is, the network device is a training node selected by the terminal as the center node and has the capability of multi-entity joint training.

[0676] In some embodiments, the terminal sending the third request to the network device indicates that the terminal as the center node allows the network device to participate in the training of the first model in the multi-entity joint training manner.

[0677] In step S2602, the network device 102 sends a sixth message to the terminal 101.

[0678] In some embodiments, the terminal receives the sixth message. For example, the terminal receives the sixth message sent by the network device.

[0679] In some embodiments, the network device sends the sixth message to the terminal, and the sixth message comprises at least one of the following:

[0680] a sixth indication for instructing the network device to train the first model online;

[0681] a seventh indication for instructing the terminal to send first data and / or a first signal to the network device, wherein the first data and / or the first signal are used by the network device to collect the first training data, and the first training data are used to train the first model.

[0682] In some embodiments, one or more of the sixth indication and the seventh indication can be sent simultaneously or separately. The present disclosure does not limit this.

[0683] For example, assuming that in step S2601, the fifth message sent by the terminal to the network device includes the third request, in step S2602, the sixth indication can be included in the sixth message sent by the network device to the terminal, and the sixth indication can be a response of the network device to the third request. This is a passive way of the network device participating in the training of the first model online in the multi-entity co-training manner according to the request of the terminal.

[0684] For example, assuming that in step S2601, the fifth message sent by the terminal to the network device does not include the third request but includes the performance information of the first model, in step S2602, the sixth indication can be included in the sixth message sent by the network device to the terminal, and the sixth indication can be sent by the network device according to the performance information of the first model reported by the terminal. This is an active way of the network device participating in the training of the first model online in the multi-entity co-training manner.

[0685] For example, step S2601 can be omitted, and in step S2602, the sixth indication can be included in the sixth message sent by the network device to the terminal, and the sixth indication can be sent by the network device to the terminal actively, for example, the network device sends the sixth indication to the terminal actively when detecting that the first model training time reaches a preset value, or the network device sends the sixth indication to the terminal actively when detecting that the performance information of the first model does not meet the requirements. This is an active way of the network device participating in the training of the first model online in the multi-entity co-training manner.

[0686] It should be noted that in the above examples, the seventh indication sent by the network device to the terminal can be that the network device will participate in the training of the first model online in the multi-entity co-training manner by default.

[0687] Step S2603, the network device 102 collects the first training data.

[0688] In some embodiments, the network device starts to collect the first training data when sending the sixth message to the terminal.

[0689] In some embodiments, the network device collects the first training data, wherein the first training data comprises at least one of:

[0690] the first data sent by the terminal;

[0691] the second data obtained by measuring the first signal sent by the terminal.

[0692] The first signal can be a pilot signal or other signal used by the network device to obtain the second data.

[0693] In some embodiments, the terminal can send the first data to the network device. In some embodiments, the terminal can send the first signal, and the network device can measure the first signal to obtain the second data.

[0694] In some embodiments, the network device can collect the first training data at multiple time points, and the first training data comprises the first data and / or the second data.

[0695] In some embodiments, the network device can obtain the first training data in other ways, such as obtaining the first training data from other network devices.

[0696] In some embodiments, the first training data comprises input samples and / or corresponding output samples of the first model.

[0697] Step S2604: The network device 102 trains the first model according to the first training data.

[0698] In some embodiments, the first model is trained according to the first training data when the first model is trained by the first node (the network device 102). When the first model is trained by multiple nodes including the first node (the network device 102), the model iteration information of the first model is interacted with the second node (the terminal 101) during the training of the first model according to the first training data.

[0699] In some embodiments, if the first model is trained by one first node, the first node can not interact the model iteration information of the first model with the second node. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102, the network device 102 can not interact the model iteration information of the first model with the terminal 101.

[0700] In some embodiments, if the first model is trained by multiple nodes including the first node, the first node can interact the model iteration information of the first model with the second node. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102 and the terminal 101, the network device 102 can interact the model iteration information of the first model with the terminal 101. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102 and one or more other network devices other than the network device 102, the network device 102 can interact the model iteration information of the first model with the terminal 101. For example, assuming that the first node is the network device 102 and the second node is the terminal 101, if the first model is trained by the network device 102, the terminal 101, and one or more other terminals other than the terminal 101, the network device 102 can interact the model iteration information of the first model with the terminal 101.

[0701] In some embodiments, during the training of the first model according to the first training data, the number of nodes training the first model is used to determine whether to interact the model iteration information of the first model with the terminal by default.

[0702] In some embodiments, during the training of the first model according to the first training data, the number of nodes training the first model is used to distinguish the model training manner, and then whether to interact the model iteration information of the first model with the terminal is determined according to the training manner of the first model by default.

[0703] In some embodiments, if the training manner of the network device for the first model is the multi-entity joint training manner, the network device can interact the model iteration information of the first model with the terminal.

[0704] In some embodiments, the first model is at least one of:

[0705] a single-sided model;

[0706] a multi-sided model including multiple partial models, one partial model corresponding to one application node;

[0707] at least one partial model allowing independent training.

[0708] The single-sided model can be a terminal-side model applied to the terminal or a network device-side model applied to the network device.

[0709] In some embodiments, the first node, i.e., the network device, is a training node, and the second node, i.e., the terminal, is a central node.

[0710] In some embodiments, the network device interacts with the terminal to implement the first model iteration information includes steps S2605 and / or S2606.

[0711] Step S2605, the network device 102 sends the first model iteration information to the terminal 101 when the second condition is met.

[0712] In some embodiments, the first model iteration information includes first gradient information and / or first model parameter iteration information. Wherein, the first gradient information is output by the first model, and the first model parameter information is obtained by the network device after training the first model locally.

[0713] In some embodiments, the gradient information includes forward gradient information and / or backward gradient information. It should be explained that the forward gradient information of the model refers to the process of calculating the output from the input layer to the output layer through the neural network, that is, forward propagation. While the backward gradient information (i.e. backward gradient information) refers to the propagation of error from the output layer to the input layer, which is used to calculate the gradient of each parameter on the loss, that is, back propagation.

[0714] In some embodiments, the second condition includes at least one of the following:

[0715] The time when the first model iteration information is reported;

[0716] Detecting that the error of the first model is in a first range;

[0717] The number of local iterations of the first model reaches a fourth threshold;

[0718] The first node (i.e. terminal) cannot continue to train the first model.

[0719] In some embodiments, one or more of the above time, the first range, and the fourth threshold are preset values indicated by the network device or specified by the protocol.

[0720] Step S2606, the terminal 101 sends the second model iteration information to the network device 102.

[0721] In some embodiments, the second model iteration information includes second model parameter iteration information.

[0722] In some embodiments, the second model parameter information is determined by the terminal according to the first gradient information and / or the first model parameter iteration information uploaded by all training nodes. For example, the terminal averages or weighted averages all the first gradient information and / or the first model parameter iteration information uploaded by all training nodes to obtain the second model parameter iteration information.

[0723] In some embodiments, the training nodes participating in training the first model can be multiple, and the training nodes can be terminals or network devices. Each training node performs operations similar to those performed by the network device 102 in this embodiment, and this embodiment will not be described again.

[0724] In step S2607, the network device 102 updates the first model according to the second model parameter iteration information.

[0725] In step S2608, the network device 102 trains the updated first model according to the first training data.

[0726] In some embodiments, steps S2604 to S2608 can be executed one or more times in a loop.

[0727] In some embodiments, the network device stops training the first model when it receives information indicating that the first model training ends sent by the terminal. The terminal can indicate that the training ends to the network device when it determines that the first model meets the performance requirement.

[0728] In some embodiments, if the application node of the first model is a terminal, the terminal can be sent the first model meeting the performance requirement or the update information of the first model when the trained first model meets the performance requirement.

[0729] In some embodiments, if the first model includes a first partial model and the application node of the first partial model is a terminal, the terminal can be sent the first partial model in the first model meeting the performance requirement or the update information of the first partial model when the trained first model meets the performance requirement.

[0730] In some embodiments, the names of information, etc. are not limited to the names described in the embodiments. The terms such as "information", "message", "signal", "signaling", "report", "configuration", "indication", "instruction", "command", "channel", "parameter", "domain", "field", "symbol", "symbol", "codebook", "codeword", "code point", "bit", "data", "program", "chip", and the like can be replaced with each other.

[0731] In some embodiments, the terms "synchronization signal (SS)", "synchronization signal block (SSB)", "reference signal (RS)", "pilot", "pilot signal" and the like can be replaced with each other.

[0732] In some embodiments, the terms "moment", "time point", "time", "time position" and the like can be replaced with each other, and the terms "time length", "time period", "time window", "window", "time" and the like can be replaced with each other.

[0733] In some embodiments, the terms "acquire", "obtain", "get", "receive", "transmit", "bidirectional transmission", "send and / or receive" and the like can be replaced with each other, and can be interpreted as receiving from other subjects, acquiring from protocols, obtaining from high layers, obtaining by self-processing, autonomously implementing and the like.

[0734] In some embodiments, the terms "send", "transmit", "report", "issue", "transmit", "bidirectional transmission", "send and / or receive" and the like can be replaced with each other.

[0735] In some embodiments, the terms "certain", "preset", "pre-set", "set", "indicated", "certain", "arbitrary", "first" and the like can be replaced with each other, and "certain A", "preset A", "pre-set A", "set A", "indicated A", "certain A", "arbitrary A", "first A" can be interpreted as A specified in advance in protocols and the like, can be interpreted as A obtained by setting, configuring or indicating and the like, and can be interpreted as certain A, certain A, arbitrary A or first A, but are not limited thereto.

[0736] In some embodiments, the determination or judgment can be performed by a value represented by 1 bit (0 or 1), can be performed by a true or false value (Boolean value) represented by true or false, or can be performed by comparison of numerical values (for example, comparison with a predetermined value), but is not limited thereto.

[0737] The model training method related to the embodiments of the present disclosure can also be referred to as a multi-entity joint training method, and the model training method shown in FIG. 2F is applicable to training at least one of the following models by the terminal:

[0738] Unilateral model;

[0739] The multi-edge model includes a plurality of sub-models, and one sub-model corresponds to one application node.

[0740] The at least one sub-model allows independent training by one training node.

[0741] The model training method according to the embodiments of the present disclosure can include at least one of steps S2601-S2608. For example, step S2605 can be implemented as an independent embodiment, step S2606 can be implemented as an independent embodiment, step S2605 and step S2606 can be implemented as independent embodiments, but are not limited thereto.

[0742] In some embodiments, the order between any two of steps S2601-S2608 can be exchanged or executed simultaneously.

[0743] In some embodiments, one or more of steps S2601-S2608 are optional, and one or more of these steps can be omitted or replaced in different embodiments.

[0744] In some embodiments, other optional implementations described before or after the corresponding description of FIG. 2F can be referred to.

[0745] In some embodiments, the model training method shown in FIG. 2F can be modified as long as it is allowed, for example, the network device performs the operation of the first node, and the terminal performs the operation of the second node, that is, the operations performed by the terminal and the network device in FIG. 2F are exchanged. For example, the terminal in FIG. 2F is replaced by another network device. For example, the network device in FIG. 2F is replaced by another terminal.

[0746] In the embodiments of FIGS. 2A-2F, the first node and the second node are taken as different types of entities as examples. In possible cases, the first node and the second node can be the same type of entity, for example, both are terminals or both are network devices. In the case where the first node and the second node are the same type of entity, the implementation of the model training is similar to the embodiments shown in FIGS. 2A-2F, and the implementation of the first node and the second node as the same type of entity can be obtained by replacing the corresponding execution subject in FIGS. 2A-2F, which will not be described here.

[0747] FIG. 3A is a flow diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 3A, the present disclosure relates to a model training method performed by a first node, and the method includes:

[0748] Step S3101, sending a first message.

[0749] The optional implementation of step S3101 can refer to the optional implementation of step S2101 in FIG. 2A, step S2301 in FIG. 2C, step S2501 in FIG. 2E, and other associated parts in the embodiments related to FIG. 2A, FIG. 2C, and FIG. 2E, which are not described here again.

[0750] In some embodiments, the first node sends the first message to the second node, but is not limited thereto, and can send the first message to other subjects.

[0751] Step S3102, receiving the second message.

[0752] The optional implementation of step S3102 can refer to the optional implementation of step S2102 in FIG. 2A, step S2302 in FIG. 2C, step S2502 in FIG. 2E, and other associated parts in the embodiments related to FIG. 2A, FIG. 2C, and FIG. 2E, which are not described here again.

[0753] In some embodiments, the first node receives the second message sent by the second node, but is not limited thereto, and can receive the second message sent by other subjects.

[0754] In some embodiments, the first node obtains the second message specified by the protocol.

[0755] In some embodiments, the first node obtains the second message from the upper layer(s).

[0756] In some embodiments, the first node processes to obtain the second message.

[0757] In some embodiments, step S3102 is omitted, and the first node autonomously implements the function indicated by the second message, or the above function is default or default.

[0758] Step S3103, collecting the first training data.

[0759] The optional implementation of step S3103 can refer to the optional implementation of step S2103 in FIG. 2A, step S2203 in FIG. 2B, step S2303 in FIG. 2C, step S2402 in FIG. 2D, step S2503 in FIG. 2E, step S2603 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2A, FIG. 2B, FIG. 2C, FIG. 2D, FIG. 2E, and FIG. 2F, which are not described here again.

[0760] Step S3104, training the first model according to the first training data.

[0761] The optional implementation of step S3104 can refer to the optional implementation of step S2104 in FIG. 2A, step S2204 in FIG. 2B, step S2504 in FIG. 2E, step S2604 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2A, FIG. 2B, FIG. 2E, and FIG. 2F, which are not described here again.

[0762] Step S3105: sending the third message.

[0763] The optional implementation of step S3105 can refer to step S2105 in FIG. 2A, the optional implementation of step S2309 in FIG. 2C, and other associated parts in the embodiments related to FIG. 2A and FIG. 2C, which are not described here again.

[0764] In some embodiments, the first node sends the third message to the second node, but is not limited thereto, and can send the third message to other subjects.

[0765] Step S3106: receiving the fourth message.

[0766] The optional implementation of step S3106 can refer to step S2106 in FIG. 2A, the optional implementation of step S2310 in FIG. 2C, and other associated parts in the embodiments related to FIG. 2A and FIG. 2C, which are not described here again.

[0767] In some embodiments, the first node receives the fourth message sent by the second node, but is not limited thereto, and can receive the fourth message sent by other subjects.

[0768] In some embodiments, the first node obtains the fourth message specified by a protocol.

[0769] In some embodiments, the first node obtains the fourth message from upper layer(s).

[0770] In some embodiments, the first node processes to obtain the fourth message.

[0771] In some embodiments, step S3106 is omitted, and the first node autonomously implements the function indicated by the fourth message, or the above function is default or default.

[0772] The model training method related to the embodiments of the present disclosure can include at least one of steps S3101 to S3106. For example, step S3101 can be implemented as an independent embodiment, step S3102 can be implemented as an independent embodiment, and step S3101 and step S3102 can be implemented as independent embodiments, but are not limited thereto.

[0773] In some embodiments, the order between any two of steps S3101-S3106 can be exchanged or performed simultaneously.

[0774] In some embodiments, one or more of steps S3101-S3106 are optional, and one or more of these steps can be omitted or replaced in different embodiments.

[0775] FIG. 3B is a flow diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 3B, the present embodiment relates to a model training method performed by a first node, and the above method comprises:

[0776] Step S3201: receiving a fifth message.

[0777] Optional implementation of step S3201 can refer to optional implementation of step S2201 in FIG. 2B, step S2601 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2B and FIG. 2F, which will not be described here.

[0778] In some embodiments, the first node receives the fifth message sent by the second node, but is not limited thereto, and can also receive the fifth message sent by other nodes.

[0779] In some embodiments, the first node obtains the fifth message as specified by a protocol.

[0780] In some embodiments, the first node obtains the fifth message from upper layer(s).

[0781] In some embodiments, the first node processes to obtain the fifth message.

[0782] In some embodiments, step S3201 is omitted, and the first node autonomously implements the function indicated by the fifth message, or the above function is default or default.

[0783] Step S3202: sending a sixth message.

[0784] Optional implementation of step S3202 can refer to optional implementation of step S2202 in FIG. 2B, step S2401 in FIG. 2D, step S2602 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2B, FIG. 2D, and FIG. 2F, which will not be described here.

[0785] In some embodiments, the first node sends the sixth message to the second node, but is not limited thereto, and can also send the sixth message to other entities.

[0786] Step S3203: collecting a first training data set.

[0787] The optional implementation of step S3203 can be referred to the optional implementation of step S2103 in FIG. 2A, step S2203 in FIG. 2B, step S2303 in FIG. 2C, step S2402 in FIG. 2D, step S2503 in FIG. 2E, step S2603 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2A, FIG. 2B, FIG. 2C, FIG. 2D, FIG. 2E, and FIG. 2F, which are not described herein.

[0788] Step S3204, training the first model according to the first training data set.

[0789] The optional implementation of step S3204 can be referred to step S2104 in FIG. 2A, step S2204 in FIG. 2B, step S2504 in FIG. 2E, the optional implementation of step S2604 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2A, FIG. 2B, FIG. 2E, and FIG. 2F, which are not described herein.

[0790] Step S3205, sending a seventh message.

[0791] The optional implementation of step S3205 can be referred to step S2205 in FIG. 2B, the optional implementation of step S2408 in FIG. 2D, and other associated parts in the embodiments related to FIG. 2B and FIG. 2D, which are not described herein.

[0792] In some embodiments, the first node sends the seventh message to the second node, but is not limited thereto, and can send the seventh message to other entities.

[0793] The model training method related to the embodiments of the present disclosure can include at least one of steps S3201 to S3205. For example, step S3201 can be implemented as an independent embodiment, step S3202 can be implemented as an independent embodiment, and step S3201 and step S3202 can be implemented as independent embodiments, but are not limited thereto.

[0794] In some embodiments, the order between any two steps of steps S3201 to S3205 can be exchanged or executed simultaneously.

[0795] In some embodiments, one or more steps of steps S3201 to S3205 are optional, and one or more steps of these steps can be omitted or replaced in different embodiments.

[0796] FIG. 3C is a flow diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 3C, the model training method related to the embodiments of the present disclosure is executed by a first node, and the above method includes:

[0797] Step S3301, inputting an input sample of the first model into the first model.

[0798] The optional implementation of step S3301 can refer to the optional implementation of step S2304 in FIG. 2C, step S2403 in FIG. 2D, and other associated parts in the embodiments related to FIG. 2C and FIG. 2D, which will not be repeated here.

[0799] In step S3302, the prediction data of the first model, the forward gradient information, and the output sample of the third model are sent.

[0800] The optional implementation of step S3302 can refer to step S2305 in FIG. 2C, the optional implementation of step S2404 in FIG. 2D, and other associated parts in the embodiments related to FIG. 2C and FIG. 2D, which will not be repeated here.

[0801] In step S3303, the backward gradient information is received.

[0802] The optional implementation of step S3303 can refer to step S2307 in FIG. 2C, the optional implementation of step S2406 in FIG. 2D, and other associated parts in the embodiments related to FIG. 2C and FIG. 2D, which will not be repeated here.

[0803] In step S3304, the model parameters of the first model are adjusted according to the backward gradient information.

[0804] The optional implementation of step S3304 can refer to step S2308 in FIG. 2C, the optional implementation of step S2407 in FIG. 2D, and other associated parts in the embodiments related to FIG. 2C and FIG. 2D, which will not be repeated here.

[0805] The model training method related to the embodiments of the present disclosure can include at least one of steps S3301-S3304. For example, step S3302 can be implemented as an independent embodiment, step S3303 can be implemented as an independent embodiment, and step S3302 and step S3303 can be implemented as independent embodiments, but are not limited thereto.

[0806] In some embodiments, the order of any two steps among steps S3301-S3304 can be exchanged or executed simultaneously.

[0807] In some embodiments, one or more steps among steps S3301-S3304 are optional, and one or more steps among these steps can be omitted or replaced in different embodiments.

[0808] In some embodiments, steps S3301-S3304 can be combined with one or more of steps S3101-S3106 in FIG. 3A. Steps S3301-S3304 can be combined with one or more of steps S3201-S3205 in FIG. 3B.

[0809] FIG. 3D is a flow diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 3D, the present embodiment relates to a model training method performed by a first node, and the method comprises:

[0810] Step S3401: sending first model iteration information.

[0811] Optional implementation of step S3401 can refer to optional implementation of step S2505 in FIG. 2E, step S2605 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2E and FIG. 2F, which will not be repeated here.

[0812] Step S3402: receiving second model iteration information.

[0813] Optional implementation of step S3402 can refer to optional implementation of step S2506 in FIG. 2E, step S2606 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2E and FIG. 2F, which will not be repeated here.

[0814] Step S3403: updating the first model according to the second model iteration information.

[0815] Optional implementation of step S3403 can refer to optional implementation of step S2507 in FIG. 2E, step S2607 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2E and FIG. 2F, which will not be repeated here.

[0816] Step S3404: training the updated first model.

[0817] Optional implementation of step S3404 can refer to optional implementation of step S2508 in FIG. 2E, step S2608 in FIG. 2F, and other associated parts in the embodiments related to FIG. 2E and FIG. 2F, which will not be repeated here.

[0818] The model training method related to the embodiments of the present disclosure can comprise at least one of steps S3401-S3404. For example, step S3401 can be implemented as an independent embodiment, step S3402 can be implemented as an independent embodiment, and steps S3401 and S3402 can be implemented as independent embodiments, but are not limited thereto.

[0819] In some embodiments, the order between any two of steps S3401-S3404 can be exchanged or performed simultaneously.

[0820] In some embodiments, one or more of steps S3401-S3404 are optional, and one or more of these steps can be omitted or replaced in different embodiments.

[0821] In some embodiments, steps S3401-S3404 can be combined with one or more of steps S3101-S3106 in FIG. 3A. Steps S3401-S3404 can be combined with one or more of steps S3201-S3205 in FIG. 3B.

[0822] FIG. 3E is a flow diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 3E, the present embodiment relates to a model training method performed by a first node, and the method comprises:

[0823] In step S3501, in the process of training the first model according to the first training data, it is determined whether to interact with a second node the model iteration information of the first model according to a model training manner, the model iteration information being information of a process quantity generated in the process of training the first model.

[0824] The model iteration information comprises at least one of:

[0825] Gradient information;

[0826] Model parameter iteration information.

[0827] Optional implementation of step S3501 can refer to optional implementation of step S2104 in FIG. 2A, step S2204 in FIG. 2B, step S2305 and step S2307 in FIG. 2C, step S2404 and step S2406 in FIG. 2D, step S2505 and step S2506 in FIG. 2E, step S2605 and step S2607 in FIG. 2F, and other related parts in the embodiments related to FIG. 2A, FIG. 2B, FIG. 2C, FIG. 2D, FIG. 2E, and FIG. 2F, which will not be repeated here.

[0828] FIG. 3F is a flow diagram of a model training method according to an embodiment of the present disclosure. As shown in FIG. 3F, the present embodiment relates to a model training method performed by a first node, and the method comprises:

[0829] Step S3601, in a case where the first model is trained by the first node, training the first model according to the first training data; in a case where the first model is trained by multiple nodes including the first node, in a process of training the first model according to the first training data, interacting model iteration information of the first model with the second node.

[0830] The optional implementation of step S3601 can refer to the optional implementation of step S2104 in FIG. 2A, step S2204 in FIG. 2B, step S2305 and step S2307 in FIG. 2C, step S2404 and step S2406 in FIG. 2D, step S2505 and step S2506 in FIG. 2E, step S2605 and step S2607 in FIG. 2F, and other associated parts in the embodiments involved in FIG. 2A, FIG. 2B, FIG. 2C, FIG. 2D, FIG. 2E, and FIG. 2F, which will not be repeated here.

[0831] In the embodiments in FIG. 3A-FIG. 3F, the first node can be a terminal, and the second node can be a terminal. Alternatively, the first node can be a terminal, and the second node can be a network device. Alternatively, the first node can be a network device, and the second node can be a network device. Alternatively, the first node can be a network device, and the second node can be a terminal.

[0832] In some embodiments, the present disclosure proposes corresponding online training workflows for different AI / ML models, realizes online training of corresponding models, or realizes fine-tuning or retraining update of the model in an online training manner. This includes a method or workflow for obtaining an AI / ML model through online training of one node, and a method or workflow for obtaining an AI / ML model through joint online training of multiple nodes, which realizes the purpose of online training of one or multiple nodes for different models.

[0833] In some embodiments, the single-sided model includes a UE-side model and a NW-side model.

[0834] In some embodiments, the training method of the UE-side model is as shown in Embodiment 1 to Embodiment 4.

[0835] Embodiment 1, UE online training UE-side model, the UE initiates the online model training request in the following steps:

[0836] Step S1: The UE initiates a request to the NW side for online training of the UE-side model. Optionally, the UE initiates a request to the NW side for online training of the UE-side model according to the monitored model performance not meeting the requirements.

[0837] Step S2: UE side collects training dataset, or obtains training dataset based on received specific message / signal, which is only for collecting training dataset and is sent by network device. Optionally, NW sends indication information to UE based on received online training request information to indicate that UE is allowed to train UE side model online, and UE starts to collect training dataset after receiving the indication information sent by NW side.

[0838] Step S3: UE side trains UE side model online, and sends indication information to NW side when certain condition is met, which indicates that online training model has ended, or whether UE side model after training can meet the requirement. If the model after training cannot meet the requirement, UE can also report the reason and / or performance of the model after training, for example, the reason is that the number of samples in training dataset is small, the maximum number of iterations for model training is too small, etc. The certain condition is one or more of the following conditions:

[0839] The performance of the model after training has met the requirement;

[0840] The collected training dataset has met the dataset size requirement;

[0841] The maximum number of iterations for training model has been reached;

[0842] The maximum duration of online training has been reached.

[0843] Step S4: If NW side receives the message indicated by UE, which represents that the model after training can meet the performance requirement, it sends information to UE to indicate that UE stops collecting training dataset, or indicates that it no longer sends specific message / signal to UE for collecting training dataset. If NW side receives the message indicated by UE, which represents that the model after training cannot meet the performance requirement and / or the reason for not meeting the performance requirement, it sends information to UE to indicate that UE performs switching model, fallback to traditional algorithm instead of AI / ML model, etc. Optionally, NW can indicate UE to continue collecting training data or configure larger number of iterations, extend the duration of online training, etc. for UE based on received indication information representing the reason for not meeting the performance requirement.

[0844] Embodiment 2, UE trains UE side model online, and the way for NW side to indicate UE to train UE side model online includes the following steps:

[0845] Step S1: UE reports monitored model performance to NW, and NW side decides whether to send indication information to UE based on the monitored model performance. Optionally, NW sends indication information to UE to indicate that UE trains UE side model online.

[0846] Step S2: The UE collects the training data set according to the received indication information.

[0847] Step S3: The UE trains the UE-side model online, and when a certain condition is met, the UE sends indication information to the NW side, which indicates that the online training model has ended, or whether the model after training can meet the requirements. If the model after training cannot meet the requirements, the UE can also report the reasons and / or the performance of the model after training that cannot meet the requirements, such as too few samples in the training data set, too small maximum iteration number for model training, etc. The certain condition is one or more of the following conditions:

[0848] The performance of the model after training has met the requirements;

[0849] The collected training data set has met the data set size requirements;

[0850] The maximum iteration number for training the model has been reached;

[0851] The maximum time length for online training has been reached.

[0852] Step S4: If the NW side receives the message indicated by the UE that the model after training has been able to meet the performance requirements, the NW sends information to the UE to indicate the UE to stop collecting the training data set, or to indicate that the UE no longer receives specific messages / signals for collecting the training data set. If the NW side receives the message indicated by the UE that the model after training cannot meet the performance requirements and / or the reasons for not meeting the performance requirements, the NW sends information to the UE to indicate the UE to perform operations such as switching models, falling back to a traditional algorithm instead of an AI / ML model, etc. Optionally, the NW can instruct the UE to continue collecting training data or configure the UE with a larger iteration number, extend the online training time length, etc. according to the received indication information from the UE indicating the reasons for not meeting the performance requirements.

[0853] Embodiment 3, the method for the UE to send a message to request the NW to train the UE-side model online includes the following steps:

[0854] Step S1: The UE requests the NW to train the UE-side model online. Optionally, the UE requests the NW to train the UE-side model online according to the monitored model performance not meeting the requirements. Optionally, the UE reports the monitored model performance to the NW.

[0855] Step S2: The NW determines whether to train the UE-side model online according to the request message and / or the model performance sent by the UE. If the UE-side model needs to be trained online, the NW sends indication information to the UE that the NW trains the UE-side model online, or the NW sends specific messages to the UE for the UE to collect the training data set.

[0856] Step S3: The UE starts collecting the training dataset according to the received indication information of the online training of the UE-side model by the NW, and sends the training dataset to the NW.

[0857] Step S4: The NW completes the online training of the UE-side model. Optionally, the trained UE-side model meets the performance requirement, the NW indicates the UE that the online training is completed, and delivers the trained UE-side model to the UE. Optionally, the NW sends information to the UE to instruct the UE to stop collecting the training dataset. Optionally, the trained UE-side model does not meet the performance requirement, the NW instructs the UE to perform switching model, fallback to a traditional algorithm instead of the AI / ML model, or delivers other AI / ML model to the UE as a replacement.

[0858] Embodiment 4, the NW online trains the UE-side model, and a method of the NW actively online training the UE-side model includes the following steps:

[0859] Step S1: The NW sends information to the UE to indicate or inform that the NW will online train the UE-side model.

[0860] Step S2: The UE collects the training dataset according to the received indication information and sends it to the NW.

[0861] Step S3: The NW completes the online training of the UE-side AI / ML model. Optionally, the trained UE-side model meets the performance requirement, the NW indicates the UE that the online training is completed, and delivers the trained UE-side model to the UE. Optionally, the NW sends information to the UE to instruct the UE to stop collecting the training dataset. Optionally, the trained UE-side model does not meet the performance requirement, the NW instructs the UE to perform switching model, fallback to a traditional algorithm instead of the AI / ML model, or delivers other AI / ML model to the UE as a replacement.

[0862] In the above-mentioned embodiments 1 to 4, the training dataset collected by the UE side can be sent to the UE side by the NW side, or can be calculated by the UE side.

[0863] In some embodiments, the training of the NW-side model is as shown in the following embodiment 5.

[0864] Embodiment 5, the NW online trains the NW-side model, and a method of the NW actively online training the NW-side model includes the following steps:

[0865] Step S1: The NW informs one or more UEs through a broadcast message that the NW will online train the NW-side model, or the NW sends indication information to one or a group of UEs to indicate that the NW will online train the NW-side model.

[0866] Step S2: The UE sends a message / signal to the NW for the NW to collect training data according to the received broadcast message or indication information. Optionally, the UE sends the training data to the NW. Optionally, the UE sends a pilot signal to the NW to facilitate the NW to collect training data.

[0867] Step S3: The NW performs online training according to the collected training data set, stops training when a certain condition is met, and sends indication information to the UE to stop online training of the NW side model. Optionally, the stopping training condition is one or more of the following conditions:

[0868] The performance of the trained model has met the requirements;

[0869] The collected training data set has met the data set size requirement;

[0870] The maximum number of iterations of the training model has been reached;

[0871] The maximum duration of online training has been reached.

[0872] Step S4: The UE stops sending training data and / or pilot signals to the NW according to the received indication information indicating that online training has stopped.

[0873] In the above embodiment 5, the training data set collected by the NW side can be sent to the NW side by the UE side, or can be measured and calculated by the NW side.

[0874] In the above embodiments 1 to 5, the training data sent by the UE side to the NW side, or the training data set sent by the NW side to the UE side, can be sent to the opposite side at one or more times.

[0875] In some embodiments, the multi-edge model includes a double-edge model, and the training method of the double-edge model is shown in the following embodiments 6-16.

[0876] Embodiment 6, the UE initiatively trains the complete double-edge model in the following steps:

[0877] Step S1: The UE initiatively initiates a request to the NW for online training of a double-edge model.

[0878] Step S2: The UE collects a training data set, or obtains a training data set based on a received specific message / signal, the specific message / signal being a message / signal for collecting a training data set only and sent by a network device. Optionally, the NW sends indication information to the UE to indicate that the UE is allowed to train the double-edge model online according to the received online training request information, and the UE starts collecting the training data set after receiving the indication information sent by the NW.

[0879] Step S3: UE trains the bilateral model online, and sends an indication information to the NW when certain conditions are met, which indicates that the online training of the bilateral model has ended, or whether the bilateral model after training can meet the requirements. If the bilateral model after training cannot meet the requirements, the UE can also report the reasons and / or the performance of the trained model, such as too few samples in the training data set, too small maximum iteration number of model training, etc. The certain conditions are one or more of the following conditions:

[0880] The performance of the trained model has met the requirements;

[0881] The collected training data set has met the data set size requirements;

[0882] The maximum iteration number of training model has been reached;

[0883] The maximum time length of online training has been reached.

[0884] Step S4: If the NW receives the message from the UE indicating that the trained bilateral model has met the performance requirements, the NW sends information to the UE to indicate that the UE stops collecting the training data set, or indicates that the UE no longer receives specific messages / signals for collecting the training data set. If the NW receives the message from the UE indicating that the trained bilateral model cannot meet the performance requirements and / or the reasons for not meeting the performance requirements, the NW sends information to the UE to indicate that the UE performs switching model, fallback to traditional algorithm instead of AI / ML model, etc. It should be noted that the UE switching model refers to the UE switching the part of the model or component applied to the UE side in the bilateral model, and the UE using the traditional algorithm to replace the part of the model or component applied to the UE side in the bilateral model. Alternatively, the NW can instruct the UE to continue collecting data or configure a larger iteration number, extend the online training time length, etc. according to the indication information received from the UE indicating the reasons for the trained model not meeting the performance requirements.

[0885] Optionally, in the case where the trained bilateral model has met the performance requirements, the part of the model or component applied to the NW in the bilateral model is sent to the NW.

[0886] Embodiment 7, the passive independent online training of the complete bilateral model by the UE includes the following steps:

[0887] Step S1: The UE reports the monitored model performance to the NW, and the NW decides whether to send an indication information to the UE according to the monitored model performance. Optionally, the NW sends an indication information to the UE to instruct the UE to train the bilateral model online. The UE can monitor the performance of the part of the model or component applied to the UE side in the bilateral model, or the UE can monitor the performance of the entire bilateral model.

[0888] Step S2: The UE collects the training data set according to the received indication information.

[0889] Step S3: The UE trains the bilateral model online, and when a certain condition is met, sends indication information to the NW, which indicates that the online training of the bilateral model has ended, or whether the bilateral model after training can meet the demand. If the bilateral model after training cannot meet the demand, the UE can also report the reason why the demand cannot be met and / or the performance of the model after training. The reason may be, for example, that the sample in the training data set is small, the maximum number of iterations of model training is too small, etc. The certain condition is one or more of the following conditions:

[0890] The performance of the model after training has met the requirement;

[0891] The collected training data set has met the data set size requirement;

[0892] The maximum number of iterations of training the model has been reached;

[0893] The maximum duration of online training has been reached.

[0894] Step S4: If the NW receives the message indicated by the UE that the model after training has been able to meet the performance requirement, it sends information to the UE to indicate that the UE stops collecting the training data set, or indicates that it no longer sends specific messages / signals to the UE for collecting the training data set. If the NW receives the message indicated by the UE that the model after training cannot meet the performance requirement and / or the reason why the performance requirement is not met, it sends information to the UE to indicate that the UE performs operations such as switching models, falling back to a traditional algorithm instead of an AI / ML model, etc. It should be noted that switching models by the UE refers to switching the part of the model or component in the bilateral model that is applied to the UE side, and replacing the model by the UE with a traditional algorithm refers to replacing the part of the model or component in the bilateral model that is applied to the UE side with a traditional algorithm. Alternatively, the NW can instruct the UE to continue collecting training data or configure a larger number of iterations for the UE, extend the online training time, etc. according to the received indication information indicating the reason why the performance requirement is not met.

[0895] Alternatively, in the case where the bilateral model after training has been able to meet the performance requirement, the part of the model or component in the bilateral model that is applied to the NW is sent to the NW.

[0896] Embodiment 8, the passive way of the NW independently training the complete bilateral model includes the following steps:

[0897] Step S1: The UE requests the NW to train the bilateral model online.

[0898] Step S2: The NW determines to train the bilateral model online according to the request message sent by the UE. Optionally, the NW sends an indication information to the UE that the NW agrees to train the bilateral model online, or the NW sends a specific message to the UE for the UE to collect the training dataset.

[0899] Step S3: The UE starts to collect the training dataset according to the received indication information of the NW training the bilateral model online, and sends to the NW.

[0900] Step S4: The NW completes the online training of the bilateral model. Optionally, the trained bilateral model meets the performance requirement, the NW indicates to the UE that the online training is completed, and delivers the part of the model or component applied to the UE side in the trained bilateral model to the UE. Optionally, the NW sends information to the UE to instruct the UE to stop collecting the dataset. Optionally, the trained model does not meet the performance requirement, the NW instructs the UE to switch the model, fallback to the traditional algorithm instead of the AI / ML model, or deliver other AI / ML model to the UE. It should be noted that the UE switching the model means that the UE switches the part of the model or component applied to the UE side in the bilateral model, and the UE using the traditional algorithm to replace the model means that the UE uses the traditional algorithm to replace the part of the model or component applied to the UE side in the bilateral model.

[0901] Embodiment 9, the NW initiatively trains the complete bilateral model independently in the following steps:

[0902] Step S1: The NW sends information to the UE to indicate or inform that the NW will train the bilateral model online.

[0903] Step S2: The UE collects the dataset according to the received indication information and sends to the NW.

[0904] Step S3: The NW completes the online training of the bilateral model. Optionally, the trained bilateral model meets the performance requirement, the NW indicates to the UE that the online training is completed, and delivers the part of the model or component applied to the UE side in the trained bilateral model to the UE. Optionally, the NW sends information to the UE to instruct the UE to stop collecting the dataset. Optionally, the trained bilateral model does not meet the performance requirement, the NW instructs the UE to switch the model, fallback to the traditional algorithm instead of the AI / ML model, or deliver other AI / ML model to the UE. It should be noted that the UE switching the model means that the UE switches the part of the model or component applied to the UE side in the bilateral model, and the UE using the traditional algorithm to replace the model means that the UE uses the traditional algorithm to replace the part of the model or component applied to the UE side in the bilateral model.

[0905] Embodiment 10, UE and NW jointly online train bilateral model, only one end of the partial model updates model parameters during online joint training, and the other end of the partial model does not update model parameters. The online training method in which only the UE side of the partial model updates model parameters includes the following steps:

[0906] Step S1: The UE initiates a request to the NW for online cooperative training of the first partial model in the bilateral model.

[0907] Step S2: The UE collects a training data set, or obtains a training data set based on a received specific message / signal, which is a message / signal for collecting a training data set only and sent by a network device. Optionally, the NW sends indication information to the UE to indicate agreement to online cooperative training of the first partial model in the bilateral model according to the received online training request information, and the UE starts collecting the training data set after receiving the indication information sent by the NW.

[0908] Step S3: The UE online trains the first partial model, and when a certain condition is met, sends indication information to the NW, which indicates that the online training model has ended, or whether the trained first partial model can meet the demand. If the trained first partial model cannot meet the demand, the UE can also report the reason and / or the performance of the trained model that cannot meet the demand, for example, the training data set is small, the maximum number of model training iterations is too small, etc. The certain condition is one or more of the following conditions:

[0909] The performance of the trained model has met the requirement;

[0910] The collected training data set has met the data set size requirement;

[0911] The maximum number of iterations of the training model has been reached;

[0912] The maximum duration of online training has been reached.

[0913] Optionally, during training, the UE sends model iteration information output by the first partial model on the UE side to the NW, such as sending forward gradient message output by the first partial model to the NW, for the second partial model on the NW side to output backward gradient information, the NW sends the backward gradient information to the UE, and the UE updates the model parameters of the first partial model.

[0914] Optionally, during training, the UE receives model iteration information output by the second partial model on the NW side sent by the NW, such as the NW sending forward gradient message output by the second partial model on the NW side to the UE, for the first partial model on the UE side to output backward gradient information, and the UE updates the model parameters of the first partial model.

[0915] Step S4: If the NW receives the message from the UE indicating that the trained first part model is able to meet the performance requirement, the NW sends information to the UE to indicate the UE to stop collecting the training dataset, or to indicate that the specific message / signal for collecting the training dataset is no longer sent to the UE. If the NW receives the message from the UE indicating that the trained first part model is not able to meet the performance requirement and / or the reason for not meeting the performance requirement, the NW sends information to the UE to indicate the UE to perform the operation of switching the first part model, fallback to the traditional algorithm instead of the first part model, etc. Optionally, the NW can indicate the UE to continue collecting data or configure the UE with a larger number of iterations, extend the online training time, etc. according to the received indication information indicating the reason for not meeting the performance requirement.

[0916] Embodiment 11, UE and NW jointly online train the bilateral model, only one end of the part model performs model parameter update during online joint training, and the part model of the other end does not update the model parameter. The online training mode in which only the part model of the UE side performs model parameter update includes the following steps:

[0917] Step S1: The UE reports the monitored performance of the first part model of the UE side to the NW, and the NW side decides whether to send indication information to the UE according to the model performance monitored by the UE. Optionally, the NW sends indication information to the UE to indicate the UE to cooperate in online training of the first part model in the bilateral model.

[0918] Step S2: The UE collects the training dataset according to the received indication information.

[0919] Step S3: The UE online trains the first part model in the bilateral model, and when a certain condition is met, sends indication information to the NW, which indicates that the online training of the model has ended, or whether the trained first part model is able to meet the requirement. If the trained first part model is not able to meet the requirement, the UE can also report the reason for not meeting the requirement and / or the performance of the trained model, and the reason can be, for example, that the training dataset sample is small, the maximum number of iterations of model training is too small, etc. The certain condition is one or more of the following conditions:

[0920] The performance of the trained model has met the requirement;

[0921] The collected training dataset has met the dataset size requirement;

[0922] The maximum number of iterations of the training model has been reached;

[0923] The maximum duration of online training has been reached.

[0924] Optionally, during the training process, the UE sends model iteration information of the first part model output on the UE side to the NW, such as sending the forward gradient message of the first part model output to the NW, and the NW sends backward gradient information of the second part model output on the NW side to the UE, and the UE updates the model parameters of the first part model.

[0925] Optionally, during the training process, the UE receives the model iteration information of the second part model output on the NW side sent by the NW, such as the NW sending the forward gradient message of the second part model output on the NW side to the UE, and the UE sending the backward gradient information of the first part model output on the UE side, and the UE updates the model parameters of the first part model.

[0926] Step S4: If the NW receives the message indicated by the UE that the trained first part model can meet the performance requirements, the NW sends information to the UE to indicate the UE to stop collecting the training data set, or to indicate that the specific message / signal for collecting the training data set is no longer sent to the UE. If the NW receives the message indicated by the UE that the trained first part model cannot meet the performance requirements and / or the reason for not meeting the performance requirements, the NW sends information to the UE to indicate the UE to perform operations such as switching the first part model, falling back to a traditional algorithm instead of the first part model, etc. Optionally, the NW can indicate the UE to continue collecting data or configure a larger number of iterations for the UE, extend the online training time, etc. according to the received indication information indicating the reason for not meeting the performance requirements.

[0927] Embodiment 12, UE and NW jointly online train the bilateral model, and only one end of the part model performs model parameter update during online joint training, and the other end of the part model does not update the model parameter. The online training method in which only the second part model on the NW side performs model parameter update includes the following steps:

[0928] Step S1: The NW informs one or more UEs through a broadcast message, or the NW sends indication information to one or more UEs to indicate that the NW will cooperate with the UE to online train the second part model in the bilateral model.

[0929] Step S2: The UE sends inference information, measurement information, or a training data set of the first part model on the UE side to the NW according to the received broadcast message or indication information, which are used as training data sets collected on the NW side. The UE can send the inference information and the measurement information at the same time, or can send the inference information and the measurement information respectively. The UE inference information is obtained by the first part AI / ML model on the UE side, the UE measurement information is calculated based on the traditional method, and the training data set is the data output by the part model on the UE side for training the model on the NW side. For example, the training data set is the backward gradient message / forward gradient message sent by the UE.

[0930] Step S3: The NW performs online training according to the collected data set, stops training when the model of online training meets the performance requirement or certain conditions, and sends indication information of stopping online training to the UE. The indication information can also be sent to each UE in a broadcast manner, or sent to a certain group of UEs or a certain UE. The stopping training conditions are one or more of the following conditions:

[0931] The performance of the trained model has met the requirement;

[0932] The collected training data set has met the data set size requirement;

[0933] The maximum number of iterations of the training model has been reached;

[0934] The maximum time of online training has been reached.

[0935] Step S4: After receiving the online training stop indication information, the UE stops sending measurement information or training data set to the NW side.

[0936] Optionally, during the training process, the NW sends model iteration information of the second part model output on the NW side to the UE, such as sending the forward gradient message of the second part model output on the NW side to the UE, for the UE side to output the backward gradient information of the first part model, the UE sends the backward gradient information of the first part model output on the UE side to the NW, and the NW updates the model parameters of the second part model.

[0937] Optionally, during the training process, the NW receives the model iteration information of the first part model output on the UE side sent by the UE, such as sending the forward gradient message of the first part model output by the UE to the NW, for the NW side to output the backward gradient information of the second part model, and the NW updates the model parameters of the second part model.

[0938] Embodiment 13, UE and NW jointly online train the bilateral model, and the partial models on both ends are updated during online joint training. Including the following steps:

[0939] Step S1: The UE initiates a request for online training model to the NW. Optionally, the NW sends indication information to the UE to indicate agreement to cooperate with the UE to train the bilateral model online according to the received online training request information.

[0940] Step S2: The UE collects a first training data set. The first training data set is input into the first part model, and the gradient information of the first part model is sent to the NW. Optionally, the first training data set is sent by the NW, including the input sample and / or output sample of the first part model.

[0941] Step S3: The NW collects a second training dataset. The second training dataset is input into the second partial model, and gradient information of the second partial model is sent to the UE. Optionally, the second training dataset is sent by the UE and includes input samples and / or output samples of the second partial model.

[0942] Step S4: The parameters of the first partial model and the second partial model are updated. When a certain condition is met, indication information is sent to the opposite side, indicating that the online training model has ended, or whether the trained model can meet the demand. If the trained model cannot meet the demand, the UE side can also report the reason and / or the performance of the trained model, for example, the training dataset sample is small, the maximum iteration number of model training is too small, etc. The certain condition is one or more of the following conditions:

[0943] The performance of the trained model has met the requirement;

[0944] The collected training dataset has met the dataset size requirement;

[0945] The maximum iteration number of the training model has been reached;

[0946] The maximum time of online training has been reached.

[0947] Step S5: If the NW side receives the message indicated by the UE side that the trained model has met the performance requirement, information is sent to the UE to indicate that the UE side stops collecting the training dataset, or to indicate that no specific message for collecting the training dataset is sent to the UE. If the NW side receives the message indicated by the UE side that the trained model cannot meet the performance requirement and / or the reason for not meeting the performance requirement, information is sent to the UE to indicate that the UE side performs switching model, fallback to traditional algorithm instead of AI / ML model, etc. Optionally, the NW can indicate the UE to continue collecting data or configure a larger iteration number, extend the online training time, etc. for the UE according to the received indication information indicating the reason for not meeting the performance requirement.

[0948] Embodiment 14, UE and NW jointly online train a bilateral model, and the partial models at both ends are updated during online joint training. It includes the following steps:

[0949] Step S1: The NW sends indication information to the UE to indicate that a bilateral model will be trained online with the UE.

[0950] Step S2: The NW collects a first training dataset. The first training dataset is input into the second partial model, and gradient information of the second partial model is sent to the UE. Optionally, the first training dataset is sent by the UE and includes input samples and / or output samples of the second partial model.

[0951] Step S3: UE collects a second training dataset. The second training dataset is input into the first part model, and the gradient information of the first part model is sent to the NW. Optionally, the second training dataset is sent by the NW, including the input sample and / or output sample of the first part model.

[0952] Step S4: The parameters of the first part model and the second part model are updated. When a certain condition is met, an indication information is sent to the opposite side, indicating that the online training model has ended, or whether the trained model can meet the demand. If the trained model cannot meet the demand, the UE side can also report the reason for not meeting the demand and / or the performance of the trained model, for example, the training dataset sample is small, the maximum iteration number of model training is too small, etc. The certain condition is one or more of the following conditions:

[0953] The performance of the trained model has met the requirement;

[0954] The collected training dataset has met the dataset size requirement;

[0955] The maximum iteration number of the training model has been reached;

[0956] The maximum time of online training has been reached.

[0957] Step S5: If the NW side receives the message indicated by the UE side that the trained model has met the performance requirement, the information is sent to the UE to indicate the UE side to stop collecting the training dataset, or to indicate that the specific message for collecting the training dataset is no longer sent to the UE. If the NW side receives the message indicated by the UE side that the trained model cannot meet the performance requirement and / or the reason for not meeting the performance requirement, the information is sent to the UE to indicate the UE side to perform the switching model, fallback to the traditional algorithm instead of the AI / ML model, etc. Optionally, according to the received indication information indicating the reason for not meeting the performance requirement, the NW can instruct the UE to continue collecting data or configure a larger iteration number for the UE, extend the online training time, etc.

[0958] In the embodiment 15, when the first part model applied to the UE side in the bilateral model can be independently trained, the implementation of the UE online training the first part model on the UE side can refer to the above-mentioned embodiments 1 to 4, which will not be repeated here.

[0959] In the embodiment 16, when the second part model applied to the NW side in the bilateral model can be independently trained, the implementation of the NW online training the second part model on the NW side can refer to the above-mentioned embodiment 5, which will not be repeated here.

[0960] In some embodiments, the multi-edge model further includes a more-than-two-edge model. The training of the more-than-two-edge model is similar to that of embodiments 6 to 16. Briefly, during the training of the more-than-two-edge model, the forward gradient information of the upstream part model is passed to the downstream part model, and the backward gradient information of the downstream part model is passed to the upstream part model.

[0961] Embodiment 17, the training node actively reports whether it has the ability to jointly train the AI model in the following steps:

[0962] Step S1: A node (such as a UE node) that can participate in distributed multi-node joint training of an AI model notifies a center node whether it has the ability to jointly train the AI model by sending first indication information or capability indication information. The center node can be a node on the NW side, or a center server node, etc. A node with the ability to jointly train the AI model is defined as a first node.

[0963] Step S2: The center node determines which nodes are the first nodes according to the received first indication information or capability indication information. The center node sends second indication information of online model training to the first nodes, which indicates that the first nodes participate in joint training of the AI model. The center node also sends first reference information to the first nodes, which is used by the nodes to obtain a data set for training the AI model. Optionally, the center node transmits the AI model or the parameters of the AI model to the first nodes.

[0964] Optionally, the center node can send the second indication information and the first reference information by broadcasting or groupcasting, or one or more of RRC / MAC-CE / DCI.

[0965] Step S3: The first nodes measure the training data set based on the received first reference information, and then train the AI model according to the received second indication information. The first nodes send the updated data (such as gradients or weights) of the training model to the center node.

[0966] Optionally, the first nodes actively send the updated data to the center node.

[0967] Optionally, the first nodes send the updated data to the center node after receiving the indication information of the updated data reported by the center node.

[0968] Optionally, the first nodes send the updated data to the center node according to an event trigger. The trigger event can be one or more of the following events:

[0969] The time to report the updated data has arrived. For example, an updated data reporting period has arrived;

[0970] The loss function of training the AI model refers to reaching a preset threshold or the number of iterations of training reaching a preset value. The preset value or threshold is configured by the center node or determined by the protocol.

[0971] The first node cannot continue to train the AI model due to hardware or software reasons.

[0972] Step S4: The center node updates the AI model based on the received update data. The center node sends the model parameters of the updated AI model to the first node. Optionally, if the center node monitors that the updated AI model meets the performance requirement, the center node sends a third indication information to the first node to stop training the AI model. The center node can send the third indication information through multicast, broadcast or RRC / MAC-CE / DCI.

[0973] Step S5: The first node updates the AI model locally according to the received model parameters, and then restarts to execute step S3 to continue training the AI model. If the first node receives the third indication information, the first node stops online training of the AI model.

[0974] Embodiment 18, the center node queries the nodes with the capability of jointly training the AI model in the following steps:

[0975] Step S1: The center node queries which nodes are the first nodes by sending one or more of broadcast / multicast messages or RRC / MAC-CE / DCI.

[0976] Step S2: Each node notifies the center node by sending the first indication information or the capability indication information according to the received query information.

[0977] Step S3: The center node determines which nodes are the first nodes according to the received first indication information or the capability indication information. The center node sends the second indication information of online model training to the first nodes, which indicates that the first nodes participate in jointly training the AI model. The center node also sends the first reference information to the first nodes, which is used for the nodes to obtain the data set of training the AI model. Optionally, the center node transmits the AI model or the parameters of the AI model to the first nodes.

[0978] Optionally, the center node can send the second indication information and the first reference information through broadcast or multicast, or one or more of RRC / MAC-CE / DCI.

[0979] Step S4: The first node measures the training data set based on the received first reference information, and trains the AI model according to the received second indication information. The first node sends the updated data (such as gradient or weight) of the trained model to the center node.

[0980] Optionally, the first node actively sends the updated data to the center node.

[0981] Optionally, the first node sends the updated data to the center node after receiving the indication information reported by the updated data sent by the center node.

[0982] Optionally, the first node sends the updated data to the center node according to the triggering event. The triggering event can be one or more of the following events:

[0983] The time of reporting the updated data is reached. For example, an updated data reporting period is reached;

[0984] The loss function of training the AI model reaches a preset threshold, or the number of iterations of training reaches a preset value. The preset value or threshold is configured by the center node or determined by the protocol;

[0985] The first node cannot continue to train the AI model due to hardware or software reasons.

[0986] Step S5: The center node updates the AI model based on the received updated data, and then sends the model parameters of the updated AI model to the first node. Optionally, if the center node monitors that the updated AI model meets the performance requirement, the center node sends third indication information to the first node to stop training the AI model. The center node can send the third indication information through multicast, broadcast or RRC / MAC-CE / DCI.

[0987] Step S6: The first node locally updates the AI model according to the received model parameters, and then restarts step S4 to continue training the AI model. If the first node receives the third indication information, the first node stops online training the AI model.

[0988] In some embodiments, the present disclosure also provides example 1: assuming that an AI / ML model has been deployed on the UE side, the AI / ML is obtained through offline training. Due to the influence of the channel change experienced by the UE, the deployed AI / ML model needs to be trained through online training to match the current channel situation, so as to improve the inference performance of the AI / ML model. The online training of the UE-side model can be achieved through the following steps.

[0989] The way of online training the UE-side model on the UE side includes the following steps:

[0990] Step S1: Assuming that the UE side monitors the performance of the AI / ML model and the UE side has the capability of online training of the AI / ML model, the UE side sends a request message to the NW side through uplink signaling such as RRC / MAC-CE / UCI, etc. L1 or L3 signaling, which indicates the request of the UE online training model.

[0991] Step S2: The NW sends a specific message / signal to the UE for data collection of the UE side according to the received request message, such as sending a specific downlink pilot signal such as CSI-RS, DMRS, etc. The specific downlink pilot can be a downlink pilot with larger power, which is different from the pilot signal used for inference, so that the UE obtains a higher quality training data set. The UE obtains the training data set by measuring the received downlink pilot. Optionally, the NW side sends an indication information to the UE indicating that the UE starts to collect the training data set.

[0992] Step S3: The UE side performs online model training based on the collected data set, and the AI / ML model obtained after training meets the model performance requirements. The model performance can be configured by the NW side to the UE, or determined by the UE and the NW through negotiation. Then, the UE sends an indication information to the NW indicating that the model training is completed and meets the performance requirements. If the trained model cannot meet the requirements, the UE side can also report the reasons why the trained model cannot meet the requirements and / or the performance of the trained model, such as too small training data set sample, too small maximum iteration number, etc.

[0993] Step S4: After the NW receives the AI / ML model trained online by the UE side, the NW sends an information to the UE indicating to stop collecting the data set, or no longer sends specific messages to the UE for collecting the data set, such as the NW side no longer sends downlink pilot signals to the UE with larger power. If the NW side receives the indication from the UE that the trained model still cannot meet the performance requirements, the NW sends an information to the UE indicating that the UE falls back to the traditional algorithm instead of the AI / ML model. If the UE indicates that the trained model does not meet the requirements because the training data set is too small, the NW sends an indication information to the UE indicating that the UE continues to collect the training data set.

[0994] The above is an example of UE side initiating online model training. For the NW side sending online model training, the steps are similar to the above steps, except that step S1 is determined by the NW side whether to perform online model training according to the model performance monitoring result. The other steps are the same, which will not be repeated here.

[0995] In some embodiments, the model of the UE side can not only be trained online by the UE side, but also be trained online by the NW side. The steps of the NW side training the UE side model online are as follows:

[0996] Step S1: Assuming that the UE side monitors the performance of the AI / ML model, the UE side sends a message to the NW side indicating that the NW side should perform online training of the model according to the monitored model performance. Optionally, the UE side reports the performance of the monitored model to the NW side through uplink signaling such as RRC / MAC-CE / UCI. The performance of the monitored model can be an intermediate performance indicator such as SGCS, or a final performance indicator such as throughput.

[0997] Step S2: The NW side determines whether to perform online training based on the current computing power or the received model performance. If the NW side determines that online training is needed, the NW side sends an indication message to the UE through one or more of RRC / MAC-CE / DCI signaling, indicating that the NW side will perform online training of the model. Alternatively, the NW side sends a specific message to the UE for collecting a training data set.

[0998] Step S3: The UE collects a training data set according to the received indication information or specific message, and sends the collected data set to the NW through one or more of RRC / MAC-CE / UCI signaling, for online model training by the NW side.

[0999] Step S4: The NW performs online training of the model based on the received data. When the trained model meets the performance requirements, the NW indicates to the UE that the online training is complete, and transfers the trained model to the UE. Optionally, the NW side sends other indication information to the UE, such as indicating the UE side to stop data set collection, or switching models or reverting to a traditional algorithm instead of an AI / ML model, or transferring other AI / ML models to the UE, etc.

[1000] For the step of the NW side actively initiating online training of the model, the steps are similar to the above steps, except that step S1 is actively sending information to the UE by the NW side to indicate online training of the NW side. The other steps are the same and will not be repeated here.

[1001] In some embodiments, the present disclosure also provides Example 2: Assuming that the UE and the NW respectively deploy Encoder and Decoder models to implement CSI compression feedback, wherein the Encoder is used to implement CSI compression feedback, and the Decoder is used to recover CSI. Referring to the schematic diagram of FIG. 4, which shows a CSI compression feedback and recovery based on a double-sided AI / ML model, the UE side compresses the downlink channel information H through a CSI generation partial model, and then quantizes it into a binary bit stream to send to the gNB. The gNB side recovers H' that is similar to the original downlink information through a CSI recovery partial model.

[1002] If the Encoder and Decoder are trained jointly by the UE and the NW side, and the Encoder on the UE side or the Decoder on the NW side needs to be trained online at the same time to update the single-sided part of the model. For the training method of updating the model parameters of the Decoder model on the NW side only, the following steps are included:

[1003] Step S1: The NW sends a broadcast message by sending a SIB containing an information field for indicating the online training of the decoder model.

[1004] Step S2: The UE sends the Encoder compressed CSI information, the channel information measured by the UE, and the forward gradient information output in the process of training the Encoder model to the NW as the data set for training the Decoder according to the received broadcast message. The UE sends these training data sets through one or more of RRC / MAC-CE / UCI signaling.

[1005] Step S3: The NW side trains the Decoder model online according to the received data set, and updates the model parameters of the Decoder model.

[1006] Step S4: If the model trained on the NW side meets the performance requirements, the NW sends a broadcast message to each UE to instruct each UE to stop collecting data sets. After receiving the online training stop indication information, the UE stops sending the channel information measured by the UE and the forward gradient information output in the process of training the Encoder model to the NW. However, the UE can still continue to send the compressed CSI to the NW for the NW to recover the CSI based on the trained Decoder.

[1007] In some embodiments, the present disclosure also provides Example 3: assuming that the center node is a gNB device, and the other nodes participating in the joint training of the AI model are UE devices. Each UE can help the gNB to train an AI / ML model online through federated learning, and the specific implementation steps are as follows:

[1008] Step S1: After accessing the gNB, the UE informs the gNB whether the UE has the capability of federated learning through the UE capability reporting mode.

[1009] Step S2: The gNB determines whether the UE can be used as the first node according to the capability reporting of the UE. The gNB sends multicast information to the UE of the first node, instructing these UEs to participate in the online training of the AI model. The gNB also sends the downlink reference signal to these UEs through multicast for measuring the data set for training the AI model. Optionally, the gNB transmits the AI model on the gNB side to the UE.

[1010] Step S3: The UE first measures a data set for training the AI model according to the received downlink reference signal. Then, after receiving the indication information sent by the gNB for online training of the AI model, the UE performs online training to update the AI model based on the measured data set or the AI model received from the gNB. If the loss function of the AI model trained by the UE reaches a preset threshold, the UE stops training the model and sends the model parameters of the trained AI model to the gNB

[1011] Step S4: The gNB averages the received model parameters reported by each UE to obtain the parameters of the updated AI model, and implements updating of the model. If the gNB monitors that the updated AI model still cannot meet the performance requirement, the gNB sends the updated model parameters to the UEs corresponding to the first nodes again.

[1012] Step S5: The UE updates the AI model on the UE side based on the received model parameters, and then re-executes step S3 to continue training the AI model. Until the UE receives the indication information sent by the gNB for stopping online training.

[1013] In the embodiments of the present disclosure, part or all of the steps, and the optional implementation manners thereof, can be combined with part or all of the steps in other embodiments, or can be combined with the optional implementation manners of other embodiments.

[1014] The embodiments of the present disclosure also propose a device for implementing any of the above methods, for example, a device including units or modules for implementing each step performed by a terminal in any of the above methods. For another example, another device is proposed, including units or modules for implementing each step performed by a network device (such as an access network device, a core network function node, a core network device, etc.) in any of the above methods.

[1015] It should be understood that the division of each unit or module in the above apparatus is only a logical function division, and all or part of them can be integrated into a physical entity or physically separated in actual implementation. In addition, the units or modules in the apparatus can be implemented in the form of processor calling software: for example, the apparatus includes a processor, the processor is connected with a memory, the memory stores instructions, and the processor calls the instructions stored in the memory to realize any of the above methods or realize the functions of each unit or module of the above apparatus, wherein the processor is a general processor such as a central processing unit (CPU) or a microprocessor, and the memory is a memory in the apparatus or a memory outside the apparatus. Alternatively, the units or modules in the apparatus can be implemented in the form of hardware circuit, and the functions of part or all of the units or modules can be realized by the design of hardware circuit. The above hardware circuit can be understood as one or more processors; for example, in one implementation, the above hardware circuit is an application-specific integrated circuit (ASIC), and the functions of part or all of the units or modules are realized by the design of the logical relationship of elements in the circuit; for another example, in another implementation, the above hardware circuit is a programmable logic device (PLD), and a field programmable gate array (FPGA) is taken as an example, which can include a large number of logic gate circuits, and the connection relationship between the logic gate circuits is configured by a configuration file, so as to realize the functions of part or all of the above units or modules. All units or modules of the above apparatus can be all implemented in the form of processor calling software, or all implemented in the form of hardware circuit, or part implemented in the form of processor calling software and the remaining part implemented in the form of hardware circuit.

[1016] In the embodiments of the present disclosure, the processor is a circuit with signal processing capability. In one implementation, the processor can be a circuit with instruction reading and running capability, such as a central processing unit (CPU), a microprocessor, a graphics processing unit (GPU) (which can be understood as a microprocessor), a digital signal processor (DSP), or the like. In another implementation, the processor can implement certain functions through a logical relationship of a hardware circuit, and the logical relationship of the hardware circuit is fixed or can be reconfigured. For example, the processor is a hardware circuit implemented by an application-specific integrated circuit (ASIC) or a programmable logic device (PLD), such as an FPGA. In the reconfigurable hardware circuit, the processor loads a configuration document to implement the hardware circuit configuration. It can be understood that the processor loads instructions to implement the functions of the above part or all units or modules. In addition, it can also be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as a neural network processing unit (NPU), a tensor processing unit (TPU), a deep learning processing unit (DPU), and the like.

[1017] FIG. 5 is a structural schematic diagram of a first node according to an embodiment of the present disclosure. As shown in FIG. 5, the first node 500 can include at least one of a transceiver module 501, a processing module 502, and the like. In some embodiments, the first node 500 is a terminal or a network device. In some embodiments, the processing module 502 is configured to determine, in a process of training a first model according to first training data, whether to interact with a second node about model iteration information of the first model according to a model training manner, the model iteration information being information of a process quantity generated in the process of training the first model, and the model iteration information including at least one of gradient information and model parameter iteration information. Optionally, the transceiver module 510 is configured to perform at least one of the transmission and / or reception and other communication steps (for example, steps S2101, S2102, S2105, S2106, S2201, S2202, S2205, S2301, S2302, S2305, S2307, S2309, S2310, S2401, S2404, S2406, S2408, S2501, S2502, S2505, S2506, S2601, S2602, S2605, S2606, but not limited to this) performed by the terminal 101 / network device 102 in any of the above methods, details are not described herein again. Optionally, the processing module 502 is configured to perform at least one of the other steps (for example, steps S2103, S2104, S2203, S2204, S2303, S2304, S2306, S2308, S2402, S2403, S2405, S2407, S2503, S2504, S2507, S2508, S2603, S2604, S2607, S2608, but not limited to this) performed by the terminal 101 / network device 102 in any of the above methods, details are not described herein again.

[1018] In some embodiments, the transceiver module can include a sending module and / or a receiving module, which can be separate or integrated together. Optionally, the transceiver module can be mutually replaced with a transceiver.

[1019] In some embodiments, the processing module can be one module or can include multiple sub-modules. Optionally, the multiple sub-modules perform all or part of the steps required to be performed by the processing module respectively. Optionally, the processing module can be mutually replaced with a processor.

[1020] FIG. 6A is a structural schematic diagram of a communication device 8100 according to an embodiment of the present disclosure. The communication device 8100 can be a network device (for example, an access network device, a core network device, etc.), a terminal (for example, a user equipment, etc.), a chip, a chip system, or a processor supporting the network device to implement any of the above methods, or a chip, a chip system, or a processor supporting the terminal to implement any of the above methods. The communication device 8100 can be used to implement the methods described in the above method embodiments, and details can be referred to the descriptions in the above method embodiments.

[1021] As shown in FIG. 6A, the communication device 8100 includes one or more processors 8101. The processor 8101 can be a general-purpose processor or a special-purpose processor, for example, a baseband processor or a central processing unit. The baseband processor can be used to process communication protocols and communication data, and the central processing unit can be used to control the communication device (for example, a base station, a baseband chip, a terminal device, a terminal device chip, a DU or a CU, etc.), execute programs, and process data of the programs. Optionally, the communication device 8100 is configured to execute any of the above methods. Optionally, the one or more processors 8101 are configured to invoke instructions to enable the communication device 8100 to execute any of the above methods.

[1022] In some embodiments, the communication device 8100 further includes one or more transceivers 8102. When the communication device 8100 includes one or more transceivers 8102, the transceiver 8102 performs at least one of the communication steps (e.g., steps S2101, S2102, S2105, S2106, S2201, S2202, S2205, S2301, S2302, S2305, S2307, S2309, S2310, S2401, S2404, S2406, S2408, S2501, S2502, S2505, S2506, S2601, S2602, S2605, S2606, but not limited to) of transmitting and / or receiving in the above-described methods, and the processor 8101 performs at least one of the other steps (e.g., steps S2103, S2104, S2203, S2204, S2303, S2304, S2306, S2308, S2402, S2403, S2405, S2407, S2503, S2504, S2507, S2508, S2603, S2604, S2607, S2608, but not limited to) in the above-described methods. In optional embodiments, the transceiver can include a receiver and / or a transmitter, which can be separate or integrated together. Optionally, the terms transceiver, transceiving unit, transceiver, transceiving circuit, interface circuit, interface, etc. can be replaced by each other, the terms transmitter, transmitting unit, transmitter, transmitting circuit, etc. can be replaced by each other, and the terms receiver, receiving unit, receiver, receiving circuit, etc. can be replaced by each other.

[1023] In some embodiments, the communication device 8100 further includes one or more memories 8103 for storing data. Optionally, all or part of the memory 8103 can also be outside the communication device 8100. In optional embodiments, the communication device 8100 can include one or more interface circuits 8104. Optionally, the interface circuit 8104 is connected to the memory 8103, and the interface circuit 8104 can be used to receive data from the memory 8103 or other devices, and can be used to send data to the memory 8103 or other devices. For example, the interface circuit 8104 can read the data stored in the memory 8103 and send the data to the processor 8101.

[1024] The communication device 8100 described in the above embodiments can be a network device or a terminal, but the scope of the communication device 8100 described in the present disclosure is not limited thereto, and the structure of the communication device 8100 can not be limited by FIG. A. The communication device can be a standalone device or can be part of a larger device. For example, the communication device can be: 1) a standalone integrated circuit (IC), or a chip, or a chip system or subsystem; (2) a set of one or more ICs, which can optionally also include storage components for storing data, programs; (3) an ASIC, such as a modem; (4) a module that can be embedded in other devices; (5) a receiver, a terminal device, a smart terminal device, a cellular phone, a wireless device, a handset, a mobile unit, a vehicle-mounted device, a network device, a cloud device, an artificial intelligence device, and the like; (6) other devices, and the like.

[1025] FIG. 6B is a structural schematic diagram of a chip 8200 according to an embodiment of the present disclosure. For the case where the communication device 8100 can be a chip or a chip system, the structural schematic diagram of the chip 8200 shown in FIG. 6B can be referred to, but is not limited thereto.

[1026] The chip 8200 includes one or more processors 8201. The chip 8200 is configured to execute any of the above methods.

[1027] In some embodiments, the chip 8200 further includes one or more interface circuits 8202. Optionally, the terms interface circuit, interface, transceiver pin, and the like can be replaced with each other. In some embodiments, the chip 8200 further includes one or more memories 8203 for storing data. Optionally, all or part of the memory 8203 can be outside the chip 8200. Optionally, the interface circuit 8202 is connected to the memory 8203, and the interface circuit 8202 can be configured to receive data from the memory 8203 or other devices, and the interface circuit 8202 can be configured to send data to the memory 8203 or other devices. For example, the interface circuit 8202 can read data stored in the memory 8203 and send the data to the processor 8201.

[1028] In some embodiments, the interface circuit 8202 performs at least one of the communication steps (for example, step S2101, step S2102, step S2105, step S2106, step S2201, step S2202, step S2205, step S2301, step S2302, step S2305, step S2307, step S2309, step S2310, step S2401, step S2404, step S2406, step S2408, step S2501, step S2502, step S2505, step S2506, step S2601, step S2602, step S2605, step S2606, but not limited to) of the above method such as transmitting and / or receiving. The interface circuit 8202 performing the communication steps in the above method such as transmitting and / or receiving refers to that the interface circuit 8202 performs data interaction between the processor 8201, the chip 8200, the memory 8203 or the transceiver device, for example. In some embodiments, the processor 8201 performs at least one of the other steps (for example, step S2103, step S2104, step S2203, step S2204, step S2303, step S2304, step S2306, step S2308, step S2402, step S2403, step S2405, step S2407, step S2503, step S2504, step S2507, step S2508, step S2603, step S2604, step S2607, step S2608, but not limited to).

[1029] The modules and / or devices described in each of the embodiments of the virtual device, the physical device, the chip, etc. can be combined or separated according to the case. Alternatively, part or all of the steps can also be performed by a plurality of modules and / or devices in cooperation, which is not limited here.

[1030] The disclosure also proposes a storage medium, and the above storage medium stores instructions, which, when running on the communication device 8100, causes the communication device 8100 to perform any of the above methods. Alternatively, the above storage medium is an electronic storage medium. Alternatively, the above storage medium is a computer readable storage medium, but is not limited to this, and it can also be a storage medium readable by other devices. Alternatively, the above storage medium can be a non-transitory storage medium, but is not limited to this, and it can also be a transitory storage medium.

[1031] The disclosure also proposes a program product, which, when executed by the communication device 8100, causes the communication device 8100 to perform any of the above methods. Alternatively, the above program product is a computer program product.

[1032] The disclosure also proposes a computer program, which, when running on a computer, causes the computer to perform any of the above methods.

Claims

1. A model training method, characterized in that, The method performed by the first node comprises: In a case where the first model is trained by the first node, training the first model according to first training data; In a case where the first model is trained by a plurality of nodes including the first node, interacting, with a second node, model iteration information of the first model in a process of training the first model according to the first training data.

2. The method of claim 1, wherein, The method further comprises: Collecting the first training data, wherein the first training data comprises at least one of: First data sent by the second node; Second data obtained by measuring a first signal sent by the second node.

3. The method of claim 2, wherein, Before collecting the first training data, the method further comprises: Sending, to the second node, a first message, the first message comprising at least one of: A first request for requesting the second node to allow the first node to train the first model online; A second request for requesting the second node for the first data and / or the first signal; Performance information of the first model; Model training capability of the first node.

4. The method according to claim 2 or 3, characterized in that, The method further comprises: Receiving a second message sent by the second node, the second message comprising at least one of: A first indication for indicating the first node to train the first model online and / or initial model parameters of the first model; A second indication for indicating the first node to collect the first training data.

5. The method according to claim 3 or 4, characterized in that, The method further comprises: In a case where a first condition is met, sending, to the second node, a third message, the third message comprising at least one of: A third indication for indicating that the trained first model meets a performance requirement; A fourth indication for indicating the second node to stop sending the first data and / or the first signal; A fifth indication for indicating that the trained first model does not meet the performance requirement and / or a reason why the trained first model does not meet the performance requirement.

6. The method of claim 5, wherein, The method further comprises: Receiving a fourth message sent by the second node, the fourth message being sent by the second node according to the fifth indication; Based on the fourth message, performing a first operation, the first operation comprising at least one of: Switching the first model to a second model, the second model having a same function as the first model; Replacing the first model with a first algorithm, the first algorithm having a same function as the first model; Increasing a number of iterations of model training and continuing to train the first model; Increasing a time length of model training and continuing to train the first model; Continuing to obtain the first training data and training the first model; Giving up training the first model.

7. The method of claim 2, wherein, Before collecting the first training data, the method further comprises: Receiving a fifth message sent by the second node, the fifth message comprising at least one of: A third request for requesting the first node to train the first model online; Performance information of the first model; Initial model parameters of the first model.

8. The method according to claim 2 or 7, characterized in that, The method further comprises: sending a sixth message to the second node, the sixth message comprising at least one of: a sixth indication indicating that the first node trains the first model online; a seventh indication indicating that the second node sends the first data and / or the first signal to the first node.

9. The method according to claim 7 or 8, characterized in that, The method further comprises: in a case where a first condition is met, sending a seventh message to the second node, the seventh message comprising at least one of: an eighth indication indicating that the second node stops sending the first data and / or the first signal; a ninth indication informing the second node that the first model training fails and / or the first node will perform a second operation; a tenth indication indicating that the second node performs the second operation; wherein the second operation comprises at least one of: switching the first model to a second model, the second model having a same function as the first model; replacing the first model with a first algorithm, the first algorithm having a same function as the first model.

10. The method according to any one of claims 1-9, characterized in that, The method further comprises: the application node of the first model is the second node, and in a case where the trained first model meets a performance requirement, the second node is sent at least one of: the trained first model; parameter information of the trained first model; update information of the trained first model.

11. The method of any one of claims 1-9, wherein: the first model comprises a first partial model, and the application node of the first partial model is the second node, and the method further comprises: in a case where the trained first model meets a performance requirement, the second node is sent at least one of: the first partial model in the trained first model; parameter information of the first partial model in the trained first model; update information of the first partial model in the trained first model.

12. The method according to any one of claims 3-6, 10 and 11, characterized in that, The first node is a terminal, and the second node is a network device.

13. The method according to any one of claims 7-11, characterized in that, The first node is a network device, and the second node is a terminal.

14. The method of any one of claims 1-13, wherein, In a case where the first model is trained by the first node, the first model comprises at least one of: a single-sided model; a multi-sided model comprising a plurality of partial models, one partial model corresponding to one application node; at least one partial model allowing independent training.

15. The method of any one of claims 1-9, 12, and 13, wherein, In a case where the first model is trained by a plurality of nodes including the first node, the first model is at least one partial model in a multi-sided model, the multi-sided model comprising a plurality of partial models, one partial model corresponding to one application node; the third model on the second node is at least one partial model in the multi-sided model other than the first model. The method further comprises: sending a sixth message to the second node, the sixth message comprising at least one of: a sixth indication indicating that the first node trains the first model online; a seventh indication indicating that the second node sends the first data and / or the first signal to the first node. The method further comprises: in a case where a first condition is met, sending a seventh message to the second node, the seventh message comprising at least one of: an eighth indication indicating that the second node stops sending the first data and / or the first signal; a ninth indication informing the second node that the first model training fails and / or the first node will perform a second operation; a tenth indication indicating that the second node performs the second operation; wherein the second operation comprises at least one of: switching the first model to a second model, the second model having a same function as the first model; replacing the first model with a first algorithm, the first algorithm having a same function as the first model. The method further comprises: the application node of the first model is the second node, and in a case where the trained first model meets a performance requirement, the second node is sent at least one of: the trained first model; parameter information of the trained first model; update information of the trained first model.

11. The method of any one of claims 1-9, wherein: the first model comprises a first partial model, and the application node of the first partial model is the second node, and the method further comprises: in a case where the trained first model meets a performance requirement, the second node is sent at least one of: the first partial model in the trained first model; parameter information of the first partial model in the trained first model; update information of the first partial model in the trained first model. The first node is a terminal, and the second node is a network device. The first node is a network device, and the second node is a terminal. In a case where the first model is trained by the first node, the first model comprises at least one of: a single-sided model; a multi-sided model comprising a plurality of partial models, one partial model corresponding to one application node; at least one partial model allowing independent training. In a case where the first model is trained by a plurality of nodes including the first node, the first model is at least one partial model in a multi-sided model, the multi-sided model comprising a plurality of partial models, one partial model corresponding to one application node; the third model on the second node is at least one partial model in the multi-sided model other than the first model.

16. The method of claim 15, wherein, The model iteration information includes gradient information, and the gradient information includes forward gradient information and / or backward gradient information. The forward gradient information is output by the first model. The backward gradient information is output by the third model. The backward gradient information is output by the first model. The forward gradient information is output by the third model.

17. The method according to claim 15 or 16, characterized in that, In the process of training the first model according to the first training data, the model parameters of the first model and the model parameters of the third model are updated, or the model parameters of the first model are updated and the model parameters of the third model are not updated.

18. The method of any one of claims 1-4, 7, 8, 10-13, wherein, In the case where the first model is trained by a plurality of nodes including the first node, the first model is at least one of: a single-sided model; a multi-sided model including a plurality of partial models, one partial model corresponding to one application node; at least one partial model allowing independent training.

19. The method of claim 18, wherein, The first node is a training node, and the second node is a center node. In the case where the second condition is met, the first model iteration information is sent to the second node, the first model iteration information including first gradient information and / or first model parameter iteration information, the first gradient information being output by the first model, and the first model parameter information being obtained after the first node trains the first model. The second model iteration information is received from the second node, the second model iteration information being determined by the second node according to the first model iteration information uploaded by all training nodes.

20. The method of claim 19, wherein, The method further includes: updating the first model according to the second model iteration information; training the updated first model according to the first training data.

21. The method of any one of claims 5, 6, 9, wherein, The first condition includes at least one of: the first model meeting a performance requirement; the data volume of the first training data reaching a first threshold; the number of iterations of training the model reaching a second threshold; the time length of training the model reaching a third threshold.

22. The method of claim 19 or 20, wherein, The second condition includes at least one of: reaching a time point of reporting the first model iteration information; detecting that the error of the first model is within a first range; the number of local iterations of the first model reaching a fourth threshold; the first node being unable to continue training the first model.

23. A communications device, characterized by The method includes: a processing module configured to train the first model according to first training data in the case where the first model is trained by the first node; in the case where the first model is trained by a plurality of nodes including the first node, interact with a second node on model iteration information of the first model in the process of training the first model according to the first training data.

24. A communications device, characterized by The method includes: one or more processors; a memory coupled to the processor, the memory having stored thereon executable instructions that, as a result of execution by the processor, cause the model training method of any one of claims 1-22 to be performed.

25. A storage medium, the storage medium storing instructions, wherein, When the instructions are executed on the communication device, the communication device is caused to perform the model training method of any one of claims 1-22.

Images