Quantization parameter determination method and related apparatus

Abstract

The present application discloses a quantization parameter determination method and a related apparatus. The quantization parameter determination method comprises: receiving first data and second data sent by a second device, wherein the first data is obtained by a first quantizer quantizing third data on the basis of quantization parameters configured in the m-th round, and the second data comprises at least one of channel state information and computing power information; determining fourth data on the basis of the first data and the second data, the fourth data being the gradient of an input layer of a first neural network model; and sending the fourth data to the second device, the fourth data being used by the second device to determine at least one first quantization parameter of the first quantizer in the (m+1)-th round. According to embodiments of the present application, quantization parameters of a quantizer can be dynamically adjusted, which is conducive to reducing communication overhead while ensuring the processing accuracy of AI tasks, or maximizes the processing accuracy of AI tasks when the given communication overhead is met.

Description

A method and related apparatus for determining quantization parameters

[0001] This application claims priority to Chinese Patent Application No. 202410445125.9, filed on April 10, 2024, entitled "A Method for Determining Quantitative Parameters and Related Apparatus", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the fields of communication and artificial intelligence technologies, and in particular to a method and apparatus for determining quantitative parameters. Background Technology

[0003] The Telecommunications Research Organisation (TRO) has proposed a new vision for 6th Generation Mobile Communication Technology (6G), pointing out that 6G networks need to possess the basic capabilities to support Artificial Intelligence (AI) tasks, such as distributed AI computing and inference, and end-to-end AI applications. As the prediction accuracy of AI models improves, the networks typically become increasingly deeper, and the memory consumption of model execution poses a significant challenge to device performance. To reduce the memory consumption of model inference and improve inference speed, existing technologies have proposed methods for quantizing models and input data, such as binarization, linear quantization, and logarithmic quantization. A quantization scheme requires determining several key parameters, which are currently typically statically configured. This means that existing quantization schemes are not compatible with constantly changing communication conditions, hindering the reduction of communication overhead. Summary of the Invention

[0004] This application provides a method and related apparatus for determining quantization parameters, which can dynamically adjust the quantization parameters of the quantizer. This can reduce communication overhead while ensuring the processing accuracy of AI tasks, or maximize the processing accuracy of AI tasks while meeting given communication overhead requirements.

[0005] In a first aspect, embodiments of this application provide a method for determining quantization parameters, applied to a first device or a chip in the first device, the method comprising:

[0006] The system receives first data and second data sent by the second device; the first data is obtained by quantizing the third data by the first quantizer through the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information.

[0007] Based on the first and second data, the fourth data is determined; the fourth data is the gradient of the input layer of the first neural network model.

[0008] Send fourth data to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in round (m+1).

[0009] As can be seen from the embodiments of this application, the first device can receive first data quantized by the first quantizer sent by the second device, and at least one of the channel state information and computing power information corresponding to the second device. Based on the first data and at least one of the channel state information and computing power information sent by the second device, the first device can determine the gradient of the input layer of its deployed neural network model and send the gradient to the second device. The second device can then calculate the gradient of the first quantizer based on the gradient, thereby dynamically updating at least one quantization parameter of the first quantizer to adapt the quantization parameters of the first quantizer to the dynamically changing computing power state and channel state, thereby minimizing the channel transmission overhead while ensuring the accuracy of the end-edge collaborative AI task.

[0010] In one possible implementation, the third data is obtained by extracting features from the fifth data using a second neural network model;

[0011] Based on the first and second data, the fourth data is determined, including:

[0012] The AI ​​task processing result is obtained by acquiring the fifth data based on the first data;

[0013] The loss value of the loss function is determined using the AI ​​task processing results of the first, second, and fifth data sets.

[0014] The fourth data is determined based on the loss value of the loss function.

[0015] In this implementation, the first device can calculate the overall loss of AI task processing and communication transmission data volume based on at least one of the channel state information and computing power information sent by the second device, combined with the first data and the AI ​​task processing result obtained based on the first data. Based on this loss, the gradient of the input layer of the neural network model on the first device side can be obtained and then sent to the second device so that the second device can perform gradient calculation and parameter update on the local model.

[0016] In one possible implementation, at least one first quantization parameter includes the number of breakpoints of the first quantizer, the value of the breakpoint, the quantization value at the breakpoint, and part or all of the approximation factor;

[0017] Here, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

[0018] In this implementation, the first quantizer supports dynamically updated quantization parameters, which can be the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and some or all of the approximation factors, making the update configuration relatively flexible.

[0019] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the method further includes, before determining the fourth data based on the first and second data:

[0020] Send at least one second quantization parameter of the first quantizer to the second device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the first device.

[0021] In this implementation, at least one second quantization parameter of the first quantizer can be pre-configured by the first device. The first device can send the configured at least one second quantization parameter to the second device, so the second device does not need to update at least one second quantization parameter during the training process.

[0022] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the method further includes, before determining the fourth data based on the first and second data:

[0023] The device receives at least one second quantization parameter from the first quantizer sent by the second device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the second device.

[0024] In this implementation, at least one second quantization parameter of the first quantizer can be pre-configured by the second device, and the second device can send the configured at least one second quantization parameter to the first device for quantization parameter synchronization.

[0025] In one possible implementation, there are one or more second devices, each second device corresponding to a set of first data, a set of fifth data, and a set of second data; the AI ​​task processing result of the fifth data is the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data;

[0026] Using the AI ​​task processing results of the first, second, and fifth data sets, the loss value of the loss function is determined, including:

[0027] Obtain the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data corresponding to each second device; wherein, the dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device;

[0028] The loss value of the loss function is determined based on the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data, and the KL divergence corresponding to each second device.

[0029] In this implementation, the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data characterizes the difference in data volume between the original input data and the quantized input data of each second device. The smaller the KL divergence, the closer the quantized data is to the reference probability distribution. By minimizing the KL divergence, the overhead of transmitting the quantized data on the k-th link can be reduced. Adding the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data in the loss function ensures the accuracy of the overall AI task. That is, the design of the loss function helps to minimize the transmission overhead of the channel while ensuring the accuracy of the overall AI task. Furthermore, the dimension of the quantized first data is constrained by at least one of channel state information and computing power information. The loss function takes at least one of channel state information and computing power information as input, and can perceive the dimension of the first data under dynamically changing computing power and / or channel states. Therefore, based on the loss value of the loss function, the quantization parameters of the quantizer can be adjusted to adapt the quantizer parameters to dynamically changing computing power and channel conditions.

[0030] In one possible implementation, there are multiple second devices, and the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are different or the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are the same.

[0031] In this implementation, the types of at least one second quantization parameter pre-configured on multiple second device sides can be the same or different, and the types of pre-configured quantization parameters are relatively flexible.

[0032] Secondly, embodiments of this application provide a method for determining quantization parameters, applied to a second device or a chip in the second device, the method comprising:

[0033] First data and second data are sent to the first device; the first data is obtained by the first quantizer quantizing the third data using the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine the fourth data; the fourth data is the gradient of the input layer of the first neural network model;

[0034] Receive the fourth data sent by the first device;

[0035] Based on the fourth data, at least one first quantization parameter of the first quantizer in round (m+1) is determined.

[0036] As can be seen from the embodiments of this application, the second device can send the first data quantized by the first quantizer, and at least one of the channel state information and computing power information corresponding to the second device to the first device. Based on the first data and at least one of the channel state information and computing power information sent by the second device, the first device can determine the gradient of the input layer of its deployed neural network model and send the gradient to the second device. The second device can then calculate the gradient of the first quantizer based on the gradient, thereby dynamically updating at least one quantization parameter of the first quantizer to adapt the quantization parameters of the first quantizer to the dynamically changing computing power state and channel state, thereby minimizing the channel transmission overhead while ensuring the accuracy of the end-edge collaborative AI task.

[0037] In one possible implementation, at least one first quantization parameter includes the number of breakpoints of the first quantizer, the value of the breakpoint, the quantization value at the breakpoint, and part or all of the approximation factor;

[0038] Here, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

[0039] In this implementation, the first quantizer supports dynamically updated quantization parameters, which can be the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and some or all of the approximation factors, making the update configuration relatively flexible.

[0040] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the method further includes, before receiving the fourth data sent by the first device:

[0041] The device receives at least one second quantization parameter from the first quantizer sent by the first device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the first device.

[0042] In this implementation, at least one second quantization parameter of the first quantizer can be pre-configured by the first device. The first device can send the configured at least one second quantization parameter to the second device, so the second device does not need to update at least one second quantization parameter during the training process.

[0043] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the method further includes, before receiving the fourth data sent by the first device:

[0044] Send at least one second quantization parameter of the first quantizer to the first device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the second device.

[0045] In this implementation, at least one second quantization parameter of the first quantizer can be pre-configured by the second device, and the second device can send the configured at least one second quantization parameter to the first device for quantization parameter synchronization.

[0046] In one possible implementation, the third data is obtained by feature extraction of the fifth data through a second neural network model; the fourth data is determined based on the loss value of the loss function, which is determined by the AI ​​task processing results of the first data, the second data, and the fifth data, and the AI ​​task processing result of the fifth data is obtained based on the first data.

[0047] In this implementation, the first device can calculate the overall loss of AI task processing and communication transmission data volume based on at least one of the channel state information and computing power information sent by the second device, combined with the first data and the AI ​​task processing result obtained based on the first data. Based on this loss, the gradient of the input layer of the neural network model on the first device side can be obtained and then sent to the second device so that the second device can perform gradient calculation and parameter update on the local model.

[0048] In one possible implementation, there are multiple second devices, and the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are different or the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are the same.

[0049] In this implementation, the types of at least one second quantization parameter pre-configured on multiple second device sides can be the same or different, and the types of pre-configured quantization parameters are relatively flexible.

[0050] Thirdly, embodiments of this application provide a quantization parameter determination device, applied to a first device or a chip in the first device, the device including a first transceiver unit and a first processing unit;

[0051] The first transceiver unit is used to receive first data and second data sent by the second device; the first data is obtained by the first quantizer quantizing the third data through the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information.

[0052] The first processing unit is used to determine the fourth data based on the first data and the second data; the fourth data is the gradient of the input layer of the first neural network model.

[0053] The first transceiver unit is also used to send fourth data to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in the (m+1)th round.

[0054] In one possible implementation, the third data is obtained by feature extraction of the fifth data using a second neural network model; regarding the determination of the fourth data based on the first and second data, the first processing unit is specifically used for:

[0055] The AI ​​task processing result is obtained by acquiring the fifth data based on the first data;

[0056] The loss value of the loss function is determined using the AI ​​task processing results of the first, second, and fifth data sets.

[0057] The fourth data is determined based on the loss value of the loss function.

[0058] In one possible implementation, at least one first quantization parameter includes the number of breakpoints of the first quantizer, the value of the breakpoint, the quantization value at the breakpoint, and part or all of the approximation factor;

[0059] Here, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

[0060] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the first transceiver unit is further configured to:

[0061] Send at least one second quantization parameter of the first quantizer to the second device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the first device.

[0062] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the first transceiver unit is further configured to:

[0063] The device receives at least one second quantization parameter from the first quantizer sent by the second device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the second device.

[0064] In one possible implementation, there are one or more second devices, each corresponding to a set of first data, a set of fifth data, and a set of second data; the AI ​​task processing result of the fifth data is the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data; in determining the loss value of the loss function using the AI ​​task processing results of the first data, second data, and fifth data, the first processing unit is specifically used for:

[0065] Obtain the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data corresponding to each second device; wherein, the dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device;

[0066] The loss value of the loss function is determined based on the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data, and the KL divergence corresponding to each second device.

[0067] In one possible implementation, there are multiple second devices, and the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are different or the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are the same.

[0068] It should be understood that since the method embodiments and the device embodiments are different presentations of the same technical concept, the content of the first aspect of the embodiments of this application should be adapted to the third aspect of the embodiments of this application simultaneously, and can achieve the same or similar beneficial effects, which will not be repeated here.

[0069] Fourthly, embodiments of this application provide a quantization parameter determination device, applied to a second device or a chip in the second device, the device including a second transceiver unit and a second processing unit;

[0070] The second transceiver unit is used to send first data and second data to the first device; the first data is obtained by the first quantizer quantizing the third data through the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine the fourth data; the fourth data is the gradient of the input layer of the first neural network model; and the unit receives the fourth data sent by the first device.

[0071] The second processing unit is used to determine at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0072] In one possible implementation, at least one first quantization parameter includes the number of breakpoints of the first quantizer, the value of the breakpoint, the quantization value at the breakpoint, and part or all of the approximation factor;

[0073] Here, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

[0074] In one possible implementation, when at least one of the first quantization parameters is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the second transceiver unit is further configured to:

[0075] The device receives at least one second quantization parameter from the first quantizer sent by the first device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the first device.

[0076] In one possible implementation, when at least one of the first quantization parameters is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the second transceiver unit is further configured to:

[0077] Send at least one second quantization parameter of the first quantizer to the first device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the second device.

[0078] In one possible implementation, the third data is obtained by feature extraction of the fifth data through a second neural network model; the fourth data is determined based on the loss value of the loss function, which is determined by the AI ​​task processing results of the first data, the second data, and the fifth data, and the AI ​​task processing result of the fifth data is obtained based on the first data.

[0079] In one possible implementation, there are multiple second devices, and the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are different or the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are the same.

[0080] It should be understood that since the method embodiments and the device embodiments are different presentations of the same technical concept, the content of the second aspect of the embodiments of this application should be adapted to the fourth aspect of the embodiments of this application simultaneously, and can achieve the same or similar beneficial effects, which will not be repeated here.

[0081] Fifthly, embodiments of this application provide a communication device, including a processor, a memory, a communication interface, and one or more programs, the one or more programs being stored in the memory and configured to, when executed by the processor, cooperate with the communication interface to implement the method in any of the embodiments of the first or second aspect described above.

[0082] In a sixth aspect, embodiments of this application provide a chip, including: a processor, configured to call and run a computer program from a memory, causing a device on which the chip is installed to perform the method as described in any of the embodiments of the first or second aspect above.

[0083] In a seventh aspect, embodiments of this application provide a computer-readable storage medium storing a computer program for execution by a device, wherein the computer program, when executed, implements the method as described in any of the embodiments of the first or second aspect above.

[0084] Eighthly, embodiments of this application provide a computer program product that, when run by a communication device, causes the communication device to perform the method as described in any of the embodiments of the first or second aspect above. Attached Figure Description

[0085] To more clearly illustrate the technical solutions in the embodiments of this application or the background art, the accompanying drawings used in the embodiments of this application or the background art will be described below.

[0086] Figure 1 is a schematic diagram of an edge-end collaborative deployment of a neural network model;

[0087] Figure 2 is a schematic diagram of a system architecture provided in an embodiment of this application;

[0088] Figure 3 is a flowchart illustrating a method for determining quantification parameters provided in an embodiment of this application;

[0089] Figure 3A is a schematic diagram of a multi-device identical quantitative configuration provided in an embodiment of this application;

[0090] Figure 3B is a schematic diagram of different quantitative configurations of multiple devices provided in an embodiment of this application;

[0091] Figure 4 is a flowchart illustrating another method for determining quantification parameters provided in an embodiment of this application;

[0092] Figure 5 is a schematic diagram of a pre-configured quantization parameter of a base station according to an embodiment of this application;

[0093] Figure 6 is a flowchart illustrating another method for determining quantification parameters provided in an embodiment of this application;

[0094] Figure 7 is a schematic diagram of a terminal pre-configured quantization parameter provided in an embodiment of this application;

[0095] Figure 8A is a schematic diagram showing the relationship between the inference error rate and wireless channel data transmission latency of an end-edge collaborative AI model provided in an embodiment of this application;

[0096] Figure 8B is a schematic diagram showing the relationship between the inference error rate and the number of end-side devices participating in the collaboration of an end-edge collaborative AI model provided in an embodiment of this application.

[0097] Figure 8C is a schematic diagram showing the relationship between the inference error rate and peak signal-to-noise ratio of an end-edge collaborative AI model provided in an embodiment of this application;

[0098] Figure 9A is a schematic diagram showing the relationship between the inference error rate and wireless channel data transmission latency of another end-edge collaborative AI model provided in an embodiment of this application.

[0099] Figure 9B is a schematic diagram showing the relationship between the inference error rate and the number of end-side devices participating in the collaboration of another end-edge collaborative AI model provided in the embodiments of this application;

[0100] Figure 9C is a schematic diagram showing the relationship between the inference error rate and peak signal-to-noise ratio of another end-edge collaborative AI model provided in the embodiments of this application;

[0101] Figure 10 is a schematic diagram of a quantization parameter determination device provided in an embodiment of this application;

[0102] Figure 11 is a schematic diagram of another quantization parameter determination device provided in an embodiment of this application;

[0103] Figure 12 is a schematic diagram of the structure of a communication device provided in an embodiment of this application. Detailed Implementation

[0104] The terms "first," "second," "third," and "fourth," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0105] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0106] The terms “component,” “module,” “system,” etc., used in this specification are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, an application running on a terminal device and the terminal device can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0107] First, a brief introduction to the relevant terms and technical background used in this application will be provided to facilitate understanding by those skilled in the art.

[0108] (1) Channel-State Information Reference Signal (CSI-RS);

[0109] (2) Distributed Information Bottleneck (DIB);

[0110] (3) Floating Point: FP;

[0111] (4) Information Bottleneck (IB);

[0112] (5) Integer data types: Integer, INT;

[0113] (6) Long Term Evolution (LTE);

[0114] (7) Medium Access Control (MAC);

[0115] (8) Machine Learning: ML;

[0116] (9) Non-access stratum: NAS;

[0117] (10) Packet Data Convergence Control (PDCP);

[0118] (11) Physical Layer: PHY;

[0119] (12) Peak signal-to-noise ratio (PSNR);

[0120] (13) Radio Access Network (RAN);

[0121] (14) Radio Link Control (RLC);

[0122] (15) Radio Resource Control (RRC);

[0123] (16) Reference Signal Received Power (RSRP);

[0124] (17) Reference Signal Received Quality (RSRQ);

[0125] (18) Service Data Adaptation Protocol: SDAP;

[0126] (19) Signal-to-Interference-to-Noise Ratio (SINR);

[0127] (20) Signal to Noise Ratio (SNR)

[0128] (21) Channel sounding reference signal: SRS;

[0129] (22) Transmission Reception Point (TRP);

[0130] (23) User Equipment (UE);

[0131] (24) New Radio (NR);

[0132] (25) The Fifth Generation Mobile Communications, 5G;

[0133] (26) Model quantization: This generally refers to converting floating-point operations of a neural network into fixed-point operations. Model quantization typically involves low-precision models, regular-precision models, and mixed-precision models. The weight values ​​of low-precision models are in FP16 (half-precision floating-point) or INT8 (8-bit fixed-point integer), but low precision usually refers to INT8. The weight values ​​of regular-precision models are in FP32 (32-bit floating-point or single-precision floating-point). Mixed-precision models use both FP32 and FP16 weight value formats. FP16 is used to save memory, but some parameters or operators must use FP32 format to maintain accuracy.

[0134] (27) Radio Access Network Protocol Layer in NR:

[0135] L1 layer (Layer 1): PHY, mainly including modulation scheme, waveform, channel code, multi-antenna technology, frame structure, duplex scheme, control and reference signals.

[0136] L2 layer (layer 2): Wireless network layer. NR's L2 layer includes MAC layer, RLC layer, PDCP layer, and SDAP layer.

[0137] Layer 3 (L3): Includes NAS layer and RRC layer. Layer 3 belongs to the control plane protocol.

[0138] For AI models in future 6G networks, relevant technical reports indicate that they can be divided into multiple parts based on specific tasks and environments. The aim is to offload computationally intensive and energy-intensive parts to the network side, while leaving data privacy-sensitive and latency-sensitive parts on the device side. The device executes specific neural network layers, transmitting intermediate data to the network. The network executes the remaining layers and feeds back the inference results to the device, as shown in Figure 1. However, for the device, the number of neural network layers is constantly increasing as accuracy requirements rise, leading to significant consumption of device memory, battery, and bandwidth with each model run. To address these issues, two main solutions exist: one is to design a more efficient network architecture to achieve acceptable accuracy with a relatively small model size; the other is to reduce network size through compression and encoding, with model quantization being a widely used compression method. Model quantization reduces model storage space, memory usage, and device power consumption, while also accelerating inference speed and improving hardware adaptability.

[0139] In existing schemes, binary quantization can simultaneously compute a large number of numbers using simple bitwise operations, with theoretical performance improvements of 5 to 128 times per bit, and it only introduces one additional quantization operation, which can benefit from the acceleration of Single Instruction Multiple Data (SIMD). Linear quantization can be divided into symmetric and asymmetric quantization, both supporting 8-bit computation, with efficiency improvements ranging from 1 to 16 times. The additional quantization / dequantization computations introduced by linear quantization are standard vector operations, so they can also be accelerated using SIMD, resulting in minimal additional computation time. Logarithmic quantization is a special type of quantization method that multiplies two powers with the same base, equivalent to adding their exponents, reducing computational intensity, and transforming addition into indexed computation. Currently, in this field, a quantization scheme is typically determined by three key parameters: the number of breakpoints, the amplitude, and the breakpoint values. A breakpoint refers to the point that divides the quantizer or quantization function into multiple value intervals, while the amplitude refers to the quantized value at the breakpoint (i.e., the value being quantized or mapped). Existing quantization schemes generally employ static configuration for the three quantization parameters mentioned above. However, the computing power and channel conditions of terminal devices are dynamically changing, making statically configured quantization parameters incompatible with these dynamic factors. Future networks will require reasonable quantization of wireless data to flexibly adjust the number of data transmission bits, reducing communication overhead, computing memory, and power consumption. This application aims to address the problem of designing a quantization scheme that adapts to real-time changes in computing power and wireless channel conditions without significantly degrading the accuracy of AI tasks.

[0140] Please refer to Figure 2, which is a schematic diagram of a system architecture provided in an embodiment of this application. As shown in Figure 2, it includes K terminal devices and network devices. The K terminal devices and network devices constitute a multi-link end-side collaborative task processing architecture, where K is greater than or equal to 1 (when K equals 1, the multi-link end-side collaborative task processing architecture degenerates into a single-link end-side collaborative task processing architecture). End-side neural network models are deployed on the K terminal device side, and edge-side neural network models are deployed on the network device side. The original data X1,…,X with target variable Y (e.g., label) are... K After feature extraction by K edge neural network models, the intermediate features Z1,…,Z are output. K For example, it can take the form of tokens or embeddings. Data Z1,…,Z K After quantization by the quantizer, the data is obtained. Data transmitted wirelessly to network devices may become corrupted due to noise or interference. And vector With vector Z k They have the same dimensionality (quantization mechanisms and wireless links do not change the data size). The network device aggregates K sets of data and performs the remaining AI tasks (e.g., classification) through a side-side neural network model to obtain the final task result. Data stream It can be viewed as a Markov chain. The edge neural network model and quantizer perform feature extraction and compression processing on the original data, respectively. The quantized data is sent into the wireless channel. The edge neural network model performs AI tasks, such as inference, based on the data received from the wireless channel to obtain the final task result.

[0141] In research on multi-link end-edge collaborative task processing without the introduction of quantization mechanisms, the researchers in this application used DIB theory for mathematical modeling and designed a mechanism for the generation, updating, and deployment of multi-link end-edge collaborative models. The theoretical formula can be expressed as:

[0142] in, This indicates the data received by the network device. The set, This indicates that the base station has received data. The mutual information between the data and the actual task result Y is considered. The larger the value of this mutual information, the more data is transmitted through the wireless channel. The more information included in the summary that leads to the correct AI task result Y, the higher the accuracy of the AI ​​task result can be understood. X represents the raw input data of the k-th device. kData with the k-th wireless link The smaller the mutual information between them, the less data is transmitted in the k-th wireless link, meaning the lower the wireless communication overhead of the k-th link. This represents the sum of data transmitted over K wireless links. The smaller the value, the lower the total wireless communication overhead of the K links. The configurable parameter β controls the ratio between the two factors, balancing the accuracy of the AI ​​task results with the total wireless communication overhead. Therefore, minimizing the above equation... This means: while minimizing the wireless communication overhead of the K links, we should maximize the accuracy of the AI ​​task results (or minimize the wireless communication overhead of the K links while ensuring the accuracy of the AI ​​task results).

[0143] The above modeling can serve as the theoretical basis for the algorithm of generating multi-link end-edge collaborative models. The input of this algorithm is: M pairs of data and labels. Where M is the size of the batch data. The dimension of the data tokens or embeddings transmitted in the wireless link is represented by N. i that is, vector or Z i The algorithm outputs K sets of edge neural network models (e.g., φ1, φ2, ..., φ) with dimensions (i = 1, 2, ..., K). K The algorithm outputs K+1 AI models, including one side neural network model (e.g., ψ) and one side neural network model. Therefore, the loss function of this algorithm can be expressed as:

[0144] The terminal equipment mentioned in this application includes, but is not limited to, UE, user unit, user station, mobile station, mobile station, remote station, remote terminal equipment, mobile terminal equipment, user terminal equipment, wireless electronic equipment, user agent, user device, cellular phone, cordless phone, session initiation protocol (SIP) phone, wireless local loop (WLL) station, personal digital assistant (PDA), handheld device with wireless communication function, computing device, processing device connected to wireless modem, vehicle communication module or other embedded communication module, wireless terminal in industrial control, wireless terminal in autonomous driving, wireless terminal in remote medical care, wireless terminal in smart grid, wearable device, terminal equipment in Internet of Things, home appliance, virtual reality device, terminal equipment in future 5G network, or terminal equipment in future evolved public land mobile network (PLMN), etc.

[0145] The network equipment mentioned in this application includes various forms of base stations, such as macro base stations, micro base stations (also known as small stations), relay stations, and access points. In systems employing different radio access technologies, the names of equipment with base station functions may differ; for example, in 5GS, it is called a gNB; in LTE systems, it is called an evolved NodeB (eNB or eNodeB); and in third-generation (3G) systems, it is called a Node B. In some deployments, network equipment can be a central unit (CU) or a distributed unit (DU). In other deployments, the CU can be divided into a CU-control plane (CP) and a CU-user plane (UP). In still other deployments, network equipment can also be a radio unit (RU), an open radio access network (ORAN) architecture, etc. This application does not limit the deployment method of the network equipment. For example, when the network device is an ORAN architecture, the network device shown in the embodiments of this application can be an access network device in ORAN, or a module in the access network device, etc. In the ORAN architecture, CU can also be called open (O)-CU, DU can also be called O-DU, CU-CP can also be called O-CU-CP, CU-UP can also be called O-CU-UP, and RU can also be called O-RU. In some deployments, the network device can also be a core network device, and so on.

[0146] Based on the DIB theory modeling, the technical solution provided in this application will be described in detail below with reference to specific implementation methods.

[0147] Please refer to Figure 3, which is a flowchart illustrating a method for determining quantization parameters provided in an embodiment of this application. As shown in Figure 3, the method includes steps 301-304:

[0148] 301: The second device sends the first data and the second data to the first device.

[0149] Accordingly, the first device receives the first data and the second data sent by the second device.

[0150] In this embodiment of the application, the second device may be a terminal device or a chip in a terminal device; the first device may be a terminal device or a network device. For example, in a conventional communication scenario, the first device may be a base station, and in a sidelink communication scenario, the first device may also be a terminal device.

[0151] The first data is obtained by quantizing the third data using the quantization parameters configured in the m-th round by the first quantizer. The third data is obtained by extracting features from the fifth data using the second neural network model (e.g., an edge-side neural network model). The second neural network model and the first quantizer are deployed on the second device. The fifth data is the input data of the second neural network model, such as X. K In the architecture shown in Figure 2, the neural network model in the Kth terminal device corresponds to X. K The third data Z is obtained by feature extraction. K Z K The first data is obtained by quantization by the quantizer on the Kth terminal device side. It should be understood that, due to the presence of channel noise, etc., the first data received by the first device is represented as follows:

[0152] For example, the first quantizer can be represented as:

[0153] Let a regular vector be denoted here. z k,i ,i=1,…,d represents vector z k Each dimension of z takes a value, i.e. k,i This represents the third data Z input to the first quantizer on the Kth terminal device side. K The value in the i-th dimension, Let T be the set of real numbers. Here, T represents the number of quantization parameter breakpoints; a k,t b represents the amplitude at the t-th breakpoint, i.e., the quantization value at the t-th breakpoint; k,t γ represents the value at the t-th breakpoint. The first quantizer is obtained by relaxing the sign function sgn in the second quantizer into a continuous function arctan, and γ is the approximation factor, characterizing the degree of approximation between the continuous function arctan and the sign function sgn.

[0154] For example, the second quantizer can be represented as:

[0155] Both the second quantizer and the first quantizer can convert z... kEach dimension of the quantizer is quantized into one of T+1 quantized values. To ensure that the quantization parameters of the quantizer can be dynamically adjusted according to the edge computing power and channel conditions without affecting the accuracy of the AI ​​task results, it is first necessary to ensure that the gradient of the loss function after quantization is calculated, i.e., without affecting the backpropagation of gradients during model training. However, the second quantizer is built based on the sign function sgn. The discontinuity of the sign function sgn will prevent gradient calculation. Therefore, it is necessary to perform a relaxation transformation on the sign function sgn based on mathematical theory to relax it into a continuous function.

[0156] The second data includes at least one of channel state information and computing power information. The computing power information of the second device may be the number of floating-point operations per second, and the channel state information may be the downlink channel measurement result or the uplink reference signal.

[0157] 302: The first device determines the fourth data based on the first and second data.

[0158] In this embodiment of the application, the fourth data is the gradient of the input layer of the first neural network model (e.g., it can be a side-side neural network model), and the first neural network model is deployed on the first device.

[0159] For example, the first device inputs the first data into the first neural network model to perform the remaining AI tasks (such as inference or prediction) and obtains the AI ​​task processing results (such as inference results or prediction values) of the original input data (i.e., the fifth data). The first device uses the AI ​​task processing results of the first data, the second data, and the fifth data to determine the loss value of the loss function, and calculates the gradient of each layer of the first neural network model based on the loss value of the loss function, thereby obtaining the fourth data.

[0160] In this implementation, the first device can calculate the overall loss of AI task processing and communication transmission data volume based on at least one of the channel state information and computing power information sent by the second device, combined with the first data and the AI ​​task processing result obtained based on the first data. Based on this loss, the gradient of the input layer of the neural network model on the first device side can be obtained and then sent to the second device so that the second device can perform gradient calculation and parameter update on the local model.

[0161] Here, the loss function can be called the first loss function. The first loss function is constructed based on the first quantizer and the second loss function. The second loss function is constructed based on the DIB theory, that is, the second loss function is... Loss function. Specifically, substitute the expression for the first quantizer into... The loss function is derived based on mathematical theories such as variational methods and the Log-sum inequality, resulting in the quantized loss function (i.e., the first loss function). The first loss function can then be expressed as:

[0162] Among them, C QML-IB Let x represent the first loss function. 1:K Let represent the set of K original input data from the second devices in each dimension, and y represent the actual value (e.g., label) of the original input data in each dimension. This represents the probability that the original input data is expected to be processed (e.g., predicted) into the true value y in each dimension. This indicates that each dimension of the expected original data is quantized. The probability, The first neural network model ψ will represent Processing (e.g., predicting) the probability of y being the true value. The processing loss of the first neural network model ψ is... This represents the KL divergence between the probability distribution of the first quantized data and the reference probability distribution of the first data after quantization. Let β represent the reference probability distribution of the first data corresponding to the k-th second device. k This represents the trade-off between the accuracy of the AI ​​task on the k-th link and the reduction of transmission overhead.

[0163] For example, the second device can be one or more (such as the K terminal devices in Figure 2), and each second device corresponds to a set of first data, a set of fifth data, and a set of second data. Based on the expression of the first loss function, it can be seen that the probability that the AI ​​task processing result of the fifth data is that the first data corresponding to one or more second devices is processed as the label (or true value) of the fifth data (or the probability that the AI ​​task processing result of the fifth data is that the first data corresponding to one or more second devices is processed correctly) is the first device. The first device can then obtain the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of that first data. Based on the probability that the first data corresponding to one or more second devices is processed as the label of the fifth data, and the KL divergence corresponding to each second device, the loss value of the loss function is determined. Specifically, the KL divergence corresponding to each second device is summed, and the summation value is added to the probability that the first data corresponding to one or more second devices is processed as the label of the fifth data, thus obtaining the loss value of the first loss function. In this context, the dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device. That is, at least one of the channel state information and computing power information sent by the second device determines the dimension of the first data sent by each second device. In other words, the quantizer learns to output the first data of what dimension under what channel state and / or computing power state, that is, to what extent the intermediate features are quantized.

[0164] In this implementation, the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data characterizes the difference in data volume between the original input data and the quantized input data of each second device. The smaller the KL divergence, the closer the quantized data is to the reference probability distribution. By minimizing the KL divergence, the overhead of transmitting the quantized data on the k-th link can be reduced. Adding the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data in the loss function ensures the accuracy of the overall AI task. That is, the design of the loss function helps to minimize the transmission overhead of the channel while ensuring the accuracy of the overall AI task. Furthermore, the dimension of the quantized first data is constrained by at least one of channel state information and computing power information. The loss function takes at least one of channel state information and computing power information as input, and can perceive the dimension of the first data under dynamically changing computing power and / or channel states. Therefore, based on the loss value of the loss function, the quantization parameters of the quantizer can be adjusted to adapt the quantizer parameters to dynamically changing computing power and channel conditions.

[0165] 303: The first device sends the fourth data to the second device.

[0166] Accordingly, the second device receives the fourth data.

[0167] 304: The second device determines at least one first quantization parameter of the first quantizer in round (m+1) based on the fourth data.

[0168] In this embodiment of the application, the second device performs gradient calculation of the local model (including the first quantizer and the second neural network model) based on the received fourth data, updates at least one first quantization parameter of the first quantizer in the (m+1)th round based on the calculated gradient, and updates the parameters of each layer of the second neural network model in the (m+1)th round.

[0169] Wherein, at least one of the first quantization parameters includes the number of breakpoints of the first quantizer. The value of the breakpoint Quantization value at the breakpoint and approximation factor In the first quantizer, the quantization parameters that can be dynamically updated can be part or all of the parameters, j = 1, ..., K. If only part of the quantization parameters in a given set are dynamically updated, then the other part of the quantization parameters can be pre-configured.

[0170] In this implementation, the first quantizer supports dynamically updated quantization parameters, which can be the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and some or all of the approximation factors, making the update configuration relatively flexible.

[0171] It should be noted that steps 301-304 describe the update mechanism of the quantization parameters of the first quantizer in a single round of training. After multiple rounds of data interaction between the first and second devices, when the first device determines that the loss value of the first loss function has converged, the first device can indicate the optimal quantization parameters to the second device. That is, the final output of this embodiment is K second neural network models, K first quantizers, and one first neural network model. The convergence of the loss value of the first loss function can be achieved by the number of training rounds reaching a preset number of rounds or by the loss value being less than or equal to a preset threshold.

[0172] For example, when the loss value of the first loss function converges, the first device can indicate a set of quantization parameters to multiple second devices. The values ​​of the set of quantization parameters configured for the multiple second devices can be the same (i.e., the K links use the same quantization configuration), or the values ​​of the set of quantization parameters configured for the multiple second devices can be different (i.e., the K links use different quantization configurations). For example, the first device can indicate the final quantization parameters of each second device to the corresponding second device through a first signaling, which can be L1 layer signaling (e.g., downlink control information signaling) or L2 layer signaling (e.g., MAC layer control unit signaling).

[0173] To better understand the scheme where K links use the same quantization configuration, the parameter update mechanism of the quantizer is briefly explained below using the end-to-edge collaborative interaction between K UEs and the base station as an example. As shown in Figure 3A, it may include the following steps:

[0174] 1: The UE reports quantized data, channel state information, and computing power information;

[0175] 2: The base station calculates the loss value of the loss function based on the quantized data, channel state information, and computing power information, and determines the gradient of the input layer of the first neural network model based on the loss value;

[0176] 3: The base station sends the input layer gradient to the UE;

[0177] 4: The UE updates the parameters of the second neural network model and the first quantizer based on this gradient.

[0178] 5: The UE and the base station repeat the update process of steps 1-4 multiple times until the value of the loss function is minimized;

[0179] 6: The base station sends the optimal quantization scheme to the UE {a * ,b * ,T * ,γ *};

[0180] 7: The UE completes the configuration of the local first quantizer.

[0181] To better understand the scheme of using different quantization configurations for K links, the following is a brief explanation of the quantizer parameter update mechanism, taking the end-to-edge collaborative interaction between K UEs and the base station as an example. As shown in Figure 3B, it may include the following steps:

[0182] 1: The UE reports quantized data, channel state information, and computing power information;

[0183] 2: The base station calculates the loss value of the loss function based on the quantized data, channel state information, and computing power information, and determines the gradient of the input layer of the first neural network model based on the loss value;

[0184] 3: The base station sends the input layer gradient to the UE;

[0185] 4: The UE updates the parameters of the second neural network model and the first quantizer based on this gradient.

[0186] 5: The UE and the base station repeat the update process of steps 1-4 multiple times until the value of the loss function is minimized;

[0187] 6-1: The base station sends the optimal quantization scheme to UE 1

[0188] …

[0189] 6-K: The base station sends the optimal quantization scheme to UE K.

[0190] 7: The UE completes the configuration of the local first quantizer.

[0191] As can be seen from the embodiments of this application, the second device can send the first data quantized by the first quantizer, and at least one of the channel state information and computing power information corresponding to the second device to the first device. Based on the first data and at least one of the channel state information and computing power information sent by the second device, the first device can determine the gradient of the input layer of its deployed neural network model and send the gradient to the second device. The second device can then calculate the gradient of the first quantizer based on the gradient, thereby dynamically updating at least one quantization parameter of the first quantizer to adapt the quantization parameters of the first quantizer to the dynamically changing computing power state and channel state, thereby minimizing the channel transmission overhead while ensuring the accuracy of the end-edge collaborative AI task.

[0192] Please refer to Figure 4, which is a flowchart illustrating another method for determining quantization parameters provided in an embodiment of this application. As shown in Figure 4, the method includes steps 401-405:

[0193] 401: The second device sends first data and second data to the first device.

[0194] 402: The first device sends at least one second quantization parameter of the first quantizer to the second device.

[0195] In this embodiment of the application, at least one second quantization parameter is: In addition to at least one first quantization parameter, another set of quantization parameters, and at least one second quantization parameter, are pre-configured by the first device. For example, the base station pre-configures the terminal device. and Base station will and The data is sent to the terminal device, which then configures the local first quantizer. and

[0196] It should be noted that there is no time restriction between steps 401 and 402. The first device may send at least one second quantization parameter to the second device before step 401.

[0197] For example, the first device can send at least one second quantization parameter of the first quantizer to the second device via the second signaling, which can be L3 layer signaling, such as RRC signaling.

[0198] 403: The first device determines the fourth data based on the first and second data.

[0199] 404: The first device sends fourth data to the second device.

[0200] 405: The second device determines at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0201] For example, if the first device has pre-configured at least one second quantization parameter for the second device, the second device only needs to update at least one parameter other than the second quantization parameter in each training round. For instance, if the first device has pre-configured at least one second quantization parameter, the second device can update at least one parameter other than the first quantization parameter in each training round. and In this case, the second device only needs to be updated. That's all.

[0202] The specific implementation methods of steps 401 and 403-405 have been described in the embodiment shown in Figure 3, and can achieve the same or similar beneficial effects.

[0203] It can be seen that at least one second quantization parameter of the first quantizer can be pre-configured by the first device. The first device can send the configured at least one second quantization parameter to the second device, so the second device does not need to update at least one second quantization parameter during the training process.

[0204] To better understand the scheme in the embodiment shown in Figure 4, the parameter update mechanism of the quantizer will be briefly explained below using the end-to-edge collaborative interaction between K UEs and the base station as an example. As shown in Figure 5, it may include the following steps:

[0205] 1: The UE reports quantized data, channel state information, and computing power information;

[0206] 2: Base station distribution and

[0207] 3: The base station calculates the loss value of the loss function based on the quantized data, channel state information, and computing power information, and determines the gradient of the input layer of the first neural network model based on the loss value;

[0208] 4: The base station sends the input layer gradient to the UE;

[0209] 5: The UE updates the first quantizer based on this gradient.

[0210] 6: The UE and the base station repeat the update process of steps 1-5 multiple times until the value of the loss function is minimized;

[0211] 7: The base station sends the optimal quantization scheme to the UE.

[0212] 8: The UE completes the configuration of the local first quantizer.

[0213] It should be noted that the types of at least one second quantization parameter pre-configured by the first quantizers on multiple second device sides may be different, or the types of at least one second quantization parameter pre-configured by the first quantizers on multiple second device sides may be the same. For example, the base station can pre-configure at least one second quantization parameter for K UEs. and The base station can also be pre-configured for UE1. and Pre-configuration for UE2 and etc.

[0214] In this implementation, the types of at least one second quantization parameter pre-configured on multiple second device sides can be the same or different, and the types of pre-configured quantization parameters are relatively flexible.

[0215] Please refer to Figure 6, which is a flowchart illustrating another method for determining quantization parameters provided in an embodiment of this application. As shown in Figure 6, the method includes steps 601-605:

[0216] 601: The second device sends first data and second data to the first device.

[0217] 602: The second device sends at least one second quantization parameter of the first quantizer to the first device.

[0218] In this embodiment of the application, at least one second quantization parameter is: In addition to at least one first quantization parameter, another portion of the quantization parameters, at least one second quantization parameter, is pre-configured by the second device. For example: the terminal device is pre-configured. and Will and The data is reported to the base station for end-to-side synchronization. Therefore, the quantization scheme ultimately configured by the base station for the terminal device may not include... and

[0219] It should be noted that there is no time restriction between steps 601 and 602, and the second device may send at least one second quantization parameter to the first device before step 601.

[0220] 603: The first device determines the fourth data based on the first and second data.

[0221] 604: The first device sends the fourth data to the second device.

[0222] 605: The second device determines at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0223] The specific implementation methods of steps 601 and 603-605 have been described in the embodiment shown in Figure 3, and can achieve the same or similar beneficial effects.

[0224] It can be seen that at least one second quantization parameter of the first quantizer can be pre-configured by the second device, and the second device can send the configured at least one second quantization parameter to the first device for quantization parameter synchronization.

[0225] To better understand the scheme in the embodiment shown in Figure 6, the parameter update mechanism of the quantizer will be briefly explained below using the end-to-edge collaborative interaction between K UEs and the base station as an example. As shown in Figure 7, it may include the following steps:

[0226] 1: The UE reports quantized data, channel state information, and computing power information;

[0227] 2: UE reports pre-configured and

[0228] 3: The base station calculates the loss value of the loss function based on the quantized data, channel state information, and computing power information, and determines the gradient of the input layer of the first neural network model based on the loss value;

[0229] 4: The base station sends the input layer gradient to the UE;

[0230] 5: The UE updates the first quantizer based on this gradient.

[0231] 6: The UE and the base station repeat the update process of steps 1-5 multiple times until the value of the loss function is minimized;

[0232] 7: The base station sends the optimal quantization scheme to the UE.

[0233] 8: The UE completes the configuration of the local first quantizer.

[0234] It should be noted that the types of at least one second quantization parameter pre-configured by the first quantizers on multiple second device sides are different, or the types of at least one second quantization parameter pre-configured by the first quantizers on multiple second device sides are the same.

[0235] To measure the impact of the quantization mechanism proposed in this application's embodiments on the performance of the AI ​​model, this application analyzes the loss function before and after quantization. The approximate KL divergence after quantization is: Its divergence with the actual KL divergence D KL The theoretical upper bound of the difference, after derivation, is:

[0236] Where C and δ are constants related to the wireless channel, and d is an intermediate quantity. The dimension of quantization. Based on this derivation, it can be concluded that the difference between the loss function before and after quantization is bounded, that is, the impact of the quantization scheme design on the performance of the edge AI model is bounded and controllable.

[0237] Regarding the quantization mechanism provided in this application, relevant personnel conducted tests on the MNIST (Mixed National Institute of Standards and Technology database, a computer vision dataset) and CIFAR (Canadian Institute for Advanced Research, an image dataset widely used in the field of computer vision). Figures 8A-8C show the test results on the MNIST dataset. Among them, line 1 represents the test results of the end-edge collaborative model based on the information bottleneck theory and the dynamic quantization mechanism designed in this scheme; line 2 represents the test results of the end-edge collaborative model based on the information bottleneck theory (without the dynamic quantization mechanism); line 3 represents the test results of the end-edge collaborative model based on the information bottleneck theory and the linear quantization scheme in the prior art; and line 4 represents the test results of the end-edge collaborative model based on the information bottleneck theory (without the linear quantization scheme) in the prior art. Figure 8A shows the relationship between the inference error rate of the end-edge collaborative AI model and the wireless channel data transmission delay; Figure 8B shows the relationship between the inference error rate of the end-edge collaborative AI model and the number of end-side devices participating in the collaboration; and Figure 8C shows the relationship between the inference error rate of the end-edge collaborative AI model and the peak signal-to-noise ratio (PSNR). Figures 9A-9C show the test results on the CIFAR dataset. Line 1 represents the test results of the end-edge collaborative model based on the information bottleneck theory with a dynamic quantization mechanism; line 2 represents the test results of the end-edge collaborative model based on the information bottleneck theory (without a dynamic quantization mechanism); line 3 represents the test results of the existing end-edge collaborative model based on the information bottleneck theory with a linear quantization scheme; and line 4 represents the test results of the existing end-edge collaborative model based on the information bottleneck theory (without a linear quantization scheme). Figure 9A shows the relationship between the inference error rate of the end-edge collaborative AI model and the wireless channel data transmission latency; Figure 9B shows the relationship between the inference error rate of the end-edge collaborative AI model and the number of end-side devices participating in the collaboration; and Figure 9C shows the relationship between the inference error rate of the end-edge collaborative AI model and PSNR. Simulation results show that the addition of quantization leads to a performance loss in the AI ​​model. However, the end-edge collaborative model generation scheme based on DIB theory and dynamic quantization mechanism in this application outperforms the existing optimal scheme without quantization. This indicates that the DIB+dynamic quantizer design can still guarantee the high-quality completion of the AI ​​task, and the design of dynamically updating the quantizer does not affect the AI ​​task performance of the original end-edge model designed based on DIB theory.Meanwhile, this dynamic quantization mechanism adapts to the dynamically changing end-edge computing power and wireless channel status, and is more suitable for future wireless networks than existing linear and nonlinear fixed quantizers.

[0238] The methods of the embodiments of this application have been described above, and the apparatus of the embodiments of this application is provided below.

[0239] Please refer to Figure 10, which is a schematic diagram of a quantization parameter determination device provided in an embodiment of this application. As shown in Figure 10, the device includes a first transceiver unit 1001 and a first processing unit 1002. Wherein:

[0240] The first transceiver unit 1001 is used to receive first data and second data sent by the second device; the first data is obtained by the first quantizer quantizing the third data through the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information.

[0241] The first processing unit 1002 is used to determine the fourth data based on the first data and the second data; the fourth data is the gradient of the input layer of the first neural network model.

[0242] The first transceiver unit 1001 is also used to send fourth data to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in the (m+1)th round.

[0243] As can be seen, in the device shown in Figure 10, the device can receive first data quantized by the first quantizer sent by the second device, and at least one of the channel state information and computing power information corresponding to the second device. Based on the first data and at least one of the channel state information and computing power information sent by the second device, the device can determine the gradient of the input layer of its deployed neural network model and send the gradient to the second device. The second device can then calculate the gradient of the first quantizer based on the gradient, thereby dynamically updating at least one quantization parameter of the first quantizer to adapt the quantization parameters of the first quantizer to the dynamically changing computing power and channel states. This minimizes the channel transmission overhead while ensuring the accuracy of the end-edge collaborative AI task.

[0244] In one possible implementation, the third data is obtained by feature extraction of the fifth data using a second neural network model; regarding the determination of the fourth data based on the first and second data, the first processing unit 1002 is specifically used for:

[0245] The AI ​​task processing result is obtained by acquiring the fifth data based on the first data;

[0246] The loss value of the loss function is determined using the AI ​​task processing results of the first, second, and fifth data sets.

[0247] The fourth data is determined based on the loss value of the loss function.

[0248] In one possible implementation, at least one first quantization parameter includes the number of breakpoints of the first quantizer, the value of the breakpoint, the quantization value at the breakpoint, and part or all of the approximation factor;

[0249] Here, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

[0250] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the first transceiver unit 1001 is further configured to:

[0251] Send at least one second quantization parameter of the first quantizer to the second device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the first device.

[0252] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the first transceiver unit 1001 is further configured to:

[0253] The device receives at least one second quantization parameter from the first quantizer sent by the second device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the second device.

[0254] In one possible implementation, there are one or more second devices, each second device corresponding to a set of first data, a set of fifth data, and a set of second data; the AI ​​task processing result of the fifth data is the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data; in determining the loss value of the loss function using the AI ​​task processing results of the first data, second data, and fifth data, the first processing unit 1002 is specifically used for:

[0255] Obtain the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data corresponding to each second device; wherein, the dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device;

[0256] The loss value of the loss function is determined based on the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data, and the KL divergence corresponding to each second device.

[0257] In one possible implementation, there are multiple second devices, and the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are different or the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are the same.

[0258] It should be noted that the implementation of each unit described in FIG10 can also correspond to the descriptions of the embodiments shown in FIG3 to FIG7. Furthermore, the beneficial effects of the quantization parameter determination device described in FIG10 can be described in the corresponding descriptions of the embodiments shown in FIG3 to FIG7, and will not be repeated here.

[0259] Please refer to Figure 11, which is a schematic diagram of another quantization parameter determination device provided in an embodiment of this application. As shown in Figure 11, the device includes a second transceiver unit 1101 and a second processing unit 1102. Wherein:

[0260] The second transceiver unit 1101 is used to send first data and second data to the first device; the first data is obtained by the first quantizer quantizing the third data through the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine the fourth data; the fourth data is the gradient of the input layer of the first neural network model; and the first device receives the fourth data sent by the first device.

[0261] The second processing unit 1102 is used to determine at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0262] As can be seen from the device shown in Figure 11, the device can send first data quantized by the first quantizer to the first device, and at least one of the channel state information and computing power information corresponding to the second device. Based on the first data and at least one of the channel state information and computing power information sent by the second device, the first device can determine the gradient of the input layer of its deployed neural network model and send the gradient to the second device. The device can then calculate the gradient of the first quantizer based on the gradient, thereby dynamically updating at least one quantization parameter of the first quantizer to adapt the quantization parameters of the first quantizer to the dynamically changing computing power and channel states, thus minimizing the channel transmission overhead while ensuring the accuracy of the end-edge collaborative AI task.

[0263] In one possible implementation, at least one first quantization parameter includes the number of breakpoints of the first quantizer, the value of the breakpoint, the quantization value at the breakpoint, and part or all of the approximation factor;

[0264] Here, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

[0265] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the second transceiver unit 1101 is further configured to:

[0266] The device receives at least one second quantization parameter from the first quantizer sent by the first device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the first device.

[0267] In one possible implementation, when at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the second transceiver unit 1101 is further configured to:

[0268] Send at least one second quantization parameter of the first quantizer to the first device; the at least one second quantization parameter is another part of the quantization parameters other than at least one first quantization parameter, including the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor; the at least one second quantization parameter is pre-configured by the second device.

[0269] In one possible implementation, the third data is obtained by feature extraction of the fifth data through a second neural network model; the fourth data is determined based on the loss value of the loss function, which is determined by the AI ​​task processing results of the first data, the second data, and the fifth data, and the AI ​​task processing result of the fifth data is obtained based on the first data.

[0270] In one possible implementation, there are multiple second devices, and the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are different or the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are the same.

[0271] It should be noted that the implementation of each unit described in FIG11 can also refer to the corresponding descriptions of the embodiments shown in FIG3 to FIG7. Furthermore, the beneficial effects of the quantization parameter determination device described in FIG11 can be referred to the corresponding descriptions of the embodiments shown in FIG3 to FIG7, and will not be repeated here.

[0272] Based on the description of the above method and device embodiments, this application also provides a communication device. Please refer to FIG12, which is a schematic diagram of the structure of a communication device provided in this application embodiment. The communication device includes at least a processor 1201, a memory 1202, and a communication interface 1203, which are interconnected via a bus 1204. This communication device can be used to execute relevant steps of the transport block segmentation method. The communication device can be a transmitting device or a chip in a transmitting device in a wireless communication system, or a receiving device or a chip in a receiving device, such as a terminal device or a network device. The processor 1201 in the communication device is used to read the computer program code stored in the memory 1202 and execute the method of any one of the embodiments shown in FIG3 to FIG7.

[0273] The memory 1202 includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or compact disc read-only memory (CD-ROM), and is used to store related computer programs and data.

[0274] The processor 1201 can be one or more central processing units (CPUs). When the processor 1201 is a CPU, the CPU can be a single-core CPU or a multi-core CPU.

[0275] For example, when the communication device is the first device, the processor 1201 in the communication device can be used to read one or more programs stored in the memory 1202 and perform the following operations:

[0276] The system receives first data and second data sent by the second device; the first data is obtained by quantizing the third data by the first quantizer through the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information.

[0277] Based on the first and second data, the fourth data is determined; the fourth data is the gradient of the input layer of the first neural network model.

[0278] Send fourth data to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in round (m+1).

[0279] For example, when the communication device is a second device, the processor 1201 in the communication device can be used to read one or more programs stored in the memory 1202 and perform the following operations:

[0280] First data and second data are sent to the first device; the first data is obtained by the first quantizer quantizing the third data using the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine the fourth data; the fourth data is the gradient of the input layer of the first neural network model;

[0281] Receive the fourth data sent by the first device;

[0282] Based on the fourth data, at least one first quantization parameter of the first quantizer in round (m+1) is determined.

[0283] It should be noted that the implementation of each operation can also correspond to the description of the method in any of the embodiments shown in Figures 3 to 7.

[0284] It should be noted that although the communication device shown in Figure 12 only illustrates the processor 1201, memory 1202, communication interface 1203, and bus 1204, those skilled in the art should understand that in specific implementations, the communication device may also include other components necessary for normal operation. Furthermore, depending on specific needs, those skilled in the art should understand that the communication device may also include hardware components for implementing other additional functions. Moreover, those skilled in the art should understand that the communication device may only include the components necessary for implementing the embodiments of this application, and not necessarily all the components shown in Figure 12.

[0285] This application also provides a chip, including: a processor, configured to call and run a computer program from a memory, causing a device with the chip installed to perform the method described in any of the embodiments shown in Figures 3 to 7 above. This chip may be a chip in a communication device.

[0286] This application also provides a computer-readable storage medium (memory) storing a computer program that, when executed, implements the method described in any of the embodiments shown in Figures 3 to 7. It is understood that the computer-readable storage medium here may include built-in storage media in a device, or it may include extended storage media supported by the device. The computer-readable storage medium provides storage space containing the device's operating system. Furthermore, one or more computer programs suitable for loading and execution by the device's processor are also stored in this storage space. It should be noted that the computer-readable storage medium here may be high-speed RAM or non-volatile memory, such as at least one disk storage device; optionally, it may also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0287] This application also provides a computer program product, which includes computer program code. When the computer program code is run by a communication device, the method flow described in any one of the embodiments in Figures 3 to 7 is implemented.

[0288] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0289] It should be understood that the processor mentioned in the embodiments of this application can be a CPU, or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0290] It should also be understood that the memory mentioned in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be ROM, Programmable Read-Only Memory (PROM), EPROM, Electrically Erasable Programmable Read-Only Memory (EEPROM), or flash memory. Volatile memory can be RAM, which is 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 Synchronous Dynamic Random Access Memory (ESDRAM), Synchlink Dynamic Random Access Memory (SLDRAM), and Direct Rambus RAM (DR RAM).

[0291] It should be noted that when the processor is a general-purpose processor, DSP, ASIC, FPGA, or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, the memory (storage module) is integrated into the processor.

[0292] It should be noted that the memories described herein are intended to include, but are not limited to, these and any other suitable types of memories.

[0293] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0294] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely exemplary. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0295] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0296] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0297] In this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. In the textual description of this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0298] The steps in the method of this application embodiment can be adjusted, combined, or deleted according to actual needs.

[0299] The modules in the device of this application embodiment can be merged, divided, and deleted according to actual needs.

[0300] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for determining quantization parameters, characterized in that, Applied to a first device; the method includes: The system receives first data and second data sent by a second device; the first data is obtained by a first quantizer quantizing third data using quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information. Based on the first data and the second data, a fourth data is determined; the fourth data is the gradient of the input layer of the first neural network model. The fourth data is sent to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in the (m+1)th round.

2. The method according to claim 1, characterized in that, The third data is obtained by feature extraction of the fifth data through the second neural network model; The determination of the fourth data based on the first data and the second data includes: The AI ​​task processing result of the fifth data is obtained based on the first data; The loss value of the loss function is determined using the AI ​​task processing results of the first data, the second data, and the fifth data. The fourth data is determined based on the loss value of the loss function.

3. The method according to claim 1 or 2, characterized in that, The at least one first quantization parameter includes the number of breakpoints of the first quantizer, the value of the breakpoint, the quantization value at the breakpoint, and part or all of the approximation factor; Wherein, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

4. The method according to claim 3, characterized in that, When the at least one first quantization parameter is a subset of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the method further includes the following steps before determining the fourth data based on the first data and the second data: Send at least one second quantization parameter of the first quantizer to the second device; the at least one second quantization parameter is the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and another part of the quantization parameters of the approximation factor other than the at least one first quantization parameter; the at least one second quantization parameter is pre-configured by the first device.

5. The method according to claim 3, characterized in that, When the at least one first quantization parameter is a subset of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the method further includes the following steps before determining the fourth data based on the first data and the second data: The device receives at least one second quantization parameter of the first quantizer sent by the second device; the at least one second quantization parameter is the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and another part of the quantization parameters in the approximation factor other than the at least one first quantization parameter; the at least one second quantization parameter is pre-configured by the second device.

6. The method according to claim 2, characterized in that, The second device may be one or more, and each second device corresponds to a set of the first data, a set of the fifth data, and a set of the second data; the AI ​​task processing result of the fifth data is the probability that the first data corresponding to one or more second devices is processed into the label of the fifth data; The process of determining the loss value of the loss function using the AI ​​task processing results of the first data, the second data, and the fifth data includes: Obtain the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data corresponding to each second device; wherein, the dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device; The loss value of the loss function is determined based on the probability that the first data corresponding to one or more of the second devices is processed into the label of the fifth data, and the KL divergence corresponding to each of the second devices.

7. The method according to claim 4 or 5, characterized in that, The second device may be multiple, and the types of the at least one second quantization parameter pre-configured by the first quantizer on the multiple second device sides may be different or the types of the at least one second quantization parameter pre-configured by the first quantizer on the multiple second device sides may be the same.

8. A method for determining quantization parameters, characterized in that, Applied to a second device; the method includes: First data and second data are sent to a first device; the first data is obtained by a first quantizer quantizing third data using quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine fourth data; the fourth data is the gradient of the input layer of the first neural network model; Receive the fourth data sent by the first device; Based on the fourth data, at least one first quantization parameter of the first quantizer in the (m+1)th round is determined.

9. The method according to claim 8, characterized in that, The at least one first quantization parameter includes the number of breakpoints of the first quantizer, the value of the breakpoint, the quantization value at the breakpoint, and part or all of the approximation factor; Wherein, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

10. The method according to claim 9, characterized in that, When the at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the method further includes, before receiving the fourth data sent by the first device: The device receives at least one second quantization parameter of the first quantizer sent by the first device; the at least one second quantization parameter is the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and another part of the quantization parameters of the approximation factor other than the at least one first quantization parameter; the at least one second quantization parameter is pre-configured by the first device.

11. The method according to claim 9, characterized in that, When the at least one first quantization parameter is a portion of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, the method further includes, before receiving the fourth data sent by the first device: Send at least one second quantization parameter of the first quantizer to the first device; the at least one second quantization parameter is the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and another part of the quantization parameters of the approximation factor other than the at least one first quantization parameter; the at least one second quantization parameter is pre-configured by the second device.

12. The method according to any one of claims 8-11, characterized in that, The third data is obtained by feature extraction of the fifth data through the second neural network model; the fourth data is determined based on the loss value of the loss function, which is determined by the AI ​​task processing results of the first data, the second data, and the fifth data, and the AI ​​task processing results of the fifth data are obtained based on the first data.

13. The method according to claim 10 or 11, characterized in that, The second device may be multiple, and the types of the at least one second quantization parameter pre-configured by the first quantizer on the multiple second device sides may be different or the types of the at least one second quantization parameter pre-configured by the first quantizer on the multiple second device sides may be the same.

14. A device for determining quantitative parameters, characterized in that, The device is applied to a first equipment; the device includes a first transceiver unit and a first processing unit. The first transceiver unit is used to receive first data and second data sent by the second device; the first data is obtained by the first quantizer quantizing the third data through the quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information. The first processing unit is configured to determine fourth data based on the first data and the second data; the fourth data is the gradient of the input layer of the first neural network model. The first transceiver unit is further configured to send the fourth data to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in the (m+1)th round.

15. A device for determining quantitative parameters, characterized in that, Applied to a second device, the device includes a second transceiver unit and a second processing unit. The second transceiver unit is used to send first data and second data to the first device; the first data is obtained by quantizing third data by a first quantizer through quantization parameters configured in the m-th round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine fourth data; the fourth data is the gradient of the input layer of the first neural network model; and the second transceiver unit receives the fourth data sent by the first device. The second processing unit is configured to determine at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

16. A communication device, characterized in that, It includes a processor, a memory, a communication interface, and one or more programs, said one or more programs being stored in the memory and configured to, when executed by the processor, cooperate with the communication interface to implement the method as described in any one of claims 1-7 or 8-13.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program for execution by the device, which, when executed, implements the method of any one of claims 1-7 or 8-13.

18. A computer program product, characterized in that, When the computer program product is run by a communication device, the communication device performs the method as described in any one of claims 1-7 or 8-13.

Images