Neural network parameter transmission method based on hybrid digital-analog and related device

Abstract

The invention discloses a neural network parameter transmission method based on hybrid digital-analog and a related device, and the method comprises the steps: carrying out the blocking of a target neural network parameter, and obtaining a first parameter block group; calculating importance indexes of parameter blocks of the first parameter block group; starting from the parameter block with the maximum importance index, calculating the performance loss of hybrid digital-analog transmission in the digital transmission mode so as to determine a second parameter block group of the digital transmission mode; calculating a power amplification factor of the parameter blocks of the third parameter block group in the analog transmission mode; and according to the number of the parameter blocks of the second parameter block group and each power amplification factor, transmitting each parameter block of the first parameter block group in a hybrid digital-analog transmission mode. According to the method, the neural network parameters are subjected to block processing, the performance loss is calculated, the parameters are reasonably distributed to the digital transmission channel and the analog transmission channel in a block mode according to the performance loss, distribution of transmission resources is achieved, and the problems of the cliff effect and the low performance upper limit are effectively solved.

Description

Technical Field

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

[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 tasks such as computer vision, machine reasoning and natural language processing. In order to achieve optimal performance, these edge devices need to select appropriate neural network parameters based on the characteristics of the input image or video, environmental conditions and accuracy requirements. Although some neural network models have generalization capabilities and can adapt to different application scenarios, adjusting neural network parameters according to different input scenarios can still help improve performance and user experience after task completion. However, storing neural network parameters for all applicable scenarios will occupy a large amount of storage resources on the edge device and is not conducive to subsequent upgrades of the model. Therefore, consider transmitting neural network parameters from the server to the edge device.

[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 decode correctly and the received signal will have serious errors. This error makes the received signal unusable, and this phenomenon is also called the "cliff effect". In contrast, analog transmission avoids complete errors in decoded data by using analog modulation schemes, but because the transmitted data is not encoded and compressed, the coding efficiency is low and the transmission performance cannot be guaranteed.

[0004] The introduction of Hybrid Digital-Analog Transmission (HDA) effectively solves the drawbacks of the above traditional solutions, but the existing hybrid digital-analog transmission technology cannot solve the problem of resource allocation. That is, when hybrid digital-analog transmission technology faces the neural network parameter as a source of information, how to reasonably allocate resources to it has become a technical problem that technicians in this field need to solve urgently. Summary of the invention

[0005] In view of the above problems, the present invention provides a neural network parameter transmission method based on hybrid digital-analog and related devices that overcome the above problems or at least partially solve the above problems.

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

[0007] Divide the target neural network parameters into blocks 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 the parameter block with the largest importance index, iteratively calculating the performance loss of the hybrid digital-analog transmission under the digital transmission mode, and determining the second parameter block group using the digital transmission mode according to the performance loss;

[0010] Calculating a power amplification factor of each parameter block of a 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] According to the number of parameter blocks included in the second parameter block group and the power amplification factors, each parameter block in the first parameter block group is transmitted in a hybrid digital-analog transmission manner.

[0012] Optionally, in some optional implementations, the target neural network parameters are divided into blocks to obtain a first parameter block group, including:

[0013] According to the level and channel of each target neural network parameter, the target neural network parameters at 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 includes multiple parameter blocks, and one parameter block includes at least one target neural network parameter.

[0014] Optionally, in some optional implementations, calculating the importance index of each parameter block in the first parameter block group includes:

[0015] 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;

[0016] 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;

[0017] For any parameter block in the first parameter block group, an importance index of the parameter block is calculated based on the standard deviation and average gradient of the parameter block.

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

[0019] 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 a loss function of the neural network model;

[0020] The loss function is differentiated to obtain the average gradient corresponding to the parameter block.

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

[0022] For any parameter block in the first parameter block group, the product of the standard deviation and the average gradient of the parameter block is calculated to obtain an importance index of the parameter block.

[0023] Optionally, in some optional implementations, starting from the parameter block with the largest importance index, sequentially iteratively calculating the performance loss of the hybrid digital-analog transmission under the digital transmission mode, and determining the second parameter block group using the digital transmission mode according to the performance loss, includes:

[0024] Starting from the parameter block with the largest importance index, the parameter blocks are sorted in sequence, wherein the parameter block with the largest importance index is ranked first;

[0025] According to the importance index of the first parameter module, the first performance loss is calculated;

[0026] A second performance loss is calculated according to the importance index of the first parameter module and the importance index of the second parameter module;

[0027] If the second performance loss is greater than the first performance loss, determining the first parameter module as a second parameter block group using a digital transmission method;

[0028] If the second performance loss is not greater than the first performance loss, a third performance loss is calculated according to 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;

[0029] If the third performance loss is greater than the second performance loss, determining that the first parameter module and the second parameter module are together a second parameter block group using a digital transmission method;

[0030] If the third performance loss is not greater than the second performance loss, a fourth performance loss is calculated according to 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;

[0031] This cycle is repeated until the second parameter block group using the digital transmission method is determined.

[0032] Optionally, in some optional implementations, the calculating the power amplification factor of each parameter block of the third parameter block group in the analog transmission mode includes:

[0033] quantizing each parameter block of the second parameter block group to obtain a quantized second parameter block group;

[0034] Performing entropy coding on each parameter block of the second parameter block group after quantization to obtain an average number of bits consumed, wherein one parameter block corresponds to one number of bits;

[0035] For any layer, according to the number of bits of the parameter blocks in the second parameter block group at the layer, calculate the digital bandwidth and digital energy consumed by the digital transmission of the layer;

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

[0037] According to each of the simulated energies, a power amplification factor of each parameter block of the third parameter block group under the simulated transmission mode is calculated.

[0038] Optionally, in some optional implementations, transmitting each parameter block in the first parameter block group in a hybrid digital-analog transmission manner according to the number of parameter blocks included in the second parameter block group and each of the power amplification factors, comprises:

[0039] quantizing, source coding, channel coding and modulating each parameter block of the second parameter block group in the first parameter block group to obtain a digital signal of a corresponding digital transmission channel;

[0040] Power-amplify each parameter block of the third parameter block group in the first parameter block group to obtain an analog signal corresponding to the analog transmission channel;

[0041] Combining the digital signal with the analog signal to obtain actual neural network parameter information in the channel;

[0042] The neural network parameter information and metadata are sent 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 of the power amplification factors.

[0043] In a second aspect, a neural network parameter transmission device based on hybrid digital-analog includes: a parameter blocking unit, an index calculation unit, a digital blocking unit, an analog blocking unit and a hybrid transmission unit;

[0044] The parameter block unit is used to block the target neural network parameters to obtain a first parameter block group;

[0045] 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 a larger importance index has a greater impact on the transmission performance, and one parameter block corresponds to one importance index;

[0046] 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 largest importance index, and determine the second parameter block group using the digital transmission mode according to the performance loss;

[0047] The analog block unit is used to calculate the power amplification factor 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;

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

[0049] In a third aspect, a computer-readable storage medium stores a program, which, when executed by a processor, implements any of the above-mentioned neural network parameter transmission methods based on hybrid digital and analog.

[0050] In a fourth aspect, an electronic device comprises at least one processor, and at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other via the bus; and the processor is used to call program instructions in the memory to execute any of the above-mentioned neural network parameter transmission methods based on hybrid digital and analog.

[0051] By means of the above technical scheme, a neural network parameter transmission method and related devices based on hybrid digital-analog transmission provided by the present invention can block the target neural network parameters to obtain a first parameter block group; calculate the importance index of each parameter block in the first parameter block group, wherein the parameter block with a larger importance index has a greater impact on the transmission performance, and one parameter block corresponds to one importance index; starting from the parameter block with the largest importance index, iteratively calculate the performance loss of hybrid digital-analog transmission under the digital transmission mode in sequence, and determine the second parameter block group using the digital transmission mode based on the performance loss; calculate the power amplification factor 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; according to the number of parameter blocks included in the second parameter block group and each of the power amplification factors, transmit each parameter block in the first parameter block group in a hybrid digital-analog transmission mode. It can be seen from this that the present invention realizes the allocation of transmission resources by processing the neural network parameters in blocks, calculating the performance loss and reasonably allocating the parameter blocks to the digital transmission channel and the analog transmission channel according to the performance loss, thereby effectively solving the cliff effect problem of digital transmission and the low performance upper limit problem of analog transmission.

[0052] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0054] Figure 1 A flow chart showing a first method for transmitting neural network parameters based on hybrid digital-analog circuits provided by the present invention is shown;

[0055] Figure 2 A flow chart showing a second method for transmitting neural network parameters based on hybrid digital-analog networks provided by the present invention is shown;

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

[0057] Figure 4The first simulation experiment data diagram provided by the present invention is shown;

[0058] Figure 5 A second simulation experiment data diagram provided by the present invention is shown;

[0059] Figure 6 A third simulation experiment data diagram provided by the present invention is shown;

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

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

[0062] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0063] like Figure 1 As shown, the present invention provides a neural network parameter transmission method based on hybrid digital-analog, including: S100, S200, S300, S400 and S500;

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

[0065] For example, in some optional embodiments, the S100 includes: step 1.1;

[0066] Step 1.1, according to the level and channel of each target neural network parameter, the target neural network parameters at 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 includes multiple parameter blocks, and one parameter block includes at least one target neural network parameter.

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

[0068] Optionally, the neural network parameters of each layer are divided into blocks according to the different channels of the neural network parameters of each layer, and each block is represented by the symbol In other words, The parameter blocks of each channel of each layer can be characterized, and the present invention does not impose any limitation on this.

[0069] Optionally, a neural network is a model composed of different layers (hierarchies), each layer performs a specific function according to different parameters, thereby realizing layer-by-layer abstraction and processing of input data. For example, the hierarchy of a neural network is divided into three layers: input layer, hidden layer, and output layer. The hidden layer includes four parts: convolution layer, activation layer, pooling layer, and fully connected layer, which is not limited by the present invention.

[0070] S200, calculating the 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;

[0071] Optionally, in some optional embodiments, the S200 includes: step 2.1, step 2.2 and step 2.3;

[0072] Step 2.1: 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;

[0073] Optionally, the standard deviation σ of each block li It can be directly used through the aforementioned w li Calculated,

[0074] The calculation formula is: Where N is the number of neural network parameters in the block group, that is, w li The number of neural network parameters in , All w li The average value of .

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

[0076] Optionally, the present invention can use a mini-batch (a subset of the mini-batch gradient descent algorithm, which can be a parameter block containing multiple neural network parameters as a mini-batch, that is, a parameter block as a mini-batch) as the input of the neural network model to calculate the corresponding loss function; the loss function is derived to obtain the gradient of each neural network parameter included in the corresponding parameter block, and one neural network parameter corresponds to one gradient; for any parameter block, the average gradient g of the parameter block is calculated using the gradient calculation of all neural network parameters in the parameter block li , the average gradient g of the parameter block li =The average value of the gradients corresponding to the neural network parameters of this 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 a loss function of the neural network model;

[0079] Step 3.2: Derivative the loss function to obtain the average gradient corresponding to the parameter block.

[0080] Step 2.3: For any parameter block in the first parameter block group, calculate the importance index of the parameter block based on the standard deviation and 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 of the parameter block and the average gradient to obtain the importance index of the parameter block.

[0083] Optionally, for any parameter block, the product of the standard deviation and the average gradient can be used as an importance index for the parameter block. It should be noted that the importance index reflects the degree of influence of the transmission of the corresponding parameter block on the performance loss. The parameter block with a large influence can consider using digital transmission. Generally speaking, the parameter block with a large importance index has a greater impact on the performance loss after transmission, and digital transmission can be considered to improve its transmission accuracy, and the present invention does not limit this.

[0084] Optionally, after calculating the importance index of each parameter block, the present invention can sort the parameter blocks in order from large to small according to the importance index, so as to facilitate the subsequent determination of the parameter blocks that can be digitally transmitted in order from large to small according to the importance index. The present invention does not impose any restrictions on this.

[0085] S300, starting from the parameter block with the largest importance index, iteratively calculating the performance loss of the hybrid digital-analog transmission under the digital transmission mode, and determining a second parameter block group using the digital transmission mode according to the performance loss;

[0086] Optionally, in some optional implementations, the S300 includes: step 5.1, step 5.2, step 5.3, step 5.4, step 5.5, step 5.6 and step 5.7;

[0087] Step 5.1, starting from the parameter block with the largest importance index, the parameter blocks are sorted in sequence, wherein the parameter block with the largest importance index is ranked first;

[0088] Optionally, as described above, the present invention can sort the parameter blocks in descending order of importance index, so as to facilitate the subsequent determination of parameter blocks that can be digitally transmitted starting from the parameter block with the largest importance index. The present invention does not impose any restrictions on this.

[0089] Step 5.2, according to the importance index of the first parameter module, calculate the first performance loss;

[0090] Optionally, as mentioned above, after the parameter blocks are sorted, the first parameter module is the parameter block with the largest importance index. Since the larger the importance index, the greater the impact on the performance loss after transmission. Therefore, the present invention can start from the parameter block with the largest importance index and obtain the number of parameter blocks for digital transmission through iterative solution. As shown in the following formula: Where k is the number of parameter blocks for a given digital transmission, and k starts from zero to calculate the performance loss of hybrid digital-analog transmission under digital transmission of different blocks, and stops iteration when the loss is greater than the previous one. l represents the quantization step size, It represents the standard deviation of the quantization error after quantization of block i, N o represents the channel noise, E a represents the energy of analog transmission, and M represents the number of blocks transmitted. When the transmission bandwidth is sufficient, the formula M=O is satisfied. At this time, the above formula for calculating the performance loss of mixed digital-analog transmission is simplified to:

[0091] Specifically, the present invention uses iterative calculation to solve the specific value of k, k starts from 0, and calculates D hyb (k,Q l ); when calculating to the k'th (k'>0), when the inequality is satisfied: D hyb (k',Q l )<D hyb (k'-1,Q l ), then continue to iterate to solve the performance loss D of the hybrid digital-analog transmission at k'+1 hyb (k'+1,Q l ); when the inequality is satisfied: D hyb (k'+1,Q l )>D hyb (k',Q l ), the calculation is stopped and k' is taken as the number of parameter blocks for digital transmission.

[0092] Step 5.3, calculating the second performance loss according to the importance index of the first parameter module and the importance index of the second parameter module;

[0093] Optionally, in combination with the above formula, it can be known that the present invention needs to calculate at least two performance losses before making a comparison. Therefore, the present invention needs to iteratively calculate the second performance loss, which is not limited by the present invention.

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

[0095] Step 5.5: If the second performance loss is not greater than the first performance loss, a third performance loss is calculated according to 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, determine that the first parameter module and the second parameter module are together a second parameter block group using a digital transmission method;

[0097] Step 5.7: If the third performance loss is not greater than the second performance loss, 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, and the cycle is repeated until the second parameter block group using the digital transmission method is determined.

[0098] Optionally, in combination with the above formula, it can be seen that the present invention calculates the performance loss through continuous iteration, and then compares the size of the performance loss, and repeats this cycle until the number of parameter blocks for digital transmission is determined, that is, the second parameter block group using digital transmission is determined. The present invention will not go into details about this, and please refer to the above formula for details.

[0099] S400, calculating the power amplification factor of each parameter block of the third parameter block group in 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 includes: step 6.1, step 6.2, step 6.3, step 6.4 and step 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, performing entropy coding on each parameter block of the quantized second parameter block group to obtain an average number of bits consumed, wherein one parameter block corresponds to one bit number;

[0103] Step 6.3: For any layer, according to the number of bits of the parameter blocks at the layer in the second parameter block group, calculate the digital bandwidth and digital energy consumed by the digital transmission of the layer;

[0104] Optionally, in general, the distribution of neural network parameters approximately satisfies the Gaussian distribution w li ~N(0,λ li ), uniform quantization is used to compress the parameters, where the quantization step is Q l , the quantized parameter blocks The expression is: in, Indicates the floor symbol, λ li Represents the variance of the block.

[0105] Optional, quantized parameters It 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 coding is:

[0107] The total bandwidth (digital bandwidth) and total energy consumption (digital energy) of the digital transmission of layer l are: Among them, r is the code rate after channel coding and modulation in digital transmission, P s It is the minimum power required to correctly transmit a symbol in digital transmission. is the total bandwidth for digital transmission at layer l, is the total energy consumption of digital transmission in layer l.

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

[0109] Step 6.5: Calculate the power amplification factor of each parameter block of the third parameter block group under the analog transmission mode according to each of the simulated energies.

[0110] Optionally, combined with the above process of calculating the total bandwidth and total energy consumption of digital transmission, the total bandwidth (analog bandwidth) and total energy consumption (analog energy) of analog transmission of layer l are: Among them, B l is the preset value (representing the total bandwidth of the mixed digital-analog transmission at layer l). l is the preset value (representing the total energy consumption of mixed digital-analog transmission at layer l), is the total bandwidth of the analog transmission at layer l, It is the total energy consumption of simulated transmission in layer l.

[0111] The power amplification factor of each parameter block in the analog transmission is calculated by the following formula: Among them, E a is the total energy consumed for simulated transmission at all levels, O represents the output dimension of the level (the output dimension is also called the channel), K represents the size of the convolution kernel in the convolutional neural network, l is the level (number) of the neural network model where the corresponding neural network parameters are located, i and j are the input channels (numbers) of the level of the neural network model where the corresponding neural network parameters are located, and the number j is applicable to the sum calculation in the square root, which is distinguished from the number i outside the sum calculation.

[0112] S500: Transmit each parameter block in the first parameter block group in a hybrid digital-analog transmission manner according to the number of parameter blocks included in the second parameter block group and each of the power amplification factors.

[0113] For example, in some optional embodiments, the S500 includes: step 7.1, step 7.2, step 7.3 and step 7.4;

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

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

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

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

[0118] Optionally, the transmitter sends the symbols and metadata encoded by the digital transmission and the analog transmission respectively, and the receiver reconstructs the neural network parameters through the received symbols and metadata, specifically:

[0119] The data transmitted by the transmitter is w li After digital transmission and analog transmission encoding, it becomes the signal actually transmitted in the channel. The metadata includes the number of blocks k for digital transmission and the power amplification factor s. li Among them, the process of digital transmission includes quantization, source coding, channel coding and modulation, while the process of analog transmission includes power amplification.

[0120] The receiving end receives the data transmitted by the digital transmission channel through demodulation, channel decoding, source decoding, and inverse quantization, while the analog transmission data is reconstructed using the following formula: ,in 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 digital transmission channel are combined to obtain the actual neural network parameter information.

[0121] Optionally, in combination with the foregoing, the present invention models the reconstruction loss of the transmission neural network parameters by processing the neural network parameters in blocks, and uses Taylor expansion to approximate the performance loss. In order to reasonably allocate the transmission parameters to the digital transmission channel and the analog transmission channel, the present invention proposes an importance index based on the performance loss, which is used to sort the parameters in blocks, determine the number of blocks for digital transmission and analog transmission, and realize the allocation of transmission resources.

[0122] Optionally, in order to more clearly describe the overall implementation process of the present invention, the present invention provides the following Figure 2Please refer to the above explanation for the overall flow chart shown, which will not be elaborated in this article.

[0123] The simulation experiment data of the present invention are provided below, which confirms the effectiveness of the present invention, and can effectively solve the problem of the cliff effect of digital transmission and the problem of low performance upper limit of analog transmission.

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

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

[0126] From the attached Figure 4 It can be observed that when the SNR is 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 under the hybrid digital-analog transmission designed by the present invention is similar to that of digital transmission, and the performance error is only less than 0.1% compared to the model performance under lossless transmission. The classification accuracy of the reconstructed neural network performance under analog transmission is 0.5% lower than that of lossless transmission under high signal-to-noise ratio. Compared with the existing FederatedAirNet, the performance loss of the neural network after transmission is smaller.

[0127] When SNR is less than 8dB, the performance of the reconstruction models of several transmission schemes has declined. The performance degradation is more obvious in the reconstruction neural network under digital transmission. After SNR≦6dB, the performance of its reconstruction model drops off like a cliff, which is a typical cliff effect in digital transmission. When SNR=6dB, the classification accuracy of the reconstruction model performance under digital transmission is only 30%, the data bit error rate after channel decoding is high, and the source decoding end cannot decode correctly. In this case, the reconstructed model basically cannot work normally. The performance degradation of analog transmission and hybrid digital-analog transmission is relatively gentle, and there is no cliff-like drop compared to digital transmission. The performance of the neural network after these two transmissions decreases slowly as the signal-to-noise ratio decreases. Compared with analog transmission and existing hybrid digital-analog transmission, the hybrid digital-analog transmission proposed in the present invention has a 5% performance improvement when SNR=5dB, and an 8% performance improvement when SNR=0dB. Since the hybrid digital-analog transmission scheme proposed in the present invention performs reasonable power optimization according to the modeled performance loss in analog coded transmission, it has better performance under the same signal-to-noise ratio than traditional analog transmission and FederatedAirNet using a single power allocation method.

[0128] In order to verify the applicability of hybrid digital-analog transmission in the transmission of other model parameters, the present invention also simulates and tests other neural networks including AlexNet and ResNet50 as well as super-resolution neural networks EDSR and ESRGAN. Figure 5 It can be seen from the experimental results shown that the hybrid digital-analog transmission scheme proposed in the present invention is also applicable to the transmission of other convolutional neural network parameters. For classification tasks, the performance of the AlexNet and ResNet50 models after hybrid digital-analog transmission reconstruction is excellent. When SNR>8dB, the classification accuracy is close to lossless transmission. When SNR<8dB, a cliff effect occurs in digital transmission, and the performance of the hybrid transmission scheme proposed in the present invention has obvious advantages over other transmission methods. For super-resolution models, the performance after hybrid digital-analog transmission is less than 0.2dB lower than that of digital transmission and lossless transmission when SNR>15dB. At the same time, the performance loss is smaller than other transmission methods at low signal-to-noise ratios.

[0129] Since the present invention simulates and tests the transmission performance under different schemes without bandwidth restrictions, the present invention analyzes the bandwidth requirements of various transmission schemes. Figure 6Statistics of the number of symbols transmitted for different transmission modes when transmitting models. The number of symbols actually required to be transmitted is used here to represent the bandwidth requirement for fully transmitting all parameters of a model at the same delay. Since there is no restriction on the transmission bandwidth of the transmission model, it can be seen that hybrid digital-analog transmission and digital transmission require more transmission bandwidth than analog transmission. However, even when both hybrid digital-analog transmission and digital transmission are higher than the analog transmission bandwidth, hybrid digital-analog transmission only requires a smaller bandwidth increment to achieve performance better than analog transmission and avoid the cliff effect in digital transmission.

[0130] like Figure 7 As shown, the present invention provides a neural network parameter transmission device based on hybrid digital-analog, comprising: 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 block unit 100 is used to block the target neural network parameters to obtain a first parameter block group;

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

[0133] The digital block unit 300 is used to iteratively calculate the performance loss of the hybrid digital-analog transmission under the digital transmission mode starting from the parameter block with the largest importance index, and determine the second parameter block group using the digital transmission mode according to the performance loss;

[0134] The analog block unit 400 is used to calculate the power amplification factor of each parameter block of the third parameter block group in 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;

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

[0136] Optionally, in some optional implementations, the parameter blocking unit 100 includes: a parameter blocking subunit;

[0137] The parameter block subunit is used to divide the target neural network parameters at the same level and the same channel into the same parameter block according to the level and channel where each target neural network parameter is located, so as to obtain the first parameter block group, wherein the first parameter block group includes multiple parameter blocks, and one parameter block includes at least one target neural network parameter.

[0138] Optionally, in some optional implementations, 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 used to calculate the 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 used to calculate 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;

[0141] The importance index calculation subunit is used to calculate 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.

[0142] Optionally, in certain 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 used 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 used to derive the loss function to obtain the average gradient corresponding to the parameter block.

[0145] Optionally, in certain optional implementations, the importance index calculation subunit includes: a product calculation subunit;

[0146] The product calculation subunit is used to calculate the 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 includes: 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 used to sort the parameter blocks in sequence starting 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 used to calculate a first performance loss according to the importance index of the first parameter module;

[0150] The second loss calculation subunit is used to calculate the second performance loss according to the importance index of the first parameter module and the importance index of the second parameter module;

[0151] The first result subunit is used to determine the first bit parameter module as a second parameter block group using a digital transmission method 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 according to 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 used to determine that the first parameter module and the second parameter module are together a second parameter block group using a digital transmission method if the third performance loss is greater than the second performance loss;

[0154] The fourth result sub-unit is used to calculate the fourth performance loss according to 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 repeat this cycle until the second parameter block group using the digital transmission method is determined.

[0155] Optionally, in some optional implementations, the analog block unit 400 includes: 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 used 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 used to perform entropy coding on each parameter block of the quantized second parameter block group to obtain an average number of bits consumed, wherein one parameter block corresponds to one bit number;

[0158] The digital transmission calculation subunit is used to calculate, for any layer, the digital bandwidth and digital energy consumed by the digital transmission of the layer according to the number of bits of the parameter blocks in the second parameter block group at the layer;

[0159] The analog transmission calculation subunit is used to calculate the analog bandwidth and analog energy consumed by the analog transmission of any level according to the digital bandwidth, digital energy, preset total bandwidth and 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 used to calculate the power amplification factor of each parameter block of the third parameter block group under the analog transmission mode according to each simulated energy.

[0161] Optionally, in some optional implementations, the hybrid transmission unit 500 includes: a digital signal subunit, an analog signal subunit, a signal merging subunit, and a signal sending subunit;

[0162] The digital signal subunit is used to quantize, source code, channel code and modulate each parameter block of the second parameter block group in the first parameter block group to obtain a digital signal corresponding to the digital transmission channel;

[0163] The analog signal subunit is used to power-amplify each parameter block of the third parameter block group in the first parameter block group to obtain an analog signal corresponding to the analog transmission channel;

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

[0165] The signal sending subunit is used to send the neural network parameter information and metadata 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.

[0166] The present invention provides a computer-readable storage medium having a program stored thereon, and when the program is executed by a processor, the method for transmitting neural network parameters based on hybrid digital-analog networks described in any one of the above is implemented.

[0167] like Figure 8As shown, the present invention provides an electronic device 70, which includes at least one processor 701, and at least one memory 702 and a bus 703 connected to the processor 701; wherein the processor 701 and the memory 702 communicate with each other through the bus 703; the processor 701 is used to call the program instructions in the memory 702 to execute any of the above-mentioned neural network parameter transmission methods based on hybrid digital and analog.

[0168] In the present invention, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0169] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

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

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

Claims

1. A neural network parameter transmission method based on hybrid digital-analog, characterized in that: include: Divide the target neural network parameters into blocks to obtain a first parameter block group; 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; Starting from the parameter block with the largest importance index, iteratively calculating the performance loss of the hybrid digital-analog transmission under the digital transmission mode, and determining the second parameter block group using the digital transmission mode according to the performance loss; Calculating a power amplification factor of each parameter block of a 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; According to the number of parameter blocks included in the second parameter block group and the power amplification factors, each parameter block in the first parameter block group is transmitted in a hybrid digital-analog transmission manner.

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

3. The method according to claim 1, characterized in that: The calculating 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, an 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, for any parameter block in the first parameter block group, an average gradient of each target neural network parameter included in the parameter block comprises: 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 a loss function of the neural network model; The loss function is differentiated to obtain the average gradient corresponding to the parameter block.

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

6. The method according to claim 1, characterized in that The method starts from the parameter block with the largest importance index, iteratively calculates the performance loss of the hybrid digital-analog transmission under the digital transmission mode, and determines the second parameter block group using the digital transmission mode according to the performance loss, including: Starting from the parameter block with the largest importance index, the parameter blocks are sorted in sequence, wherein the parameter block with the largest importance index is ranked first; According to the importance index of the first parameter module, the first performance loss is calculated; A second performance loss is calculated according to 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, determining the first parameter module as a second parameter block group using a digital transmission method; If the second performance loss is not greater than the first performance loss, a third performance loss is calculated according to 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, determining that the first parameter module and the second parameter module are together a second parameter block group using a digital transmission method; If the third performance loss is not greater than the second performance loss, a fourth performance loss is calculated according to 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 cycle is repeated until the second parameter block group using the digital transmission method is determined.

7. The method according to claim 1, characterized in that The calculating of the power amplification factor of each parameter block of the third parameter block group in the analog transmission mode includes: quantizing each parameter block of the second parameter block group to obtain a quantized second parameter block group; Performing entropy coding on each parameter block of the quantized second parameter block group to obtain an average number of bits consumed, wherein one parameter block corresponds to one number of bits; For any layer, according to the number of bits of the parameter blocks in the second parameter block group at the layer, calculate the digital bandwidth and digital energy consumed by the digital transmission of the layer; For any level, according to the digital bandwidth, digital energy, preset total bandwidth and preset total energy of the level, the analog bandwidth and analog energy consumed by the level for analog transmission are calculated, wherein one level corresponds to one preset total bandwidth and one preset total energy; According to each of the simulated energies, a power amplification factor of each parameter block of the third parameter block group under the simulated transmission mode is calculated.

8. The method according to claim 1, characterized in that The transmitting each parameter block in the first parameter block group in a hybrid digital-analog transmission manner according to the number of parameter blocks included in the second parameter block group and each of the power amplification factors, comprises: quantizing, source coding, channel coding and modulating each parameter block of the second parameter block group in the first parameter block group to obtain a digital signal of a corresponding digital transmission channel; Power-amplify each parameter block of the third parameter block group in the first parameter block group to obtain an analog signal of a corresponding analog transmission channel; Combining the digital signal with the analog signal to obtain actual neural network parameter information in the channel; The neural network parameter information and metadata are sent 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 of the power amplification factors.

9. A neural network parameter transmission device based on hybrid digital-analog, characterized in that: include: Parameter block unit, index calculation unit, digital block unit, analog block unit and hybrid transmission unit; The parameter block unit is used to block the target neural network parameters to obtain a first parameter block 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 a larger 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 largest importance index, and determine the second parameter block group using the digital transmission mode according to the performance loss; The analog block unit is used to calculate the power amplification factor 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; The hybrid transmission unit is used to transmit each parameter block in the first parameter block group in a hybrid digital-analog transmission manner according to the number of parameter blocks included in the second parameter block group and each power amplification factor.

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

Images