Semantic-based AI model parameter transmission method, generation method and equipment

Abstract

The invention provides a semantic-based AI model parameter transmission method, a semantic-based AI model parameter generation method and equipment, and the transmission method comprises the steps that a target AI model deployment request is sent by a receiving end side, and the request comprises task description data used for describing local data set features of end side equipment and downstream task demand data of a target AI model; matching the task description data with description data corresponding to a local preset target data set, and screening target parameter semantics from a preset model coding parameter semantic database corresponding to the target data set according to a matching result, or generating the target parameter semantics based on the database; and transmitting the target parameter semantics to an end side, so that the end side decodes the target parameter semantics to generate a corresponding target AI model parameter. According to the invention, the communication transmission amount can be reduced, and the stable performance can be maintained under low SNR. In addition, the generation of the target AI model parameters does not need to upload local data of the target AI model parameters at an end side, and personalized adaptation can be realized only by lightweight fine tuning, so that privacy can be protected.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of artificial intelligence and communication technology, and in particular to a semantic-based AI model parameter transmission method, generation method and device. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, AI models have been widely deployed in end-side environments such as mobile devices and smart Internet of Things terminals. These end-side AI models need to be continuously updated to adapt to dynamic environmental changes, user preferences or emerging tasks. However, end-side devices are usually limited by computing resources and cannot perform large-scale model training and extensive fine-tuning, which hinders the further development of end-side AI.

[0003] To solve this problem, a common strategy is to use the computing resources of edge servers for model training and transmit the optimized AI model parameters to the end-side devices, thereby sharing the end-side computing load and achieving dynamic adaptation. However, this process faces two major challenges: First, AI model parameters are high-dimensional structured tensor data, and their transmission requires extremely high network bandwidth, resulting in waste of communication resources. Traditional communication systems are divided into digital transmission and analog transmission, the former is prone to "cliff effect" (in digital communication, cliff effect refers to when the signal-to-noise ratio is lower than a certain critical value, the quality of the received signal will deteriorate sharply, resulting in the complete failure of the model parameters reconstructed at the receiving end) under low signal-to-noise ratio (SNR). The latter has low coding efficiency and unstable performance. In recent years, as a new paradigm, semantic communication can reduce transmission redundancy and improve efficiency by extracting and transmitting core semantic information rather than raw bit streams, and has shown advantages in image, speech, video and other fields, but its application in AI model parameter transmission is still in its infancy. Second, privacy issues further complicate model deployment. End-side devices periodically collect data related to device usage, user behavior and environmental status, but due to privacy concerns, end-side devices are reluctant to disclose the collected data to edge servers to prevent sensitive information from being leaked. Therefore, the data distribution used by the model for training on the edge server is significantly mismatched with the data distribution actually generated (used) by the end-side device, resulting in the model failing to effectively complete the inference task after being deployed on the end-side. The end-side needs to retrain or fine-tune the model to achieve personalized adaptation, but the end-side device has limited computing resources, and traditional full-fine-tuning methods are time-consuming and resource-intensive, which cannot meet the efficiency requirements of AI applications. SUMMARY

[0004] In view of this, the embodiments of the present application provide a semantic-based AI model parameter transmission method, generation method and device to eliminate or improve one or more defects in the prior art.

[0005] The present application provides a semantic-based AI model parameter transmission method, which comprises: receive a target AI model deployment request sent by an end-side device, wherein the target AI model deployment request contains task description data describing characteristics of a local dataset of the end-side device and downstream task requirement data for a target AI model; match the task description data with description data corresponding to a local preset target dataset, and according to a matching result, filter target parameter semantics that conform to the downstream task requirement data from a preset model encoding parameter semantic library corresponding to the target dataset, or generate target parameter semantics that conform to the downstream task requirement data based on the model encoding parameter semantic library; transmit the target parameter semantics to the end-side device, so that the end-side device decodes the target parameter semantics to generate target AI model parameters applied to a downstream task.

[0006] In some embodiments of the present application, before the target AI model deployment request sent by the receiving end-side device, there are also: perform semantic encoding on a plurality of AI model parameters obtained by training on the target dataset, to obtain encoding parameter semantics for representing model weight distribution rules and / or feature extraction capabilities; aggregate each of the encoding parameter semantics and associate them with corresponding model performance labels to form the model encoding parameter semantic library.

[0007] In some embodiments of the present application, the matching of the task description data with description data corresponding to a local preset target dataset, and according to a matching result, filtering target parameter semantics that conform to the downstream task requirement data from a preset model encoding parameter semantic library corresponding to the target dataset, or generating target parameter semantics that conform to the downstream task requirement data based on the model encoding parameter semantic library, comprises: matching the task description data with description data corresponding to a local preset target dataset; if the task description data matches the description data corresponding to the local preset target dataset, then from the model performance labels corresponding to each of the encoding parameter semantics in the model encoding parameter semantic library, filtering encoding parameter semantics in the model encoding parameter semantic library whose model performance labels conform to the downstream task requirement data as target parameter semantics; if the task description data does not match the description data corresponding to the local preset target dataset, then using a kernel density estimation method to model the distribution of the encoding parameter semantics in the model encoding parameter semantic library, and based on the downstream task requirement data, sampling from the obtained probability density function to generate target parameter semantics.

[0008] In some embodiments of the present application, the semantic encoding of the plurality of AI model parameters trained on the target data set obtains an encoded parameter semantic for representing model weight distribution regularity and / or feature extraction capability, including: serializing and converting each of the plurality of AI model parameters trained on the target data set to obtain a respective serialized parameter corresponding to each of the AI model parameters; extracting, by using a semantic extraction model based on a multi-scale attention mechanism, a respective original parameter semantic for representing model weight distribution regularity and / or feature extraction capability corresponding to each of the AI model parameters from each of the respective serialized parameters; jointly encoding, by using a preset source and channel joint encoding model, each of the original parameter semantics to obtain a respective encoded parameter semantic corresponding to each of the original parameter semantics.

[0009] In some embodiments of the present application, the end-side device is configured to decode the target parameter semantic data by using a preset source and channel joint decoding model to generate target AI model parameters applied to downstream tasks; Correspondingly, before the receiving end-side device sends the target AI model deployment request, the method further includes: jointly training, based on a loss function, the source and channel joint decoding model and the source and channel joint encoding model in an end-to-end manner; wherein the loss function includes a layer normalization-based reconstruction loss function and a contrastive loss function.

[0010] The layer normalization-based reconstruction loss function is shown in formula (1): In formula (1), is the reconstruction loss, denotes the reconstruction value of the i-th parameter in the j-th layer, denotes the target value of the i-th parameter in the j-th layer, denotes the mean value of the parameters in the j-th layer, denotes the standard deviation of the parameters in the j-th layer, denotes the total number of layers of the model, denotes the number of parameters in the j-th layer, denotes the total number of parameters of all layers, used for global normalization.

[0011] ​​​​​​​In some embodiments of the present application, the contrast loss function is optimized by constructing positive sample pairs and negative sample pairs. The positive sample pairs are two functionally equivalent views obtained by different enhancement operations on the same AI model parameters. The negative sample pairs are combinations between enhanced views from different AI model parameters.

[0012] In some embodiments of the present application, the endpoint device is further configured to perform lightweight fine-tuning on the target AI model parameters, wherein the lightweight fine-tuning comprises updating part of the layers or part of the parameters in the target AI model parameters; and / or, updating the target AI model parameters using less than a preset number of fine-tuning rounds.

[0013] Another aspect of the present application provides a semantic-based AI model parameter generation method, which comprises: sending a target AI model deployment request to a server, wherein the target AI model deployment request contains task description data describing features of a local data set and downstream task requirement data for a target AI model; so that the server performs: matching the task description data with description data corresponding to a local preset target data set, and according to the matching result, screening target parameter semantics conforming to the downstream task requirement data from a preset model encoding parameter semantic library corresponding to the target data set, or generating target parameter semantics conforming to the downstream task requirement data based on the model encoding parameter semantic library; receiving the target parameter semantics sent by the server, and decoding the target parameter semantics to generate target AI model parameters applied to downstream tasks.

[0014] In some embodiments of the present application, the generation method further comprises: performing lightweight fine-tuning on the target AI model parameters, wherein the lightweight fine-tuning comprises updating part of the layers or part of the parameters in the target AI model parameters; and / or, updating the target AI model parameters using less than a preset number of fine-tuning rounds.

[0015] A third aspect of the present application provides an electronic device comprising a processor and a memory; the processor implements the semantic-based AI model parameter transmission method or the semantic-based AI model parameter generation method when executing a running program stored in the memory.

[0016] A fourth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the aforementioned semantic-based AI model parameter transmission method or the semantic-based AI model parameter generation method.

[0017] A fifth aspect of the present application provides a computer program product comprising a computer program which, when executed by a processor, implements the aforementioned semantic-based AI model parameter transmission method or the aforementioned semantic-based AI model parameter generation method.

[0018] The semantic-based AI model parameter transmission method provided by the present application comprises: receiving a target AI model deployment request sent by an end-side device, wherein the target AI model deployment request comprises task description data for describing characteristics of a local data set of the end-side device and downstream task requirement data for a target AI model; matching the task description data with description data corresponding to a local preset target data set, and according to a matching result obtained, screening a target parameter semantic that meets the downstream task requirement data from a preset model encoding parameter semantic library corresponding to the target data set, or generating the target parameter semantic that meets the downstream task requirement data based on the model encoding parameter semantic library; and transmitting the target parameter semantic to the end-side device, so that the end-side device decodes the target parameter semantic to generate a target AI model parameter applied to a downstream task. The present application can significantly reduce the communication transmission amount by transmitting the target AI model parameter through semantic communication, and maintain stable performance at a low SNR, avoiding the "cliff effect". In addition, the end-side device does not need to upload the local data, and the generated model parameter is fine-tuned, the precision of which is comparable to or exceeds that of the traditional from-scratch training or large-scale fine-tuning method, which can prevent sensitive information from being leaked, and is suitable for end-side AI deployment scenarios such as mobile devices, smart Internet of Things terminals, edge computing, etc.

[0019] Additional advantages, objects, and features of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following or can be learned by practice of the application. The objects and other advantages of the application can be realized and attained by the structure particularly pointed out in the description and claims hereof as well as the appended drawings.

[0020] It will be understood by those skilled in the art that the objects and advantages of the present application are not limited to the above specifically described, and the above and other objects capable of being achieved by the present application will be more clearly understood according to the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0021] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the principles of the application. The components in the drawings are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the application. For purposes of clarity and understanding, some portions of the drawings can be exaggerated, such as relative sizes of some parts, to show details of certain features. In the drawings: Figure 1 The first flowchart of the semantic-based AI model parameter transmission method in an embodiment of the present application.

[0022] Figure 2 The second flowchart of the semantic-based AI model parameter transmission method in an embodiment of the present application.

[0023] Figure 3 The third flowchart of the semantic-based AI model parameter transmission method in an embodiment of the present application.

[0024] Figure 4 The fourth flowchart of the semantic-based AI model parameter transmission method in an embodiment of the present application.

[0025] Figure 5 The first flowchart of the semantic-based AI model parameter generation method in an embodiment of the present application.

[0026] Figure 6 The parameter semantic communication framework of the semantic-based AI model parameter transmission method in an embodiment of the present application.

[0027] Figure 7 The second flowchart of the semantic-based AI model parameter generation method in an embodiment of the present application.

[0028] Figure 8 The flowchart of the semantic-based AI model parameter transmission and generation method realized by the server and the terminal-side device double-end interaction in an embodiment of the present application.

[0029] It should also be noted that, in order to avoid obscuring the present application due to unnecessary details, only the processing steps closely related to the scheme according to the present application are shown in the drawings, and other details not closely related to the present application are omitted. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the embodiments and drawings. Herein, the illustrative embodiments of the present application and their descriptions serve to explain the present application, but do not limit the present application.

[0031] It should be noted that, in order not to obscure the application with unnecessary details, only structures and / or processing steps closely related to the solution according to the application are shown in the drawings, and other details not closely related to the application are omitted.

[0032] It should be emphasized that the term "comprises / comprising" when used in this text, refers to the presence of the features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0033] It should be noted that, unless otherwise specified, the term "connected" in this text can not only mean direct connection, but also indirect connection with the presence of intermediates.

[0034] In the following, embodiments of the application will be described with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts or the same or similar steps.

[0035] It should be noted that the pre-trained model is transmitted to the edge server side at the edge server side. The traditional full fine-tuning refers to retraining the entire pre-trained model parameters with the end-side dataset, which can achieve performance adaptation, but has the problems of large computational overhead and high resource consumption. The AI model parameter transmission method, generation method and device based on semantics proposed in the application only need to rely on the end-side dataset to train the generated model parameters, and can achieve task performance comparable to full fine-tuning without iterating too many rounds, greatly reducing the computational resource consumption of the end-side device in the model adaptation process, and more in line with the limited hardware resource constraints in the end-side scenario.

[0036] Based on this, in order to solve the problems of high communication transmission of the existing AI model parameter transmission technology, and the "cliff effect" caused by unstable performance under low SNR, and the problem that in the existing AI model parameter generation technology, the end-side device is unwilling to disclose the collected data to the edge server due to privacy problems, in order to prevent sensitive information from being leaked, so that the data distribution used by the model in the edge server training is significantly mismatched with the data distribution actually generated (used) by the end-side device, resulting in that the model cannot effectively complete the inference task after being deployed on the end-side, and the end-side needs to retrain or fine-tune the model to achieve individual adaptation, but the end-side device has limited computing resources, and the traditional full fine-tuning method is time-consuming and resource-intensive, which cannot meet the efficiency requirements of AI applications, the embodiments of the application respectively provide an AI model parameter transmission method based on semantics, an AI model parameter generation method based on semantics, an entity device, a computer readable storage medium and a computer program product.

[0037] The embodiments will be described in detail below.

[0038] The embodiment of the application provides a semantic-based AI model parameter transmission method which can be executed by a transmission device in a server, referring to Figure 1 The semantic-based AI model parameter transmission method specifically comprises the following contents: Step 100: receiving a target AI model deployment request sent by an end-side device, wherein the target AI model deployment request comprises task description data used for describing characteristics of a local data set of the end-side device and downstream task demand data of a target AI model; In step 100, the downstream task can be a specific application scenario of the target AI model, such as image classification.

[0039] Step 200: matching the task description data with description data corresponding to a local preset target data set, and according to a matching result, screening a target parameter semantic conforming to the downstream task demand data from a preset model coding parameter semantic library corresponding to the target data set, or generating the target parameter semantic conforming to the downstream task demand data based on the model coding parameter semantic library; In step 200, the target parameter semantic refers to semantic information capable of representing internal attributes (weight distribution law) and external functional attributes (feature extraction capability) of the target AI model. The target parameter semantic conforming to the downstream task demand data can be generated based on the model coding parameter semantic library by using a traditional statistical-based sampling method or by training an artificial intelligence model to learn a probability distribution of the parameter semantic library.

[0040] Step 300: transmitting the target parameter semantic to the end-side device, so that the end-side device decodes the target parameter semantic to generate a target AI model parameter applied to a downstream task.

[0041] In step 300, the target parameter semantic is transmitted to the end-side device through a channel.

[0042] In one or more embodiments of the application, the end-side device is provided with an information device, and the device comprises a source-channel joint decoding module and a multi-scale attention mechanism structure semantic decoding module. The source-channel joint decoding module is used for decoding a signal transmitted through a channel into a target parameter semantic. The multi-scale attention mechanism structure semantic decoding module is used for reconstructing the target parameter semantic into a target AI model parameter.

[0043] In summary, the semantic-based AI model parameter transmission method and semantic-based AI model parameter generation method provided in this application can transmit target AI model parameters through semantic communication, significantly reducing communication volume and maintaining stable performance under low SNR, thus avoiding the "cliff effect." Furthermore, the edge device does not need to upload local data, and the generated model parameters, after minimal fine-tuning, achieve accuracy comparable to or exceeding traditional de novo training or extensive fine-tuning methods. This prevents the leakage of sensitive information and is suitable for edge AI deployment scenarios such as mobile devices, smart IoT terminals, and edge computing.

[0044] To reduce communication throughput while still transmitting target AI model parameters, this application provides a semantic-based AI model parameter transmission method, see [link to relevant documentation]. Figure 2 Before the target AI model deployment request sent by the receiving end device, the following is also included: Step A00: Semantically encode the parameters of multiple AI models trained on the target dataset to obtain encoded parameter semantics that characterize the model weight distribution pattern and / or feature extraction capability; Step B00: Summarize the semantics of each of the encoding parameters and associate them with the corresponding model performance labels to form the model encoding parameter semantic library.

[0045] In one or more embodiments of this application, the model performance label may be model accuracy.

[0046] To further reduce communication throughput and achieve the transmission of target AI model parameters, a semantic-based AI model parameter transmission method is provided in this application embodiment, see [link to relevant documentation]. Figure 3 Step 200 specifically includes: Step 210: Match the task description data with the description data corresponding to the locally preset target dataset; Step 220: If the task description data matches the description data corresponding to the local preset target dataset, then according to the model performance label corresponding to each of the encoding parameter semantics in the model encoding parameter semantic library, the encoding parameter semantics whose model performance label matches the downstream task requirement data are selected as the target parameter semantics. If the task description data does not match the description data corresponding to the local preset target dataset, a kernel density estimation method is used to model the distribution of the encoding parameter semantics in the model encoding parameter semantic library, and samples are taken from the probability density function obtained from the modeling based on the downstream task requirement data to generate the target parameter semantics.

[0047] In one or more embodiments of the present application, when the task description data does not match the description data corresponding to the locally preset target data set, a target parameter semantic generated based on a kernel density estimation method integrates the knowledge of multiple AI models that have been trained.

[0048] Specifically, a new parameter semantic is generated based on a kernel density estimation (KDE) method. KDE is a non-parametric statistical method that does not require a mathematical form of a preset data distribution, and can adaptively fit a probability density function through the distribution characteristics of the data itself, with the characteristics of high flexibility, simple implementation, and strong determinacy. The core idea is that for each data point, a Gaussian kernel is used to "smear" the point, and all kernel functions are superimposed to form the overall probability density estimation. In the generation of the target parameter semantic, the present application embodiment performs KDE modeling on each dimension of the target parameter semantic KDE modeling is performed separately: The kernel function is selected as a Gaussian kernel function, and the formula is: wherein, denotes the Gaussian kernel function; denotes a random variable; exp(·) denotes an exponential function.

[0049] The smoothness and symmetry thereof can effectively capture the local distribution characteristics of the data, while avoiding the introduction of excessive model bias.

[0050] The bandwidth parameter h is the only hyperparameter in KDE, and the bandwidth controls the "width" of the kernel function. Too small bandwidth will cause the estimation result to overfit (noise fluctuation), and too large bandwidth will cause over-smoothing (loss of local details). The optimal bandwidth is selected by cross-validation method to balance the bias and variance.

[0051] Density estimation process: for the probability density of the target parameter semantic in the i-th dimension , the probability density function estimation expression is: wherein, The formula indicates that the probability density of any point is the weighted sum of all anchor point samples through the Gaussian kernel function, and the weight is determined by the distance between the sample and , that is, the closer the distance, the greater the contribution. Wherein, K is the Gaussian kernel function, M is the number of anchor point samples, is the probability density of the target parameter semantic in the i-th dimension, is the anchor point sample.

[0052] ​​Through the above steps, the embodiment of the application can obtain the marginal distribution of each dimension of the target parameter semantics. When constructing, samples are independently extracted from the KDE distribution of each dimension , and the samples of each dimension are spliced to generate complete target parameter semantics , where D represents the dimension of the parameter semantics.

[0053] The target parameter semantics is obtained through a channel . The received vector is sent to the trained source-channel joint decoding module and the semantic decoding module based on the multi-scale attention mechanism structure for decoding to obtain the target AI model parameter , where represents the parameter semantic decoding module with parameters , and represents the channel-source joint decoding module with parameters . The target AI model parameter is loaded into the target AI model, and the local data set is used for lightweight fine-tuning to adapt to the local data set.

[0054] In order to further reduce the communication transmission amount and realize the transmission of the target AI model parameter, in an AI model parameter transmission method based on semantics provided in an embodiment of the application, referring to Figure 4 , step A00 specifically includes: Step A10: using a parameter preprocessing module, converting a plurality of AI model parameters trained on the target data set into serialized representation respectively to obtain respective serialized parameters corresponding to each of the AI model parameters; In one or more embodiments of the application, the serialized parameter can be a model parameter converted into an ordered vector sequence so that the attention mechanism can process it.

[0055] Step A20: using a semantic extraction model based on a multi-scale attention mechanism, extracting respective original parameter semantics corresponding to each of the AI model parameters from each of the serialized parameters, the original parameter semantics being used to represent the model weight distribution rule and / or feature extraction capability; In one or more embodiments of the application, the original parameter semantics can be the model transferable feature extraction capability.

[0056] Step A30: using a preset source and channel joint encoding model, jointly encoding the source and channel of each of the original parameter semantics to obtain respective encoded parameter semantics corresponding to each of the original parameter semantics.

[0057] In one or more embodiments of the application, the encoded parameter semantics can be a symbol stream transmitted through a channel.

[0058] In order to further reduce the amount of communication transmission and realize the transmission of target AI model parameters, in a semantic-based AI model parameter transmission method provided in this application embodiment, the end-side device is used to decode the target parameter semantic data using a preset source and channel joint decoding model to generate target AI model parameters for application to downstream tasks. Correspondingly, before the target AI model deployment request sent by the receiving end device, the following is also included: Based on the loss function, the source and channel joint decoding model and the source and channel joint coding model are jointly trained end-to-end; wherein, the loss function includes a reconstruction loss function based on layer normalization and a contrast loss function.

[0059] The reconstruction loss function based on layer normalization is shown in Equation (1): In formula (1), To rebuild the losses, Indicates the first Layer The reconstructed values ​​of each parameter, Indicates the first Layer The target value of each parameter Indicates the first The mean of the layer parameters, Indicates the first Standard deviation of layer parameters This indicates the total number of layers in the model. Indicates the first The number of parameters in the layer This represents the total number of parameters across all layers and is used for global normalization.

[0060] In one or more embodiments of this application, a reconstruction loss function based on layer normalization is introduced. The mean squared error is calculated after normalizing the weights of each layer to address reconstruction bias caused by uneven distribution. The design employs a contrastive loss function that maximizes the similarity between positive sample pairs while minimizing the similarity between negative sample pairs, thereby bringing similar parameters closer together in the semantic space. For batches of... The parameter samples are randomly augmented twice to form two views. and This enhancement operation is based on parameter characteristics and mainly adopts methods such as structure preservation arrangement (utilizing the equivalence of neuron position exchange) and random erasure (masking part of the weight values). It not only ensures the correlation between views (originating from the same original sample), but also introduces differences through perturbation, forcing the model to learn the core features in the weight space that are unrelated to the specific perturbation.

[0061] In order to further reduce the communication transmission amount and realize the transmission of the target AI model parameter, in an AI model parameter transmission method based on semantics provided in an embodiment of the present application, the contrast loss function is optimized by constructing positive sample pairs and negative sample pairs. The positive sample pair is two functionally equivalent views of the same AI model parameter after different enhancement operations. The negative sample pair is a combination between enhanced views from different AI model parameters.

[0062] In one or more embodiments of the present application, the positive sample pair refers to two views of the same model parameter after enhancement, and the numerical values or local structures of the two views change, but the functional equivalence of the original network is maintained. The acquisition method is to apply noise enhancement and arrangement enhancement to the model parameter. The two views of the positive sample pair do not require the same enhancement method, and can be generated by the same enhancement or different enhancement combination. The core only needs to ensure that the two views come from the same model parameter. The negative sample pair refers to a combination between enhanced views from different model parameters, that is, A1 and A2 are enhanced from model parameter 1, and B1 and B2 are enhanced from model parameter 2. Any two pairs are paired, for example, A1 and B1, A1 and B2, A2 and B1, and A2 and B2. These combinations correspond to model parameters (model parameter 1 and model parameter 2) with different training configurations (such as hyperparameters, initialization seeds), representing “dissimilar” samples, and the core difference does not disappear due to the enhancement operation, so it can be used as a negative sample pair for contrast learning to guide the model to distinguish different model parameters.

[0063] In order to further reduce the communication transmission amount and realize the transmission of the target AI model parameter, in an AI model parameter transmission method based on semantics provided in an embodiment of the present application, the end-side device is further configured to perform light fine-tuning on the target AI model parameter, wherein the light fine-tuning includes: updating part of layers or part of parameters in the target AI model parameter; and / or, updating the target AI model parameter using less than a preset number of fine-tuning rounds.

[0064] It can be understood that the preset number can be set according to actual application requirements and is dynamic, depending on variables such as scenarios and models.

[0065] An AI model parameter generation method based on semantics, which can be executed by an information device arranged in a client device, is also provided in an embodiment of the present application, as shown in Figure 5 The AI model parameter generation method based on semantics specifically includes the following contents: Step 10: sending a target AI model deployment request to the server, wherein the target AI model deployment request contains task description data describing characteristics of the local data set and downstream task requirement data for the target AI model; so that the server performs: matching the task description data with the description data of the local preset target data set, and according to the matching result, screening the target parameter semantics that meet the downstream task requirement data from the preset model coding parameter semantic library corresponding to the target data set, or generating the target parameter semantics that meet the downstream task requirement data based on the model coding parameter semantic library; Step 20: receiving the target parameter semantics sent by the server, and decoding the target parameter semantics to generate the target AI model parameters applied to the downstream task.

[0066] In order to reduce the consumption of fine-tuning and realize the generation of target AI model parameters, in an AI model parameter transmission method based on semantics provided in an embodiment of the present application, the method further comprises: Lightweight fine-tuning of the target AI model parameters, wherein the lightweight fine-tuning comprises: updating part of the layers or part of the parameters in the target AI model parameters; and / or using less than a preset number of fine-tuning rounds to update the target AI model parameters.

[0067] It can be understood that the preset number can be set according to actual application requirements, and is dynamic, depending on variables such as scenarios and models.

[0068] In a specific example of a semantic-based AI model parameter transmission and generation scenario, the complete flow of the semantic-based AI model parameter transmission and generation method is illustrated, and an AI model parameter transmission architecture of the semantic-based AI model parameter transmission and generation method is designed, as shown in Figure 6 , wherein the transmission method specifically comprises: First, the original AI model parameters are converted into a serialized representation. In addition, sinusoidal / cosine position encoding or learnable position embedding is added to preserve the relative position of the parameters in the original tensor. The preprocessed parameter representation is , represents the parameters of the i-th layer.

[0069] The parameter semantic coding is composed of multiple multi-scale attention mechanism blocks. This structure can not only pay attention to the dependency relationship between layers during the training process, but also pay attention to the weight distribution pattern structure within the layer, which can effectively extract the "parameter semantics". The semantic information of the parameters can reflect the intrinsic properties (weight distribution law) and external functional properties (feature extraction ability) of the model. The specific process can be represented as: ​ where, denotes the parameter semantic, is a semantic encoding module with parameters , denotes the learnable weights of the parameter semantic encoding module.

[0070] The JSCC encoder ensures the effective transmission of the parameter semantic information over the physical channel, and the encoded symbol stream can be represented as: where, denotes the encoded symbol stream, is a source-channel joint encoding module with parameters , denotes the learnable weights of the JSCC encoding module network. Please note that the tanh activation function is introduced at the output end of the parameter semantic transmitting module in the embodiments of the present application, and the output value of each dimension is strictly constrained within the interval (-1, 1) by tanh.

[0071] The vector after parameter semantic encoding and JSCC encoding will be transmitted through the wireless channel, and the signal received by the receiver can be represented as: where, corresponds to the Rayleigh fading channel, which obeys the complex Gaussian distribution; is the additive white Gaussian noise, which obeys the complex Gaussian distribution . For the additive white Gaussian noise channel (AWGN), . In order to realize the end-to-end training of the encoding module and the decoding module, the channel must support backpropagation. For simplicity, mainly consider the AWGN channel and the Rayleigh fading channel.

[0072] At the receiving end, the received vector is sent to the source-channel joint decoding module and the parameter semantic decoding module for decoding, which can be represented as: where, is a joint channel-source decoding module neural network with parameters , is a parameter semantic decoding module with parameters , and its purpose is to reconstruct the parameter semantic into the model parameter , denotes the parameters of the layer, and then the model parameter is loaded into the model to perform downstream tasks (classification, recognition, etc.).

[0073] When training the end-to-end system, a reconstruction loss function based on layer normalization is introduced. After normalizing the weights of each layer, the mean square error is calculated to solve the reconstruction bias caused by uneven distribution. Specifically, it is shown in formula (1): in, Indicates the first Layer The reconstructed values ​​of each parameter, Indicates the first Layer The target value of each parameter Indicates the first The mean of the layer parameters, Indicates the first Standard deviation of layer parameters This indicates the total number of layers in the model. Indicates the first The number of parameters in the layer This represents the total number of parameters across all layers and is used for global normalization.

[0074] Simultaneously, a contrastive loss function is designed to maximize the similarity between positive sample pairs while minimizing the similarity between negative sample pairs, making similar parameters closer in semantic space. For batches with a number of... The parameter samples are randomly augmented twice to form two views. and This enhancement operation is based on parameter characteristics and mainly adopts methods such as structure preservation arrangement (utilizing the equivalence of neuron position exchange) and random erasure (masking part of the weight values). It not only ensures the correlation between views (originating from the same original sample), but also introduces differences through perturbation, forcing the model to learn the core features in the weight space that are unrelated to the specific perturbation.

[0075] Each view is semantically encoded by the module and JSCC encoding module The parameter semantics are obtained after processing. and Then through the projection head Mapping to low-dimensional embedding and The introduction of the projection head aims to transform the semantics of the parameters into a space more suitable for calculating similarity, further reducing redundant information interference and enhancing the effectiveness of contrastive learning.

[0076] To quantify the similarity between two view embeddings, a cosine similarity metric is used: where T denotes the transpose operation, the metric eliminates the influence of scale difference by normalizing vector inner product, and reflects the direction consistency of feature distribution more accurately, which is suitable as the similarity index of the intrinsic structure of parameter samples.

[0077] Contrastive loss The core of the contrastive loss is to maximize the similarity of positive sample pairs (two views of the same original sample) while minimizing the similarity of negative sample pairs (views of different original samples), which is defined as: where when , is 1, otherwise 0. The numerator term measures the similarity of positive sample pairs, and the temperature parameter T controls the steepness of the similarity distribution (the smaller T is, the more concentrated the distribution is, and the higher the discrimination of the similarity difference is); the denominator term is the sum of the similarities of all negative sample pairs and the current sample .

[0078] The contrastive loss does not require accurate reconstruction of parameter values, but focuses on capturing abstract associations related to model characteristics, providing more transferable representations for downstream tasks.

[0079] In summary, the total loss function can be represented as: where the contrastive loss optimizes the discriminability of parameter semantics, and the reconstruction loss optimizes weight reconstruction, both of which work together to improve the quality of parameter semantics. is used to balance the contribution ratio of and .

[0080] Based on the optimized parameter semantic emission module, the "parameter semantics" of the target data set are extracted to form a target data set exclusive parameter semantic library. Next, the distribution of "parameter semantics" in space is analyzed, assuming that the dimension of the "parameter semantics" vector is D, and each dimension element is. Since the output end of the parameter semantic emission module introduces a tanh activation function, the output value of each dimension is strictly constrained within the interval (-1, 1). At this time, all possible values of parameter semantics in space form a "high-dimensional box", that is, a set of high-dimensional vectors that satisfy the following conditions: where denotes the mathematical symbol of the high-dimensional box, denotes the set of all vectors composed of D real numbers, and geometrically, the parameter semantics The norm represents the Euclidean distance of a vector to the origin of a high-dimensional space, and thus can be calculated by The distribution of the parameter semantics in the space is analyzed. It is verified that all the parameter semantics vectors of the parameters are The norms are concentrated in the interval of 10 to 12, which indicates that although the tanh activation function restricts the parameter semantics vectors in a high-dimensional box, the distribution of the parameter semantics vectors does not uniformly fill the entire box, but tends to concentrate on a "high-dimensional hypersphere". The main reason is that the core goal of the contrast loss in the loss function is to make the parameter semantics of similar models close to each other in the space, and the parameter semantics of dissimilar models far away from each other. In order to maximize the distance between classes and minimize the distance within the class, the parameter semantics naturally gathers on the shell of the hypersphere in the high-dimensional space, because the surface area of the hypersphere grows extremely fast with the dimension, which is the optimal topological structure for the distribution of samples.

[0081] In addition, experiments show that the parameter semantics vectors of similar models are close to each other in the space, and the parameter semantics vectors of different models tend to be uniformly dispersed.

[0082] In summary, the parameter semantics contains key features such as the distribution of the parameters of the model, and the parameter semantics not only distributes on the shell of the hypersphere in the high-dimensional space, but also presents a regular arrangement on the hypersphere according to the model attributes, rather than a random distribution.

[0083] According to this rule, a new "generated parameter semantics" can be generated from the "parameter semantics library". The "generated parameter semantics" is based on the distribution characteristics and probability density of the "parameter semantics library", and extracts a new "parameter semantics" from the high-probability effective area. The "generated parameter semantics" has structural consistency with the original parameter semantics, but is not a direct copy, and has "regularity" and "novelty".

[0084] The specific process of generating the model parameters is shown in Figure 7 A new parameter semantics is generated by using a method based on kernel density estimation (KDE). KDE is a non-parametric statistical method that does not need to pre-set the mathematical form of the data distribution, and can adaptively fit the probability density function through the distribution characteristics of the data itself, and has the characteristics of high flexibility, simple implementation and strong determinacy. The core idea is that for each data point, a Gaussian kernel is used to "smear" the point, and then all the kernel functions are superimposed to form the overall probability density estimation. In the construction of the "parameter semantics", the KDE modeling is performed on each dimension of the "parameter semantics" separately. Kernel function selection: Gaussian kernel function is used, and the formula is: ​The smoothness and symmetry can effectively capture the local distribution characteristics of data while avoiding introducing too much model bias.

[0085] Bandwidth parameter (h): As the only hyperparameter in KDE, bandwidth controls the "width" of the kernel function. Too small bandwidth will cause overfitting (noise fluctuation) of the estimation result, and too large bandwidth will cause over-smoothing (loss of local details). The optimal bandwidth is selected by cross-validation method to balance bias and variance.

[0086] Density estimation process: for the "parameter semantics" of the first dimension , the probability density function estimation expression is: , the probability density function estimation expression is: The formula shows that the probability density of any point is the weighted sum of all anchor samples through the Gaussian kernel function, and the weight is determined by the distance between the sample and , that is, the closer the distance, the greater the contribution.

[0087] Through the above steps, the marginal distribution of each dimension of the "parameter semantics" can be obtained by the embodiment of the application. When constructing, samples are independently extracted from the KDE distribution of each dimension, and then the samples of each dimension are spliced to generate a complete "generated parameter semantics" .

[0088] The "generated parameter semantics" is obtained through the channel . The received vector is sent to the trained source-channel joint decoding module and the semantic decoding module based on the multi-scale attention mechanism structure for decoding , to obtain the target "generated model parameter" . Load the target AI model parameter into the target AI model, and use the local data set for lightweight fine-tuning to adapt to the local data set. There is no need to disclose the local data set to the edge server, which realizes privacy protection, and lightweight fine-tuning can effectively reduce the consumption of computing resources.

[0089] The embodiment of the application also provides an electronic device, which can be a server or an edge server. The electronic device can include a processor, a memory, a receiver and a transmitter. The processor is used to execute the above-mentioned semantic-based AI model parameter transmission method.

[0090] ​The electronic device can be an end-side device such as a mobile phone, a computer, a vehicle-mounted device, and the like. The electronic device can include a processor, a memory, a receiver, and a transmitter. The processor is configured to execute the method for generating AI model parameters based on semantics. The processor and the memory can be connected by a bus or other means. The receiver can be connected to the processor and the memory by wired or wireless means.

[0091] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination thereof.

[0092] The memory is a non-transitory computer-readable storage medium, which can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as program instructions / modules corresponding to the method for transmitting AI model parameters based on semantics or the method for generating AI model parameters based on semantics. The processor executes various functions of the processor and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, that is, implements the method for transmitting AI model parameters based on semantics or the method for generating AI model parameters based on semantics.

[0093] The memory can include a program storage area and a data storage area. The program storage area can store an operating system and at least one application required by a function. The data storage area can store data created by the processor, and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory remotely disposed relative to the processor, which can be connected to the processor through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0094] The one or more modules are stored in the memory and executed by the processor to implement the method for transmitting AI model parameters based on semantics or the method for generating AI model parameters based on semantics.

[0095] In some embodiments of the present application, the user equipment can include a processor, a memory, and a transceiver which can include a receiver and a transmitter, the processor, the memory, the receiver and the transmitter can be connected through a bus system, the memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver to transceive signals.

[0096] As an implementation manner, the functions of the receiver and the transmitter in the present application can be implemented by a transceiving circuit or a dedicated transceiving chip, and the processor can be implemented by a dedicated processing chip, a processing circuit or a general-purpose chip.

[0097] As another implementation manner, the server provided by the embodiments of the present application can be implemented by using a general-purpose computer. That is, the program codes for implementing the functions of the processor, the receiver and the transmitter are stored in the memory, and the general-purpose processor implements the functions of the processor, the receiver and the transmitter by executing the codes in the memory.

[0098] In order to further illustrate the various embodiments of the above-mentioned server-executed semantic-based AI model parameter transmission method and the end-side device-executed semantic-based AI model parameter generation method, the present application further provides an application example of a semantic-based AI model parameter transmission and generation method implemented by the interaction between the server and the end-side device, which is described in detail in the following Figure 8 , and specifically includes the following contents: S1: The end-side device sends a target AI model deployment request to the server; S2: The server receives the target AI model deployment request, matches the task description data with the description data corresponding to the locally preset target data set, and if the task description data matches the description data corresponding to the locally preset target data set, filters the model performance tags corresponding to each of the encoding parameter semantics in the model encoding parameter semantic library according to the model performance tags corresponding to each of the encoding parameter semantics in the model encoding parameter semantic library, and selects the encoding parameter semantics in the model encoding parameter semantic library that are consistent with the downstream task requirement data as target parameter semantics; if the task description data does not match the description data corresponding to the locally preset target data set, models the distribution of the encoding parameter semantics in the model encoding parameter semantic library based on a kernel density estimation method, and samples from the probability density function obtained by modeling based on the downstream task requirement data to generate target parameter semantics; S3: The server sends the target parameter semantics to the end-side device; S4: The end-side device receives the target parameter semantics and decodes the target parameter semantics to generate target AI model parameters applied to the downstream task.

[0099] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the steps of the aforementioned AI model parameter transmission method based on semantics or the AI model parameter generation method based on semantics. The computer readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable disk, a CD-ROM, or any other form of storage known in the art.

[0100] The embodiments of the present application also provide a computer program product, which includes a computer program. The computer program is executed by a processor to implement the steps of the aforementioned AI model parameter transmission method based on semantics or the AI model parameter generation method based on semantics.

[0101] Those skilled in the art should understand that the exemplary components, systems and methods described in connection with the embodiments disclosed herein can be implemented in hardware, software or a combination thereof. The choice of hardware or software implementation is dependent on the particular application and design constraints imposed on the solution. Skilled persons can use different methods to implement the described functions for each particular application, but such implementation should not be considered beyond the scope of the present application. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, etc. When implemented in software, the elements of the present application are program or code segments used to perform the required tasks. The program or code segments can be stored in a machine readable medium or transmitted through a data signal carried in a carrier wave in a transmission medium or communication link.

[0102] It should be noted that the present application is not limited to the specific configurations and processes described above and shown in the drawings. For the sake of brevity, detailed descriptions of well-known methods are omitted. In the above embodiments, several specific steps are described and shown as examples. However, the method processes of the present application are not limited to the specific steps described and shown, and skilled persons can make various changes, modifications and additions, or change the order of the steps, after understanding the spirit of the present application.

[0103] In the present application, the features described and / or illustrated for one embodiment can be used in the same way or in a similar way in one or more other embodiments, and / or in combination with or instead of features of other embodiments.

[0104] The above descriptions are only the preferred embodiments of the present application, and are not intended to limit the present application. The embodiments of the present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of protection of the present application.

Claims

1. A semantic-based AI model parameter transmission method, characterized in that, The transmission method includes: The system receives a target AI model deployment request sent by an edge device, wherein the target AI model deployment request includes task description data describing the characteristics of the local dataset of the edge device and downstream task requirement data for the target AI model. The task description data is matched with the description data corresponding to the local preset target dataset. Based on the matching result, the target parameter semantics that meet the downstream task requirements are selected from the preset model encoding parameter semantic library corresponding to the target dataset. Alternatively, the target parameter semantics that meet the downstream task requirements are generated based on the model encoding parameter semantic library. The target parameter semantics are transmitted to the edge device so that the edge device can decode the target parameter semantics to generate target AI model parameters for downstream tasks. 2.The semantic-based AI model parameter transmission method of claim 1, wherein, Before the target AI model deployment request sent by the receiving end device, the following is also included: Semantic encoding is performed on the parameters of multiple AI models trained on the target dataset to obtain encoded parameter semantics that characterize the model weight distribution pattern and / or feature extraction capability. The semantics of each of the encoding parameters are summarized and associated with the corresponding model performance labels to form the model encoding parameter semantic library. 3.The semantic-based AI model parameter transmission method of claim 2, wherein, The step of matching the task description data with the description data corresponding to a locally preset target dataset, and based on the matching result, filtering target parameter semantics that meet the downstream task requirements from a preset model encoding parameter semantic library corresponding to the target dataset, or generating target parameter semantics that meet the downstream task requirements based on the model encoding parameter semantic library, includes: The task description data is matched with the description data corresponding to the locally preset target dataset; If the task description data matches the description data corresponding to the local preset target dataset, then based on the model performance labels corresponding to each of the encoded parameter semantics in the model encoding parameter semantic library, the encoded parameter semantics whose model performance labels match the downstream task requirement data are selected as target parameter semantics. If the task description data does not match the description data corresponding to the local preset target dataset, a kernel density estimation method is used to model the distribution of the encoding parameter semantics in the model encoding parameter semantic library, and samples are taken from the probability density function obtained from the modeling based on the downstream task requirement data to generate the target parameter semantics. 4.The semantic-based AI model parameter transmission method of claim 2, wherein, The step of semantically encoding the parameters of multiple AI models trained on the target dataset to obtain encoded parameter semantics used to characterize the model weight distribution pattern and / or feature extraction capability includes: The parameters of multiple AI models trained on the target dataset are converted into serialization representations to obtain the serialization parameters corresponding to each AI model parameter. A semantic extraction model based on a multi-scale attention mechanism is adopted to extract the original parameter semantics corresponding to each AI model parameter from each of the serialization parameters, which are used to characterize the model weight distribution pattern and / or feature extraction capability. Using a pre-defined source and channel joint coding model, the semantics of each of the original parameters are jointly coded by the source and channel to obtain the coded parameter semantics corresponding to each of the original parameter semantics. 5.The semantic-based AI model parameter transmission method of claim 4, wherein, The end-side device is used to decode the target parameter semantic data using a preset source and channel joint decoding model to generate target AI model parameters for application to downstream tasks; Correspondingly, before the target AI model deployment request sent by the receiving end device, the following is also included: Based on the loss function, the source and channel joint decoding model and the source and channel joint coding model are jointly trained end-to-end; wherein, the loss function includes a reconstruction loss function based on layer normalization and a contrastive loss function; The reconstruction loss function based on layer normalization is shown in Equation (1): In formula (1), To rebuild the losses, Indicates the first Layer The reconstructed values ​​of each parameter, Indicates the first Layer The target value of each parameter Indicates the first The mean of the layer parameters, Indicates the first Standard deviation of layer parameters This indicates the total number of layers in the model. Indicates the first The number of parameters in the layer This represents the total number of parameters across all layers and is used for global normalization. 6.The semantic-based AI model parameter transmission method of claim 5, wherein, The contrastive loss function is optimized by constructing positive and negative sample pairs; The positive sample pair refers to two functionally equivalent views obtained by performing different enhancement operations on the same AI model parameters; The negative sample pairs are combinations of augmented views from different AI model parameters. 7.The semantic-based AI model parameter transmission method of claim 5, wherein, The edge device is also used to perform lightweight fine-tuning of the target AI model parameters, wherein the lightweight fine-tuning includes: updating some layers or some parameters in the target AI model parameters; and / or updating the target AI model parameters using fewer than a preset number of fine-tuning rounds. 8.A method for generating parameters of an AI model based on semantics, the method comprising: include: A target AI model deployment request is sent to the server, wherein the target AI model deployment request includes task description data for describing the features of the local dataset and downstream task requirement data for the target AI model; so that the server performs the following: matching the task description data with the description data corresponding to the local preset target dataset, and according to the matching result, filtering target parameter semantics that match the downstream task requirement data from the preset model encoding parameter semantic library corresponding to the target dataset, or generating target parameter semantics that match the downstream task requirement data based on the model encoding parameter semantic library; The target parameter semantics are received from the server, and the target parameter semantics are decoded to generate target AI model parameters for application to downstream tasks. 9.The semantic-based AI model parameter generation method of claim 8, wherein, Also includes: Lightweight fine-tuning of the target AI model parameters includes: updating some layers or some parameters in the target AI model parameters; and / or updating the generated model parameters using fewer than a preset number of fine-tuning rounds.

10. An electronic device, comprising: It includes a processor and a memory; when the processor executes the running program stored in the memory, it implements the semantic-based AI model parameter transmission method as described in claims 1 to 7 or the semantic-based AI model parameter generation method as described in claim 8 or 9.

Images