Update processing method and device, equipment, readable storage medium and product

Abstract

The invention provides an update processing method and device, equipment, a readable storage medium and a product, and relates to the technical field of communication. The method is executed by first equipment and comprises the following steps: acquiring a personal data set of a target user; determining a general data set according to the personal data set and a target model; wherein a first parameter is added to an important layer of the target model; and updating the first parameter according to the personal data set and the general data set. According to the scheme of the application, the problem of leakage risk existing in an end-cloud collaborative optimization scheme adopted by the target model is solved.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to an update processing method, device, equipment, readable storage medium and product. Background Art

[0002] As computing resources on the end-side increase, large models are expected to be deployed on the end-side to achieve a better user experience with low latency, high reliability, personalization, low cost, and privacy protection, and to enrich terminal product functions and services. Among them, the optimization of large models on the end-side is achieved through end-cloud collaboration, that is, the cloud side is responsible for large model training and personalization adjustment, and the end side is responsible for large model reasoning.

[0003] However, the optimization solution of end-cloud collaboration is prone to leakage risks because it requires transmitting data and models to the cloud side. Summary of the invention

[0004] The purpose of this application is to provide an update processing method, device, equipment, readable storage medium and product to avoid the leakage risk of the end-cloud collaborative optimization solution.

[0005] To achieve the above object, an embodiment of the present application provides an update processing method, which is executed by a first device and includes:

[0006] Obtain the target user’s personal data set;

[0007] Determine a general data set according to the personal data set and the target model; wherein a first parameter is added to an important layer of the target model; the important layer is a layer in a pre-built model of the target model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons;

[0008] The first parameter is updated according to the personal data set and the general data set.

[0009] Optionally, obtaining the personal data set of the target user includes:

[0010] Collecting first interaction data of the target user;

[0011] Acquire context information of the first interaction data;

[0012] Integrate the first interaction data based on the context information to obtain second interaction data;

[0013] The personal data set is generated based on the second interaction data.

[0014] Optionally, determining a common data set according to the personal data set and the target model includes:

[0015] Inputting the second interaction data into the target model to obtain third interaction data output by the target model;

[0016] The general data set is generated based on the second interaction data and the third interaction data.

[0017] Optionally, the updating the first parameter according to the personal data set and the general data set includes:

[0018] Adding the personal data set and the general data set to the training set of the target model;

[0019] Training a first parameter of the target model based on the training set;

[0020] The second parameter of the target model remains unchanged during the training process.

[0021] Optionally, the method further comprises:

[0022] Obtaining input data of the target user;

[0023] The input data is input into the target model after the first parameter is updated to obtain target data.

[0024] Optionally, in the target model after the first parameter is updated,

[0025] The important layer processes the input according to the updated first parameter and the fixed second parameter;

[0026] Non-significant layers process the input according to a fixed second parameter.

[0027] To achieve the above object, an embodiment of the present application provides an update processing method, which is executed by a second device and includes:

[0028] Identify important layers of pre-built models;

[0029] Add the first parameter in each important layer to obtain the target model;

[0030] The important layer is a layer in the pre-built model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons.

[0031] Optionally, determining the important layers of the pre-built model includes:

[0032] Based on the test data set, obtaining a first result output by each layer in the pre-built model and a second result output by each layer in the test model; wherein the test model is obtained after adding a third parameter to each layer of the pre-built model;

[0033] The important layer is determined based on a similarity between the first result and the second result.

[0034] Optionally, the method further comprises:

[0035] Before obtaining the second result based on the test data set, the test model is trained using the test data set and the target data set.

[0036] To achieve the above object, an embodiment of the present application provides an update processing device, including:

[0037] The first acquisition module is used to acquire the personal data set of the target user;

[0038] A first processing module is used to determine a general data set according to the personal data set and the target model; wherein a first parameter is added to an important layer of the target model; the important layer is a layer in a pre-built model of the target model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons;

[0039] The second processing module is used to update the first parameter according to the personal data set and the general data set.

[0040] Optionally, the first acquisition module is further used for:

[0041] Collecting first interaction data of the target user;

[0042] Acquire context information of the first interaction data;

[0043] Integrate the first interaction data based on the context information to obtain second interaction data;

[0044] The personal data set is generated based on the second interaction data.

[0045] Optionally, the first processing module is further used for:

[0046] Inputting the second interaction data into the target model to obtain third interaction data output by the target model;

[0047] The general data set is generated based on the second interaction data and the third interaction data.

[0048] Optionally, the second processing module is further used for:

[0049] Adding the personal data set and the general data set to the training set of the target model;

[0050] Training a first parameter of the target model based on the training set;

[0051] The second parameter of the target model remains unchanged during the training process.

[0052] Optionally, the device further comprises:

[0053] A second acquisition module, used to acquire input data of the target user;

[0054] The fifth processing module is used to input the input data into the target model after updating the first parameter to obtain target data.

[0055] Optionally, in the target model after the first parameter is updated,

[0056] The important layer processes the input according to the updated first parameter and the fixed second parameter;

[0057] Non-significant layers process the input according to a fixed second parameter.

[0058] To achieve the above object, an embodiment of the present application provides an update processing device, including:

[0059] a third processing module for determining important layers of the pre-built model;

[0060] A fourth processing module is used to add the first parameter to each important layer to obtain a target model;

[0061] The important layer is a layer in the pre-built model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons.

[0062] Optionally, the third processing module is further used for:

[0063] Based on the test data set, obtaining a first result output by each layer in the pre-built model and a second result output by each layer in the test model; wherein the test model is obtained after adding a third parameter to each layer of the pre-built model;

[0064] The important layer is determined based on a similarity between the first result and the second result.

[0065] Optionally, the device further comprises:

[0066] A sixth processing module is used to train the test model using the test data set and the target data set before obtaining the second result based on the test data set.

[0067] To achieve the above-mentioned purpose, an embodiment of the present application provides an update processing device, including a transceiver, a processor, a memory, and a program or instruction stored in the memory and executable on the processor; when the processor executes the program or instruction, the update processing method as described above is implemented.

[0068] To achieve the above objectives, an embodiment of the present application provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps in the update processing method as described above are implemented.

[0069] To achieve the above objective, an embodiment of the present application provides a computer program product, including computer instructions, which implement the steps of the update processing method described above when executed by a processor.

[0070] The beneficial effects of the above technical solution of the present application are as follows:

[0071] In the method of the embodiment of the present application, after the first device obtains the personal data set of the target user, it can determine the general data set in combination with the target model, and then update the first parameter added to the important layer in the target model according to the personal data set and the general data set. In this way, there is no need to update the parameters of all layers of the target model. It is only necessary to add the first parameter to the important layer of the target model (the layer with an output similarity greater than a preset threshold in the pre-built model based on the target model, and the layer with a greater impact on the output result), and update the first parameter. Because the important layer is the transformer layer that must be calculated in the large model reasoning, it has an important impact on the reasoning result. In contrast, the calculation of the non-important layer has little effect on the reasoning result, and the calculation can be omitted to reduce computing resources and reasoning delay. The first parameter is a new connection parameter generated by one or more hidden layer neurons. Compared with the update of all parameters of the original model, the model adjustment can be realized faster, which significantly reduces the computing power and data requirements for model optimization, does not require cloud collaboration, and reduces the risk of leakage. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is one of the flowcharts of the update processing method according to an embodiment of the present application;

[0073] Figure 2 This is a schematic diagram of the update process in the embodiment of the present application;

[0074] Figure 3 This is a reasoning flow chart in the embodiment of this application;

[0075] Figure 4 This is a system architecture diagram of the first device in the embodiment of the present application;

[0076] Figure 5This is a second flowchart of the update processing method of an embodiment of the present application;

[0077] Figure 6 This is one of the module structure diagrams of the update processing device according to an embodiment of the present application;

[0078] Figure 7 This is a second schematic diagram of the module structure of the update processing device according to an embodiment of the present application;

[0079] Figure 8 A structural diagram of an update processing device according to an embodiment of the present application;

[0080] Fig. 9 This is a structural diagram of an update processing device according to another embodiment of the present application. DETAILED DESCRIPTION

[0081] In order to make the technical problems, technical solutions and advantages to be solved by the present application clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0082] It should be understood that the references to "one embodiment" or "an embodiment" throughout the specification mean that the specific features, structures, or characteristics associated with the embodiment are included in at least one embodiment of the present application. Therefore, the references to "in one embodiment" or "in an embodiment" appearing throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0083] In the various embodiments of the present application, it should be understood that the size of the serial numbers of the following processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0084] Additionally, the terms "system" and "network" are often used interchangeably herein.

[0085] In the embodiments provided in the present application, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined according to A. However, it should also be understood that determining B according to A does not mean determining B only according to A, and B can also be determined according to A and / or other information.

[0086] For ease of understanding, some contents involved in the embodiments of the present application are described below:

[0087] The end-side big model (referred to as the big model in this application), also known as the personalized big model, can provide more personalized services based on the user's personal data and behavioral habits, thereby enhancing the differentiation of terminal products.

[0088] The storage-computing integrated architecture reduces the data transfer process through in-memory computing, thereby improving performance and efficiency, providing high-performance, energy-efficient computing infrastructure for terminals, supporting local updates and deployment of large models on the terminal side, and improving product differentiation. The storage-computing integrated architecture can be divided into analog storage-computing solutions and digital storage-computing solutions. Among them, the analog storage-computing solution uses resistive variable devices to realize calculations through analog quantities such as voltage and current, without the need for additional digital multiplication and addition calculation units. The digital storage-computing solution uses logical computing devices such as multipliers and addition trees to realize matrix multiplication and addition calculations, and achieves minimum transmission by placing logical computing devices close to the storage array without loss of calculation accuracy.

[0089] Among them, the storage and computing integrated architecture can be implemented as a storage and computing integrated chip.

[0090] Smart speakers / smart mid-screen: The integrated storage and computing chip enables smart speakers to run and update large models locally, thereby providing faster and more accurate voice recognition and natural language processing capabilities. Users can interact with smart speakers more naturally and smoothly, and receive more precise services. For example, smart speakers can deeply understand users' preferences and needs, and provide users with personalized recommendations, such as songs and articles, to enhance the user experience.

[0091] Smart cockpit: The big model on the end can help the smart cockpit achieve applications such as safety improvement, complex voice interaction, and autonomous driving assistance. Based on the personalized big model, the smart cockpit can provide personalized voice recognition, entertainment information, and driving assistance services for thousands of people to meet the personalized needs of passengers and drivers.

[0092] Home robots: The large end-side model can help robots achieve real-time interaction and decision-making, complex task processing, and diverse application expansion. By learning users' living habits and preferences, home robots can provide personalized services such as home cleaning, elderly care, and children's education.

[0093] The optimization solution for end-to-end collaboration of large models on the end-side is that, first of all, the use of personal data to optimize large models poses privacy leakage risks at both the data and model levels. Secondly, the number of parameters in large models on the end-side is as high as hundreds of millions to tens of billions, and the transmission of models to the cloud side requires high network bandwidth, especially in environments with poor network conditions. Therefore, the optimization solution for end-to-end collaboration of large models on the end-side is subject to data leakage, network bandwidth limitation, and cloud-side resource crowding.

[0094] like Figure 1 As shown, an update processing method in an embodiment of the present application is executed by a first device, including:

[0095] Step 101, obtaining a personal data set of a target user;

[0096] Step 102, determining a general data set according to the personal data set and the target model; wherein a first parameter is added to an important layer of the target model; the important layer is a layer in a pre-built model of the target model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons;

[0097] Step 103: Update the first parameter according to the personal data set and the general data set.

[0098] Here, the output similarity of a layer of the pre-built model refers to the similarity between the output result of the layer and the target of the layer.

[0099] Thus, according to the above steps 101-103, after the first device obtains the personal data set of the target user, it can determine the general data set in combination with the target model, and then update the first parameter added to the important layer in the target model according to the personal data set and the general data set. In this way, there is no need to update the parameters of all layers of the target model, only the first parameter needs to be added to the important layer of the target model (determined based on the similarity between the output of each layer of the pre-built model of the target model and the target, and the layer with a greater impact on the output result, that is, the layer with an output similarity greater than a preset threshold) and the first parameter is updated. Because the important layer is the transformer layer that must be calculated in the large model reasoning, it has an important impact on the reasoning result. In contrast, the calculation of the non-important layer has little effect on the reasoning result, and the calculation can be omitted to reduce computing resources and reasoning delay. The first parameter is a new connection parameter generated by one or more hidden layer neurons. Compared with the update of all parameters of the original model, the model adjustment can be realized faster, which significantly reduces the computing power and data requirements for model optimization, does not require cloud collaboration, and reduces the risk of leakage.

[0100] The number of the first parameters is smaller than a specific value to achieve a small number of parameter updates.

[0101] Optionally, in this embodiment, the first device is a terminal device, and the target model is a large model on the terminal side.

[0102] It should be noted that, in this embodiment, the basic structure of the end-side large model (hereinafter referred to as the large model) is a transformer, which consists of three parts: encoding, multiple layers (also called transformation layers (transformer block)) and decoding, wherein the transformer block consists of self-attention and a feedforward neural network (also called a forward neural network). The self-attention module is used to process the context, the feedforward neural network is used for prediction, the feedforward neural network can be regarded as a special cross-attention mechanism, the network connection parameters are equivalent to the knowledge base, the integrated context information is input into the knowledge base query, and the integrated related information is output. Therefore, adjusting the feedforward neural network can adjust the internal knowledge base of the model used for prediction from the source domain to the target domain, and training the newly added parameters through personalized data can adjust the internal knowledge base of the model from the public domain to the user domain. Therefore, optionally, in this embodiment, the first parameter added to the important layer is to add the first parameter to the feedforward neural network of the important layer. The first parameter can be a small number of parameters.

[0103] Of course, the target model also includes a fixed second parameter.

[0104] As an optional implementation method, the first parameter is a new connection parameter generated by adding a small number of hidden layer neurons to the hidden layer of the feedforward network (FFN) in each layer of the large model on the end, and the second parameter is the original network parameter of the large model on the end.

[0105] In addition, optionally, in this embodiment, the step of obtaining the personal data set of the target user includes:

[0106] Collecting first interaction data of the target user;

[0107] The personal data set is generated based on the first interaction data.

[0108] That is, the collected first interaction data can be directly used as the data in the personal data set.

[0109] On the other hand, optionally, in this embodiment, the step of obtaining the personal data set of the target user includes:

[0110] Collecting first interaction data of the target user;

[0111] Acquire context information of the first interaction data;

[0112] Integrate the first interaction data based on the context information to obtain second interaction data;

[0113] The personal data set is generated based on the second interaction data.

[0114] Here, after collecting the first interaction data, the context information of the first interaction data is further obtained to obtain second interaction data with higher quality through information integration, and then the second interaction data is used as the data in the personal data set. After the first parameter is updated according to the personal data set, the target model can increase its understanding of the target user data or behavioral habits, thereby improving the personalization of the target model.

[0115] It should be known that the first interaction data is the interaction data X input by the target user. Corresponding to the first interaction data, the result Y output by the end-side large model can also be collected. Therefore, optionally, the collected data set in this embodiment includes one or more groups of {X, Y}.

[0116] Optionally, the acquiring context information of the first interaction data includes:

[0117] The context information is retrieved from a knowledge base based on the first interaction data.

[0118] The knowledge base may be constructed based on the historical behavior habits of the target user, or may be an Internet knowledge base.

[0119] After the context information P is acquired, the second interaction data X' is obtained by integrating X and P. Therefore, optionally, the personal data set in this embodiment includes one or more sets of {X', Y}.

[0120] Optionally, in this embodiment, determining the general data set according to the personal data set and the target model includes:

[0121] Inputting the second interaction data into the target model to obtain third interaction data output by the target model;

[0122] The general data set is generated based on the second interaction data and the third interaction data.

[0123] That is, after obtaining the second interaction data X', X' is input into the target model to obtain output third interaction data Z, so that the general data set includes one or more sets of {X', Z}.

[0124] Thus, in this embodiment, one or more sets of {X', Y} and one or more sets of {X', Z} are used to update the first parameter of the target model.

[0125] Optionally, the updating the first parameter according to the personal data set and the general data set includes:

[0126] Adding the personal data set and the general data set to the training set of the target model;

[0127] Training a first parameter of the target model based on the training set;

[0128] The second parameter of the target model remains unchanged during the training process.

[0129] That is to say, the data of the personal data set and the general data set can be added to the original training set of the target model, or the personal data set and the general data set can be used as the new training set of the target model to retrain the target model and calculate the difference between the output and the target. However, during the training process, the second parameter is fixed and only the first parameter is updated.

[0130] The training process may use error back propagation to train the first parameter.

[0131] Optionally, the second parameter of the target model is deployed in an analog compute-in-memory (CIM) array, and the newly added parameters of the large model are deployed in a digital CIM array to avoid the problem of limited erase and write times of the analog CIM.

[0132] Optionally, in this embodiment, the method further includes:

[0133] Obtaining input data of the target user;

[0134] The input data is input into the target model after the first parameter is updated to obtain target data.

[0135] That is, when the target model is applied, the update of the first parameter is completed through the above steps, and the input data of the target user is input into the target model (also called the new target model) after the first parameter update, and the output result of the target model is the target data.

[0136] For example, the first device is a smart speaker (the smart speaker applies a target model), and the target user asks question 1 to the smart speaker. Question 1 is the input data. After question 1 is input into the new target model, the target model processes question 1, generates and outputs target data.

[0137] Specifically, the processing process of the new target model is: layer by layer calculation, block layer output Xi, for the feedforward neural network calculation of non-important block layers, only (also known as the general large model parameter) is calculated based on the second parameter; for the feedforward neural network calculation of important block layers, it is calculated based on the first parameter and the second parameter. After that, the two parts of the output results are integrated and then the subsequent transformation operation is performed.

[0138] After all layers are calculated, they are decoded and output as new tokens, that is, the output data (predicted word segmentation) of the target model will become new tokens. It is determined whether the token is an end symbol. If not, the new token is added to the input, and the new target model performs the above processing again.

[0139] It should also be known that before the target user's input data is input into the new target model, it can be matched in the knowledge base to obtain its context information, and the input data and its context information are integrated to input the integrated expanded input data into the new target model.

[0140] The data of the input target model and the new target model may be first segmented and encoded and then input.

[0141] In addition, it should be noted that the important layers of the target model are determined by the second device, and the second device determines the important layers of the pre-built model; the first parameter is added to each important layer to obtain the target model. Here, the pre-built model can also be understood as the source target model.

[0142] In this embodiment, considering that the target model is composed of multiple transformer blocks, the shallow transformer infers simple concepts, and the deep transformer infers complex concepts, the influence of different layers on the domain migration of the large model is determined by weight analysis, and the second device determines whether it is an important layer based on the similarity between the output of each layer of the pre-built model and the target. Therefore, optionally, the determination of the important layers of the pre-built model includes:

[0143] Based on the test data set, obtaining a first result output by each layer in the pre-built model and a second result output by each layer in the test model; wherein the test model is obtained after adding a third parameter to each layer of the pre-built model;

[0144] The important layer is determined based on a similarity between the first result and the second result.

[0145] Here, the first result is the layer output result and the second result is the layer target.

[0146] In this way, by comparing the output results of each layer in the pre-built model and the test model with the same input, it is possible to determine the important layers and the unimportant layers. For example, a preset threshold is set to determine whether the output similarity is greater than the preset threshold. The layers whose output similarity is greater than the preset threshold are important layers, and the layers whose output similarity is not greater than the preset threshold are unimportant layers. The selected set of important layers is recorded as L.

[0147] The similarity calculation can be implemented by using algorithms such as cosine similarity and Pearson correlation coefficient.

[0148] The test data set includes one or more U, where U is constructed input data.

[0149] The test model is a model obtained by adding the third parameter to each layer of the pre-built model. Specifically, a third parameter is added to each layer of the feedforward neural network, and the third parameter is a small amount. The third parameter can be a new connection parameter generated by adding one or more hidden layer neurons to the hidden layer of the feedforward neural network FFN, and the first parameter is the valid parameter left after the third parameter is trained (that is, the increase of this parameter affects the end-side large model reasoning).

[0150] Optionally, before obtaining the second result based on the test data set, the second device trains the test model using the test data set and the target data set.

[0151] Here, the target dataset includes one or more V', where V' is the target data constructed corresponding to U. One or more sets of {U, V'} can be called the target domain dataset. Inputting U into the pre-built model can obtain the output result V, and one or more sets of {U, V} are the source domain datasets. V'≠V, and {U, V'} obeys a certain distribution.

[0152] Generally speaking, the second device determines the important layers of the pre-built model as a one-time operation, which is completed before the model is shipped, and does not need to occupy the end-side computing resources. Of course, the first device can also execute the update processing method of the second device to process the pre-built model to obtain the target model.

[0153] So far, from the above content, it can be known that an optional implementation method is to update the target model (update the first parameter) as follows Figure 2 As shown:

[0154] S1, the second device screens the blocks that have a greater impact on the migration of large model domains, which can also be understood as determining the important layers of the pre-built model.

[0155] S2, the second device adds a small amount of first parameters to the feedforward neural network in the important layer set L screened out in S1.

[0156] S3, the first device constructs a training set through interaction data and knowledge base: collects first interaction data X, obtains context information P of X through matching through knowledge base, integrates X and P to obtain X', inputs X' into the target model to obtain a general data set; determines the training set based on the personal data set and the general data set.

[0157] S4, using the training set to update the first parameter added in S2.

[0158] An optional implementation method is to update the inference process of the target model as follows: Figure 3 As shown:

[0159] S1', obtain input data, match relevant context information from the knowledge base and integrate it with the input data to obtain personalized input. Specifically, send the input data P of the target user into the knowledge base for matching to obtain relevant context information (prompt information) Q, and integrate the input data with the context information to obtain personalized input P'.

[0160] S2', segment and encode P' and input it into the target model.

[0161] S3', the target model is calculated layer by layer, the output results of each layer are integrated, and then the subsequent transformation operation is performed. Specifically, for the feedforward neural network calculation of non-important layers, only the second parameter is used for calculation; for the feedforward neural network calculation of important block layers, the second parameter and the updated first parameter are used for calculation, and the output results of the two parts are integrated.

[0162] S4', after all layers are calculated, decode and output a new token.

[0163] S5', determine whether the output token is the end character, if not, return to S3'.

[0164] In summary, the update processing method of the embodiment of the present application adjusts the model by adding a small number of parameters to the partial structure of the large model, and only updates the FFN module parameters of the important layers, which significantly reduces the computing power and data requirements for fine-tuning the large model. Among them, for personal interaction data, the large model is fine-tuned by adding contextual information, so that the end-side large model can better understand personal data or behavioral habits.

[0165] It should also be noted that if Figure 4 As shown, the reasoning component included in the first device in the present application executes the above-mentioned update processing method, and the reasoning component is divided into: data collection sub-component, knowledge base sub-component, prompt sub-component, general reasoning sub-component, update sub-component, and output sub-component. Among them, the data collection sub-component is used for personal user information collection, mainly including known user information and user data collected during the interaction process. The known user information can be user identity information, address information, etc. The user data collected during the interaction process can be the first interaction data, the result of the output of the large model on the end side corresponding to the first interaction data, etc. The knowledge base sub-component is used to process the collected data, and the refined user context information is vectorized and stored in the knowledge base. The prompt sub-component stores and retrieves the knowledge base, retrieves relevant knowledge from the knowledge base according to the input information, and transmits the knowledge to the general reasoning sub-component after merging the input. The general reasoning sub-component stores and calculates the second parameter and outputs the general domain reasoning result. The update sub-component stores and calculates the first parameter for model personalization and outputs the personalized domain reasoning result. The output sub-component integrates the calculation results of the second parameter and the first parameter, and outputs the final reasoning value.

[0166] Among them, the prompt subcomponent and the reasoning subcomponent adopt the analog storage and computing integrated computing architecture. The update subcomponent adopts the digital storage and computing integrated computing architecture with two-way propagation. In this way, the hybrid computing architecture of digital storage and computing integrated and analog storage and computing integrated is integrated. The analog storage and computing integrated is deployed and calculated according to the second parameter, and the digital storage and computing integrated is deployed and calculated according to the first parameter, and the first parameter is updated online, thereby achieving high-performance and low-power large model end-side reasoning and update.

[0167] like Figure 5 As shown, an update processing method in an embodiment of the present application is executed by a second device, including:

[0168] Step 501, determining the important layers of the pre-built model;

[0169] Step 502, adding the first parameter to each important layer to obtain a target model;

[0170] The important layer is a layer in the pre-built model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons.

[0171] In this way, after the second device confirms the important layer of the pre-built model and adds the first parameter, it can obtain the target model with the first parameter added to the important layer, so that after the first device obtains the personal data set of the target user, it can determine the general data set in combination with the target model, and then update the first parameter added to the important layer in the target model according to the personal data set and the general data set. In this way, there is no need to update the parameters of all layers of the target model. It is only necessary to add the first parameter to the important layer of the target model (determined based on the similarity between the output of each layer of the pre-built model of the target model and the target, the layer with a greater impact on the output result, and the layer with an output similarity greater than a preset threshold) and update the first parameter. Because the important layer is a transformer layer that must be calculated in the reasoning of the large model, it has an important impact on the reasoning result. In contrast, the calculation of the non-important layer has little effect on the reasoning result, and the calculation can be omitted to reduce computing resources and reasoning delay. The first parameter is a new connection parameter generated by one or more hidden layer neurons. Compared with the update of all parameters of the original model, the model adjustment can be realized faster, which significantly reduces the computing power and data requirements for model optimization, and does not require cloud collaboration, reducing the risk of leakage.

[0172] Optionally, determining the important layers of the pre-built model includes:

[0173] Based on the test data set, obtaining a first result output by each layer in the pre-built model and a second result output by each layer in the test model; wherein the test model is obtained after adding a third parameter to each layer of the pre-built model;

[0174] The important layer is determined based on a similarity between the first result and the second result.

[0175] Optionally, the method further comprises:

[0176] Before obtaining the second result based on the test data set, the test model is trained using the test data set and the target data set.

[0177] It should be noted that this method is implemented in conjunction with the above-mentioned update processing method executed by the first device. The implementation method of the above-mentioned method embodiment is applicable to this method and can also achieve the same technical effect.

[0178] like Figure 6 As shown, an update processing device in an embodiment of the present application includes:

[0179] A first acquisition module 610 is used to acquire a personal data set of a target user;

[0180] A first processing module 620 is used to determine a general data set based on the personal data set and the target model; wherein a first parameter is added to an important layer of the target model; the important layer is a layer in a pre-built model of the target model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons;

[0181] The second processing module 630 is configured to update the first parameter according to the personal data set and the general data set.

[0182] Optionally, the first acquisition module is further used for:

[0183] Collecting first interaction data of the target user;

[0184] Acquire context information of the first interaction data;

[0185] Integrate the first interaction data based on the context information to obtain second interaction data;

[0186] The personal data set is generated based on the second interaction data.

[0187] Optionally, the first processing module is further used for:

[0188] Inputting the second interaction data into the target model to obtain third interaction data output by the target model;

[0189] The general data set is generated based on the second interaction data and the third interaction data.

[0190] Optionally, the second processing module is further used for:

[0191] Adding the personal data set and the general data set to the training set of the target model;

[0192] Training a first parameter of the target model based on the training set;

[0193] The second parameter of the target model remains unchanged during the training process.

[0194] Optionally, the device further comprises:

[0195] A second acquisition module, used to acquire input data of the target user;

[0196] The fifth processing module is used to input the input data into the target model after updating the first parameter to obtain target data.

[0197] Optionally, in the target model after the first parameter is updated,

[0198] The important layer processes the input according to the updated first parameter and the fixed second parameter;

[0199] Non-significant layers process the input according to a fixed second parameter.

[0200] After the device obtains the personal data set of the target user, it can determine the general data set in combination with the target model, and then update the first parameter added to the important layer in the target model according to the personal data set and the general data set. In this way, there is no need to update the parameters of all layers of the target model. It is only necessary to add the first parameter to the important layer of the target model (determined based on the similarity between the output of each layer of the pre-built model of the target model and the target, the layer with a greater impact on the output result, and the layer with an output similarity greater than a preset threshold) and update the first parameter. Because the important layer is the transformer layer that must be calculated in the large model reasoning, it has an important impact on the reasoning result. In contrast, the calculation of the non-important layer has little effect on the reasoning result, and the calculation can be omitted to reduce computing resources and reasoning delay. The first parameter is a new connection parameter generated by one or more hidden layer neurons. Compared with the update of all parameters of the original model, the model adjustment can be realized faster, which significantly reduces the computing power and data requirements for model optimization, does not require cloud collaboration, and reduces the risk of leakage.

[0201] It should be noted that the apparatus of the embodiment of the present application is applied to the first device, and the embodiment of the method executed by the first device is applicable to the apparatus and can achieve the same technical effect.

[0202] like Figure 7 As shown, an update processing device in an embodiment of the present application includes:

[0203] The third processing module 710 is used to determine the important layers of the pre-built model;

[0204] A fourth processing module 720 is used to add the first parameter to each important layer to obtain a target model;

[0205] The important layer is a layer in the pre-built model of the target model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons.

[0206] Optionally, the third processing module is further used for:

[0207] Based on the test data set, obtaining a first result output by each layer in the pre-built model and a second result output by each layer in the test model; wherein the test model is obtained after adding a third parameter to each layer of the pre-built model;

[0208] The important layer is determined based on a similarity between the first result and the second result.

[0209] Optionally, the device further comprises:

[0210] A sixth processing module is used to train the test model using the test data set and the target data set before obtaining the second result based on the test data set.

[0211] After confirming the important layers of the pre-built model and adding the first parameter, the device can obtain the target model with the first parameter added to the important layers, so that after the first device obtains the personal data set of the target user, it can determine the general data set in combination with the target model, and then update the first parameter added to the important layers in the target model according to the personal data set and the general data set. In this way, there is no need to update the parameters of all layers of the target model, only the first parameter needs to be added to the important layers of the target model (determined based on the similarity between the output of each layer of the pre-built model of the target model and the target, the layer with a greater impact on the output result, and the layer with an output similarity greater than a preset threshold) and the first parameter is updated. Because the important layer is a transformer layer that must be calculated in the reasoning of the large model, it has an important impact on the reasoning result. In contrast, the calculation of the non-important layer has little effect on the reasoning result, and the calculation can be omitted to reduce computing resources and reasoning delay. The first parameter is a new connection parameter generated by one or more hidden layer neurons. Compared with the update of all parameters of the original model, the model adjustment can be realized faster, which significantly reduces the computing power and data requirements for model optimization, and does not require cloud collaboration, reducing the risk of leakage.

[0212] It should be noted that the apparatus of the embodiment of the present application is applied to the second device, and the embodiment of the method executed by the second device is applicable to the apparatus and can achieve the same technical effect.

[0213] like Figure 8 As shown, an update processing device of an embodiment of the present application includes a transceiver 810, a processor 800, a memory 820, and a program or instruction stored in the memory 820 and executable on the processor 800; when the processor 800 executes the program or instruction, the update processing method performed by the first device is implemented.

[0214] The transceiver 810 is used to receive and send data under the control of the processor 800 .

[0215] Among them, Figure 8 In the embodiment, the bus architecture may include any number of interconnected buses and bridges, specifically one or more processors represented by processor 800 and various circuits of memory represented by memory 820 are linked together. The bus architecture may also link together various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface. The transceiver 810 may be a plurality of components, namely, a transmitter and a receiver, providing a unit for communicating with various other devices on a transmission medium. For different user devices, the user interface 830 may also be an interface capable of externally and internally connecting required devices, and the connected devices include but are not limited to a keypad, a display, a speaker, a microphone, a joystick, and the like.

[0216] The processor 800 is responsible for managing the bus architecture and general processing, and the memory 820 can store data used by the processor 800 when performing operations.

[0217] A network side device according to another embodiment of the present application, such as Fig. 9 As shown, it includes a transceiver 910, a processor 900, a memory 920, and a program or instruction stored in the memory 920 and executable on the processor 900; when the processor 900 executes the program or instruction, the update processing method executed by the first device is implemented.

[0218] The transceiver 910 is used to receive and send data under the control of the processor 900 .

[0219] Among them, Fig. 9In the embodiment, the bus architecture may include any number of interconnected buses and bridges, specifically linking together various circuits of one or more processors represented by processor 900 and memory represented by memory 920. The bus architecture may also link together various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface. The transceiver 910 may be a plurality of components, i.e., including a transmitter and a receiver, providing a unit for communicating with various other devices on a transmission medium. The processor 900 is responsible for managing the bus architecture and general processing, and the memory 920 may store data used by the processor 900 when performing operations.

[0220] A readable storage medium according to an embodiment of the present application stores a program or instruction thereon. When the program or instruction is executed by a processor, it implements the update processing method executed by the first device as described above, or the steps in the update processing method executed by the second device, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0221] The processor is the processor in the update processing device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0222] An embodiment of the present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-mentioned update processing method performed by the first device, or the various processes of the update processing method embodiment performed by the second device, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0223] It should be further explained that many functional components described in this specification are referred to as modules in order to more particularly emphasize the independence of their implementation methods.

[0224] In the embodiment of the present application, module can be implemented with software so that it can be executed by various types of processors. For example, an executable code module of an identification can include one or more physical or logical blocks of computer instructions, for example, it can be constructed as an object, process or function. Nevertheless, the executable code of the identified module does not need to be physically located together, but can include different instructions stored in different positions, and when these instructions are logically combined together, it constitutes a module and realizes the specified purpose of the module.

[0225] In fact, executable code module can be a single instruction or many instructions, and can even be distributed on a plurality of different code segments, distributed among different programs, and distributed across a plurality of memory devices. Similarly, operating data can be identified in the module, and can be implemented and organized in the data structure of any appropriate type according to any appropriate form. The operating data can be collected as a single data set, or can be distributed in different locations (including on different storage devices), and can only be present on a system or network as an electronic signal at least in part.

[0226] When a module can be implemented by software, considering the level of existing hardware technology, a person skilled in the art can build a corresponding hardware circuit to implement the corresponding function of the module that can be implemented by software without considering the cost. The hardware circuit includes a conventional very large scale integration (VLSI) circuit or gate array and existing semiconductors such as logic chips, transistors, or other discrete components. The module can also be implemented by a programmable hardware device, such as a field programmable gate array, a programmable array logic, a programmable logic device, etc.

[0227] The above exemplary embodiments are described with reference to the accompanying drawings, and many different forms and embodiments are feasible without departing from the spirit and teachings of the present application. Therefore, the present application should not be constructed as a limitation of the exemplary embodiments proposed herein. More specifically, these exemplary embodiments are provided so that the present application will be perfect and complete, and the scope of the present application will be conveyed to those who are familiar with the technology. In these figures, the component sizes and relative sizes may be exaggerated for clarity. The terms used here are only based on the purpose of describing specific exemplary embodiments and are not intended to be limiting. As used herein, unless the text clearly indicates otherwise, the singular forms "one", "an" and "the" are intended to include these multiple forms. It will be further understood that the terms "comprising" and / or "including" when used in this specification indicate the presence of the features, integers, steps, operations, components and / or components, but do not exclude the presence or increase of one or more other features, integers, steps, operations, components, components and / or their groups. Unless otherwise indicated, when stated, a range of values ​​includes the upper and lower limits of that range and any subranges therebetween.

[0228] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. An update processing method, characterized in that: The method is executed by a first device and includes: Obtain the target user’s personal data set; Determine a general data set according to the personal data set and the target model; wherein a first parameter is added to an important layer of the target model; the important layer is a layer in a pre-built model of the target model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons; The first parameter is updated according to the personal data set and the general data set.

2. The method according to claim 1, characterized in that The obtaining of the personal data set of the target user includes: Collecting first interaction data of the target user; Acquire context information of the first interaction data; Integrate the first interaction data based on the context information to obtain second interaction data; The personal data set is generated based on the second interaction data.

3. The method according to claim 2, characterized in that The determining of the general data set according to the personal data set and the target model includes: Inputting the second interaction data into the target model to obtain third interaction data output by the target model; The general data set is generated based on the second interaction data and the third interaction data.

4. The method according to claim 3, characterized in that: The updating of the first parameter according to the personal data set and the general data set includes: Adding the personal data set and the general data set to the training set of the target model; Training a first parameter of the target model based on the training set; The second parameter of the target model remains unchanged during the training process.

5. The method according to claim 2, characterized in that: Also includes: Obtaining input data of the target user; The input data is input into the target model after the first parameter is updated to obtain target data.

6. The method according to claim 1, characterized in that In the target model after the first parameter is updated, The important layer processes the input according to the updated first parameter and the fixed second parameter; Non-significant layers process the input according to a fixed second parameter.

7. An update processing method, characterized in that: The method is executed by the second device and includes: Identify important layers of pre-built models; Add the first parameter in each important layer to obtain the target model; The important layer is a layer in the pre-built model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons.

8. The method according to claim 7, characterized in that The important layers of the pre-built model are determined, including: Based on the test data set, obtaining a first result output by each layer in the pre-built model and a second result output by each layer in the test model; wherein the test model is obtained after adding a third parameter to each layer of the pre-built model; The important layer is determined based on a similarity between the first result and the second result.

9. The method according to claim 8, characterized in that Also includes: Before obtaining the second result based on the test data set, the test model is trained using the test data set and the target data set.

10. An update processing device, characterized in that: include: The first acquisition module is used to acquire the personal data set of the target user; A first processing module is used to determine a general data set according to the personal data set and the target model; wherein a first parameter is added to an important layer of the target model; the important layer is a layer in a pre-built model of the target model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons; The second processing module is used to update the first parameter according to the personal data set and the general data set.

11. An update processing device, characterized in that: include: a third processing module for determining important layers of the pre-built model; A fourth processing module is used to add the first parameter to each important layer to obtain a target model; The important layer is a layer in the pre-built model whose output similarity is greater than a preset threshold, and the first parameter is a new connection parameter generated by one or more hidden layer neurons.

12. An update processing device, comprising: A transceiver, a processor, a memory, and a program or instruction stored in the memory and executable on the processor; characterized in that when the processor executes the program or instruction, the update processing method as described in any one of claims 1 to 6 is implemented, or the update processing method as described in any one of claims 7 to 9 is implemented.

13. A readable storage medium having a program or instruction stored thereon, characterized in that: When the program or instruction is executed by the processor, the steps in the update processing method as described in any one of claims 1 to 6 are implemented, or the steps in the update processing method as described in any one of claims 7 to 9 are implemented.

14. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implement the steps in the update processing method according to any one of claims 1 to 6, or the steps in the update processing method according to any one of claims 7 to 9.

Images