A semantic communication method and system based on LSTM network automatic training

Abstract

The application discloses a semantic communication method and system based on LSTM network automatic training, constructs an LSTM network as a main body framework of a meta-optimizer, and carries out activation on the input of the meta-optimizer through two different weight matrices and a Relu function; the processed input enters forgetting gates, input gates and output gates of the LSTM network respectively, and the output of the LSTM network is obtained; the obtained output is processed through a weight matrix and a limiting factor, and a predicted learning rate value is obtained; a semantic communication main task optimizes parameters of a semantic encoder-decoder by using the predicted learning rate of the meta-optimizer, and completes semantic communication. The application can automatically train the semantic encoder-decoder and make it meet convergence requirements, solve technical problems that cannot be universally trained and actually deployed on a large scale due to heterogeneous factors such as semantic communication tasks, data types, data modalities and encoder-decoder networks, and improve the generalization ability of the application to different semantic communication tasks.

Description

Technical Field

[0001] This invention belongs to the fields of wireless communication technology and intelligent communication technology, specifically relating to a semantic communication method and system based on automated training of LSTM networks. Background Technology

[0002] Semantic communication, as a new communication paradigm, mainly utilizes artificial intelligence technology to extract, transmit, and reconstruct the semantics of transmitted information. It can not only restore the original semantics of information under low signal-to-noise ratio conditions, but also greatly compress the amount of data to solve the problem of data expansion in the future 6G.

[0003] According to the specific process of semantic communication, the sender first performs semantic encoding and channel coding on the information. Semantic encoding is used to extract the semantic meaning contained within the information, while channel coding is used to compress the data volume and mitigate the impact of poor channel conditions on the transmitted information. Specifically, it is represented as follows:

[0004] (1)

[0005] in, The parameter is Channel encoder, The parameter is semantic encoder, It represents the information to be transmitted. The encoded information contains sufficient semantics and has a certain degree of robustness to noise.

[0006] The encoded information will then be transmitted via a wireless channel, as shown below:

[0007] (2)

[0008] in, For channel coefficients, It is additive white Gaussian noise (AWGN), meaning the receiver receives information that is contaminated by noise.

[0009] Finally, the receiving end extracts and reconstructs the semantics of the received information using a channel decoder and a semantic decoder, as shown below:

[0010] (3)

[0011] in, The parameter is semantic decoder, The parameter is Channel decoder, This represents the decoded information.

[0012] In summary, a codec for a semantic communication system contains four sets of parameters: These parameters are obtained through training with a shared knowledge background between the sender and receiver. Traditional training processes not only require setting specific network structures for the codec based on the semantic communication task, but also require comprehensive consideration of various factors such as data type (e.g., text, image, audio and video), data features (e.g., text format, image resolution, video clarity), semantic communication task (e.g., image transmission, image classification), and codec network structure to set reasonable optimizers and learning rate strategies for training.

[0013] However, as the number of participants in semantic communication increases, each pair of users has differences in semantic information data characteristics, semantic information data modalities, semantic communication main tasks, and codec architecture. Setting up optimizers and learning rate strategies for each pair of users separately would consume a lot of training resources and manpower, making it difficult to deploy semantic communication systems on a large scale.

[0014] Meanwhile, existing research often focuses on end-to-end semantic communication architecture design, which is not conducive to the large-scale application of semantic communication devices, and cannot meet the future vision of "Internet of Everything" and "Intelligent Internet of Everything". Summary of the Invention

[0015] The technical problem to be solved by the present invention is to provide a semantic communication method and system based on automated training of LSTM network to address the shortcomings of the prior art. This method and system solve the problems of the current semantic communication system, which consumes a lot of training resources and manpower, as well as the inability to universally train and deploy on a large scale due to the heterogeneity of semantic communication tasks, data types, data modalities, codec networks and other factors.

[0016] The present invention adopts the following technical solution:

[0017] A semantic communication method based on automated training of LSTM networks includes the following steps:

[0018] S1. Construct an LSTM network as the main architecture of the meta-optimizer. The input of the meta-optimizer is activated by the ReLU function after passing through two different weight matrices.

[0019] S2. The input processed in step S1 enters the forget gate, input gate, and output gate of the LSTM network respectively to obtain the output of the LSTM network.

[0020] S3. After processing the output obtained in step S2 with the weight matrix and the constraint factor, the predicted learning rate value is obtained.

[0021] S4. The main task of semantic communication is to optimize the parameters of the semantic codec using the learning rate predicted by the meta-optimizer in step S3, thereby completing semantic communication.

[0022] Preferably, in step S1, the input to the meta-optimizer is the training loss value of the semantic communication task and the historical information from the previous stage. The training process of the semantic communication task is as follows:

[0023] S101. Initialize the encoder / decoder parameters of the semantic communication system, the weight matrix parameters of the meta-optimizer, the initial learning rate, the cell parameters of the main LSTM of the meta-optimizer, and set the number of meta-learning rounds T.

[0024] S102. Randomly divide the training knowledge background of semantic communication into main task knowledge background and meta-learning knowledge background according to a set ratio.

[0025] S103. The input to the semantic communication task is randomly sampled from the knowledge background of the main task and preprocessed.

[0026] S104. The data preprocessed in step S103 is used as input to sequentially enter the semantic encoder, channel encoder, wireless channel, channel decoder and semantic decoder to obtain the output of the semantic communication main task.

[0027] S105. After calculating the training loss using the input and output, perform gradient descent according to the initial learning rate to optimize the encoder / decoder parameters. ;

[0028] S106. After every T rounds of training, the training loss value of the current round is input into the meta-optimizer. The meta-optimizer internally uses the meta-learning knowledge background to perform the calculations required in Requirement 1, predicts and records the learning rate that should be adopted for the current semantic communication main task, and updates the parameters of the meta-optimizer. ;

[0029] S107. Update the learning rate of the main task to the learning rate predicted by the meta-optimizer, and repeat steps S102-S107 until the training loss in step S105 no longer decreases or reaches the total number of training rounds.

[0030] Preferably, the parameters of the meta-optimizer are updated. for:

[0031]

[0032] in, Represents the Adam optimizer; LR for the Adam optimizer; It is the predicted value of the semantic communication codec parameters for the next time step, calculated using the learning rate estimated by the currently unupdated meta-optimizer.

[0033] Preferably, the predicted values ​​of the semantic communication codec parameters for the next time step are calculated using the learning rate estimated by the currently unupdated meta-optimizer. for:

[0034]

[0035] in, This represents the input-output relationship of the meta-optimizer. The loss value of the input meta-optimizer, The cell state of the meta-optimizer (containing historical training information). This provides parameter information for the meta-optimizer. The Loss value represents the gradient descent information about parameter w on the semantic communication main task network.

[0036] Preferably, codec parameters for:

[0037]

[0038] in, This represents the input-output relationship of the meta-optimizer. The loss value of the input meta-optimizer, The cell state of the meta-optimizer (containing historical training information). This provides parameter information for the meta-optimizer at the next time step. The Loss value represents the gradient descent information about parameter w on the semantic communication main task network.

[0039] Preferably, in step S3, the predicted learning rate value is obtained as follows:

[0040] The two inputs pass through a fully connected layer with a ReLU activation function. Perform preprocessing, then import into LSTM and obtain the output. ;use Fully connected layer with activation function right Perform a linear transformation to obtain the predicted value. Finally, multiply the predicted value by the scaling factor. Ensure that the final predicted learning rate value is within Interval.

[0041] Preferably, in step S4, an online approximation technique is used to jointly update the meta-optimizer parameters. Semantic communication codec parameters It will be updated Several steps ( (Updated later) Once; and the overall knowledge background of semantic communication is divided into the knowledge background of the main semantic communication task and the knowledge background of meta-learning. and The parameters are trained on their respective knowledge backgrounds.

[0042] Preferably, in step S4, during the semantic communication main task and meta-learning task stages, the channel adaptive training method is as follows:

[0043] Suppose that the semantic communication task operates in a real-world scenario with a signal-to-noise ratio (SNR) ranging from mdB to ndB, where n is greater than or equal to m. Set the channel SNR of the main semantic communication task to mdB and train it using the main task's knowledge background. In the meta-learning phase, adjust the channel SNR to ndB and train it using the meta-learning knowledge background. The main task's knowledge background will be much larger than the meta-learning knowledge background, and the meta-learning phase will only be performed once after multiple rounds of training.

[0044] Preferably, the parameters of the semantic codec for:

[0045]

[0046]

[0047] in, The codec parameters will be updated in the next time step. To validate the loss value within the context of existing knowledge, This represents the input-output relationship of the meta-optimizer. This represents the input-output relationship of the gradient descent function. This is based on the total number of training rounds.

[0048] Secondly, embodiments of the present invention provide a semantic communication system based on automated training of an LSTM network, comprising:

[0049] The module constructs an LSTM network as the main architecture of the meta-optimizer. The input of the meta-optimizer is processed by two different weight matrices and then activated by the ReLU function.

[0050] The network module, after being processed by the construction module, enters the forget gate, input gate, and output gate of the LSTM network to obtain the output of the LSTM network;

[0051] The prediction module will process the output with weight matrix and constraint factor to obtain the predicted learning rate value;

[0052] The optimization module, the main task of semantic communication, will use the learning rate predicted by the meta-optimizer to optimize the parameters of the semantic codec and complete the semantic communication.

[0053] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the semantic communication method based on LSTM network automated training described above.

[0054] Fourthly, embodiments of the present invention provide a computer-readable storage medium including a computer program, which, when executed by a processor, implements the steps of the semantic communication method based on LSTM network automated training described above.

[0055] Compared with the prior art, the present invention has at least the following beneficial effects:

[0056] A semantic communication method based on automated training of LSTM network is proposed. It makes full use of the unidirectional sequence memory capability of LSTM network to calculate and weightedly fuse historical training information with current training information within the network to cope with the dynamic changes in the training process. In addition, LSTM network, as a meta-optimizer component, has the characteristics of lightweight and plug-and-play, and is isolated from the main semantic communication task, which meets the design concept of high cohesion and low coupling.

[0057] Furthermore, the knowledge background of semantic communication is divided into training knowledge background and validation knowledge background. The specific division ratio is related to channel quality conditions, the size of the knowledge background, and the semantic communication task. This method adopts a 9:1 division ratio. In addition, this method adopts a meta-learning scheme to update the meta-optimizer parameters every T rounds. The setting of the T parameter is related to training cost, training performance, and the size of the knowledge background. This method sets T to the number of iterations required to train a complete knowledge background in one round, in order to balance training cost and training performance.

[0058] Furthermore, the fully connected layer W1 can concatenate the loss value and historical information and perform dimensional transformation to meet the input dimensionality requirements of the LSTM network. The LSTM network fully utilizes its ability to process unidirectional sequence information, combining the current training information and historical training information by assigning weights to generate the output buffer value. The fully connected layer W3 performs dimensional transformation on the output buffer to meet the output format requirements of a single learning rate floating-point value. The y-factor is selectively set; if prior training knowledge exists, it can be set to ensure that the final predicted learning rate value is within the specified range. .

[0059] Furthermore, online approximation techniques can reduce communication latency between the semantic communication system process and the meta-optimizer training process. The loss value of the semantic communication system process is directly transmitted to the meta-optimizer training process via shared memory. Compared to traditional offline techniques, this method reduces the time and space costs of reading from and writing to the cache. In addition, setting the meta-learning rounds and knowledge background partitioning strategies helps reduce training resource consumption and meets the adaptive signal-to-noise ratio requirement described in claim 8.

[0060] Furthermore, the study found that training in high SNR regions converges more easily, while training in low SNR regions is more difficult to converge due to greater noise. Therefore, this method sets the channel SNR of the semantic communication main task to mdB to ensure the robustness of the semantic communication system in low SNR regions, and adopts the main task knowledge background for training. In the meta-learning stage, the channel SNR is adjusted to ndB and the meta-learning knowledge background is adopted, so that when the semantic communication main task adopts the learning rate predicted by ASCNet, the codec parameters are more biased towards high SNR regions. This method enables the codec to have high performance across the entire SNR range rather than tending towards a certain SNR value. This also explains the division strategy of the main task knowledge background and the meta-learning knowledge background (i.e., the former is much larger than the latter), and the meta-learning stage is only performed once after multiple rounds of training. This will make full use of training resources, reduce training costs, and meet the essential concept of meta-learning.

[0061] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0062] In summary, this invention optimizes the training of semantic communication codecs, enabling automatic training and convergence without manual parameter tuning. Furthermore, the meta-optimizer utilizes knowledge context for parameter updates, addressing the technical challenges of universal training and large-scale deployment caused by heterogeneous factors such as semantic communication tasks, data types, data modalities, and codec networks, thereby enhancing its generalization ability across different semantic communication tasks.

[0063] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the following description of the relative embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 Automated training scenarios for semantic communication that enable meta-learning optimizers;

[0066] Figure 2 This is a simulation diagram showing how the loss changes with epoch in this invention;

[0067] Figure 3 This is a simulation diagram of the performance of the present invention at a signal-to-noise ratio of 0dB;

[0068] Figure 4 This is a simulation diagram of the performance of the present invention at a signal-to-noise ratio of 6dB;

[0069] Figure 5 This is a simulation diagram showing the predicted LR variation with epoch according to the present invention.

[0070] Figure 6 A schematic diagram of a computer device provided in an embodiment of the present invention;

[0071] Figure 7 This is a block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0074] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0075] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0076] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0077] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0078] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0079] This invention provides a semantic communication method based on automated training of LSTM networks, which aims to optimize the training of semantic communication codecs so that they can ignore various differences, automatically train and optimize to meet accuracy requirements, and improve their generalization ability for heterogeneous tasks, so as to ultimately meet the automated training of multiple semantic communication devices.

[0080] Technical issues and comparison algorithms:

[0081] The problem of training a semantic communication codec can be formulated as the following nonconvex optimization problem:

[0082] (4)

[0083] in, This represents the training loss function Loss in the context of shared knowledge. Indicate the knowledge background of those participating in the training; Indicates the parameters of the codec to be optimized, such as Specific parameters in; This indicates the sampling of data in each round of training.

[0084] Commonly used optimizers for semantic communication include stochastic gradient descent (SGD) and its variants, such as Momentum, Adagrad, Adadelta, RMSprop, and Adam. Generally, these algorithms are represented as:

[0085] (5)

[0086] in, For updated codec parameters, express exist gradient at, Represents historical gradient information, The optimizer in the current iteration Hyperparameters, such as the learning rate.

[0087] For the sake of simplicity, this invention will be described in the form of vanilla SGD, that is:

[0088] (6)

[0089] The following are commonly used learning rate update strategies in the training of current semantic communication codecs:

[0090] Fixed strategy: ;

[0091] Multi-step descent strategy: , Given;

[0092] Index decline strategy: ;

[0093] SGDR strategy: .

[0094] in, This represents the initial learning rate. This represents the learning rate at the t-th iteration. Specify the learning rate setting range for the SGDR strategy. Indicates how many rounds were executed. Indicates in After that, SGDR restarted. These represent the decay factors for the multi-step descent strategy and the exponential descent strategy, respectively.

[0095] This invention discloses a semantic communication method based on automated training of LSTM networks. It introduces an LSTM neural network-assisted meta-learning method, parameterizing the learning rate strategy through the explicit mapping effect of LSTM and naming it ASCNet (Auto Semantic Communication Net). Subsequent descriptions will use ASCNet to represent the meta-optimizer, LR to represent the learning rate, and Loss to represent the loss function value. The method includes the following steps:

[0096] First, the functional expression of ASCNet is as follows: ;

[0097] The process of learning LR using ASCNet is represented as follows:

[0098] (8)

[0099] in, , , The output is the LR calculated in the t-th iteration. These are the parameters of ASCNet (such as the weight matrix and biases). It is the loss of the sample in the t-th iteration. These represent the output and state of the LSTM cell at the t-th iteration, respectively. This represents the dimension of the state vector.

[0100] In each iteration of stochastic gradient descent It can learn an explicit Loss-LR dependency, allowing the network to adapt to the current input. and historical training information stored in the network. Adaptive prediction LR.

[0101] For each iteration step, the entire calculation process is represented as follows:

[0102]

[0103]

[0104]

[0105]

[0106] (9)

[0107] in, , , These are the input gate, forget gate, and output gate of the LSTM, respectively. ReLU and tanh are three activation functions; the input to ASCNet is historical information. and current training loss ;Parameters of ASCNet , .

[0108] The input-output calculation process of ASCNet can be understood as follows:

[0109] S1. The two inputs pass through a fully connected layer with a ReLU activation function. Perform preprocessing, then import into LSTM and obtain the output. ;

[0110] S2, Utilization Fully connected layer with activation function right Perform a linear transformation to obtain the predicted value. Finally, multiply the predicted value by the scaling factor. Guarantee that the final predicted LR is located in Interval.

[0111] ASCNet, with LSTM as its backbone network, can handle long-term information dependency problems well, and therefore can learn LR strategies that are adapted to the current knowledge background and semantic communication main task.

[0112] On the one hand, unlike directly learning the LR strategy itself, this technique uses parameters as follows: The ASCNet is used to learn the LR policy. This parameterized meta-learner helps to extract the underlying methods, i.e., how to design appropriate LR policies for the training problem of semantic communication codecs, rather than just a fixed mathematical formula. On the other hand, ASCNet is very lightweight. It only uses the loss to learn the LR policy and only trains once after multiple rounds of training on the main semantic communication task, without consuming too many training resources, and is relatively easier to train and better controlled.

[0113] Design of ASCNet step-by-step meta-learning algorithm and channel adaptation algorithm:

[0114] This study performs meta-training on ASCNet to improve its generalization performance. The optimization problem can be expressed as finding the semantic communication encoder-decoder parameters that minimize the loss on the meta-learned knowledge background Val. ,in Generated by ASCNet prediction, the specific formula is as follows:

[0115] (10)

[0116]

[0117] in, Represents the updated codec parameters at the next moment. Depends on current codec parameters Historical information Current Loss and ASCNet parameters .

[0118] This invention uses an online approximation technique to jointly update ASCNet parameters. Semantic communication codec parameters This means learning and updating the LR policy synchronously during model training. The training of both involves the following two main issues:

[0119] Q1. Because semantic communication knowledge backgrounds often involve large amounts of data and complex encoder-decoder structures, if... Performing incremental optimizations will consume a lot of computing power and increase the latency of semantic communication systems.

[0120] Regarding Q1, this invention will be updated Several steps ( (Updated later) Once. Furthermore, the overall knowledge background of semantic communication is divided into the knowledge background of the main semantic communication task and the knowledge background of meta-learning. and The parameters are trained separately within their respective knowledge contexts to isolate them and improve performance.

[0121] Q2. As shown in equation (2), the channel state of semantic communication Closely related to codec training, even small differences in channel states can lead to significant differences in codec parameters. Codecs trained in low signal-to-noise ratio (SNR) ranges perform poorly in high SNR ranges, and vice versa. Traditional semantic communication systems often incorporate significant network redundancy or randomly select SNR for optimization in each round to address this issue. This increases network load and randomness, making it unsuitable for practical deployment of semantic communication devices.

[0122] Regarding Q2, this invention proposes a channel-adaptive training method based on the two-stage training of the semantic communication main task and the meta-learning task. Specifically:

[0123] 1. Suppose that the semantic communication task is applied in a real-world scenario with a signal-to-noise ratio range of mdB to ndB, where n is greater than or equal to m.

[0124] 2. Set the channel signal-to-noise ratio of the semantic communication main task to mdB to meet the robustness of the semantic communication system in the low signal-to-noise ratio region, and use the knowledge background of the main task for training.

[0125] 3. In the meta-learning stage, the channel signal-to-noise ratio is adjusted to ndB and meta-learning knowledge background is adopted. This makes the codec parameters more biased towards the high signal-to-noise ratio region when the semantic communication main task adopts the learning rate predicted by ASCNet. This ensures that the codec has high performance throughout the entire signal-to-noise ratio range rather than tending towards a certain signal-to-noise ratio value.

[0126] 4. Since training in high signal-to-noise ratio regions is easier to converge, while training in low signal-to-noise ratio regions is more affected by noise and is more difficult to converge, the knowledge background of the main task will be much larger than the knowledge background of meta-learning. The meta-learning stage will only be carried out once after multiple rounds of training. This will make full use of training resources, reduce training costs, and meet the essential concept of meta-learning.

[0127] The ASCNet step-by-step meta-learning method is represented as follows:

[0128] S101. Initialize the encoder / decoder parameters of the semantic communication system, the weight matrix parameters of the meta-optimizer, the initial learning rate, the cell parameters of the main LSTM of the meta-optimizer, and set the number of meta-learning rounds T.

[0129] S102. The training knowledge background of semantic communication is randomly divided into the main task knowledge background and the meta-learning knowledge background according to a certain proportion.

[0130] S103. The input to the semantic communication task is randomly sampled from the knowledge background of the main task and preprocessed.

[0131] S104. The preprocessed data is sequentially fed into the semantic encoder, channel encoder, wireless channel (low signal-to-noise ratio threshold), channel decoder, and semantic decoder to obtain the output of the semantic communication main task.

[0132] S105. After calculating the training loss using the input and output, perform gradient descent according to the initial learning rate to optimize the encoder / decoder parameters. ;

[0133] S106. After every T rounds of training, the training loss value of the current round is input into the meta-optimizer ASCNet. The meta-optimizer internally uses the meta-learning knowledge background (high signal-to-noise ratio threshold) to calculate and record the learning rate that should be adopted for the current semantic communication main task. Finally, the parameters of ASCNet are updated. ;

[0134] S107. The learning rate of the main task will be updated to the learning rate predicted by the meta-optimizer. Repeat steps S102 to S107 until the training loss in step S105 no longer decreases or reaches the total number of training rounds.

[0135] It will be updated if the requirements are met; otherwise, it will remain unchanged. The update will use the semantic codec parameters from the previous iteration. and current ASCNet parameters The loss of the meta-learning knowledge background is minimized using formula (10), and the specific process can be expressed as follows:

[0136] (11)

[0137] in, This represents the Adam optimizer, whose input is the loss relative to the ASCNet parameters in the context of meta-learning knowledge. The gradient; LR for the Adam optimizer; It is the predicted value of the semantic communication codec parameters for the next time step, calculated using the learning rate estimated by the currently unupdated ASCNet. The calculation process is as follows:

[0138] (12)

[0139] Finally, the updated ASCNet parameters Guide semantic communication codec parameters The update process is as follows:

[0140] (13)

[0141] ASCNet can be gradually optimized during the learning process and dynamically adjusted based on the training dynamics of the semantic codec. Furthermore, the channel adaptation strategy enables the semantic communication codec to exhibit high performance across the entire signal-to-noise ratio range.

[0142] Those skilled in the art will understand that various aspects of the present invention can be implemented as systems, methods, or program products. Therefore, various aspects of the present invention can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, collectively referred to herein as a "circuit," "module," or "platform."

[0143] In another embodiment of the present invention, a semantic communication system based on automated training of LSTM network is provided. This system can be used to implement the above-mentioned semantic communication method based on automated training of LSTM network. Specifically, the semantic communication system based on automated training of LSTM network includes a construction module, a network module, a prediction module, and an optimization module.

[0144] The module constructs an LSTM network as the main architecture of the meta-optimizer. The input of the meta-optimizer is activated by the ReLU function after passing through two different weight matrices.

[0145] The network module, after being processed by the construction module, enters the forget gate, input gate, and output gate of the LSTM network to obtain the output of the LSTM network;

[0146] The prediction module will process the output with weight matrix and constraint factor to obtain the predicted learning rate value;

[0147] The optimization module, the main task of semantic communication, will use the learning rate predicted by the meta-optimizer to optimize the parameters of the semantic codec and complete the semantic communication.

[0148] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, the computer program including program instructions, and the processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to implement a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a semantic communication method based on LSTM network automated training, including:

[0149] An LSTM network is constructed as the main architecture of the meta-optimizer. The input of the meta-optimizer is activated by the ReLU function after passing through two different weight matrices. The processed input is then fed into the forget gate, input gate, and output gate of the LSTM network to obtain the output of the LSTM network. The obtained output is processed by the weight matrix and the constraint factor to obtain the predicted learning rate value. The main semantic communication task will use the learning rate predicted by the meta-optimizer to optimize the parameters of the semantic encoder and decoder to complete the semantic communication.

[0150] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device; it can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that more specific examples (a non-exhaustive list) of the computer-readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0151] Computer-readable storage media also include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0152] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0153] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the semantic communication method based on LSTM network automated training in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor in the following steps:

[0154] An LSTM network is constructed as the main architecture of the meta-optimizer. The input of the meta-optimizer is activated by the ReLU function after passing through two different weight matrices. The processed input is then fed into the forget gate, input gate, and output gate of the LSTM network to obtain the output of the LSTM network. The obtained output is processed by the weight matrix and the constraint factor to obtain the predicted learning rate value. The main semantic communication task will use the learning rate predicted by the meta-optimizer to optimize the parameters of the semantic encoder and decoder to complete the semantic communication.

[0155] Please see Figure 6 The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the semantic communication method based on LSTM network automated training in this embodiment. To avoid repetition, details are omitted here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the semantic communication system based on LSTM network automated training in this embodiment. To avoid repetition, details are omitted here.

[0156] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 6This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0157] The processor 61 may be a central processing unit (CPU), or other general-purpose processors, CPUs, graphics processing units (GPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, quantum computing-based data processing logic units, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0158] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or RAM of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0159] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0160] Any references to memory, databases, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0161] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0162] Please see Figure 7 The terminal device 600 is an electronic device, which takes the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0163] The storage unit stores program code, which can be executed by the processing unit 610 to perform the steps described in the method section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit 610 can perform actions such as... Figure 1 The steps are shown in the figure.

[0164] Storage unit 620 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include a read-only memory (ROM) 6203.

[0165] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0166] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0167] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0168] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0169] Taking the main task of semantic communication as transmitting image semantics and performing image classification at the receiver through image semantics as an example, the signal-to-noise ratio range of semantic communication is set to (0dB, 6dB), the knowledge background adopts the CIFAR-10 dataset, the knowledge background is divided into 10:1 ratios, the semantic communication codec adopts the ResNet architecture, the wireless channel adopts the AWGN channel, and the comparison method adopts the existing scheme described in the background of this invention.

[0170] Figure 2 The graph illustrates the trend of the loss function as a function of training epochs when using different optimization algorithms during training. As we can see from the graph, when training with the ASCNet meta-learning method, the training loss exhibits a gradual decreasing trend. This indicates that ASCNet can effectively guide the updating of model parameters, allowing the model to gradually approach its optimal state.

[0171] In particular, when compared with the best-performing Multistep Stochastic Gradient Descent (MSGD), ASCNet exhibits very similar performance. In the early stages of training, the training loss of both methods decreases rapidly, indicating their ability to quickly capture the fundamental features of the data. As training progresses, ASCNet maintains a low training loss, and in the later stages, its decreasing trend is almost identical to that of MSGD, demonstrating very stable performance.

[0172] The ASCNet meta-learning method achieves training accuracy comparable to Multistep SGD primarily due to its advanced adaptive learning rate adjustment mechanism. By monitoring dynamic changes during training in real time, ASCNet intelligently adjusts the learning rate of each parameter, thus enabling fine-tuning of model details while maintaining training speed. Furthermore, ASCNet's meta-optimization framework also considers improving model generalization ability, dynamically adjusting the learning rate and model structure to ensure high accuracy even when facing unknown data.

[0173] In short, Figure 2 This study visually demonstrates the superior performance of the ASCNet meta-learning method during training, showcasing performance comparable to state-of-the-art optimization algorithms. This finding offers a novel approach: introducing meta-learning methods to optimize the training process of semantic communication codecs can potentially lead to higher communication efficiency and more reliable communication services.

[0174] In today's rapidly developing field of digital communication, semantic communication, as an emerging communication mode, is gradually becoming a research hotspot. It aims to achieve more efficient and accurate information transmission by capturing and transmitting semantic information in signals rather than simply symbols or data. To further improve the performance of semantic communication systems, researchers have proposed various codecs and optimization techniques. This invention will focus on a meta-optimizer called ASCNet and a novel channel adaptation algorithm, demonstrating their performance under different signal-to-noise ratio conditions through experimental data.

[0175] First, through Figure 3 and Figure 4 This paper visually demonstrates the performance comparison between the semantic communication codec trained with ASCNet and existing comparative algorithms under signal-to-noise ratios (SNR) of 0dB and 6dB, respectively. As shown in the figure, the codec trained with ASCNet exhibits superior performance compared to the comparative algorithms under both SNR conditions. Particularly in low SNR environments, ASCNet helps the codec better adapt to harsh channel conditions, thereby reducing information loss and noise interference, and improving the accuracy of downstream classification.

[0176] Furthermore, an innovative channel-adaptive method is proposed, which enables the semantic communication codec to achieve optimal performance across the entire signal-to-noise ratio range through a single training iteration. The core idea of ​​this method is to dynamically adjust the codec parameters during training, allowing it to automatically optimize according to different channel conditions, thereby maintaining high communication quality under various signal-to-noise ratio environments.

[0177] In summary, the combined effect of the ASCNet meta-optimizer and channel adaptation algorithm significantly improves the performance of the semantic communication codec. This improvement is particularly pronounced when facing harsh channel environments. Through these innovative technologies, it is foreseeable that future semantic communication systems will provide more stable and reliable communication services while maintaining high efficiency.

[0178] Figure 5 The graph shows the curve of the learning rate (LR) predicted by ASCNet as a function of the number of training iterations (Epochs). It is clearly visible from the graph that the learning rate exhibits a significant real-time dynamic decreasing trend throughout the training process. This phenomenon indicates that ASCNet can intelligently adjust the learning rate according to the progress of training to achieve more effective learning results.

[0179] Specifically, at the beginning of training, ASCNet assigns a high learning rate, allowing for rapid parameter updates and accelerating the model's learning of data features. As training progresses, the learning rate gradually decreases, which helps the model fine-tune within the parameter space, continuously approaching the optimal solution. Furthermore, this dynamic adjustment of the learning rate helps avoid overfitting, ensuring the model has better generalization ability.

[0180] It's worth noting that the learning rate (LR) decline curve predicted by ASCNet is not monotonically decreasing, but rather exhibits some fluctuation. This fluctuation reflects ASCNet's sensitivity to training complexity at different stages, allowing it to appropriately increase the learning rate when needed to overcome training bottlenecks, and then decrease the learning rate for fine-tuning. This strategy enables ASCNet to demonstrate greater flexibility and adaptability during training.

[0181] therefore, Figure 5 This paper clearly reveals how ASCNet optimizes the training process through intelligent prediction of the learning rate. This method not only accelerates the model's convergence speed but also results in a model with better performance and stability. By gaining a deeper understanding of this mechanism of ASCNet, we can further optimize the design of semantic communication codecs and drive them to achieve higher levels of performance in practical applications.

[0182] In summary, this invention provides a semantic communication method and system based on automated training of LSTM networks. Taking the main task of image classification semantic communication as an example, the meta-optimizer proposed in this invention exhibits the best convergence performance in terms of training loss loss compared to traditional optimizers. In terms of transmission classification accuracy, it achieves a 3% to 5% improvement over traditional optimizers in both 0dB and 6dB signal-to-noise ratio environments.

[0183] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0184] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0185] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0186] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0187] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0188] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0189] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0190] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0191] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0192] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0193] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A semantic communication method based on automated training of LSTM networks, characterized in that, Includes the following steps: S1. Construct an LSTM network as the main architecture of the meta-optimizer. The input of the meta-optimizer is activated by the ReLU function after passing through two different weight matrices. The input of the meta-optimizer is the training loss value of the semantic communication task and the historical information of the previous stage. The training process of the semantic communication task is as follows: S101. Initialize the encoder / decoder parameters of the semantic communication system, the weight matrix parameters of the meta-optimizer, the initial learning rate, the cell parameters of the main LSTM of the meta-optimizer, and set the number of meta-learning rounds T. S102. Randomly divide the training knowledge background of semantic communication into main task knowledge background and meta-learning knowledge background according to a set ratio. S103. The input to the semantic communication task is randomly sampled from the knowledge background of the main task and preprocessed. S104. The data preprocessed in step S103 is used as input to sequentially enter the semantic encoder, channel encoder, wireless channel, channel decoder and semantic decoder to obtain the output of the semantic communication main task. S105. After calculating the training loss using the input and output, perform gradient descent according to the initial learning rate to optimize the encoder / decoder parameters. Specifically: in, This represents the input-output relationship of the meta-optimizer. The loss value of the input meta-optimizer, For the cell state of the meta-optimizer, This provides parameter information for the meta-optimizer at the next time step. The Loss value represents the gradient descent information of the parameter w on the semantic communication main task network; S106. After every T rounds of training, the training loss value of the current round is input into the meta-optimizer. The meta-optimizer internally uses the meta-learning knowledge background to calculate and record the learning rate that should be adopted for the current semantic communication main task, and updates the parameters of the meta-optimizer accordingly. Specifically: in, Represents the Adam optimizer; LR for the Adam optimizer; It is the predicted value of the semantic communication codec parameters for the next time step, calculated using the learning rate estimated by the currently unupdated meta-optimizer; The predicted values ​​of the semantic communication codec parameters for the next time step, calculated using the learning rate estimated by the currently unupdated meta-optimizer. for: in, This represents the input-output relationship of the meta-optimizer. The loss value of the input meta-optimizer, For the cell state of the meta-optimizer, This provides parameter information for the meta-optimizer. The Loss value represents the gradient descent information of the parameter w on the semantic communication main task network; S107. Update the learning rate of the main task to the learning rate predicted by the meta-optimizer, and repeat steps S102-S107 until the training loss in step S105 no longer decreases or reaches the total number of training rounds. S2. The input processed in step S1 enters the forget gate, input gate, and output gate of the LSTM network respectively to obtain the output of the LSTM network. S3. After processing the output obtained in step S2 with the weight matrix and the constraint factor, the predicted learning rate value is obtained. S4. The main task of semantic communication is to optimize the parameters of the semantic codec using the learning rate predicted by the meta-optimizer in step S3, thereby completing semantic communication.

2. The semantic communication method based on LSTM network automated training according to claim 1, characterized in that, In step S3, the predicted learning rate value is obtained as follows: The two inputs pass through a fully connected layer with a ReLU activation function. Perform preprocessing, then import into LSTM and obtain the output. ;use Fully connected layer with activation function right Perform a linear transformation to obtain the predicted value. Finally, multiply the predicted value by the scaling factor. Ensure that the final predicted learning rate value is within Interval.

3. The semantic communication method based on automated training of LSTM networks according to claim 1, characterized in that, In step S4, an online approximation technique is used to jointly update the meta-optimizer parameters. Semantic communication codec parameters It will be updated Steps ( (Updated later) once; Furthermore, the overall knowledge background of semantic communication is divided into the knowledge background of the main semantic communication task and the knowledge background of meta-learning. and The parameters are trained on their respective knowledge backgrounds.

4. The semantic communication method based on LSTM network automated training according to claim 1, characterized in that, In step S4, during the semantic communication main task and meta-learning task stages, the channel adaptive training method is as follows: Suppose that the semantic communication task is applied in a real-world scenario with a signal-to-noise ratio (SNR) range of mdB to ndB, where n is greater than or equal to m; set the channel SNR of the main semantic communication task to mdB and train it using the knowledge background of the main task. During the meta-learning phase, the channel signal-to-noise ratio is adjusted to ndB and a meta-learning knowledge background is adopted; the knowledge background of the main task will be much larger than the knowledge background of the meta-learning, and the meta-learning phase is only carried out once after multiple rounds of training.

5. The semantic communication method based on LSTM network automated training according to claim 4, characterized in that, Semantic codec parameters for: in, The codec parameters will be updated in the next time step. To validate the loss value within the context of existing knowledge, This represents the input-output relationship of the meta-optimizer. This represents the input-output relationship of the gradient descent function. This is based on the total number of training rounds.

6. A semantic communication system based on automated training of LSTM networks, characterized in that, include: The module constructs an LSTM network as the main architecture of the meta-optimizer. The input to the meta-optimizer is processed through two different weight matrices and then activated by the ReLU function. The input to the meta-optimizer consists of the training loss value of the semantic communication task and the historical information from the previous stage. The training process of the semantic communication task is as follows: Initialize the encoder / decoder parameters of the semantic communication system, the weight matrix parameters of the meta-optimizer, the initial learning rate, the cell parameters of the main LSTM of the meta-optimizer, and set the number of meta-learning rounds T; The training knowledge background of semantic communication is randomly divided into main task knowledge background and meta-learning knowledge background according to a set ratio. The input to the semantic communication task is randomly sampled from the knowledge background of the main task and then preprocessed. The preprocessed data is fed into the semantic encoder, channel encoder, wireless channel, channel decoder and semantic decoder in sequence to obtain the output of the semantic communication main task. After calculating the training loss using the input and output, gradient descent is performed according to the initial learning rate to optimize the encoder-decoder parameters. Specifically: in, This represents the input-output relationship of the meta-optimizer. The loss value of the input meta-optimizer, For the cell state of the meta-optimizer, This provides parameter information for the meta-optimizer at the next time step. The Loss value represents the gradient descent information of the parameter w on the semantic communication main task network; After every T training rounds, the training loss value of the current round is input into the meta-optimizer. Internally, the meta-optimizer uses the background knowledge of meta-learning to calculate and record the learning rate that should be adopted for the current semantic communication main task, and then updates the parameters of the meta-optimizer. Specifically: in, Represents the Adam optimizer; LR for the Adam optimizer; It is the predicted value of the semantic communication codec parameters for the next time step, calculated using the learning rate estimated by the currently unupdated meta-optimizer; The predicted values ​​of the semantic communication codec parameters for the next time step, calculated using the learning rate estimated by the currently unupdated meta-optimizer. for: in, This represents the input-output relationship of the meta-optimizer. The loss value of the input meta-optimizer, For the cell state of the meta-optimizer, This provides parameter information for the meta-optimizer. The Loss value represents the gradient descent information of the parameter w on the semantic communication main task network; Update the learning rate of the main task to the learning rate predicted by the meta-optimizer, and repeat the steps until the training loss no longer decreases or the total number of training rounds is reached. The network module, after being processed by the construction module, enters the forget gate, input gate, and output gate of the LSTM network to obtain the output of the LSTM network; The prediction module will process the output with weight matrix and constraint factor to obtain the predicted learning rate value; The optimization module, the main task of semantic communication, will use the learning rate predicted by the meta-optimizer to optimize the parameters of the semantic codec and complete the semantic communication.

Images