Communication apparatus and method for machine learning model training

Abstract

A method for configuring at least one first node training machine learning (ML) model includes providing training assistance information by a second node, where the training assistance information is used for a first node to jointly train with the second node to train a plurality of ML models having a common portion.

Description

Technical Field

[0001] The present application relates to the field of wireless communication systems, and more specifically to communication devices and methods for machine learning (ML) model training. For example, the present application is related to the new study item description (SID) of artificial intelligence / machine learning (AI / ML) for the 18th version of the new radio (NR) air interface, which was established in December 2022 at the 3rd generation partnership project (3GPP) radio access network (RAN) plenary meeting 94e, led by RAN1 and started in May 2022. Specifically, the present application relates to enhanced channel state information (CSI) report feedback, beam management and / or positioning, in which multiple ML models can have a common part. For enhanced CSI feedback, these ML models can have a common CSI generation part or a common CSI reconstruction part. Background Art

[0002] AI / ML is applied to 3GPP RAN1, which decided to study multiple use cases, namely CSI feedback enhancement, beam management, and positioning. According to 3GPP SID, although specific AI / ML algorithms and models may be studied for evaluation, AI / ML algorithms and models are implementation-specific and are not expected to be standardized. ML models will go through multiple stages. First, the ML model can be trained online or offline, and then the model is deployed. After deployment, the ML model is activated and enters the inference phase. During the inference phase, the ML model (whether it has a common part or not) will be monitored. There are some issues, such as how to determine whether the ML model of the user equipment (UE) has failed, and how to deactivate the common part of the ML model when a model failure occurs. Therefore, the activation and deactivation mechanism of multiple ML models with a common part needs to be discussed.

[0003] If the AI / ML based CSI compression sub-use case supports ML models with a common part, a method to save signaling needs to be designed. If multiple ML models are configured for relevant UEs one by one in an intuitive way, it will be a very tedious task. In addition, the problem of joint training of multiple ML models with a common part needs to be discussed. How to update the common model part, or only update one non-common model part? Some rules need to be designed for this purpose so that multiple ML models with a common part can be trained simultaneously. The ML model for CSI feedback is a two-sided model. Currently, there is no solution to the problem of joint training of CSI feedback ML models with a common part.

[0004] Therefore, there is a need to design new model training methods for machine learning (ML) enhancement in communication devices and physical layer (PHY) and media access control layer (MAC). These methods can solve the problems in the prior art, simplify the management of multiple ML models with common parts, provide methods for training multiple ML models with common parts, reduce system overhead, improve communication performance and / or provide high reliability. Summary of the invention

[0005] The purpose of the present application is to provide a communication device and method for machine learning (ML) model training, which can solve the problems in the prior art, simplify the management of multiple ML models with a common part, provide a method for training multiple ML models with a common part, reduce system signaling overhead, and provide good communication performance and / or high reliability.

[0006] In a first aspect of the present application, a method for configuring at least one first node to train a machine learning (ML) model includes providing training auxiliary information through a second node, wherein the training auxiliary information is used for the first node and the second node to jointly train multiple ML models having a common part.

[0007] In a second aspect of the present application, a first node includes a memory, a transceiver, and a processor coupled to the memory and the transceiver. The processor is configured to execute the above method.

[0008] In a third aspect of the present application, a communication system includes a memory, a transceiver, and a processor coupled to the memory and the transceiver. The processor is configured to execute the above method.

[0009] In a fourth aspect of the present application, a second node includes a memory, a transceiver, and a processor coupled to the memory and the transceiver. The processor is configured to execute the above method.

[0010] In a fifth aspect of the present application, a non-transitory machine-readable storage medium stores instructions thereon, and when a computer executes the instructions, the instructions cause the computer to execute the above method.

[0011] In a sixth aspect of the present application, a chip includes a processor configured to call and run a computer program stored in a memory so that a device equipped with the chip executes the above method.

[0012] In a seventh aspect of the present application, a computer-readable storage medium stores a computer program, wherein the computer program enables a computer to execute the above method.

[0013] In an eighth aspect of the present application, a computer program product comprises a computer program, wherein the computer program enables a computer to execute the above method.

[0014] In a ninth aspect of the present application, a computer program is provided, wherein the computer program enables a computer to execute the above method.

[0015] In the tenth aspect of the present application, a communication system for configuring at least one first node to train a machine learning (ML) model includes a first node and a second node, wherein the second node provides training auxiliary information to the first node, and the training auxiliary information is used for the first node and the second node to jointly train multiple ML models having a common part. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or related technologies, the following briefly introduces the drawings to be described in the embodiments. Obviously, the drawings are only some embodiments of the present application, and ordinary technicians in this field can obtain other drawings based on these drawings without paying any price.

[0017] Figure 1 is a schematic diagram showing an example of a basic autoencoder model for enhancing CSI feedback according to an embodiment of the present application.

[0018] Figure 2 It is a block diagram showing one or more user equipment (UE) and a base station (such as a gNB) in a communication network system according to an embodiment of the present application.

[0019] Figure 3 It is a flowchart showing a method for configuring at least one first node to train a machine learning (ML) model according to an embodiment of the present application.

[0020] Figure 4is a schematic diagram showing an example of a basic autoencoder model for enhancing CSI feedback according to an embodiment of the present application.

[0021] Figure 5 It is a schematic diagram showing an example of a RAN intelligent function framework according to an embodiment of the present application.

[0022] Figure 6 It is a schematic diagram showing an example in which corresponding encoders of two UEs share a common decoder on the gNB side according to an embodiment of the present application.

[0023] Figure 7 is a schematic diagram showing an example of forward propagation and back propagation of an ML model with a common part according to an embodiment of the present application.

[0024] Figure 8 is a schematic diagram showing an example of forward propagation and back propagation of an ML model with a common part according to an embodiment of the present application.

[0025] Fig. 9 is a schematic diagram showing an example of forward propagation and back propagation of an ML model with a common part according to an embodiment of the present application.

[0026] Fig.10 is a block diagram of a system for wireless communication according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] The technical content, structural features, objectives and effects of the present application are described in detail below in conjunction with the accompanying drawings. Specifically, the terms in the embodiments of the present application are only used to describe the purpose of specific embodiments, and are not intended to limit the present application.

[0028] Introducing AI / ML into the physical layer (PHY) and media access control layer (MAC) to improve system performance. 3GPP RAN1 decided to study multiple use cases, namely CSI feedback compression, beam management and positioning. ML learning models can be trained online or offline.

[0029] The machine learning model deployed in at least one cell can have a common part, which is a generalized form of the ML model. The common part can process data from or send data to multiple non-common parts. Managing such ML models requires unified signaling, otherwise it will be very time-consuming and wasteful of resources if these ML models are configured one by one. To solve this problem, the ML model with a common part can be managed through broadcast signaling.

[0030] Some embodiments of the present application discuss cases of CSI feedback enhancement. Figure 1According to an embodiment of the present application, a schematic diagram of a basic autoencoder model for enhancing CSI feedback is exemplarily shown. Figure 1 In some embodiments, the basic autoencoder model is shown as follows: the encoder compresses the original CSI-RS value (referred to as original CSI) / maximum eigenvector and reports its output to the gNB. The gNB decompresses it. The new CSI report is the CSI report generated after the CSI feedback is enhanced by the AI / ML model.

[0031] In addition, AI / ML is introduced into the PHY layer and MAC layer to enhance system performance. In 3GPP RAN1, it was decided to study multiple use cases, namely CSI feedback compression, beam management and positioning. ML learning models can be trained online or offline.

[0032] On the UE side, the input is compressed and output to the channel to achieve ML-enhanced CSI feedback. The input can be a (maximum) eigenvector or a channel matrix. The compressed output is input to the decoder and reconstructed on the gNB side. Several types of ML model training methods are discussed, including training on the UE side and passing the ML model to the gNB; training on the gNB side and passing the model to the UE; joint training by the UE and gNB; and separate training on the UE and gNB respectively. Some embodiments of the present application mainly discuss the joint training of the UE and gNB.

[0033] Figure 2It is shown that in some embodiments, according to an embodiment of the present application, at least one first node 10 for communication is provided in a communication network system 40, such as at least one user equipment (UE), a second node 20, such as a base station 20 (e.g., a gNB), and at least one first node 30, such as at least one user equipment (UE). The communication network system 40 includes at least one first node 10, such as at least one user equipment (UE), a second node 20, such as a base station 20 (e.g., a gNB), and at least one first node 30, such as at least one user equipment (UE). The first node 10 may include a memory 12, a transceiver 13, and a processor 11 coupled to the memory 12 and the transceiver 13. The second node 20 may include a memory 22, a transceiver 23, and a processor 21 coupled to the memory 22 and the transceiver 23. The third node 30 may include a memory 32, a transceiver 33, and a processor 31 coupled to the memory 32 and the transceiver 33. The processor 11, 21 or 31 may be configured to implement the functions, processes and / or methods described in this specification. Each layer of the wireless interface protocol may be implemented in the processor 11, 21 or 31. The memory 12, 22 or 32 is operatively coupled to the processor 11, 21 or 31 and stores various information used to operate the processor 11, 21 or 31. The transceiver 13, 23 or 33 is operatively coupled to the processor 11, 21 or 31 for sending and / or receiving wireless signals.

[0034] The processor 11, 21 or 31 may include an application-specific integrated circuit (ASIC), other chipsets, logic circuits and / or data processing devices. The memory 12, 22 or 32 may include a read-only memory (ROM), a random access memory (RAM), a flash memory, a memory card, a storage medium and / or other storage devices. The transceiver 13, 23 or 33 may include a baseband circuit for processing radio frequency signals. When the embodiment is implemented in software, the technology described herein may be implemented by modules (such as processes, functions, etc.) that perform the functions. These modules may be stored in the memory 12, 22 or 32 and executed by the processor 11, 21 or 31. The memory 12, 22 or 32 may be integrated in the processor 11, 21 or 31, or it may be implemented externally and communicate with the processor 11, 21 or 31 in a known manner.

[0035] Figure 3A method 300 according to an embodiment of the present application is shown, which is used to configure machine learning (ML) model training of at least one first node. In some embodiments, the method 300 includes: step 302, the second node provides training auxiliary information to the first node, and the training auxiliary information is used for joint training of the first node and the second node to train multiple ML models with a common part. This can solve the problems in the prior art, simplify the management of multiple ML models with a common part, provide a training method, reduce system signaling overhead, and provide good communication performance and / or high reliability.

[0036] In some embodiments, processor 11 and / or processor 21 is used to execute method 300 to configure machine learning (ML) model training of at least one first node. This can solve the problems in the prior art, simplify the management of multiple ML models with common parts, provide a training method, reduce system signaling overhead, and provide good communication performance and / or high reliability.

[0037] In some embodiments, the first node is a user equipment (UE), the second node is a base station, and at least one third node is another UE. The encoder of the UE and the encoder of at least one other UE (CSI generation part) share a common decoder in the base station, and the encoder of the UE and the encoder of at least one other UE may correspond to the channel state information (CSI) generation part or the CSI reconstruction part, and the common decoder of the base station corresponds to the other part. In some embodiments, the training auxiliary information is included in the common signaling or broadcast signaling of the UE group, and the signaling may be sent via downlink control information (DCI) 2_0 or DCI 2_x, or included in the system information block (SIB) and / or the master information block (MIB).

[0038] In some embodiments, the training auxiliary information includes at least one of the following: activation / enabling of ML model training, a period of forward propagation data reporting, a period of backward propagation data reporting, deactivation / disabling of ML model training, and identification information. In some embodiments, activation / enabling of ML model training, quantization level of forward propagation, quantization level of backward propagation, and deactivation / disabling of ML model training are DCI fields in DCI 2_0 or DCI 2_x. In some embodiments, the forward propagation data reporting period and / or the backward propagation data reporting period are configured by the second node in the form of bits, lists, or tables, or the forward propagation data reporting period and / or the backward propagation data reporting period are default periods.

[0039] In some embodiments, the identification information includes at least one of the following: a cell identifier (ID) or a radio network temporary identifier (RNTI) of the second node, wherein the UE group common signaling is scrambled by a slot format indication radio network temporary identifier (SFI-RNTI) of DCI 2_0 or a new RNTI of DCI 2_x. In some embodiments, at least one first node and a second node are deployed with an ML model having a common portion, wherein the at least one first node forms a group. In some embodiments, the grouping rule includes at least one of the following: beam, wherein at least one first node connected to the same beam of a second node is grouped together and paired with a common ML model portion; cell residence time of the first node, wherein at least one first node connected to the second node whose residence time exceeds a specific time is grouped together and paired with a common ML model portion; distance, wherein at least one first node connected to the second node and located within the same distance range is grouped together and paired with a common ML model portion; and channel condition, wherein at least one first node having the same channel condition is grouped together and paired with a common ML model portion, wherein the same channel condition includes signal to noise ratio (SNR), delay and / or Doppler shift within the same range.

[0040] In some embodiments, during the training process, forward propagation and backward propagation form a loop for the encoder of at least one first node and the common decoder of the second node. In some embodiments, each at least one first node contributes to the forward propagation, and after the second node calculates the loss function, the backward propagation begins and sends the updated parameters to each first node and at least one third node. In some embodiments, at least one first node reports forward propagation data and true values ​​in batches, and the data comes from a single measurement, a single input of the encoder, multiple measurements, or multiple inputs of the encoder.

[0041] In some embodiments, during training, at least one first node reports forward propagation in batches, and backpropagation is applied to all at least one first node, even if some at least one first node has not reported any forward propagation since the last propagation. In some embodiments, a forward propagation reported by one of the at least one first nodes triggers one backpropagation for all related first nodes. In some embodiments, for an ML model with a common CSI reconstruction portion, if no forward propagation data is sent in a training cycle, the related at least one first node does not expect to receive any data for gradient descent or backpropagation data.

[0042] In some embodiments, the quantization of the forward propagation data may be scalar quantization or vector quantization, and the quantization level of the scalar quantization or vector quantization is indicated by the training auxiliary information. In some examples, the quantization of the backward propagation data may be scalar quantization or vector quantization, and the quantization level of the scalar quantization or vector quantization is indicated by the training auxiliary information. Generally, the quantization of the forward propagation data and the backward propagation data should be kept to a small number of bits, otherwise the training overhead will be very large. Therefore, in order to reduce the training overhead, the quantization of the forward propagation data should be coarse quantization or a low quantization level, such as 8-bit scalar quantization of real numbers. The quantization of the backward propagation should have high precision, otherwise the ML model may not converge smoothly or quickly. In some examples, the quantization in the training auxiliary information contains a bit indicating the quantization type, for example, the quantization type is {0 / 1} scalar quantization or {1 / 0} vector quantization, which is applicable to forward propagation and backward propagation, respectively. If the quantization type is not configured, the default value is vector quantization or scalar quantization.

[0043] In some examples, quantization is uniformly configured via broadcast-type signaling through training auxiliary information. In addition, quantization can be reconfigured to different quantization levels or quantization methods via DCI / MAC-CE / RRC signaling. In the process of training ML models with common parts, the quantization level or quantization method may or may not be fixed.

[0044] In some embodiments, at least one first node reports forward propagation data in batches, and performs backward propagation on the reporting first node and at least one third node. In some embodiments, if one of the at least one first node has not reported any forward propagation data since the last propagation, backward propagation is not performed. In certain embodiments, the training of the ML model portion located at at least one first node is performed separately and independently of the other first nodes, and at least one first node takes turns to perform training. In some embodiments, the training type of the ML model with a common portion is configured by the second node. In some embodiments, the second node configures the ML model training type with a common portion to at least one first node via radio resource configuration (RRC) signaling, media access control-control element (MAC-CE), or DCI field.

[0045] Figure 4 is a schematic diagram showing an example of a basic autoencoder model for enhancing CSI feedback according to an embodiment of the present application. Figure 5 It is a schematic diagram illustrating an example of a functional framework of RAN intelligence according to an embodiment of the present application. Figure 4 and Figure 5 It is shown that in some embodiments, the ML model needs to be monitored during the inference phase. The functional framework of RAN intelligence is provided in RAN3 and can be further modified for RAN1. The ML model will be monitored after deployment to check whether it is working properly. Typically, the performance of the ML model is compared with a standard. If the ML model does not work properly, the UE will switch to another ML model or fall back to a non-AI working mode. The monitored ML model will be retrained.

[0046] Embodiment 1:

[0047] Figure 6 This is a schematic diagram illustrating an example in which encoders of two UEs share a common decoder on the gNB side according to an embodiment of the present application. Figure 6 It is shown that, as an example, there are several first nodes, which can be UEs, and one second node, which can be a gNB. The encoders of UE1 and UE2 share a common decoder on the gNB side. The decoder refers to the CSI reconstruction part, and the encoder refers to the CSI generation part. When the gNB configures the encoders of multiple UEs (UE1 and UE2 here), the configuration information should be a broadcast signal transmitted through the downlink. The configuration information is the training auxiliary information.

[0048] For example, the training auxiliary information is included in DCI 2_0 or new DCI 2_x. For example, the training auxiliary information includes at least one of the following: activation / enabling of ML model training, forward propagation data reporting period, backward propagation data sending period, deactivation / disabling of ML model training, quantization information, and identification information of the second node (base station).

[0049] Activation / enablement of ML model training is a field in DCI 2_0 or the new DCI 2_x. Deactivation / disablement of ML model training is also a field in DCI 2_0 or the new DCI 2_x. In some examples, this is a one-bit field in DCI 2_x, where "1" indicates activation of the training of the relevant multiple ML models and "0" indicates deactivation of the training of the relevant multiple ML models. The forward propagation data reporting period is configured by the gNB through a list or table. Forward propagation is a step in ML model training. For the ML model for CSI feedback, data is input and passed through the calculation of each layer. This is forward propagation.

[0050] In DCI format 2_0, the bitmap corresponds one-to-one to a set of values ​​for the forward propagation data reporting period, where the value "0" indicates that the reporting period is available for reception and the value "1" indicates that the reporting period is not available for reception. It is set through the availablePeriod-FP-PerCell field, which remains valid or invalid in DCI format 2_x / 2_0 until the end of the training duration / mode unless reconfigured.

[0051] Backpropagation is a step in the training of ML models. The gradient is calculated from the loss function and propagated backward from the output layer to the input layer. In DCI format 2_0, the bitmap corresponds to a set of values ​​for the period of providing backward propagation data to the UE, where the value "0" indicates that the reporting period is available for reception and the value "1" indicates that the reporting period is not available for reception. It is set through the availablePeriod-BP-PerCell field, which remains valid or invalid in DCI format 2_x / 2_0 until the end of the training duration / mode unless reconfigured. In another example, the reporting period of forward propagation data or backward propagation is configured by the gNB through several bits. For example, the case of a two-bit field is as follows:

[0052] Two-digit field example:

[0053] 00 01 10 11 5ms 10ms 20ms 40ms

[0054] In another example, if availablePeriod-FP-PerCell is not configured, there is a default reporting period, and the default period is one of the following: {5ms, 10ms, 20ms, 30ms, 40ms, 60ms, 100ms}. In another example, if availablePeriod-BP-PerCell is not configured, there is a default reporting period: {5ms, 10ms, 20ms, 30ms, 40ms, 60ms, 100ms}. In another example, for the CSI feedback ML model with a common part, the default forward propagation data reporting is periodic or semi-persistent. The field in DCI 2_x can be provided to the UE. If the UE receives DCI 2_x and decodes the field as activated, the reporting of the forward propagation data will be activated when the reporting mode is semi-periodic. Otherwise, the reporting of the forward propagation data will be deactivated.

[0055] In another example, for the CSI feedback ML model with a common part, the default backpropagation data reporting is periodic or semi-periodic. The field in DCI 2_x can be provided to the UE. If the UE receives DCI 2_x and decodes the field as activated, the reception of backpropagation data will be activated when the reporting mode is semi-periodic. Otherwise, the reception of backpropagation data will be deactivated.

[0056] In another example, for the CSI feedback ML model with common part, the default forward data reporting is aperiodic. The reporting of the forward data is triggered by DCI. In another example, the default backward data reporting is aperiodic. In some examples, this depends on the request of the UE. If the UE has not received the backward data after exceeding a time window / threshold, the UE will request the backward data from the gNB. The time window / threshold is configured by the gNB through RRC signaling. If not configured, there is a default value.

[0057] The training assistance information is identified by the identification information of the second node, including at least one of the following: cell ID or RNTI. For example, the common signaling of the UE group is scrambled by the SFI-RNTI of DCI_2.0 or the new RNTI of DCI 2_x. SFI-RNTI: Slot Format Indicator Radio Network Temporary Identifier; RNTI: Radio Network Temporary Identifier.

[0058] In some examples, UEs deployed with ML models having a common portion form a group and are provided with UE group common signaling. In some examples, the grouping rule includes at least one of the following.

[0059] Beam: UEs connected to the same gNB beam are grouped together.

[0060] UE residence time: UEs that are connected to the gNB for more than a certain period of time are grouped together.

[0061] Distance: UEs connected to the gNB that are within the same distance range are grouped together.

[0062] UEs with similar channel conditions are grouped together, for example, with similar signal-to-noise ratios (signal-to-noise ratios within a range).

[0063] After completing the UE grouping, these UEs are deployed with the ML model with the common part.

[0064] Embodiment 2:

[0065] In another example, the training auxiliary information may be included in the SIB / MIB. For example, the training auxiliary information includes at least one of the following: enabling / activating ML model training, deactivating / disabling ML model training, and identification information of the second node.

[0066] For example, the SIB(x) / MIB may contain a field ML-model-training-common{0 / 1}, where {0} indicates enabling / activating ML model training. For example, this may be applied to CSI feedback training. {1} indicates disabling / disabling ML model training. For example, this may be applied to CSI feedback training. In this way, all UEs in the current cell will be provided with this information, and this information can be used to configure the training of the ML model.

[0067] Embodiment 3:

[0068] Figure 7 is a schematic diagram according to an embodiment of the present application, illustrating an example of forward propagation and backward propagation of an ML model with a common part during training. Figure 7 It shows that in order to make the ML model converge smoothly and quickly, the data is input into the ML model in batches during the training process. Therefore, the forward propagation step is performed first, and the loss function is calculated based on the output of the decoder. Then the backward propagation is performed. Figure 7 As shown, the three ML models share a common CSI reconstruction part. The forward propagation and backward propagation form a cycle.

[0069] Each UE participates in the forward propagation. After the loss function is calculated, the backward propagation starts and the updated parameters are sent to each UE. First, all UEs report the compressed CSI and the ground truth. Then, after receiving enough data, the base station calculates the value of the loss function, then calculates the gradient of the decoder and transmits the gradient back to each UE.

[0070] In some examples, the UE reports forward propagation data and true values ​​in batches. The true value is the original CSI (channel matrix after ideal or non-ideal channel estimation). The batch size is configured by the base station to at least one UE through broadcast-like signaling. In some examples, the UE reports forward propagation data and true values ​​from a single measurement or a single input of the encoder. In other examples, the UE reports forward propagation data and true values ​​from multiple measurements or multiple inputs of the encoder. The reported data and / or the specific number of inputs are configured by the gNB through RRC signaling, MAC-CE or DCI fields. Backward propagation is performed on participating UEs in batches.

[0071] Embodiment 4:

[0072] Figure 8 is a schematic diagram according to an embodiment of the present application, illustrating an example of forward propagation and backward propagation of an ML model with a common part during training. Figure 8 It is shown that in some examples, UEs report forward propagation data in batches and backpropagate to all UEs, even if some UEs have not reported any forward propagation data since the last propagation. A forward propagation report from one UE can trigger backpropagation for all participating UEs. For ML models with common CSI reconstruction part, if a participating UE does not send any forward propagation data in a certain training cycle, then the UE does not expect to receive any backpropagation data or gradient descent data.

[0073] Embodiment 5:

[0074] Fig. 9 is a schematic diagram according to an embodiment of the present application, illustrating an example of forward propagation and backward propagation of an ML model with a common part during training. Fig. 9 It shows that in some examples, the UE reports the forward propagation data in batches, and the backward propagation is performed for the reporting UE. Other UEs that have not reported any forward propagation data do not participate in the backward propagation. From another perspective, the training of the ML model part on the UE side is performed separately and has nothing to do with other UEs. UE1, UE2 and UE3 are trained sequentially. In some examples, the training methods in Example 3, Example 4 and Example 5 are named training type a, training type b and training type c, respectively. The gNB will configure a specific training type to the UE that shares the common ML model part. The configuration is provided to the UE through RRC signaling, MAC-CE or DCI field.

[0075] In summary, in some embodiments of the present application, a method for training multiple ML models with a common part is provided. The common part may be a CSI generation part or a CSI reconstruction part. A broadcast-like signaling, such as a field in group common signaling (DCI 2_x) or SIB / MIB, is designed for training multiple ML models for CSI feedback. This method can complete the training of multiple ML models with less signaling overhead. In addition, a solution for updating the CSI generation part based on the update of the CSI reconstruction part is provided, so that multiple ML models with a common part can be correctly trained. Certain embodiments of the present application have at least one of the following inventive effects: reducing the training overhead of multiple ML models with a common part; the training of multiple ML models with a common part should have certain constraints; updating a certain ML model part does not necessarily require the update of another model part.

[0076] Fig.10 7 is a block diagram of an exemplary wireless communication system 700 according to an embodiment of the present application. The embodiments described herein may be implemented into a system using any appropriately configured hardware and / or software. Fig.10 A system 700 is shown, which includes a radio frequency (RF) circuit 710, a baseband circuit 720, an application circuit 730, a memory / storage device 740, a display 750, a camera 760, a sensor 770, and an input / output (I / O) interface 780, which are coupled to each other at least as shown. The application circuit 730 may include circuits, such as but not limited to one or more single-core or multi-core processors. The processor may include any combination of general-purpose processors and special-purpose processors, such as a graphics processor, an application processor. The processor may be coupled to a memory / storage device and configured to execute instructions stored in the memory / storage device to enable various applications and / or operating systems to run on the system.

[0077] Training is not limited to reporting from the UE to the gNB. Training can involve a third node. The third node can be a UE, a gNB, or a server. If the third node is another gNB, all signaling and data will be transmitted to it for model training. The method in the disclosed embodiment is applicable. After the model training is completed, the ML model (the common part) will be transmitted to the second node (gNB). In this way, the signaling overhead between the sending node and the receiving node is reduced. In some examples, the gNB or UE can be the sending node and the third node can be the receiving node.

[0078] While the present disclosure has been described in connection with what is considered to be the most practical and preferred embodiment, it is to be understood that the present disclosure is not limited to the disclosed embodiment, but is intended to cover various arrangements made without departing from the scope of the broadest interpretation of the appended claims.

Claims

1. A method for configuring at least one first node to train a machine learning ML model, characterized in that: include: Training auxiliary information is provided by the second node, wherein the training auxiliary information is used for joint training of the first node and the second node to train multiple ML models having a common portion.

2. The method according to claim 1, characterized in that: The first node is a user equipment UE, the second node is a base station, and at least one third node is at least one other UE. The encoder of the UE and the encoder (CSI generation part) of the at least one other UE share a common decoder in the base station, the encoder of the UE and the encoder of the at least one other UE refer to one of a channel state information CSI generation part and a CSI reconstruction part, and the common decoder of the base station refers to the other of the CSI generation part and the CSI reconstruction part.

3. The method according to claim 1 or 2, characterized in that: The training auxiliary information is included in UE group common signaling or broadcast signaling, the UE group common signaling or the broadcast signaling is included in downlink control information DCI 2_0 or DCI2_x, or the training auxiliary information is included in a system information block SIB and / or a master information block MIB.

4. The method according to any one of claims 1 to 3, characterized in that The training auxiliary information includes at least one of: activation / enabling of ML model training, forward propagation data reporting period, backward propagation data reporting period, deactivation / disabling of the ML model training, quantization information, and identification information.

5. The method according to claim 4, characterized in that Activating / enabling ML model training and deactivating / disabling the ML model training are DCI fields in DCI 2_0 or DCI 2_x.

6. The method according to claim 4, characterized in that The period for reporting forward propagation data and / or the period for reporting reverse propagation data is configured by the second node and has a bit, a list or a table, or the period for reporting forward propagation data and / or the period for reporting reverse propagation data is a default period.

7. The method according to claim 4, wherein the identification information comprises at least one of the following: a cell identifier ID or a radio network temporary identifier RNTI of the second node, wherein the UE group common signaling indicates a radio network temporary identifier SFI-RNTI or a new RNTI scrambled for DCI 2_x through a time slot format for DCI_2.

0.

8. The method according to any one of claims 1 to 7, characterized in that The at least one first node forms a group with the second node on which a common portion of the ML model is deployed.

9. The method according to claim 8, characterized in that The grouping rules of the grouping include at least one of the following: beams, wherein the at least one first node of the same beam connected to the second node are grouped together and paired with a common ML model portion; a cell remaining time of the first nodes, wherein the at least one first node connected to the second node for longer than a time is grouped together and paired with the common ML model portion; distance, wherein the at least one first node connected to the second node and within the same distance range as the second node are grouped together and paired with the common ML model portion; as well as Channel conditions, wherein the at least one first node having the same channel conditions are grouped together and paired with the common ML model portion, wherein the same channel conditions include signal-to-noise ratio (SNR), time delay and / or Doppler within the same range.

10. The method according to any one of claims 1 to 9, characterized in that During the training process, a forward propagation and backward propagation cycle is formed between the encoder of the at least one first node and the common decoder of the second node.

11. The method according to claim 10, characterized in that Each of the at least one first node contributes to the forward propagation, and the second node starts the backward propagation after calculating the loss function, and sends the updated parameters to each of the first node and the at least one third node.

12. The method according to claim 10 or 11, characterized in that: The at least one first node reports forward propagation data and true values ​​in batches, from a single measurement, a single input to the encoder, multiple measurements, or multiple inputs to the encoder.

13. The method according to any one of claims 1 to 9, characterized in that During the training process, the at least one first node reports the forward propagation in batches, and even if some of the first nodes have not reported the forward propagation since the last propagation, the backward propagation will still be performed on all participating at least one first node.

14. The method according to claim 13, characterized in that One or more reports of the forward propagation from one of the at least one first nodes trigger a backward propagation to all participating at least one first nodes.

15. The method according to claim 13 or 14, characterized in that For the ML model with a common CSI reconstruction part, if one of the at least one first nodes does not send forward propagation data in a training cycle, it is not expected to receive any data or backward propagation data for gradient descent.

16. The method according to any one of claims 1 to 9, characterized in that The at least one first node batch reports forward propagation, and the backward propagation is performed only on one of the first node and the at least one third node that reports.

17. The method according to claim 16, characterized in that One of the at least one first node has not reported any forward propagation data since the last propagation, and thus does not perform backpropagation.

18. The method according to claim 16 or 17, characterized in that The training of the ML model portion located at the at least one first node is performed separately and independently from another node in the at least one first node, and the at least one first node performs training sequentially.

19. The method according to any one of claims 1 to 18, characterized in that The training type of the ML model having the common portion is configured by the second node.

20. The method according to claim 19, characterized in that The training type of the ML model with the common part is configured by the second node to the at least one first node through radio resource configuration RRC signaling, medium access control-control element MAC-CE or DCI field.

21. A communication system, characterized in that: include: Memory; Transceiver; as well as a processor coupled to the memory and the transceiver; The processor is configured to execute the method according to any one of claims 1 to 20.

22. A communication system for configuring at least one first node to train a machine learning (ML) model, characterized in that: include: A first node and a second node, wherein the second node provides training auxiliary information to the first node, wherein the training auxiliary information is used for the first node and the second node to jointly train a plurality of ML models having a common portion.

Images