Artificial intelligence model parameter acquisition method and apparatus, storage medium, and chip system

Abstract

The application discloses an AI model parameter acquisition method and device, a storage medium and a chip system, and belongs to the technical field of communication. The method comprises the following steps: receiving n symbol sequences sent by a network device. Wherein, after receiving one symbol sequence, the symbol sequence is demodulated to obtain a first soft information sequence; a second soft information sequence with a preset length is generated according to the first soft information sequence; the second soft information sequence is decoded according to side information related to the update parameters of a target NN layer in a target AI model to obtain a target coding block, and the target coding block comprises the update parameters of the target NN layer, and the side information is determined according to the original NN parameters of the target AI model. In the application, the side information related to the update parameters of the target NN layer is input as additional prior knowledge for decoding, so that the decoding performance can be greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, and particularly relates to an AI model parameter acquisition method and device, a storage medium and a chip system. BACKGROUND

[0002] In a wireless communication system, an artificial intelligence (AI) model can be used to enhance system performance. These AI models are usually trained in a specific training environment. However, as the actual use environment changes, the deployed AI model may not be able to adapt to the new environment, thereby affecting the system performance. To address this problem, a network device can update the parameters of the AI model by fine-tuning the AI model, and transmit the updated parameters to a terminal device. However, an AI model usually includes a large number of high-precision neural network (NN) parameters, which are easily affected by channel noise during wireless transmission, which brings great challenges to the decoding of the terminal device. SUMMARY

[0003] The present application provides an AI model parameter acquisition method, device, storage medium and chip system, which can improve decoding performance. The technical solution is as follows:

[0004] In a first aspect, an AI model parameter acquisition method is provided. The method can be executed by a terminal device, or by a component (such as a circuit, a chip or a chip system, etc.) configured in the terminal device, or by a logic module or software capable of realizing all or part of the functions of the terminal device, and the present application does not limit this. The following is described by taking the terminal device as an example.

[0005] The method includes: receiving, by a terminal device, n symbol sequences sent by a network device, n being a positive integer. Wherein, the terminal device demodulates each received symbol sequence to obtain a first soft information sequence, and then generates a second soft information sequence of a preset length according to the first soft information sequence. Then, the second soft information sequence is decoded according to side information related to the update parameters of a target NN layer in a target AI model to obtain a target code block, the target code block including the update parameters of the target NN layer, and the side information is determined according to the original NN parameters of the target AI model.

[0006] The side information is a core element of the Slepian-Wolf theory, which refers to the auxiliary information related to the to-be-decoded source that the decoding end already has, and the decoding end can realize efficient reconstruction by combining the side information with the received data.

[0007] In the present application, the terminal device can determine the side information related to the update parameters of the target NN layer according to the NN original parameters of the target AI model. The terminal device inputs the side information related to the update parameters of the target NN layer as additional priori knowledge and jointly decodes the soft information sequence obtained by demodulation, so as to greatly improve the decoding performance. In this way, even under poor channel conditions, the decoding can be accurately completed, and the decoding success rate is high, thereby helping to realize the robust transmission and accurate recovery of the NN parameters.

[0008] In a possible implementation, before the terminal device receives the n symbol sequences sent by the network device, the terminal device can also receive control information sent by the network device. The control information includes a target AI model identifier, a target NN layer identifier, and a total number of blocks, and the total number of blocks is n. In this way, the terminal device can determine which NN layer in which AI model needs to be updated according to the control information.

[0009] In a possible implementation, after the terminal device receives the control information sent by the network device, the terminal device can also obtain the NN original parameters of the target AI model; input the binary quantization sequence of the NN original parameters of the target AI model and the parameters of the hidden Markov model into the FBA module to obtain the output data of the FBA module; and estimate the soft information sequence of the binary quantization sequence of the NN update parameters according to the output data of the FBA module to obtain a target soft information sequence, where the target soft information sequence includes the side information related to the update parameters of the target NN layer.

[0010] The hidden Markov model is used to represent the correlation between the binary quantization sequence of the NN original parameters of the target AI model and the binary quantization sequence of the NN update parameters of the target AI model.

[0011] The binary quantization sequence of the NN original parameters of the target AI model is an observable past sequence, and the target of the FBA module is to estimate the binary quantization sequence of the NN update parameters of the target AI model through the hidden Markov model and the known observable sequence. The FBA module will continuously correct the parameters of the hidden Markov model in the iteration process. When the sample of the observable sequence is long enough, the influence of the error of the initial input state can be ignored. Based on this, the FBA module can output after reaching a preset number of iterations. The output data based on the FBA can accurately estimate the soft information sequence of the binary quantization sequence of the NN update parameters.

[0012] In a possible implementation, before receiving the n symbol sequences sent by the network device, the terminal device can also obtain the soft information sequence corresponding to the target NN layer from the target soft information sequence, and the target soft information sequence is determined according to the NN original parameter of the target AI model; the soft information sequence corresponding to the target NN layer is divided into n soft information sequences, and the n soft information sequences are the side information. Correspondingly, the operation of the terminal device coding the second soft information sequence according to the side information related to the update parameter of the target NN layer in the target AI model to obtain the target coding block can be: generating a fourth soft information sequence according to the second soft information sequence and a third soft information sequence, the third soft information sequence being one soft information sequence corresponding to the second soft information sequence in the n soft information sequences; inputting the fourth soft information sequence into a channel decoding module to obtain a target coding block output by the channel decoding module.

[0013] In the present application, the second soft information sequence obtained based on demodulation and the third soft information sequence obtained based on estimation can be jointly coded as input, which can enhance the coding performance.

[0014] In a possible implementation, the operation of the terminal device generating a fourth soft information sequence according to the second soft information sequence and the third soft information sequence can be: adding the soft information of the first Q bits in the second soft information sequence and the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the first Q bits in the fourth soft information sequence, the first Q bits being the punctured bits; taking the soft information of the remaining bits in the second soft information sequence except the first Q bits as the soft information of the remaining bits in the fourth soft information sequence except the first Q bits.

[0015] In the related art, the punctured bits are assumed to be unknown information, that is, the coding input is set to 0. Such input setting will lose part of the coding block information, resulting in certain degradation of the coding performance. In the present application, the coding input of the punctured bits can be set based on the third soft information sequence obtained based on estimation, which can enhance the coding performance.

[0016] In a possible implementation, the operation of the terminal device generating a fourth soft information sequence according to the second soft information sequence and the third soft information sequence can be: taking the soft information of the first Q bits in the second soft information sequence as the soft information of the first Q bits in the fourth soft information sequence, the first Q bits being the punctured bits; adding the soft information of the Q+1th bit to the Kth bit in the second soft information sequence and the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the Q+1th bit to the Kth bit in the fourth soft information sequence, the Q+1th bit to the Kth bit being the information bits; taking the soft information of the remaining bits in the second soft information sequence except the first K bits as the soft information of the remaining bits in the fourth soft information sequence except the first K bits.

[0017] In the present application, for the information bits from the (Q+1)th bit to the Kth bit, joint decoding input can be performed based on the second soft information sequence obtained by demodulation and the third soft information sequence obtained by estimation, so as to enhance the decoding performance.

[0018] In a possible implementation, the operation of generating, by the terminal device, the fourth soft information sequence according to the second soft information sequence and the third soft information sequence can be: adding the soft information of the first Q bits in the second soft information sequence and the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the first Q bits in the fourth soft information sequence, the first Q bits being puncturing bits; adding the soft information of the (Q+1)th bit to the Kth bit in the second soft information sequence and the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the (Q+1)th bit to the Kth bit in the fourth soft information sequence, the (Q+1)th bit to the Kth bit being information bits; and taking the soft information of the remaining bits except the first K bits in the second soft information sequence as the soft information of the remaining bits except the first K bits in the fourth soft information sequence.

[0019] In the present application, not only the decoding input of the puncturing bits can be set based on the third soft information sequence obtained by estimation, but also joint decoding input of the information bits from the (Q+1)th bit to the Kth bit can be performed based on the second soft information sequence obtained by demodulation and the third soft information sequence obtained by estimation, so as to significantly enhance the decoding performance.

[0020] In a second aspect, a communication apparatus is provided, which includes a processing module and a communication module. The communication module is configured to receive n symbol sequences sent by a network device, n being a positive integer. The processing module is configured to, for each received symbol sequence, demodulate the symbol sequence to obtain a first soft information sequence; generate a second soft information sequence of a preset length according to the first soft information sequence; and decode the second soft information sequence according to side information related to updated parameters of a target neural network (NN) layer in a target AI model to obtain a target code block, the target code block including the updated parameters of the target NN layer, the side information being determined according to original parameters of the NN in the target AI model.

[0021] The second aspect is a device-side implementation corresponding to the first aspect. The explanations, supplements and beneficial effects of the first aspect also apply to the second aspect, and will not be repeated here.

[0022] In a third aspect, a communication apparatus is provided, which includes a processor. The processor is coupled with a memory and can be configured to execute instructions or data in the memory to implement the method in any possible implementation manner of any aspect described above. Optionally, the communication apparatus further includes the memory. Optionally, the communication apparatus further includes a communication interface, and the processor is coupled with the communication interface.

[0023] In an implementation manner, the communication interface can be a transceiver, or an input / output interface.

[0024] In another implementation, the communication device is a chip configured in the terminal device. When the communication device is a chip configured in the terminal device, the communication interface can be an input / output interface.

[0025] In a fourth aspect, a computer program product is provided, which includes a computer program (also referred to as code or instructions), which, when executed, causes a computer to perform the method in any possible implementation of any of the aspects.

[0026] In a fifth aspect, a computer-readable storage medium is provided, which stores a computer program (also referred to as code or instructions), which, when executed on a computer, causes the computer to perform the method in any possible implementation of any of the aspects.

[0027] In a sixth aspect, an embodiment of the present application provides a chip system, which includes one or more processors for calling and executing instructions stored in a memory, so that the method in any possible implementation of any of the aspects is performed. The chip system can be composed of a chip, or can include a chip and other discrete devices.

[0028] The chip system can include an input circuit or interface for sending information or data, and an output circuit or interface for receiving information or data.

[0029] In a seventh aspect, a communication system is provided, which includes the terminal device and the network device described above. Optionally, the communication system can further include other devices in communication with the terminal device and / or the network device. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a schematic diagram of a communication system provided by an embodiment of the present application.

[0031] Figure 2 is a schematic diagram of another communication system provided by an embodiment of the present application.

[0032] Figure 3 is a flowchart of an AI model parameter acquisition method provided by an embodiment of the present application.

[0033] Figure 4 is a schematic diagram of a hidden Markov model provided by an embodiment of the present application.

[0034] Figure 5 is a schematic diagram of data processing on the network device side provided by an embodiment of the present application.

[0035] Figure 6is a schematic diagram of data processing at a terminal device side provided by an embodiment of the present application.

[0036] Figure 7 is a schematic block diagram of a communication apparatus provided by an embodiment of the present application.

[0037] Figure 8 is a schematic block diagram of a communication apparatus provided by an embodiment of the present application. DETAILED DESCRIPTION

[0038] In the following description, for the purposes of explanation and not limitation, specific details are set forth, such as particular system configurations, techniques, etc., in order to provide a thorough understanding of the embodiments of the application. However, it will be apparent to those skilled in the art that the application can be practiced in other embodiments that depart from these specific details.

[0039] It should be understood that, when used in the specification and the appended claims of the present application, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. The terms "comprising", "including", "containing", "have" and their conjugates mean "including but not limited to", unless otherwise expressly specified.

[0040] It should be understood that the "one or more" referred to in the present application means one, two or more than two, and the "multiple" referred to in the present application means two or more than two. In the description of the present application, unless otherwise specified, " / " means or, for example, A / B can mean A or B. "And / or" in this paper is only a description of the relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone.

[0041] In order to clearly describe the technical solutions of the present application, "first", "second" and the like are used to distinguish the same items or similar items with basically the same function and role. Those skilled in the art can understand that "first", "second" and the like do not limit the quantity and execution order, and "first", "second" and the like do not necessarily mean different.

[0042] The phrases "one embodiment," "an implementation," "some implementations," "one version," "an implementation," "some implementations," "one version," or the like as used herein do not necessarily refer to the same implementation, although they can. Such phrases as used herein are understood to enable at least one embodiment of the present application while still being broad enough to include relevant embodiments of the present application. The terms "application" and "invention" used herein generally relate to one or more embodiments of the present application, regardless of

[0043] Embodiments of the present application can be applied to various communication systems, such as a global system for mobile communications (GSM) system, a general packet radio service (GPRS) system, a wireless local area network (WLAN) system (e.g., a wireless fidelity (Wi-Fi) system), a long term evolution (LTE) system, an LTE frequency division duplex (FDD) system, an LTE time division duplex (TDD) system, a sidelink communication system, a universal mobile telecommunications system (UMTS), a worldwide interoperability for microwave access (WiMAX) communication system, a non-terrestrial network (NTN) communication system, a 4th generation (4G) mobile communication system, a 5th generation (5G) mobile communication system or a new radio access technology (NR) system, a 6th generation (6G) mobile communication system, and the like. The 5G mobile communication system can include a non-standalone (NSA) and / or standalone (SA). It can be understood that embodiments of the present application can also be applied to future communication systems, and embodiments of the present application are not limited thereto.

[0044] Figure 1is a schematic diagram of a communication system 100 provided by an embodiment of the present application. The communication system 100 can include a network device, such as the network device 110 shown in Figure 1 . The communication system 100 can also include a terminal device, such as the terminal device 120 shown in Figure 1 . The network device 110 and the terminal device 120 can communicate through a wireless link. Figure 1 One network device 110 and one terminal device 120 are exemplarily shown. Alternatively, the communication system 100 can also include multiple network devices and / or multiple terminal devices.

[0045] The network device in the embodiments of the present application can be a device on the network side, such as an access network device, a core network device, etc.

[0046] The access network device is also sometimes referred to as an access node. The access network device has a wireless transceiving function and is used to communicate with the terminal device. Exemplarily, the access network device can be a base station, an evolved NodeB (eNodeB), a transmission reception point (TRP), a next generation NodeB (gNB) in a 5G mobile communication system, an access network device or a module of an access network device in an open RAN (ORAN) system, a satellite in an NTN communication system, a base station in a future mobile communication system, or an access node (AP) in a Wi-Fi system, etc. The access network device can also be a module or unit capable of realizing part of the function of a base station, such as a macro base station, a micro base station, an indoor station, a relay node, or a donor node, etc. The multiple access network devices in the communication system 100 can be the same type of base station or different types of base stations. The embodiments of the present application do not limit the specific technology and specific device form adopted by the access network device.

[0047] The core network device has functions of data processing, session management, network interconnection, operation administration and maintenance (OAM), and the like. The core network device can implement user access authentication, service bearer establishment, data interaction with external networks, and the like, and complete network configuration monitoring, resource scheduling optimization, fault maintenance, and the like through the OAM function. For example, the core network device can be a mobility management entity (MME) network element, a serving gateway (SGW) network element, a packet data network gateway (PGW) network element, and the like in a 4G mobile communication system, or an access and mobility management function (AMF) network element, a session management function (SMF) network element, a user plane function (UPF) network element, and the like in a 5G mobile communication system. For example, the core network device can also be an operation and maintenance management system, such as a network management platform (NMS) integrated with an OAM function, an automated operation and maintenance module, or a stand-alone operation and maintenance server cooperating with the core network. For example, the core network device can also be a virtualized network element in a network function virtualization (NFV) architecture, such as a cloud-deployed network function module, and the like. Optionally, a lightweight OAM capability can be integrated in the virtualized network element. The core network device can also be a new core network entity in a future communication system. The plurality of core network devices in the communication system 100 can be deployed in a centralized or distributed architecture, and each core network device can bear the same or different types of network functions. The embodiments of the present application do not limit the specific technologies and specific device forms adopted by the core network device.

[0048] In the embodiments of the present application, the device for implementing the function of the network device can be a network device or a device capable of supporting the network device to implement the function, such as a processor, a circuit, a chip, or a chip system. The device can be installed in the network device or used in connection with the network device. In the embodiments of the present application, the device for implementing the function of the network device is taken as an example to describe the technical solutions provided by the present application.

[0049] The terminal device in the embodiments of the present application can be a wireless terminal device capable of receiving network device scheduling and indication information. The wireless terminal device can be a device that provides voice and / or data connectivity for a user, or a handheld device with a wireless connection function, or other processing devices connected to a wireless modem. For example, the terminal device can communicate with one or more core networks or the Internet through a radio access network (RAN). The terminal device can also be referred to as a terminal, user equipment (UE), mobile station, mobile terminal, etc. The terminal device can be widely used in various scenarios, such as device-to-device (D2D), vehicle-to-everything (V2X) communication, machine-type communication (MTC), internet of things (IOT), ultra-reliable low-latency communication (URLLC), virtual reality, augmented reality, industrial control, automatic driving, remote medical treatment, smart power grid, smart furniture, smart office, smart wear, smart transportation, smart city, or satellite communication, etc. The terminal device can be a mobile phone, a tablet computer, a computer with wireless transceiver function, a wearable device, a vehicle, an aircraft (such as a drone, a helicopter, an airplane), a hot air balloon, a ship, a robot, a mechanical arm, or a smart home device, etc. The embodiments of the present application do not limit the form of the terminal device.

[0050] In the embodiments of the present application, the device for implementing the function of the terminal device can be a terminal device, or a device capable of supporting the terminal device to implement the function, such as a processor, a circuit, a chip, a chip system, etc. The device can be installed in the terminal device or used in connection with the terminal device. In the embodiments of the present application, the device for implementing the function of the terminal device is taken as an example to describe the technical solutions provided by the present application.

[0051] The application scenarios related to the embodiments of the present application are described below.

[0052] In a wireless communication system, AI models can be used to enhance system performance. For example, in a 6G mobile communication system, in the application scenarios of AI air interface or AI enhanced mobility management, in which communication and AI are deeply integrated, a terminal device can deploy multiple special AI models to perform key inference tasks such as channel state information (CSI) prediction / compression, beamforming optimization, and resource scheduling allocation. To enable these AI models to achieve the best system performance in a specific wireless environment (such as a specific channel model, interference condition, and user distribution), the model training process needs to be pre-configured with a series of key parameters and targets, including setting the training channel environment parameters (such as a specific signal-to-noise ratio range), selecting the appropriate optimization target (such as a specific loss function), and constructing a training dataset with sufficient representativeness (covering the typical characteristics and dynamic changes of the target scenario).

[0053] However, when the actual operating environment of the deployed AI model changes significantly, the AI model may not adapt to the changed environment, resulting in a sharp degradation of model performance. For example, a CSI prediction model is trained under a set signal-to-noise ratio range, but is deployed in an environment with a significantly different actual signal-to-noise ratio, or the input data distribution processed by the model during inference is significantly different from the data distribution learned during the training phase. For such cases where the inference environment does not match the training environment, the common solution is to fine-tune the model, that is, to incrementally train a small amount of samples in the new environment based on the original model parameters. However, due to the inherent resource constraints of terminal devices, such as insufficient computing power, limited storage space, and limited energy supply, terminal devices cannot independently complete large-scale local fine-tuning of model parameters. Therefore, as shown in Figure 2

[0054] AI models usually include a large number of high-precision NN parameters, which are easily affected by channel noise during wireless transmission. If the NN parameters received by the terminal device are not accurate, the model performance will be severely degraded. However, it is particularly difficult to ensure high-fidelity recovery of NN parameters in low signal-to-noise ratio channel conditions. If the traditional automatic repeat request (ARQ) mechanism is relied on to ensure the reliability of NN parameter transmission, it will cause huge signaling overhead and wireless resource consumption due to frequent retransmission. This not only easily causes wireless link congestion, but also significantly prolongs the end-to-end delay of model updating, making the terminal device unable to obtain the updated AI model in a timely manner, and ultimately hindering the effective adaptation of AI models in dynamic environments.

[0055] ​To this end, an embodiment of the present application provides an AI model parameter acquisition method, which provides a NN parameter decoding algorithm with side information. Specifically, the terminal device can extract side information for the fine-tuned NN parameter based on the correlation between the pre-fine-tuned NN parameter and the fine-tuned NN parameter, and then decode based on the side information to enhance the decoding performance. In this way, the recovery accuracy and robustness of the NN parameter under poor wireless channel conditions such as low signal-to-noise ratio, high interference, and fast time variation can be significantly improved. The side information is a core element of the Slepian-Wolf theory, which refers to the auxiliary information related to the to-be-decoded source that the decoding end already has. The decoding end can achieve efficient reconstruction by combining the side information with the received data. In some embodiments, the AI model life cycle management (LCM) framework of 6G can be applied to the AI model life cycle management (LCM) framework of 6G.

[0056] Optionally, the AI model in the embodiment of the present application can also be referred to as a machine learning (ML) model. Optionally, the decoding in the embodiment of the present application can also be referred to as decoding.

[0057] The technical principle related to the AI model parameter acquisition method provided by the embodiment of the present application is described below:

[0058] Based on the current model fine-tuning algorithm (including but not limited to low-rank adaptation (LoRA) fine-tuning algorithm, etc.), the relationship between the original NN parameter matrix and the fine-tuned NN parameter matrix can be modeled as: , is the NN update parameter accumulated during model fine-tuning. It is assumed that is the binary quantization sequence of , is the binary quantization sequence of . The terminal device has the AI model locally deployed, so the terminal device has . Based on this, a correlation transmission model with side information can be modeled based on the Slepian-Wolf theory to regard the original sequence as the target sequence . The correlation transmission model can be expressed as: where is the correlation channel noise, which is related to . Since the correlation channel noise The distribution characteristics of the target sequence are difficult to accurately estimate, so the embodiment of the application is based on the correlation transmission model, and the side information of the target sequence is constructed according to the known original sequence The side information of the target sequence is constructed The constructed side information is input as additional prior knowledge for decoding, which can greatly enhance the decoding performance. Compared with the independent decoding algorithm, the embodiment of the application can realize lossless data transmission with a higher code rate.

[0059] The AI model parameter acquisition method provided by the embodiment of the application will be described in detail below in conjunction with the corresponding flowchart. It can be understood that the AI model parameter acquisition method is illustrated by mainly taking different devices (such as terminal devices and network devices) as the interactive execution subject in the illustrative flowchart, but the embodiment of the application does not limit the interactive execution subject. For example, the devices (such as terminal devices and network devices) in the illustrative flowchart can also be chips, chip systems, or processors that support the devices to implement the AI model parameter acquisition method, and can also be logic modules or software that can implement all or part of the functions of the devices.

[0060] Here, it is uniformly stated that the messages or signaling interactions involved in the interactive process of the embodiment of the application can adopt messages or signaling in standards or newly introduced messages or signaling, and the embodiment of the application does not limit this.

[0061] The explanations of some technical terms in the embodiment of the application can refer to the explanations in the 3rd generation partnership project (3rd generation partnership project, 3GPP) standard protocol. It should be understood that the technical terms in the embodiment of the application are only used as examples and are not limited. As technology evolves, technical terms will also change, and other technical terms should also apply to the embodiment of the application in the case of the same technical meaning.

[0062] It can be understood that the following Figure 3 terminal devices described in the embodiments can be any terminal devices described in the above Figure 1 embodiments, or can be devices (such as processors, chips, or chip systems, etc.) in terminal devices. The following Figure 3 network devices described in the embodiments can be any network devices (such as access network devices or core network devices) described in the above Figure 1 embodiments, or can be devices (such as processors, chips, or chip systems, etc.) in network devices.

[0063] Figure 3 is a flowchart of an AI model parameter acquisition method provided by the embodiment of the application. As Figure 3 shown, the AI model parameter acquisition method can include the following steps:

[0064] Step 301: The network device obtains the NN update parameter of the target AI model.

[0065] The target AI model is an AI model that needs to be updated. The NN update parameter is used to update the target AI model. The NN update parameter can be part or all of the NN parameters in the target AI model.

[0066] The NN update parameter can include one or more update parameters in one or more NN layers in the target AI model. The update parameter of an NN layer is used to update the NN layer.

[0067] The network device can be an access network device or a core network device, and the embodiments of the present application do not limit this.

[0068] The NN update parameter can be generated by the network device or received by the network device from other devices. For example, the network device is a base station. The base station can receive an updated target AI model sent by a core network device (such as a cloud server or an edge server, etc.), and the base station can determine the NN update parameter of the target AI model according to the target AI model before updating and the updated target AI model. Alternatively, the base station can receive the NN update parameter of the target AI model sent by the core network device. Of course, the NN update parameter can also be obtained by the network device in other ways, and the embodiments of the present application do not limit this.

[0069] Step 302: For the update parameter of any NN layer (hereinafter referred to as the target NN layer) in the NN update parameter, the network device obtains the binary quantization sequence of the update parameter of the target NN layer.

[0070] It should be noted that in the case where the NN update parameter only includes the update parameter of one NN layer, the network device can send the update parameter of the NN layer to the terminal device. In the case where the NN update parameter includes the update parameters of multiple NN layers, the network device can sequentially send the update parameter of each NN layer in the multiple NN layers to the terminal device in sequence.

[0071] For the update parameter of a certain NN layer (i.e. the target NN layer) that needs to be sent currently, the network device can first perform planarization processing and quantization processing on the update parameter of the target NN layer to obtain the corresponding binary quantization sequence, and then perform block sending. For example, the binary quantization sequence can be encapsulated as a single transport block (TB).

[0072] Optionally, before the network device transmits the binary quantization sequence of the update parameter of the target NN layer in blocks, the network device can first transmit control information to the terminal device according to the binary quantization sequence. The control information can include a target AI model identifier (ID), a target NN layer identifier, and a total number of blocks. After receiving the control information, the terminal device can know the target AI model identifier, the target NN layer identifier, and the total number of blocks.

[0073] The target AI model identifier is used to identify the target AI model. The target NN layer identifier is used to identify the target NN layer. The total number of blocks is the number of blocks into which the binary quantization sequence is divided when being transmitted. Here, it is assumed that the total number of blocks is n, where n is a positive integer.

[0074] That is, before the network device needs to transmit the update parameter of an NN layer to the terminal device, the network device can first transmit control information related to the NN layer to the terminal device, so that the terminal device can accurately decode the NN parameter according to the control information subsequently.

[0075] For example, the network device can transmit the control information to the terminal device in an AI model update control signaling. Of course, the network device can also transmit the control information to the terminal device in other messages or signalings, which are not limited in the embodiments of the present application.

[0076] Step 303: The network device divides the binary quantization sequence of the update parameter of the target NN layer into n blocks.

[0077] Optionally, the network device can segment the binary quantization sequence in sequence according to a first preset length to obtain n blocks, and the n blocks exist in order. The length of each block in the n blocks is the first preset length. The first preset length can be set in advance, and the unit of the first preset length is bit. Here, it is assumed that the first preset length is L.

[0078] As can be easily understood, one or more blocks of the first preset length can be continuously cut from the binary quantization sequence in sequence. In some cases, when the length of the remaining bits in the binary quantization sequence (assuming r) is insufficient (i.e., less than the first preset length), the entire remaining bits are taken as the effective part in the nth block (i.e., the last block) in the n blocks, and L-r bits are filled after the effective part, which can usually be filled with 0 or filled according to the agreement. Based on this, the n blocks can be directly cut from the binary quantization sequence; or the first n-1 blocks in the n blocks are directly cut from the binary quantization sequence, the first bit to the rth bit in the nth block in the n blocks are cut from the binary quantization sequence, and the r+1th bit to the Lth bit in the nth block are filled.

[0079] Step 304: The network device respectively performs channel encoding on each of the n sub-blocks to obtain n first encoded blocks.

[0080] The length of each of the n first encoded blocks is a second preset length. The second preset length can be preset, and the unit of the second preset length can be bit. Here, it is assumed that the second preset length is K.

[0081] One first encoded block is obtained by adding a check bit after one sub-block. The first encoded block is a systematic code, that is, the first L bits in the first encoded block are information bits, and the (L+1)th to Kth bits in the first encoded block are check bits. The information bits include the update parameters of the NN layer. The check bits are redundant data generated by a channel encoding algorithm.

[0082] Optionally, for any one of the n sub-blocks, the network device can input the sub-block into a channel encoding module, and output a first encoded block from the channel encoding module.

[0083] The channel encoding module is configured to perform channel encoding. For example, the channel encoding module can be a module using a low density parity check code (LDPC) encoding algorithm, that is, an LDPC encoding module; or the channel encoding module can be a module using a Turbo encoding algorithm, that is, a Turbo encoding module. Of course, the channel encoding module can also be a module using other channel encoding algorithms, which are not limited in the embodiments of the present application.

[0084] Step 305: The network device respectively performs puncturing operation on each of the n first encoded blocks to obtain n second encoded blocks.

[0085] The length of each of the n second encoded blocks is a third preset length. The third preset length can be preset, and the unit of the third preset length can be bit. Here, it is assumed that the third preset length is I.

[0086] The puncturing operation is configured to delete part of the information bits in the first encoded block, and retain all the check bits. Here, it is assumed that the puncturing operation is configured to delete the first Q bits in the first encoded block to match the transmission rate of the current channel. In this case, Q

[0087] In some embodiments, the above sub-blocks, first encoded blocks and second encoded blocks can be collectively referred to as code blocks (CBs).

[0088] Step 306: The network device respectively modulates each of the n second encoded blocks to obtain n symbol sequences.

[0089] The n symbol sequences are complex-valued symbol sequences.

[0090] Optionally, the network device can send downlink control information (DCI) to the terminal device, and the DCI can include the time-frequency position of each second coded block in the plurality of second coded blocks.

[0091] Step 307: The network device sends the n symbol sequences to the terminal device.

[0092] Step 308: The terminal device receives the n symbol sequences sent by the network device, wherein, for each received symbol sequence, the symbol sequence is demodulated to obtain a first soft information sequence.

[0093] The terminal device can separate the symbol sequence corresponding to each second coded block from the received symbol sequence according to the time-frequency position of each second coded block indicated by the DCI sent by the network device, and can determine the block index of each symbol sequence.

[0094] The soft information can be a log-likelihood ratio (LLR), which can be an estimated value. LLR is used to quantify the reliability confidence of a bit. LLR>0 indicates that the probability of judging a bit as "0" is higher, and LLR<0 indicates that the probability of judging a bit as "1" is higher. Wherein, the larger the |LLR| is, the higher the confidence is, the smaller the |LLR| is, the greater the noise influence is, and the easier the decision is to make mistakes. LLR=0 indicates that the confidence is neutral, representing "no information".

[0095] The length of the first soft information sequence is the same as the length of the second coded block, that is, both are a third preset length.

[0096] Optionally, the terminal device can demodulate the symbol sequence to obtain the first soft information sequence by the following formula.

[0097]

[0098] Wherein, is the symbol sequence corresponding to the second coded block The LLR of the kth bit in the symbol sequence is the symbol value of the ith bit in the symbol sequence . is the channel noise variance, which determines the received signal quality.

[0099] Step 309: The terminal device generates a second soft information sequence according to the first soft information sequence.

[0100] The length of the second soft information sequence is the same as that of the first encoded block, i.e., both are the second preset length. The first L bits in the second soft information sequence are the soft information of the information bits, and the (L+1)th to Kth bits in the second soft information sequence are the soft information of the check bits.

[0101] Optionally, the terminal device can further add Q bits at the head of the first soft information sequence, and set the Q bits to 0 to obtain the second soft information sequence. In this case, the first Q bits in the second soft information sequence (i.e., the punctured bits in the information bits) are 0, and the (Q+1)th to Kth bits in the second soft information sequence are the first soft information sequence.

[0102] Step 310: The terminal device decodes the second soft information sequence according to the side information related to the updated parameters of the target NN layer in the AI model to obtain a target encoded block, and the target encoded block includes the updated parameters of the target NN layer.

[0103] The side information related to the updated parameters of the target NN layer is determined according to the original NN parameters of the target AI model. The side information related to the updated parameters of the target NN layer is side information for the updated parameters. In this case, performing side information assisted joint decoding on the second soft information sequence can accurately recover the updated parameters of the target NN layer.

[0104] The terminal device is deployed with an AI model, and thus has the original NN parameters of the target AI model. The terminal device can determine the side information related to the updated parameters of the target NN layer based on this. In the embodiments of the present application, the side information related to the updated parameters of the target NN layer is input as additional prior knowledge for decoding, which can greatly improve the decoding performance. Moreover, since the decoding gain of the embodiments of the present application is higher, there is a certain performance redundancy, so even in poor channel conditions, decoding can be accurately completed, and the decoding success rate is high. In this way, robust transmission and accurate recovery of the NN parameters can be achieved. In addition, due to the improvement of the decoding performance, the sending end can adopt a more aggressive encoding strategy to further improve the transmission efficiency.

[0105] In some embodiments, the terminal device receives control information sent by the network device before receiving the n symbol sequences. According to the target AI model identifier, the target NN layer identifier and the total number of blocks in the control information, the terminal device can know that the updated parameters of the target NN layer in the target AI model are to be received next, and can know that the n symbol sequences correspond to the updated parameters.

[0106] In this case, the terminal device can obtain the NN original parameter of the target AI model. The binary quantization sequence of the NN original parameter of the target AI model and the parameters of the hidden Markov model are input into a forward-backward algorithm (FBA) module to obtain output data of the FBA module. Then, according to the output data of the FBA module, a soft information sequence of the binary quantization sequence of the NN update parameter of the target AI model is estimated to obtain a target soft information sequence. The target soft information sequence includes side information related to the update parameter of the target NN layer. That is, the side information related to the update parameter of the target NN layer is the soft information sequence of the binary quantization sequence of the update parameter of the target NN layer, that is, the soft information sequence corresponding to the target NN layer. The soft information sequence of one binary quantization sequence includes the soft information of each bit in the binary quantization sequence.

[0107] The hidden Markov model is used to represent the correlation between the binary quantization sequence of the NN original parameter of the target AI model and the binary quantization sequence of the NN update parameter of the target AI model.

[0108] For example, as shown in the figure, Figure 4 the embodiment of the present application models the correlation between the binary quantization sequence of the NN original parameter of the target AI model and the binary quantization sequence of the NN update parameter of the target AI model as a hidden Markov model, Figure 4 For example, a binary hidden Markov model is used. Wherein, is an observable sequence, is a prediction sequence.

[0109] Suppose and have a length of M, and the state of the mth bit is , the initial input of the FBA module is: . Wherein, is the initial parameter of the hidden Markov model. is the initial hidden state probability; is the state transition probability matrix, represents the corresponding state transition probability; is the observation probability matrix, represents the probability of being 0 or 1 under the corresponding hidden state.

[0110] As the binary quantization sequence of the NN original parameter, the past sequence is observable, and the target of the FBA module is to estimate by the hidden Markov model and the known ​​​The FBA module corrects the parameters of the hidden Markov model in an iterative process When the observable sequence sample is long enough, the influence of the error of the initial input state can be ignored. Based on this, the FBA module can output after reaching a preset number of iterations: wherein represents the transition probability in the corresponding state, represents the probability of , and represents the probability of

[0111] . Thus, the LLR of the mth bit in can be determined by the following formula: LLRm=ln(Pm / 1-Pm), where LLR is an estimated value.

[0112]

[0113] wherein represents the joint probability of the following two events given the current hidden state : 1. the system remains in the same state: the hidden state at the next moment is still (that is, ); and 2. a specific bit is generated: the observed bit output at the current moment is . represents the joint probability of the following two events given the current hidden state : 1. the system remains in the same state: the hidden state at the next moment is still (that is, ); and 2. a specific bit is generated: the observed bit output at the current moment is . represents the probability of . represents the probability of .

[0114] In this way, the LLR sequence of can be estimated, that is, the target soft information sequence is obtained.

[0115] It should be noted that there can be one or more parameters of the NN layer in the target AI model that need to be updated. In this case, before the network device sends each symbol sequence corresponding to the updated parameters of a NN layer to the terminal device, the network device can first send control information carrying the identifier of the target AI model, the identifier of the NN layer, and the total number of blocks corresponding to the updated parameters of the NN layer to the terminal device. In the embodiments of the present application, the terminal device does not need to calculate the target soft information sequence every time it receives the control information, but can calculate the target soft information sequence when it first receives the control information related to the update of the target AI model, and does not need to calculate the target soft information sequence again when it subsequently receives the control information related to the update of the target AI model, so as to save computing resources.

[0116] In this case, the terminal device has obtained the target soft information sequence according to the original parameters of the NN of the target AI model before receiving the symbol sequence corresponding to the updated parameters of any one of the NN layers in the target AI model. The target soft information sequence includes the soft information sequence of the binary quantization sequence of the updated parameters of each of the one or more NN layers in the target AI model, that is, the target soft information sequence includes the soft information sequence corresponding to each of the one or more NN layers. In this case, the side information related to the updated parameters of the target NN layer can be the soft information sequence corresponding to the target NN layer.

[0117] It should be noted that since the control information received by the terminal device includes the identifier of the target AI model, the identifier of the target NN layer, and the total number of blocks (i.e., n), after the terminal device receives the control information and obtains the target soft information sequence, the terminal device can also obtain the soft information sequence corresponding to the target NN layer from the target soft information sequence, and divide the soft information sequence corresponding to the target NN layer into n soft information sequences. In this way, the terminal device can subsequently perform decoding processing based on the n soft information sequences.

[0118] The terminal device can segment the soft information sequence corresponding to the target NN layer in order according to the first preset length to obtain n soft information sequences, and the n soft information sequences exist in order. In this case, the length of each soft information sequence in the n soft information sequences is the first preset length, that is, L.

[0119] It is easily understood that one or more soft information sequences of the first preset length can be continuously cut from the soft information sequence corresponding to the target NN layer in sequence. In some cases, when the length of the remaining bits (assuming r) in the soft information sequence corresponding to the target NN layer is insufficient (i.e., less than the first preset length), the entire remaining bits are taken as the effective part in the nth soft information sequence of the n soft information sequences, and L-r bits are padded after the effective part, which can usually be filled with 0. Based on this, the n soft information sequences can be directly cut from the soft information sequence corresponding to the target NN layer; or the first n-1 soft information sequences of the n soft information sequences are directly cut from the soft information sequence corresponding to the target NN layer, the first r bits of the nth soft information sequence of the n soft information sequences are cut from the soft information sequence corresponding to the target NN layer, and the r+1th to Lth bits of the nth soft information sequence are padded.

[0120] In this case, the side information related to the update parameters of the target NN layer can be the n soft information sequences. The operation of step 309 can be: generating a fourth soft information sequence according to the second soft information sequence and a third soft information sequence, the third soft information sequence being one of the n soft information sequences corresponding to the second soft information sequence; inputting the fourth soft information sequence into the channel decoding module to obtain the target coded block output by the channel decoding module.

[0121] The n soft information sequences correspond one-to-one to n symbol sequences corresponding to the update parameters of the target NN layer. Since the n soft information sequences exist in order, the terminal device can obtain one of the n soft information sequences as the third soft information sequence according to the block index of one of the symbol sequences corresponding to the second soft information sequence.

[0122] The channel decoding module is used for channel decoding. For example, the channel decoding module can be an LDPC decoding module, a Turbo decoding module, etc., which is not limited in the embodiments of the present application. For example, the LDPC decoding module can use a belief propagation (BP) decoding algorithm. The Turbo decoding module can use a maximum a posteriori probability iterative decoding algorithm.

[0123] It should be noted that the channel encoding module should correspond to the channel decoding module. That is, when the channel encoding module in the network device is an LDPC encoding module, the channel decoding module in the terminal device can be an LDPC decoding module; when the channel encoding module in the network device is a Turbo encoding module, the channel decoding module in the terminal device can be a Turbo decoding module.

[0124] The length of the fourth soft information sequence is the same as the length of the second soft information sequence. The operation of generating the fourth soft information sequence according to the second soft information sequence and the third soft information sequence can include any one of the following three manners:

[0125] In the first manner, the terminal device adds the soft information of the first Q bits in the second soft information sequence and the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the first Q bits in the fourth soft information sequence, and the first Q bits are punctured bits; and the soft information of the remaining bits in the second soft information sequence except the soft information of the first Q bits is taken as the soft information of the remaining bits in the fourth soft information sequence except the first Q bits.

[0126] It should be noted that in the related art, the punctured bits are assumed to be unknown information, that is, the decoding input is set to 0, and such input setting will lose part of the coded block information, resulting in certain degradation of decoding performance. In the embodiments of the present application, the decoding input of the punctured bits can be set based on the third soft information sequence obtained by estimation, so that the decoding performance can be enhanced.

[0127] In the second manner, the terminal device takes the soft information of the first Q bits in the second soft information sequence as the soft information of the first Q bits in the fourth soft information sequence, and the first Q bits are punctured bits; adds the soft information of the Q+1th bit to the Kth bit in the second soft information sequence and the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the Q+1th bit to the Kth bit in the fourth soft information sequence, and the Q+1th bit to the Kth bit are information bits; and takes the soft information of the remaining bits in the second soft information sequence except the soft information of the first K bits as the soft information of the remaining bits in the fourth soft information sequence except the first K bits.

[0128] In the embodiments of the present application, for the information bits of the Q+1th bit to the Kth bit, joint decoding input can be performed based on the second soft information sequence obtained by demodulation and the third soft information sequence obtained by estimation, so that the decoding performance can be enhanced.

[0129] In the third manner, the terminal device adds the soft information of the first Q bits in the second soft information sequence and the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the first Q bits in the fourth soft information sequence, and the first Q bits are punctured bits; adds the soft information of the Q+1th bit to the Kth bit in the second soft information sequence and the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the Q+1th bit to the Kth bit in the fourth soft information sequence, and the Q+1th bit to the Kth bit are information bits; and takes the soft information of the remaining bits in the second soft information sequence except the soft information of the first K bits as the soft information of the remaining bits in the fourth soft information sequence except the first K bits.

[0130] In this embodiment, not only can the decoding input of the punched bit be set based on the estimated third soft information sequence, but the joint decoding input of the information bits from the Q+1th bit to the Kth bit can also be performed based on the demodulated second soft information sequence and the estimated third soft information sequence, thereby significantly enhancing the decoding performance.

[0131] For example, the operation of obtaining the fourth soft information sequence in the third method can be represented by the following formula:

[0132]

[0133] It should be noted that the channel decoding module generally has a pre-set maximum number of iterations. If the decoded information bits and check bits meet the requirements before the maximum number of iterations is reached, the decoding result (i.e., the coded block) can be output. If the decoded information bits and check bits still do not meet the requirements even after the maximum number of iterations is reached, decoding will fail. This application embodiment improves the decoding success rate by introducing side information for joint decoding input. Furthermore, compared to related technologies, the prior knowledge added in this application embodiment can accelerate convergence, thereby reducing the number of iterations and increasing the decoding speed.

[0134] In this embodiment, the network device can send n symbol sequences corresponding to the update parameters of the target NN layer. Each received symbol sequence can be decoded to obtain a target coding block, ultimately resulting in n target coding blocks corresponding one-to-one with the n symbol sequences. Based on these n target coding blocks, the update parameters of the target NN layer can be determined. Furthermore, the network device can sequentially send symbol sequences corresponding to the update parameters of each of the one or more NN layers in the target AI model that need updating to the terminal device. Based on these sequences, the terminal device can determine the update parameters of each of the one or more NN layers, ultimately obtaining the NN update parameters of the target AI model. The terminal device can then update the target AI module accordingly.

[0135] To facilitate understanding, the following will be combined with... Figure 5 and Figure 6 An example is provided to illustrate the process of transmitting AI model parameters:

[0136] like Figure 5As shown, on the network device side, for the update parameters of the NN layer in the AI ​​model, the update parameters are planarized and quantized to obtain a binary quantized sequence; the binary quantized sequence is divided into blocks to obtain n blocks; each of the n blocks is channel-coded to obtain n first coded blocks; each of the n first coded blocks is punctured to obtain n second coded blocks. Each of the n second coded blocks is modulated to obtain n symbol sequences, which are then sent to the terminal device.

[0137] like Figure 6 As shown, on the terminal device side, after receiving one symbol sequence from n symbol sequences, the symbol sequence is demodulated to obtain an LLR sequence. This LLR sequence is then jointly decoded with the LLR sequence related to the NN update parameters estimated based on the original NN parameters, yielding a decoding result. Each decoded result is processed to obtain n blocks. Based on these n blocks and the AI ​​model ID, NN layer ID, and total number of blocks in the AI ​​model update control signaling sent by the network device, the NN layer in the updatable AI model is determined, and its parameters are updated according to these n blocks.

[0138] In this embodiment, the terminal device can determine the side information related to the update parameters of the NN layer based on the original NN parameters of the AI ​​model. This terminal device can use the side information related to the update parameters of the NN layer as additional prior knowledge and perform joint decoding input with the soft information sequence obtained from demodulation, thereby significantly improving decoding performance. Thus, even under poor channel conditions, decoding can be completed relatively accurately with a high success rate, contributing to robust transmission and accurate recovery of the NN parameters.

[0139] It should be understood that Figures 1 to 6 The flowcharts or scene diagrams shown are for illustrative purposes only and are not intended to limit the embodiments of this application to the examples illustrated. In fact, those skilled in the art can interpret the embodiments based on... Figures 1 to 6 The examples in the document can be transformed into equivalent ways to obtain more implementations.

[0140] The above text combined Figures 1 to 6 This document describes in detail the AI ​​model parameter acquisition method provided in the embodiments of this application. The following will combine... Figures 7 to 8 The device embodiments of this application are described in detail below. It should be understood that the communication device of this application embodiment can execute the various AI model parameter acquisition methods described in the foregoing embodiments of this application. That is, the specific working processes of the various products below can be referred to the corresponding processes in the foregoing method embodiments.

[0141] In the embodiments above, the terminal device can perform some or all of the steps in the embodiments; the network device can perform some or all of the steps in the embodiments. These steps or operations are only examples, and the embodiments of the present application can also perform other operations or variations of the various operations. In addition, the various steps can be performed in different orders presented in the embodiments, and it is possible that not all operations in the embodiments of the present application are performed. The magnitude of the serial number of each step does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0142] Figure 7 is a schematic block diagram of a communication apparatus 700 provided by an embodiment of the present application. As shown in Figure 7 , the communication apparatus 700 can include a communication module 720. The communication module 720 can implement a corresponding communication function, which can be an internal communication function of the communication apparatus 700, or a communication function of the communication apparatus 700 and other apparatuses. Alternatively, the communication module 720 can also be referred to as a communication interface or a transceiver module. Alternatively, the communication apparatus 700 further includes a processing module 710. The processing module 710 can implement a corresponding processing function.

[0143] Alternatively, the communication apparatus 700 further includes a storage module, which can be used to store instructions and / or data; the processing module 710 can read the instructions and / or data in the storage module, so that the communication apparatus 700 implements the foregoing method embodiments.

[0144] In a possible design, the communication apparatus 700 can correspond to the terminal device in the foregoing method embodiments, or a component (such as a circuit, a chip or a chip system, etc.) configured in the terminal device. The communication apparatus 700 can be used to perform the steps or processes performed by the terminal device in any of the foregoing method embodiments.

[0145] For example, the communication module 720 is configured to receive n symbol sequences sent by a network device, where n is a positive integer;

[0146] The processing module 710 is configured to, for each received symbol sequence, demodulate the symbol sequence to obtain a first soft information sequence; generate a second soft information sequence of a preset length according to the first soft information sequence; and perform decoding on the second soft information sequence according to side information related to updated parameters of a target NN layer in a target AI model to obtain a target code block, where the target code block includes the updated parameters of the target NN layer, and the side information is determined according to original parameters of the NN in the target AI model.

[0147] The communication module 720 is configured to receive control information sent by the network device, where the control information includes a target AI model identifier, a target NN layer identifier, and a total number of blocks, and the total number of blocks is n.

[0148] The processing module 710 is configured to obtain the NN original parameters of the target AI model, input a binary quantization sequence of the NN original parameters of the target AI model and parameters of a hidden Markov model into the FBA module, obtain output data of the FBA module, and estimate a soft information sequence of a binary quantization sequence of the NN updated parameters according to the output data of the FBA module, to obtain a target soft information sequence, where the hidden Markov model is used to represent a correlation between the binary quantization sequence of the NN original parameters of the target AI model and the binary quantization sequence of the NN updated parameters of the target AI model, and the target soft information sequence includes the side information.

[0149] The processing module 710 is configured to obtain a soft information sequence corresponding to the target NN layer from the target soft information sequence, and determine the target soft information sequence according to the NN original parameters of the target AI model, and divide the soft information sequence corresponding to the target NN layer into n soft information sequences, where the n soft information sequences are the side information.

[0150] The processing module 710 is configured to generate a fourth soft information sequence according to a second soft information sequence and a third soft information sequence, the third soft information sequence is one of the n soft information sequences corresponding to the second soft information sequence, and input the fourth soft information sequence into the channel decoding module to obtain a target coding block output by the channel decoding module.

[0151] The processing module 710 is configured to add the soft information of the first Q bits in the second soft information sequence and the soft information of corresponding bits in the third soft information sequence one by one to obtain the soft information of the first Q bits in the fourth soft information sequence, the first Q bits are puncturing bits, and the soft information of the remaining bits in the second soft information sequence except the first Q bits is used as the soft information of the remaining bits in the fourth soft information sequence except the first Q bits.

[0152] The processing module 710 is configured to use the soft information of the first Q bits in the second soft information sequence as the soft information of the first Q bits in the fourth soft information sequence, the first Q bits are puncturing bits, add the soft information of the Q+1th bit to the Kth bit in the second soft information sequence and the soft information of corresponding bits in the third soft information sequence one by one to obtain the soft information of the Q+1th bit to the Kth bit in the fourth soft information sequence, the Q+1th bit to the Kth bit are information bits, and the soft information of the remaining bits in the second soft information sequence except the first K bits is used as the soft information of the remaining bits in the fourth soft information sequence except the first K bits.

[0153] For example, the processing module 710 is used to add the soft information of the first Q bits in the second soft information sequence to the corresponding soft information of the third soft information sequence one by one to obtain the soft information of the first Q bits in the fourth soft information sequence, where the first Q bits are punch bits; add the soft information of the (Q+1)th to the Kth bits in the second soft information sequence to the corresponding soft information of the third soft information sequence one by one to obtain the soft information of the (Q+1)th to the Kth bits in the fourth soft information sequence, where the (Q+1)th to the Kth bits are information bits; and use the soft information of the remaining bits in the second soft information sequence other than the first K bits as the soft information of the remaining bits in the fourth soft information sequence other than the first K bits.

[0154] The above are merely examples; for detailed steps or procedures, please refer to the descriptions in the foregoing embodiments.

[0155] Figure 8 This is a schematic block diagram of a communication device 800 provided in an embodiment of this application. The communication device 800 may be a chip, chip system, or processor, etc., in a terminal device or network device that implements the above-described methods. The communication device 800 can be used to implement the methods described in the above-described method embodiments, and for details, please refer to the descriptions in the above-described method embodiments.

[0156] like Figure 8 As shown, the communication device 800 may include one or more processors 810, which may also be referred to as processing units or processing modules, and can implement certain control functions. The processor 810 may be a general-purpose processor or a dedicated processor, such as a baseband processor or a central processing unit. The baseband processor can be used to process communication protocols and communication data, while the central processing unit can be used to control the communication device 800 (e.g., a base station, baseband chip, user, user chip), execute software programs, and process data from the software programs.

[0157] In an alternative design, the processor 810 may also store instructions and / or data that can be executed by the processor 810 to cause the communication device 800 to perform the methods described in the above method embodiments.

[0158] In another alternative design, the communication device 800 may include a communication interface 820 for implementing receiving and transmitting functions. For example, the communication interface 820 may be a transceiver circuit, interface, interface circuit, or transceiver. The transceiver circuit, interface, interface circuit, or transceiver for implementing receiving and transmitting functions may be separate or integrated. The aforementioned transceiver circuit, interface, interface circuit, or transceiver may be used for reading and writing code / data, or it may be used for transmitting or relaying signals.

[0159] Optionally, one or more memories 830 can be included in the communications device 800, on which instructions can be stored, which can be run on the processor 810, so that the communications device 800 performs the methods described in the above method embodiments. Optionally, the memories 830 can also store data. Optionally, the processor 810 can also store instructions and / or data. The processor 810 and the memories 830 can be separately arranged, or can be integrated together.

[0160] It should be understood that, in a possible design, each step in the method embodiments provided in the present application can be completed by integrated logic circuits of hardware in the processor or instructions in the form of software. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being completed by a hardware processor, or being completed by a combination of hardware and software modules in the processor. The software modules can be located in a storage medium which is mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads information in the memory, and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0161] In an implementation manner, the communications device 800 can correspond to the terminal device in the above method embodiments, and can be used to execute each step and / or process executed by the terminal device in the above method embodiments. The processor 810 can be used to execute the instructions stored in the memories 830, and when the processor 810 executes the instructions stored in the memories, the processor 810 is used to execute each step and / or process of the above method embodiments corresponding to the terminal device.

[0162] In another implementation manner, the communications device 800 can correspond to the network device in the above method embodiments, and can be used to execute each step and / or process executed by the network device in the above method embodiments. The processor 810 can be used to execute the instructions stored in the memories 830, and when the processor 810 executes the instructions stored in the memories, the processor 810 is used to execute each step and / or process of the above method embodiments corresponding to the network device.

[0163] It should be understood that the above-mentioned processing device can be one or more chips. For example, the processing device can be a field programmable gate array (FPGA), can be an application specific integrated circuit (ASIC), can also be a system on chip (SoC), can also be a central processor unit (CPU), can also be a network processor (NP), can also be a digital signal processor (DSP), can also be a micro controller unit (MCU), can also be a programmable logic device (PLD) or other integrated chip.

[0164] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically EPROM (EEPROM) or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM) and direct rambus RAM (DR RAM). It should be noted that the memory of the system and method described herein is intended to include, but not limited to, these and any other suitable types of memory.

[0165] According to the method provided in the embodiments of the present application, the present application further provides a chip system, which comprises one or more processors, and is configured to call and run instructions stored in a memory, so that the method provided in the embodiments of the present application is executed. The chip system can be composed of a chip, or can comprise a chip and other discrete devices.

[0166] The chip system can comprise an input circuit or interface configured to send information or data, and an output circuit or interface configured to receive information or data.

[0167] According to the method provided in the embodiments of the present application, the present application further provides a communication system, which comprises the network device and the terminal device described above.

[0168] According to the method provided in the embodiments of the present application, the present application further provides a computer program product, which comprises computer program codes, and when the computer program codes are executed on a computer, the computer is caused to execute each step or process of the network device and the terminal device in any of the method embodiments.

[0169] According to the method provided in the embodiments of the present application, the present application further provides a computer readable storage medium, which stores program codes, and when the program codes are executed on a computer, the computer is caused to execute each step or process of the network device and the terminal device in any of the method embodiments.

[0170] The computer readable storage medium can be the volatile memory or the non-volatile memory described above, or can comprise the volatile memory and the non-volatile memory at the same time.

[0171] In the embodiments of the present application, each term and English abbreviation is an exemplary example given for convenience of description, and should not constitute any limitation on the present application. The present application does not exclude the possibility of defining other terms capable of achieving the same or similar functions in the existing or future protocols.

[0172] In the above embodiments, all or part of the embodiments can be realized by software, hardware, firmware or any combination thereof. When realized by software, all or part of the embodiments can be realized in the form of a computer program product. The computer program product comprises one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated.

[0173] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the division of the above-described device embodiment is only a logical function division, and there can be another division manner for actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, or the among different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0174] It should be understood that, in various embodiments of the present application, the sequence of the processes does not mean the execution sequence, and the execution sequence of the processes should be determined according to the functions and the inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0175] In summary, the above description is only optional embodiments of the technical solutions of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An AI model parameter acquisition method, characterized by, The method is applied to a terminal device and comprises the following steps: receiving n symbol sequences sent by a network device, wherein n is a positive integer; demodulating each received symbol sequence to obtain a first soft information sequence; generating a second soft information sequence with a preset length according to the first soft information sequence; decoding the second soft information sequence according to side information related to updated parameters of a target neural network (NN) layer in a target artificial intelligence (AI) model to obtain a target encoding block, wherein the target encoding block comprises the updated parameters of the target NN layer, and the side information is determined according to original parameters of the target AI model.

2. The method of claim 1, wherein, Before the step of receiving n symbol sequences sent by a network device, the method further comprises the following step: receiving control information sent by the network device, wherein the control information comprises a target AI model identifier, a target NN layer identifier and a total number of blocks, and the total number of blocks is n.

3. The method of claim 2, wherein, After the step of receiving the control information sent by the network device, the method further comprises the following steps: obtaining the original parameters of the target AI model; inputting a binary quantization sequence of the original parameters of the target AI model and parameters of a hidden Markov model into a forward-backward algorithm (FBA) module to obtain output data of the FBA module, wherein the hidden Markov model is used to represent the correlation between the binary quantization sequence of the original parameters of the target AI model and a binary quantization sequence of updated parameters of the target AI model; estimating a soft information sequence of the binary quantization sequence of the updated parameters of the target AI model according to the output data of the FBA module to obtain a target soft information sequence, wherein the target soft information sequence comprises the side information.

4. The method according to any one of claims 1 to 3, characterized in that, Before the step of receiving n symbol sequences sent by a network device, the method further comprises the following steps: obtaining a soft information sequence corresponding to the target NN layer from a target soft information sequence, wherein the target soft information sequence is determined according to the original parameters of the target AI model; dividing the soft information sequence corresponding to the target NN layer into n soft information sequences, wherein the n soft information sequences are the side information. The step of decoding the second soft information sequence according to side information related to updated parameters of a target neural network (NN) layer in a target artificial intelligence (AI) model to obtain a target encoding block comprises the following steps: generating a fourth soft information sequence according to the second soft information sequence and a third soft information sequence, wherein the third soft information sequence is one soft information sequence corresponding to the second soft information sequence in the n soft information sequences; inputting the fourth soft information sequence into a channel decoding module to obtain a target encoding block output by the channel decoding module.

5. The method of claim 4, wherein, The step of generating a fourth soft information sequence according to the second soft information sequence and a third soft information sequence comprises the following steps: adding the soft information of the first Q bits in the second soft information sequence to the soft information of the corresponding bits in the third soft information sequence one by one to obtain the soft information of the first Q bits in the fourth soft information sequence, wherein the first Q bits are puncturing bits; taking the soft information of the remaining bits in the second soft information sequence except the first Q bits as the soft information of the remaining bits in the fourth soft information sequence except the first Q bits.

6. The method of claim 4, wherein, The generating the fourth soft information sequence according to the second soft information sequence and the third soft information sequence comprises: soft information of the first Q bits in the second soft information sequence is taken as soft information of the first Q bits in the fourth soft information sequence, the first Q bits being puncturing bits; soft information of the Q+1th bit to the Kth bit in the second soft information sequence and soft information of corresponding bits in the third soft information sequence are added one by one to obtain soft information of the Q+1th bit to the Kth bit in the fourth soft information sequence, the Q+1th bit to the Kth bit being information bits; soft information of the rest bits in the second soft information sequence except the first K bits is taken as soft information of the rest bits in the fourth soft information sequence except the first K bits.

7. The method of claim 4, wherein, The generating the fourth soft information sequence according to the second soft information sequence and the third soft information sequence comprises: soft information of the first Q bits in the second soft information sequence and soft information of corresponding bits in the third soft information sequence are added one by one to obtain soft information of the first Q bits in the fourth soft information sequence, the first Q bits being puncturing bits; soft information of the Q+1th bit to the Kth bit in the second soft information sequence and soft information of corresponding bits in the third soft information sequence are added one by one to obtain soft information of the Q+1th bit to the Kth bit in the fourth soft information sequence, the Q+1th bit to the Kth bit being information bits; soft information of the rest bits in the second soft information sequence except the first K bits is taken as soft information of the rest bits in the fourth soft information sequence except the first K bits.

8. A communication device, characterized by The communication device comprises at least one processor coupled with a memory, the memory storing a program or instructions, and the processor executes the program or instructions to enable the communication device to perform the method in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program or instructions, and the computer program or instructions, when executed, enable a computer to perform the method in any one of claims 1 to 7.

10. A chip system, characterized by The chip system comprises one or more processors for calling and running instructions stored in a memory, so that the method in any one of claims 1 to 7 is executed.

Images