Hybrid digital-analog based neural network parameter transmission method and related apparatus

Abstract

The application discloses a neural network parameter transmission method based on mixed number mode and a related device. The method comprises the following steps: dividing target neural network parameters into blocks to obtain a first parameter block group; calculating importance indexes of the parameter blocks in the first parameter block group; starting from the parameter block with the largest importance index, calculating the performance loss of mixed number mode transmission under a digital transmission mode, thereby determining a second parameter block group of the digital transmission mode; calculating power amplification factors of the parameter blocks in the third parameter block group under an analog transmission mode; and transmitting each parameter block in the first parameter block group in the mixed number mode according to the number of the parameter blocks in the second parameter block group and the power amplification factors. The neural network parameters are processed by block division, the performance loss is calculated, and the parameter blocks are reasonably allocated to digital transmission channels and analog transmission channels according to the performance loss, so that the allocation of transmission resources is realized, and the problems of cliff effect and low performance upper limit are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and in particular to a neural network parameter transmission method based on hybrid digital-analog and related apparatus. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, deep neural networks have made significant breakthroughs in computer vision, natural language processing and other fields. The continuous improvement of hardware computing power has enabled more and more edge devices to deploy neural network models to perform computer vision, machine reasoning and natural language processing tasks. In order to achieve the best performance, these edge devices need to select appropriate neural network parameters according to the characteristics of the input image or video, environmental conditions and precision requirements. Although some neural network models have generalization ability and can adapt to different application scenarios, adjusting neural network parameters according to different input scenarios can still help improve the performance after completing the task and the user experience. However, storing all applicable neural network parameters will occupy a large amount of storage resources of the edge device, and is also not conducive to the subsequent upgrade of the model. Therefore, it is considered to transmit the neural network parameters from the server to the edge device end.

[0003] Traditional transmission schemes include digital transmission and analog transmission. Digital transmission encodes and compresses the transmission data, and converts the transmission data into actual transmission symbols through channel coding and modulation. This transmission method can ensure the accuracy of transmission under high signal-to-noise ratio (SNR), but under low signal-to-noise ratio, the receiving end will not be able to correctly decode, and the received signal will produce serious errors. Such errors make the received signal unable to be used normally, and this phenomenon is also called "cliff effect". In contrast, analog transmission uses an analog modulation scheme to avoid completely incorrect decoding of data, but because the transmission data is not compressed by encoding, the encoding efficiency is low, and the transmission performance cannot be guaranteed.

[0004] The proposal of hybrid digital-analog transmission (HDA) effectively solves the drawbacks of the above-mentioned traditional schemes, but the existing hybrid digital-analog transmission technology cannot solve the problem of resource allocation. That is, when facing the neural network parameters as the source, how to reasonably allocate resources to them becomes a technical problem that technicians in the field urgently need to solve. SUMMARY

[0005] In view of the above problems, the present application provides a neural network parameter transmission method based on hybrid digital-analog and related apparatus which overcomes the above problems or at least partially solves the above problems.

[0006] In a first aspect, a neural network parameter transmission method based on hybrid digital-analog includes:

[0007] performing block division on the target neural network parameters to obtain a first parameter block group;

[0008] calculating an importance index of each parameter block in the first parameter block group, wherein a parameter block with a larger importance index has a greater impact on transmission performance, and one parameter block corresponds to one importance index;

[0009] starting from a parameter block with the largest importance index, iteratively calculating performance loss of hybrid digital-analog transmission under a digital transmission mode, and determining a second parameter block group using the digital transmission mode according to the performance loss;

[0010] calculating a power amplification factor of each parameter block in the third parameter block group under an analog transmission mode, wherein one power amplification factor corresponds to one parameter block, and the third parameter block group is equal to the first parameter block group minus the second parameter block;

[0011] transmitting each parameter block in the first parameter block group in a hybrid digital-analog transmission mode according to the number of parameter blocks included in the second parameter block group and each power amplification factor.

[0012] Optionally, in some optional embodiments, the block division on the target neural network parameters to obtain a first parameter block group comprises:

[0013] dividing target neural network parameters at the same level and in the same channel into the same parameter block according to the level and channel of each target neural network parameter to obtain the first parameter block group, wherein the first parameter block group includes a plurality of parameter blocks, and one parameter block includes at least one target neural network parameter.

[0014] Optionally, in some optional embodiments, the calculation of the importance index of each parameter block in the first parameter block group comprises:

[0015] calculating the standard deviation of any parameter block in the first parameter block group, wherein one parameter block corresponds to one standard deviation;

[0016] calculating the average gradient of each target neural network parameter included in any parameter block in the first parameter block group, wherein one parameter block corresponds to one average gradient;

[0017] calculating the importance index of any parameter block in the first parameter block group according to the standard deviation and average gradient of the parameter block.

[0018] Optionally, in some optional embodiments, the calculating, for any parameter block in the first parameter block group, the average gradient of each target neural network parameter included in the parameter block comprises:

[0019] inputting, for any parameter block in the first parameter block group, each target neural network parameter included in the parameter block into a preset neural network model to obtain a loss function of the neural network model;

[0020] deriving the average gradient corresponding to the parameter block by derivation of the loss function.

[0021] Optionally, in some optional embodiments, the calculating, for any parameter block in the first parameter block group, the importance indicator of the parameter block according to the standard deviation and the average gradient of the parameter block comprises:

[0022] calculating, for any parameter block in the first parameter block group, the product of the standard deviation and the average gradient of the parameter block to obtain the importance indicator of the parameter block.

[0023] Optionally, in some optional embodiments, the iteratively calculating, starting from the parameter block with the largest importance indicator, the performance loss of the hybrid digital-analog transmission under the digital transmission mode and determining the second parameter block group adopting the digital transmission mode according to the performance loss comprises:

[0024] sequentially ranking each parameter block starting from the parameter block with the largest importance indicator, wherein the parameter block with the largest importance indicator is ranked first;

[0025] calculating the first performance loss according to the importance indicator of the first parameter module;

[0026] calculating the second performance loss according to the importance indicator of the first parameter module and the importance indicator of the second parameter module;

[0027] if the second performance loss is greater than the first performance loss, determining the first parameter module as the second parameter block group adopting the digital transmission mode;

[0028] if the second performance loss is not greater than the first performance loss, calculating the third performance loss according to the importance indicator of the first parameter module, the importance indicator of the second parameter module and the importance indicator of the third parameter module;

[0029] if the third performance loss is greater than the second performance loss, determining the first parameter module and the second parameter module as the second parameter block group adopting the digital transmission mode;

[0030] If the third performance loss is not greater than the second performance loss, a fourth performance loss is calculated according to the importance indicators of the first bit parameter module, the importance indicators of the second bit parameter module, the importance indicators of the third bit parameter module, and the importance indicators of the fourth bit parameter module.

[0031] The above process is repeated until the second parameter block group using digital transmission is determined.

[0032] Optionally, in some optional embodiments, the calculation of the power amplification factor of each parameter block in the third parameter block group in the analog transmission mode comprises:

[0033] The second parameter block group is quantized to obtain a quantized second parameter block group.

[0034] The parameter blocks in the quantized second parameter block group are entropy encoded to obtain the average number of bits consumed, wherein one parameter block corresponds to one number of bits.

[0035] For any level, the digital bandwidth and digital energy consumed by the level for digital transmission are calculated according to the number of bits of the parameter block at the level in the second parameter block group.

[0036] For any level, the analog bandwidth and analog energy consumed by the level for analog transmission are calculated according to the digital bandwidth, the digital energy, a preset total bandwidth, and a preset total energy of the level, wherein one level corresponds to one preset total bandwidth and one preset total energy.

[0037] The power amplification factor of each parameter block in the third parameter block group in the analog transmission mode is calculated according to the analog energies.

[0038] Optionally, in some optional embodiments, the transmission of each parameter block in the first parameter block group in the mixed digital-analog transmission mode according to the number of parameter blocks included in the second parameter block group and the power amplification factors comprises:

[0039] The parameter blocks in the second parameter block group in the first parameter block group are quantized, source encoded, channel encoded, and modulated to obtain a digital signal corresponding to a digital transmission channel.

[0040] The parameter blocks in the third parameter block group in the first parameter block group are power amplified to obtain an analog signal corresponding to an analog transmission channel.

[0041] The digital signal and the analog signal are combined to obtain actual neural network parameter information in the channel.

[0042] The neural network parameter information and the metadata are sent to a receiving end through a channel, wherein the metadata includes a number of parameter blocks included in the second parameter block group and the power amplification factors.

[0043] In a second aspect, a neural network parameter transmission device based on hybrid digital-analog is provided, which includes a parameter block unit, an index calculation unit, a digital block unit, an analog block unit and a hybrid transmission unit.

[0044] The parameter block unit is configured to block target neural network parameters to obtain a first parameter block group.

[0045] The index calculation unit is configured to calculate importance indexes of the parameter blocks in the first parameter block group, wherein a parameter block with a larger importance index has a greater impact on transmission performance, and one parameter block corresponds to one importance index.

[0046] The digital block unit is configured to iteratively calculate performance losses of hybrid digital-analog transmission in a digital transmission mode from the parameter block with the largest importance index, and determine a second parameter block group adopting the digital transmission mode according to the performance losses.

[0047] The analog block unit is configured to calculate power amplification factors of the parameter blocks in a third parameter block group in an analog transmission mode, wherein one power amplification factor corresponds to one parameter block, and the third parameter block group is equal to the first parameter block group minus the second parameter block.

[0048] The hybrid transmission unit is configured to transmit the parameter blocks in the first parameter block group in a hybrid digital-analog transmission mode according to a number of parameter blocks included in the second parameter block group and the power amplification factors.

[0049] In a third aspect, a computer readable storage medium is provided, which stores a program, and the program is executed by a processor to implement the neural network parameter transmission method based on hybrid digital-analog.

[0050] In a fourth aspect, an electronic device is provided, which includes at least one processor, at least one memory connected to the processor through a bus, wherein the processor, the memory and the bus complete communication with each other; the processor is configured to call program instructions in the memory to execute the neural network parameter transmission method based on hybrid digital-analog.

[0051] By the technical scheme, the neural network parameter transmission method and related device based on mixed digital-analog can obtain a first parameter block group by block processing target neural network parameters, calculate importance indexes of each parameter block in the first parameter block group, wherein a parameter block with a larger importance index has a greater influence on transmission performance, and one parameter block corresponds to one importance index, iteratively calculate performance loss of mixed digital-analog transmission in a digital transmission mode from a parameter block with the largest importance index, and determine a second parameter block group adopting the digital transmission mode according to the performance loss, calculate power amplification factors of each parameter block in a third parameter block group in an analog transmission mode, wherein one power amplification factor corresponds to one parameter block, and the third parameter block group is equal to the first parameter block group minus the second parameter block, and transmit each parameter block in the first parameter block group in a mixed digital-analog transmission mode according to the number of parameter blocks included in the second parameter block group and the power amplification factors. It can be seen that the neural network parameter transmission method and related device based on mixed digital-analog realize the allocation of transmission resources by block processing neural network parameters, calculating performance loss, and reasonably allocating parameter blocks to digital transmission channels and analog transmission channels according to the performance loss, and effectively solve the cliff effect problem of digital transmission and the problem of low performance upper limit of analog transmission.

[0052] The above description is only a summary of the technical scheme of the present application. In order to more clearly understand the technical means of the present application, the content of the specification can be implemented, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0053] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of the preferred embodiments and are not meant to limit the present application. Furthermore, the same reference numerals are used throughout the several drawings to designate the same or similar parts. In the drawings:

[0054] Figure 1 A flow chart of a first neural network parameter transmission method based on mixed digital-analog provided by the present application is shown;

[0055] Figure 2 A flow chart of a second neural network parameter transmission method based on mixed digital-analog provided by the present application is shown;

[0056] Figure 3 A system block diagram of a neural network parameter transmission method based on mixed digital-analog provided by the present application is shown;

[0057] Figure 4A first simulation experiment data graph provided by the application is shown;

[0058] Figure 5 A second simulation experiment data graph provided by the application is shown;

[0059] Figure 6 A third simulation experiment data graph provided by the application is shown;

[0060] Figure 7 A structure schematic diagram of a neural network parameter transmission device based on a hybrid digital-analog model provided by the application is shown;

[0061] Figure 8 A structure schematic diagram of an electronic device provided by the application is shown. DETAILED DESCRIPTION

[0062] Exemplary embodiments of the present application will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present application can be more thoroughly understood and so that the scope of the present application can be conveyed to those skilled in the art.

[0063] As Figure 1 shown, the application provides a neural network parameter transmission method based on a hybrid digital-analog model, comprising: S100, S200, S300, S400 and S500.

[0064] S100, block the target neural network parameters to obtain a first parameter block group;

[0065] For example, in some optional embodiments, the S100 comprises step 1.1.

[0066] Step 1.1, according to the level and channel of each target neural network parameter, the target neural network parameters in the same level and the same channel are divided into the same parameter block to obtain the first parameter block group, wherein the first parameter block group comprises a plurality of parameter blocks, and one parameter block comprises at least one target neural network parameter.

[0067] Optionally, it is assumed that the neural network parameters to be transmitted by the application are W=(W1, W2, …, W l ,…,W L ), the neural network parameters of the lth layer are represented by the symbol , wherein I and O respectively represent the dimensions of the input and output of the level (the dimension of the output is also referred to as the channel), K represents the size of the convolution kernel in the convolutional neural network, and if it is a fully connected neural network, K is equal to 1.

[0068] Optionally, the layer of neural network parameters is divided into blocks according to different channels of the layer of neural network parameters, and each block is denoted by a symbol , wherein i = 1, 2, …, O. That is, the parameter blocks of each channel of each layer are represented by The present application does not limit the parameter blocks of each channel of each layer.

[0069] Optionally, the neural network is a model composed of different layers (levels), and each layer performs a specific function according to different parameters, thereby realizing layer-by-layer abstraction and processing of input data. For example, the hierarchical structure of the neural network is divided into three layers: an input layer, a hidden layer, and an output layer. The hidden layer includes four parts: a convolutional layer, an activation layer, a pooling layer, and a fully connected layer, and the present application does not limit the same.

[0070] S200, calculating the importance indicators of the parameter blocks in the first parameter block group, wherein the parameter block with a larger importance indicator has a greater impact on the transmission performance, and one parameter block corresponds to one importance indicator;

[0071] Optionally, in some optional embodiments, the S200 includes steps 2.1, 2.2, and 2.3.

[0072] Step 2.1, calculating the standard deviation of any parameter block in the first parameter block group, wherein one parameter block corresponds to one standard deviation;

[0073] Optionally, the standard deviation σ li of each block can be directly calculated by the aforementioned w li .

[0074] The calculation formula is: wherein N is the number of neural network parameters in the block group, i.e., the number of w li . is the average value of all w li .

[0075] Step 2.2, calculating the average gradient of each target neural network parameter included in any parameter block in the first parameter block group, wherein one parameter block corresponds to one average gradient;

[0076] Optionally, the application can take a mini-batch (a subset of the mini-batch gradient descent algorithm, which can include multiple neural network parameters of a parameter block as a mini-batch, i.e., a parameter block as a mini-batch) as input of the neural network model, calculate the corresponding loss function, and derive the gradient of each neural network parameter included in the parameter block corresponding to the parameter block from the loss function, one neural network parameter corresponding to one gradient. li The average gradient g of the parameter block li = the average of the gradients corresponding to each neural network parameter of the parameter block.

[0077] That is, in some optional embodiments, the step 2.2 includes step 3.1 and step 3.2.

[0078] Step 3.1, for any parameter block in the first parameter block group, input each target neural network parameter included in the parameter block into a preset neural network model to obtain the loss function of the neural network model.

[0079] Step 3.2, derive the average gradient corresponding to the parameter block from the loss function.

[0080] Step 2.3, for any parameter block in the first parameter block group, calculate the importance indicator of the parameter block according to the standard deviation and the average gradient of the parameter block.

[0081] For example, in some optional embodiments, the step 2.3 includes step 4.1.

[0082] Step 4.1, for any parameter block in the first parameter block group, calculate the product of the standard deviation and the average gradient of the parameter block to obtain the importance indicator of the parameter block.

[0083] Optionally, for any parameter block, the product of the standard deviation and the average gradient can be used as the importance indicator of the parameter block. It should be noted that the importance indicator reflects the degree of influence of transmitting the corresponding parameter block on the performance loss, and the parameter block with a large degree of influence can be considered for digital transmission. Generally, the parameter block with a large importance indicator has a greater impact on the performance loss after transmission, and digital transmission can be considered to improve the transmission accuracy, which is not limited by the application.

[0084] Optionally, after the importance indicators of the parameter blocks are calculated, the application can sort the parameter blocks in descending order of the importance indicators, so as to subsequently determine the parameter blocks that can be transmitted in the digital manner in descending order of the importance indicators, and the application does not limit this.

[0085] S300, starting from the parameter block with the largest importance indicator, iteratively calculate the performance loss of the hybrid digital-analog transmission in the digital transmission manner, and determine a second parameter block group using the digital transmission manner according to the performance loss;

[0086] Optionally, in some optional embodiments, the S300 includes steps 5.1, 5.2, 5.3, 5.4, 5.5, 5.6, and 5.7.

[0087] Step 5.1, starting from the parameter block with the largest importance indicator, sort the parameter blocks in descending order of the importance indicators, wherein the parameter block with the largest importance indicator is ranked first.

[0088] Optionally, as described above, the application can sort the parameter blocks in descending order of the importance indicators, so as to subsequently determine the parameter blocks that can be transmitted in the digital manner starting from the parameter block with the largest importance indicator, and the application does not limit this.

[0089] Step 5.2, calculate the first performance loss according to the importance indicator of the first parameter block.

[0090] Optionally, as described above, after the parameter blocks are sorted, the first parameter block is the parameter block with the largest importance indicator. Since the larger the importance indicator is, the greater the impact on the performance loss after transmission is, the application can start from the parameter block with the largest importance indicator to iteratively solve the number of parameter blocks that are transmitted in the digital manner. The formula is as follows: wherein k is the number of parameter blocks that are transmitted in the digital manner, k is calculated from zero, and the performance loss of the hybrid digital-analog transmission when different blocks are transmitted in the digital manner is stopped when the loss is greater than the previous one. l denotes the quantization step size, denotes the standard deviation of the quantization error generated after the quantization of the block i, N o denotes the channel noise, E a denotes the energy of the analog transmission, M is the number of blocks that are transmitted. When the transmission bandwidth is sufficient, formula M = 0 is satisfied, and at this time, the formula for calculating the performance loss of the hybrid digital-analog transmission is simplified as follows:

[0091] Specifically, this invention uses iterative calculation to solve for the specific value of k, starting from 0, and calculates D. hyb (k,Q l When calculating to the k'th (k'>0) inequality: D hyb (k',Q l ) < D hyb (k'-1,Q l Then, continue iteratively solving for the performance loss D of hybrid digital-to-analog transmission at k'+1. hyb (k'+1,Q l When the inequality: D hyb (k'+1,Q l )>D hyb (k',Q l If k' is not found, the calculation stops, and k' is used as the number of blocks for digital transmission.

[0092] Step 5.3: Calculate the second performance loss based on the importance index of the first parameter module and the importance index of the second parameter module;

[0093] Optionally, as can be seen from the aforementioned formula, this invention requires at least two performance losses to be calculated before a comparison can be made. Therefore, this invention needs to iteratively calculate the second performance loss, but this invention does not impose any limitations on this.

[0094] Step 5.4: If the second performance loss is greater than the first performance loss, then determine the first bit parameter module as the second parameter block group using digital transmission method;

[0095] Step 5.5: If the second performance loss is not greater than the first performance loss, then the third performance loss is calculated based on the importance index of the first parameter module, the importance index of the second parameter module, and the importance index of the third parameter module.

[0096] Step 5.6: If the third performance loss is greater than the second performance loss, then determine that the first bit parameter module and the second bit parameter module together form a second parameter block group using digital transmission method;

[0097] Step 5.7: If the third performance loss is not greater than the second performance loss, then calculate the fourth performance loss based on the importance index of the first parameter module, the importance index of the second parameter module, the importance index of the third parameter module, and the importance index of the fourth parameter module. Repeat this process until the second parameter block group using digital transmission mode is determined.

[0098] Optionally, in combination with the foregoing formula, the application determines the number of parameter blocks of digital transmission by iteratively calculating the performance loss, comparing the size of the performance loss, and repeating the cycle until the number of parameter blocks of digital transmission is determined, i.e., the second parameter block group using the digital transmission mode is determined, which will not be described herein again, and please refer to the foregoing formula for details.

[0099] S400, calculating power amplification factors of each parameter block of the third parameter block group under the analog transmission mode, wherein one power amplification factor corresponds to one parameter block, and the third parameter block group is equal to the first parameter block group minus the second parameter block;

[0100] For example, in some optional embodiments, the S400 comprises steps 6.1, 6.2, 6.3, 6.4 and 6.5.

[0101] Step 6.1, quantizing each parameter block of the second parameter block group to obtain a quantized second parameter block group.

[0102] Step 6.2, entropy encoding each parameter block of the quantized second parameter block group to obtain the average number of bits consumed, wherein one parameter block corresponds to one bit number.

[0103] Step 6.3, for any level, calculating the digital bandwidth and digital energy consumed by the level for digital transmission according to the bit number of the parameter block at the level in the second parameter block group.

[0104] Optionally, in general, the distribution of neural network parameters approximately satisfies the Gaussian distribution li ~ N(0, λ li , and the parameters are compressed by uniform quantization, wherein the quantization step is Q l , and the quantized parameter block The expression is: wherein represents the floor symbol, λ li represents the variance of the block.

[0105] Optionally, the quantized parameter also satisfies the Gaussian distribution The quantized residual can be expressed as:

[0106] Optionally, the average number of bits consumed by the quantized neural network parameters after entropy encoding is:

[0107] The total bandwidth (digital bandwidth) and total energy consumption (digital energy) of the lth level digital transmission are: wherein r is the code rate after channel coding and modulation in digital transmission, P s is the minimum power required for correct transmission of a symbol in digital transmission, is the total bandwidth for digital transmission of the lth layer, is the total energy consumption for digital transmission of the lth layer.

[0108] Step 6.4, for any layer, calculating the analog bandwidth and analog energy consumed by the layer for analog transmission according to the digital bandwidth, the digital energy, the preset total bandwidth and the preset total energy of the layer, wherein one layer corresponds to one preset total bandwidth and one preset total energy;

[0109] Step 6.5, calculating the power amplification factor of each parameter block in the third parameter block group in the analog transmission mode according to each analog energy.

[0110] Optionally, in combination with the process of calculating the total bandwidth and the total energy consumption for digital transmission, the total bandwidth (analog bandwidth) and the total energy consumption (analog energy) of analog transmission of the lth layer are: wherein B l is a preset setting value (representing the total bandwidth of the lth layer for hybrid digital-analog transmission), E l is a preset setting value (representing the total energy consumption of the lth layer for hybrid digital-analog transmission), is the total bandwidth for analog transmission of the lth layer, is the total energy consumption for analog transmission of the lth layer.

[0111] The power amplification factor of each parameter block in analog transmission is calculated by the following formula: wherein E a is the total consumed energy of all layers for analog transmission, O represents the dimension of the output of the layer (the dimension of the output is also referred to as a channel), K represents the size of the convolution kernel in the convolutional neural network, l is the layer (number) of the neural network model in which the corresponding neural network parameter is located, i and j are both the input channels (numbers) of the layer of the neural network model in which the corresponding neural network parameter is located, and the number j is applicable to the summation calculation in the root sign, which is distinguished from the number i outside the summation calculation.

[0112] S500, transmitting each parameter block in the first parameter block group in a hybrid digital-analog transmission mode according to the number of parameter blocks included in the second parameter block group and each power amplification factor.

[0113] For example, in some optional embodiments, the S500 includes steps 7.1, 7.2, 7.3 and 7.4.

[0114] Step 7.1, quantizing, source coding, channel coding and modulating each parameter block in the second parameter block group in the first parameter block group to obtain a digital signal corresponding to a digital transmission channel;

[0115] Step 7.2, power amplifying each parameter block in the third parameter block group in the first parameter block group to obtain an analog signal corresponding to an analog transmission channel;

[0116] Step 7.3, merging the digital signal and the analog signal to obtain actual neural network parameter information in a channel;

[0117] Step 7.4, sending the neural network parameter information and metadata to a receiving end through a channel, wherein the metadata includes the number of parameter blocks included in the second parameter block group and each power amplification factor.

[0118] Optionally, the transmitting end sends the encoded symbols and metadata of the digital transmission and the analog transmission, and the receiving end reconstructs the neural network parameters from the received symbols and metadata, specifically:

[0119] The data w transmitted by the transmitting end li After being encoded by the digital transmission and the analog transmission, the data becomes the signal actually transmitted in the channel, and the metadata includes the number k of blocks of the digital transmission and the power amplification factor s li . The process of the digital transmission includes quantization, source coding, channel coding and modulation, and the process of the analog transmission includes power amplification.

[0120] After receiving the data transmitted by the digital transmission channel, the receiving end obtains the data through demodulation, channel decoding, source decoding and inverse quantization, and the data of the analog transmission channel is reconstructed by the following formula: , wherein is the reconstructed data information, y li is the information received by the receiving end, is the variance of the channel noise. The information of the analog transmission channel and the information of the digital transmission channel are merged to obtain the actual neural network parameter information.

[0121] Optionally, as known from the foregoing, the present application processes the neural network parameters by block processing, models the reconstruction loss of the transmitted neural network parameters, and uses Taylor expansion to approximately solve the performance loss. In order to reasonably allocate the transmission parameters to the digital transmission channel and the analog transmission channel, the present application proposes an importance index according to the performance loss, which is used for block sorting of the parameters, determining the number of blocks of the digital transmission and the analog transmission, and realizing the allocation of transmission resources.

[0122] Optionally, in order to more clearly describe the overall execution process of the present application, the present application provides a flowchart as shown inFigure 2 The overall flowchart shown is understood in the light of the foregoing explanation, which will not be repeated here.

[0123] The simulation experiment data of the present application is provided below to demonstrate the effectiveness of the present application, which can effectively solve the cliff effect problem of digital transmission and the low performance limit problem of analog transmission.

[0124] The simulation experiment uses digital transmission, analog transmission and the existing hybrid digital-analog scheme Federated AirNet as a control experiment. Among them, in the analog transmission, a uniform distribution mechanism is used to allocate power to the transmission parameters; in the digital transmission, 8-bit uniform quantization and Huffman coding are used for compression, the channel coding uses a convolution code with a code rate of 2 / 3, and the channel modulation uses QAM64. In the hybrid digital-analog transmission scheme, according to the system block diagram of the appended Figure 3 , the neural network parameters to be transmitted are encoded and transmitted, and the transmission scheme of the digital channel is the same as that of the digital transmission, wherein the occurrence symbol power of the digital transmission is set to 10dB and the quantization step is set to 8 bits per layer.

[0125] Due to the variety of neural networks, in the simulation, models under two typical tasks are used for simulation, namely classification task and super resolution task. In the classification task, ResNet18, ResNet50 and AlexNet models are tested for classification accuracy TOP1 of the classification CIFAR10 test set after transmission and reconstruction; in the super resolution (SR) task, EDSR and ESRGAN models are tested for peak signal-to-noise ratio (PSNR) of the super resolution picture after super resolution of the Set5 data set.

[0126] From the appended Figure 4 , it can be observed that under the condition of SNR greater than 8dB, the performance of the models reconstructed by several transmission methods is basically stable, but there are still slight differences. The performance of the model reconstructed by the hybrid digital-analog transmission designed by the present application is similar to that of the digital transmission, and the performance error compared with the model performance under lossless transmission is less than 0.1%. The performance of the reconstructed neural network under analog transmission is 0.5% lower than that of lossless transmission in high signal-to-noise ratio. Compared with the existing Federated AirNet, the performance loss of the transmitted neural network is smaller.

[0127] When SNR < 8dB, the performance of reconstruction models for several transmission schemes degrades. The most significant performance degradation is seen in the reconstruction neural network under digital transmission. When SNR ≤ 6dB, the reconstruction model's performance experiences a precipitous drop, a typical cliff effect in digital transmission. At SNR = 6dB, the classification accuracy of the reconstruction model under digital transmission is only 30%, the data error rate after channel decoding is high, and the source decoding end cannot decode correctly. Under these conditions, the reconstructed model is essentially unusable. In contrast, the performance degradation of analog transmission and hybrid analog-digital transmission is more gradual, without the precipitous drop seen in digital transmission. The neural network performance after these two transmissions decreases slowly with decreasing SNR. Compared to analog transmission and existing hybrid analog-digital transmission, the hybrid analog-digital transmission proposed in this invention shows a 5% performance improvement at SNR = 5dB and an 8% performance improvement at SNR = 0dB. Because the hybrid analog-digital transmission scheme proposed in this invention performs reasonable power optimization based on the performance loss after modeling in analog coded transmission, it has better performance at the same signal-to-noise ratio compared to traditional analog transmission and FederatedAirNet's single power allocation method.

[0128] To verify the applicability of hybrid analog-to-digital transmission in the transmission of other model parameters, this invention also simulated and tested other neural networks, including AlexNet and ResNet50, as well as super-resolution neural networks EDSR and ESRGAN. (See appendix...) Figure 5 The experimental results show that the hybrid analog-digital transmission scheme proposed in this invention is also applicable to transmitting other convolutional neural network parameters. For classification tasks, the AlexNet and ResNet50 models reconstructed by hybrid analog-digital transmission exhibit excellent performance; when SNR > 8dB, the classification accuracy is close to that of lossless transmission. When SNR < 8dB, digital transmission exhibits a cliff effect, while the hybrid transmission scheme proposed in this invention shows a significant performance advantage over other transmission methods. For super-resolution models, the performance of hybrid analog-digital transmission is only less than 0.2dB lower than that of digital transmission and lossless transmission when SNR > 15dB. Furthermore, at low signal-to-noise ratios, the performance loss is even smaller compared to other transmission methods.

[0129] Because this invention simulates and tests the transmission performance of different schemes without bandwidth limitations, it analyzes the bandwidth requirements of various transmission schemes. (Appendix) Figure 6The number of transmission symbols of different transmission modes in the transmission model is counted. The actual number of symbols to be transmitted is used to represent the bandwidth requirement for transmitting all parameters of a model completely under the same delay. Since there is no limit to the transmission bandwidth of the transmission model, it can be seen that compared with analog transmission, hybrid digital-analog transmission and digital transmission require more transmission bandwidth. However, even if the bandwidth of hybrid digital-analog transmission and digital transmission is higher than that of analog transmission, hybrid digital-analog transmission only needs a smaller bandwidth increment to achieve better performance than analog transmission and avoid the cliff effect in digital transmission.

[0130] As shown in Figure 7 The application provides a hybrid digital-analog based neural network parameter transmission device, which comprises a parameter blocking unit 100, an index calculation unit 200, a digital blocking unit 300, an analog blocking unit 400 and a hybrid transmission unit 500.

[0131] The parameter blocking unit 100 is configured to block target neural network parameters to obtain a first parameter blocking group.

[0132] The index calculation unit 200 is configured to calculate importance indexes of each parameter block in the first parameter blocking group, wherein a parameter block with a larger importance index has a greater impact on transmission performance, and one parameter block corresponds to one importance index.

[0133] The digital blocking unit 300 is configured to iteratively calculate performance loss of hybrid digital-analog transmission under a digital transmission mode from a parameter block with the largest importance index, and determine a second parameter blocking group adopting the digital transmission mode according to the performance loss.

[0134] The analog blocking unit 400 is configured to calculate power amplification factors of each parameter block in a third parameter blocking group under an analog transmission mode, wherein one power amplification factor corresponds to one parameter block, and the third parameter blocking group is equal to the first parameter blocking group minus the second parameter blocking group.

[0135] The hybrid transmission unit 500 is configured to transmit each parameter block in the first parameter blocking group in a hybrid digital-analog transmission mode according to the number of parameter blocks included in the second parameter blocking group and the power amplification factors.

[0136] Optionally, in some optional embodiments, the parameter blocking unit 100 comprises a parameter blocking subunit.

[0137] The parameter block subunit is configured to divide target neural network parameters at the same level and the same channel into a same parameter block according to the level and the channel of the target neural network parameters, to obtain the first parameter block group, wherein the first parameter block group includes a plurality of parameter blocks, and one parameter block includes at least one target neural network parameter.

[0138] Optionally, in some optional embodiments, the index calculation unit 200 includes a standard deviation calculation subunit, an average gradient calculation subunit, and an importance index calculation subunit.

[0139] The standard deviation calculation subunit is configured to calculate a standard deviation of any parameter block in the first parameter block group, wherein one parameter block corresponds to one standard deviation.

[0140] The average gradient calculation subunit is configured to calculate an average gradient of each target neural network parameter included in any parameter block in the first parameter block group, wherein one parameter block corresponds to one average gradient.

[0141] The importance index calculation subunit is configured to calculate an importance index of any parameter block in the first parameter block group according to the standard deviation and the average gradient of the parameter block.

[0142] Optionally, in some optional embodiments, the average gradient calculation subunit includes a loss function determination subunit and a loss function derivation subunit.

[0143] The loss function determination subunit is configured to input each target neural network parameter included in any parameter block in the first parameter block group into a preset neural network model, to obtain a loss function of the neural network model.

[0144] The loss function derivation subunit is configured to derive the loss function to obtain the average gradient corresponding to the parameter block.

[0145] Optionally, in some optional embodiments, the importance index calculation subunit includes a product calculation subunit.

[0146] The product calculation subunit is configured to calculate a product of the standard deviation and the average gradient of any parameter block in the first parameter block group, to obtain the importance index of the parameter block.

[0147] Optionally, in some optional embodiments, the digital block unit 300 comprises: a parameter block sorting subunit, a first loss calculation subunit, a second loss calculation subunit, a first result subunit, a second result subunit, a third result subunit, and a fourth result subunit.

[0148] The parameter block sorting subunit is configured to sort the parameter blocks in sequence from the parameter block with the largest importance index, wherein the parameter block with the largest importance index is ranked first.

[0149] The first loss calculation subunit is configured to calculate a first performance loss based on the importance index of the first parameter module.

[0150] The second loss calculation subunit is configured to calculate a second performance loss based on the importance index of the first parameter module and the importance index of the second parameter module.

[0151] The first result subunit is configured to determine the first parameter module as the second parameter block group using digital transmission if the second performance loss is greater than the first performance loss.

[0152] The second result subunit is configured to calculate a third performance loss based on the importance index of the first parameter module, the importance index of the second parameter module, and the importance index of the third parameter module if the second performance loss is not greater than the first performance loss.

[0153] The third result subunit is configured to determine the first parameter module and the second parameter module as the second parameter block group using digital transmission if the third performance loss is greater than the second performance loss.

[0154] The fourth result subunit is configured to calculate a fourth performance loss based on the importance index of the first parameter module, the importance index of the second parameter module, the importance index of the third parameter module, and the importance index of the fourth parameter module if the third performance loss is not greater than the second performance loss; and the process is repeated until the second parameter block group using digital transmission is determined.

[0155] Optionally, in some optional embodiments, the analog block unit 400 comprises: a parameter quantization subunit, a bit number calculation subunit, a digital transmission calculation subunit, an analog transmission calculation subunit, and an amplification factor calculation subunit.

[0156] The parameter quantization subunit is configured to quantize each parameter block of the second parameter block group to obtain a quantized second parameter block group.

[0157] The bit number calculation subunit is configured to perform entropy encoding on each parameter block in the quantized second parameter block group to obtain an average consumed bit number, wherein one parameter block corresponds to one bit number.

[0158] The digital transmission calculation subunit is configured to calculate, for any level, a digital bandwidth and a digital energy consumed by digital transmission of the level according to the bit numbers of the parameter blocks in the second parameter block group at the level.

[0159] The analog transmission calculation subunit is configured to calculate, for any level, an analog bandwidth and an analog energy consumed by analog transmission of the level according to the digital bandwidth, the digital energy, a preset total bandwidth and a preset total energy of the level, wherein one level corresponds to one preset total bandwidth and one preset total energy.

[0160] The amplification factor calculation subunit is configured to calculate, according to the analog energies, power amplification factors of the parameter blocks in the third parameter block group in the analog transmission mode.

[0161] Optionally, in some optional embodiments, the hybrid transmission unit 500 comprises a digital signal subunit, an analog signal subunit, a signal merging subunit and a signal sending subunit.

[0162] The digital signal subunit is configured to quantize, source encode, channel encode and modulate each parameter block in the second parameter block group in the first parameter block group to obtain a digital signal corresponding to a digital transmission channel.

[0163] The analog signal subunit is configured to perform power amplification on each parameter block in the third parameter block group in the first parameter block group to obtain an analog signal corresponding to an analog transmission channel.

[0164] The signal merging subunit is configured to merge the digital signal and the analog signal to obtain actual neural network parameter information in a channel.

[0165] The signal sending subunit is configured to send the neural network parameter information and metadata to a receiving end through a channel, wherein the metadata comprises a number of parameter blocks included in the second parameter block group and the power amplification factors.

[0166] The application provides a computer readable storage medium having a program stored thereon, wherein the program is executed by a processor to implement the hybrid digital-analog based neural network parameter transmission method.

[0167] As Figure 8As shown, the present application provides an electronic device 70, which comprises at least one processor 701, and at least one memory 702 connected with the processor 701, a bus 703; wherein the processor 701, the memory 702 complete mutual communication through the bus 703; the processor 701 is used to call the program instruction in the memory 702, to execute the neural network parameter transmission method based on hybrid digital model of any one described above.

[0168] In the present application, the relational terms such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises... " does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the stated elements.

[0169] Each of the embodiments in the present specification is described in a related manner, and the same or similar parts between each of the embodiments can be referred to each other, and each of the embodiments focuses on the difference from other embodiments. Especially, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the related parts can be referred to the part of the description of the method embodiments.

[0170] The above description of disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present application can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown in the present application, but will conform to the widest scope consistent with the principles and novel features disclosed in the present application.

[0171] The above description is only the preferred embodiment of the present application, and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for parameter transfer in a neural network based on a hybrid digital-analog model, characterized in that, include: The parameters of the target neural network are divided into blocks to obtain the first parameter block group; Calculate the importance index of each parameter block in the first parameter block group, wherein the parameter block with the larger the importance index has a greater impact on transmission performance, and one parameter block corresponds to one importance index; Starting with the parameter block with the highest importance index, the performance loss of hybrid digital-analog transmission under digital transmission mode is calculated iteratively, and the second parameter block group for adopting digital transmission mode is determined based on the performance loss. Calculate the power amplification factor of each parameter block in the third parameter block group under analog transmission mode, wherein one power amplification factor corresponds to one parameter block, and the third parameter block group is equal to the first parameter block group minus the second parameter block; Based on the number of parameter blocks included in the second parameter block group and each of the power amplification factors, each parameter block in the first parameter block group is transmitted in a mixed digital-to-analog transmission manner.

2. The method according to claim 1, characterized in that, The step of dividing the target neural network parameters into blocks to obtain the first parameter block group includes: According to the level and channel of each target neural network parameter, the target neural network parameters in the same level and the same channel are divided into the same parameter block to obtain the first parameter block group. The first parameter block group includes multiple parameter blocks, and each parameter block includes at least one target neural network parameter.

3. The method according to claim 1, characterized in that, The calculation of the importance index of each parameter block in the first parameter block group includes: For any parameter block in the first parameter block group, calculate the standard deviation of the parameter block, wherein one parameter block corresponds to one standard deviation; For any parameter block in the first parameter block group, calculate the average gradient of each target neural network parameter included in the parameter block, wherein one parameter block corresponds to one average gradient; For any parameter block in the first parameter block group, the importance index of the parameter block is calculated based on the standard deviation and average gradient of the parameter block.

4. The method according to claim 3, characterized in that, The step of calculating the average gradient of each target neural network parameter included in any parameter block within the first parameter block group includes: For any parameter block in the first parameter block group, the target neural network parameters included in the parameter block are input into a preset neural network model to obtain the loss function of the neural network model; The average gradient corresponding to the parameter block is obtained by taking the derivative of the loss function.

5. The method according to claim 3, characterized in that, For any parameter block in the first parameter block group, the importance index of the parameter block is calculated based on the standard deviation and average gradient of the parameter block, including: For any parameter block in the first parameter block group, calculate the product of the standard deviation and the average gradient of the parameter block to obtain the importance index of the parameter block.

6. The method according to claim 1, characterized in that, The process begins with the parameter block with the highest importance index, iteratively calculating the performance loss of hybrid digital-analog transmission under digital transmission mode, and determining the second parameter block group using digital transmission mode based on the performance loss, including: Starting with the parameter block with the highest importance index, each parameter block is sorted sequentially, with the parameter block with the highest importance index being ranked first. The first performance loss is calculated based on the importance index of the first parameter module; The second performance loss is calculated based on the importance index of the first parameter module and the importance index of the second parameter module. If the second performance loss is greater than the first performance loss, then the first bit parameter module is determined as the second parameter block group using digital transmission method; If the second performance loss is not greater than the first performance loss, then the third performance loss is calculated based on the importance index of the first parameter module, the importance index of the second parameter module, and the importance index of the third parameter module. If the third performance loss is greater than the second performance loss, then the first bit parameter module and the second bit parameter module are determined to be a second parameter block group using digital transmission. If the third performance loss is not greater than the second performance loss, then the fourth performance loss is calculated based on the importance index of the first parameter module, the importance index of the second parameter module, the importance index of the third parameter module, and the importance index of the fourth parameter module. This process is repeated until the second parameter block group, which uses digital transmission, is determined.

7. The method according to claim 1, characterized in that, The calculation of the power amplification factor of each parameter block in the third parameter block group under analog transmission mode includes: Quantize each parameter block of the second parameter block group to obtain the quantized second parameter block group; Entropy encoding is performed on each parameter block of the quantized second parameter block group to obtain the average number of bits consumed, where one parameter block corresponds to one number of bits; For any level, the digital bandwidth and digital energy consumed by the level for digital transmission are calculated based on the number of bits in the parameter block of the second parameter block group. For any level, the analog bandwidth and analog energy consumed by the level for analog transmission are calculated based on the digital bandwidth, digital energy, preset total bandwidth and preset total energy of the level. Each level corresponds to a preset total bandwidth and a preset total energy. Based on the simulated energy, the power amplification factor of each parameter block in the third parameter block group under the simulated transmission mode is calculated.

8. The method according to claim 1, characterized in that, The step of transmitting each parameter block in the first parameter block group in a mixed digital-to-analog transmission manner according to the number of parameter blocks included in the second parameter block group and each of the power amplification factors includes: Quantization, source coding, channel coding and modulation are performed on each parameter block of the second parameter block group in the first parameter block group to obtain the digital signal of the corresponding digital transmission channel. Power amplification is performed on each parameter block of the third parameter block group in the first parameter block group to obtain the analog signal of the corresponding analog transmission channel; The digital signal and the analog signal are combined to obtain the actual neural network parameter information in the channel; The neural network parameter information and metadata are sent to the receiving end through a channel, wherein the metadata includes the number of parameter blocks included in the second parameter block group and each of the power amplification factors.

9. A neural network parameter transmission device based on hybrid digital-analog converters, characterized in that, include: Parameter segmentation unit, index calculation unit, digital segmentation unit, analog segmentation unit, and hybrid transmission unit; The parameter segmentation unit is used to segment the parameters of the target neural network into blocks to obtain the first parameter segmentation group; The index calculation unit is used to calculate the importance index of each parameter block in the first parameter block group, wherein the parameter block with the larger the importance index has a greater impact on the transmission performance, and one parameter block corresponds to one importance index. The digital block unit is used to iteratively calculate the performance loss of hybrid digital-analog transmission under the digital transmission mode, starting from the parameter block with the highest importance index, and determine the second parameter block group to adopt the digital transmission mode based on the performance loss. The analog block unit is used to calculate the power amplification factor of each parameter block in the third parameter block group under the analog transmission mode, wherein one power amplification factor corresponds to one parameter block, and the third parameter block group is equal to the first parameter block group minus the second parameter block. The hybrid transmission unit is used to transmit each parameter block in the first parameter block group in a hybrid digital-to-analog transmission manner according to the number of parameter blocks included in the second parameter block group and each of the power amplification factors.

10. An electronic device, characterized in that, The electronic device includes at least one processor, and at least one memory and bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the neural network parameter transmission method based on hybrid digital model as described in any one of claims 1 to 8.

Images