Personalized intelligent terminal artificial intelligence model architecture and upgrading method

Abstract

The invention provides a personalized intelligent terminal artificial intelligence model architecture and an upgrading method. The model architecture comprises a global model module, a local personalized model module, a private personalized model module and a personalized control interface. Wherein the local personalized model module is a universal personalized implementation module of the intelligent terminal; the personality control interface inputs personality options or a personality parameter matrix, instantiates a generalization model in a local personality model module of the intelligent terminal, and generates personality, namely differentiation characteristics of the intelligent terminal; the private personalized model module is an intelligent terminal private personality realization module, and the realized personality is the unique personality of the intelligent terminal. When the intelligent terminal model is subjected to OTA upgrading, the intelligent terminal can keep original personality work by setting a personality control interface and configuring a virtual interface module. The personalized expressive force of the intelligent terminal is improved, and the workload of intelligent terminal model training is reduced.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a personalized smart terminal artificial intelligence model architecture and upgrade method. Background Art

[0002] Artificial intelligence (AI) technology is being widely used in various devices and systems, from smartphones to cars, from medical equipment to industrial automation equipment, and its influence is everywhere. End-cloud collaboration is a cutting-edge technology in the field of artificial intelligence applications today. End-cloud joint training and division of labor deployment build an AI application framework. Under this framework, terminal devices usually use local data to pre-train models and upload the training results to the cloud for aggregation. Especially for large models, it is not feasible to train directly on terminal devices. This requires cloud servers to be responsible for model training. Cloud servers are also responsible for collecting model updates from various terminals, using optimization algorithms to update the global model, and distributing the updated model back to each terminal device to form an iterative optimization closed loop. Through model training and updating led by cloud servers, public intelligence that is generally shared among different smart terminals is generally formed. In addition to public intelligence, making artificial intelligence applications more personalized and private is also a current demand. With the joint efforts of the industry, AI models are performing better and better in general tasks. However, this general AI model built on a general theoretical framework and trained with public data lacks individuality for ordinary users when used. Injecting specialization and personalization on the basis of mainstream general AI models is an important exploration direction. Through personalized labels and the integration of more segmented and diversified data, fine-tuning smart terminals for different application scenarios will better meet the needs of human-computer interaction and scenario-based, and bring a more professional and personalized experience. The "smart terminal + fine-tuning model" approach will also become a new trend in application upgrades.

[0003] The application fields of smart terminals and robots are becoming increasingly extensive, such as industry, medical treatment, autonomous driving, home, entertainment and service industries. People's requirements for robots have also been upgraded from ordinary robots with simple repetitive mechanical movements to highly intelligent, autonomous and personalized intelligent robots, providing better human-machine interaction and being able to complete tasks alone or with assistance in more occasions. In smart homes, personalized artificial intelligence collects and analyzes information such as user preferences, needs and living habits, and then accurately adjusts and optimizes the home environment and equipment based on this information to meet the personalized needs of users. In the field of self-driving cars, smart cars provide better human-machine interaction and a more comfortable and familiar driving method for drivers by learning the driver's driving style, habits and local data.

[0004] The development of artificial intelligence in a large scale, miniaturized and personalized way requires personalized intelligent terminals, that is, the AI ​​models in intelligent terminals need to be personalized. As the working environment and work tasks change, intelligent terminals need to be continuously upgraded, and AI models also need to be continuously learned and trained. However, after the personalized intelligent terminal is upgraded, how to maintain the individual characteristics of the original intelligent terminal is a problem. Sometimes, the personalized model in the intelligent terminal needs to be retrained before it can work properly. In addition, the local training of personalized models requires a large workload, resulting in problems such as long upgrade time and insufficient computing power of intelligent terminals.

[0005] After searching the prior art documents, it was found that the patent invention with application number CN201610537688 and name "Information processing method and system for intelligent robots" discloses an information processing method and system for intelligent robots. The invention configures inherent personality parameters for intelligent robots, so that there are personality differences between individual intelligent robots, which effectively avoids the convergence of intelligent robots. The personality characteristic parameter method of the patent mainly considers the personalization of anthropomorphism, and does not explain the composition of the personality of the intelligent terminal, the acquisition of personality characteristic parameters, and the upgrade of the intelligent terminal. The patent invention with application number CN111602143A and name "Mobile autonomous personal companion based on user artificial intelligence (AI) model" discloses a method for building an artificial intelligence (AI) model. The deep learning engine in the AI ​​model in the patent provides user personalization function. The robot in the patent uses local data and local AI capabilities to train the local AI model, or uploads local data to the back-end server to train the local AI model. The patent generates and stores the private AI model of the corresponding user at each autonomous personal companion to realize the unique and unique personality of the autonomous personal companion, which is not shared outside the companion. However, in order to support multiple users and personalization, the backend server needs to generate and store many local AI model versions. When the number of users is huge, the backend server and network link load will increase. In addition, model upgrades and private AI model processing for smart terminals are not considered. Summary of the invention

[0006] The present application provides a personalized smart terminal artificial intelligence model architecture and upgrade method, which can be used to solve the technical problem of the huge contradiction between the individuality and quantity of local AI models.

[0007] This application provides a personalized intelligent terminal artificial intelligence model architecture, which includes:

[0008] Global model module, local personalized model module, private personalized model module, personalized control interface;

[0009] The global model module is a common model module shared among different smart terminals, realizing common functions shared among different smart terminals;

[0010] The local personalized model module is a module for realizing the universal personality of the intelligent terminal, and forms the personality of the intelligent terminal by locally instantiating the generalized model downloaded from the back-end server;

[0011] The private personalized model module is a module for realizing the private personality of the smart terminal, and the realized personality is the unique personality of the smart terminal; the private personalized model training is completed locally, and the training sample set is also obtained locally, and the model and the training sample set are not shared between different smart terminals;

[0012] The personality control interface inputs personality options or personality parameter matrix, instantiates the generalized model in the local personality model module of the smart terminal, and generates the personality, ie, differentiated characteristics, of the smart terminal.

[0013] Furthermore, the personality control interface provides personality option input or optional personality parameter matrix input; the user selects the personality parameter matrix through the personality option input, and then sets the weight parameter matrix of the local personalization model; or directly sets the weight parameter matrix of the local personalization model through the personality parameter matrix input.

[0014] Furthermore, after the user sets the personality parameter matrix through the personality control interface, the smart terminal optimizes the personality parameters through local training and back propagation to obtain the personality of the smart terminal;

[0015] Since the training is limited to fine-tuning of the personality parameter matrix, the workload of local training of the personalized model for smart terminals is greatly reduced.

[0016] Furthermore, by adding a personalized control interface to the AI ​​model, the training of the personalized model can be deployed on the back-end server including the cloud side and the edge side, thereby minimizing the workload of the terminal model training of the personalized smart terminal, and providing an effective method for the model deployment and training of personalized smart terminals and the privacy protection of smart terminals.

[0017] Furthermore, the personalized model is a generalized model used in the local personalized model, and the training process is as follows:

[0018] (1) Aggregate the training sample sets of multiple users;

[0019] (2) Perform model training to obtain a generalized model and generate a weight parameter matrix (including weight parameters and bias parameters), including a fixed weight parameter matrix and a dynamic weight parameter matrix;

[0020] (3) Define and select personalized features, extract personality parameters, and generate a personality parameter matrix;

[0021] (4) Select typical features from the personalized features to generate personalized options; link the personalized options, the personalized parameter matrix, and the weight parameter matrix to form a personality option-personality parameter matrix-weight parameter matrix mapping.

[0022] Furthermore, the size of the personality parameter matrix and the degree of personality difference of the smart terminal are reflected in that the larger the personality parameter matrix, the more personalized the smart terminal; the determination of the size of the personality parameter matrix is ​​related to the personality complexity of the AI ​​model, and the size is indicated by [number of parameter items, number of parameter values]; the personality parameter item values ​​are orthogonal.

[0023] Furthermore, the personalized parameter matrix or personalized options representing the personalization of the smart terminal are uploaded and backed up to the back-end server through the network to prevent the smart terminal from malfunctioning and the loss of personalized parameter data.

[0024] The present application also provides a personalized smart terminal artificial intelligence model upgrade method, which is used to upgrade the architecture provided by the present application. When the smart terminal model is upgraded via OTA, the back-end server downloads the generalized model of the local personalized model generated by the back-end server training, the global model, and the private personalized interface model for upgrading the private personalized model of the smart terminal to the smart terminal through the network to complete the upgrade of the smart terminal.

[0025] When the smart terminal model is upgraded OTA, the personalized parameter values ​​of the local personalized model remain unchanged, so that the smart terminal maintains the original personalized features;

[0026] Through local training, fine-tune the setting parameters of the personality options and the personality parameters in the personality parameter matrix.

[0027] The weight parameter matrix is ​​fine-tuned to achieve accurate smart terminal personality. This method can conveniently and well implement OTA upgrade of smart terminal model without changing the original personalized performance of the smart terminal.

[0028] Furthermore, when the smart terminal model is OTA upgraded, the generalized model used for the local personalized model undergoes significant changes, and the interface protocol of the personalized control interface changes, such as the increase in the personalized parameter matrix and personalized options;

[0029] After the smart terminal model is OTA upgraded, in order to ensure that the local personalized model runs according to the personality before the model upgrade as much as possible, the original corresponding items in the personality parameter matrix or personality options remain unchanged, and the new items in the personality parameter matrix or personality options take the recommended values ​​or default values ​​of the upgraded model provided by the back-end server, so that the smart terminal can maintain the original personalized characteristics as much as possible; through local training, the new personality options and new personality parameter matrix parameters are trained, and the original personality options and personality parameter matrix parameters are fine-tuned to achieve accurate smart terminal personality.

[0030] Furthermore, when the smart terminal model is upgraded OTA, the backend server provides a virtual interface module for the original private personalized model to run; because of the existence of the virtual interface module, the private personalized model still runs normally after the upgrade;

[0031] In the work of smart terminals, the private personalized model and the virtual interface module are trained together and gradually integrated. The virtual interface module is gradually digested by the private personalized model, and the private personalized model finally completes the self-upgrade of the smart terminal model OTA upgrade. The private personalized model is trained and optimized to improve the accuracy of the unique personality of the smart terminal.

[0032] The present invention has the following beneficial effects:

[0033] (1) The present invention proposes a personalized smart terminal artificial intelligence (AI) model architecture, which improves the expressiveness of the smart terminal personality by introducing a personalized control interface into the local personalized model of the smart terminal and introducing a private personalized model into the smart terminal, simplifies the training of the AI ​​model personality, and reduces the workload of smart terminal model training.

[0034] (2) The present invention proposes a method for upgrading the personalized smart terminal artificial intelligence (AI) model architecture, which solves the problems of the inability to maintain the personality during the OTA upgrade of the personalized smart terminal model and the inability of the smart terminal to run immediately. Through this upgrade method, the personalized smart terminal does not need to be retrained, and the original personality of the smart terminal can be maintained to the greatest extent and run immediately.

[0035] The present invention is widely used in fields requiring personalized application of artificial intelligence, such as companion robots and smart cars. The present invention is novel, highly operable, has a wide range of applications, and has great application significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a personalized intelligent terminal artificial intelligence (AI) model architecture diagram provided by an embodiment of the present invention;

[0037] Figure 2 It is a schematic diagram of the system architecture of the intelligent terminal and the back-end server;

[0038] Figure 3-1 It is a schematic diagram of a local personalized model implemented by instantiation in an intelligent terminal provided by an embodiment of the present invention;

[0039] Figure 3-2 It is a flowchart of a training method for a generalization model provided by an embodiment of the present invention;

[0040] Figure 3-3 It is a decomposition diagram of a local personalized model module provided by an embodiment of the present invention;

[0041] Figure 4-1 It is a schematic diagram of an upgrade of a personalized smart terminal artificial intelligence (AI) model architecture provided by an embodiment of the present invention;

[0042] Figure 4-2 is a functional schematic diagram of a virtual interface module provided by an embodiment of the present invention;

[0043] Figure 5 It is a schematic diagram of a method for upgrading a local personalized model of a personalized smart terminal provided by an embodiment of the present invention;

[0044] Figure 6 It is a schematic diagram of a personalized AI model architecture of an intelligent robot provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0046] As a first aspect of the present invention, the present invention provides a personalized intelligent terminal artificial intelligence model architecture, including: a global model module, a local personalized model module, a private personalized model module, and a personalized control interface;

[0047] Among them, the global model module is a common model module shared among different smart terminals, realizing common functions shared by different smart terminals;

[0048] The local personalized model module is a universal personality realization module for intelligent terminals, which forms the personality of intelligent terminals by locally instantiating the generalized model downloaded from the back-end server;

[0049] The private personalized model module is a private personality implementation module for smart terminals. The personality it implements is the unique personality of the smart terminal. This module is a private personalized model module for smart terminals. Its model training is completed locally, and the training sample set is also obtained locally. The model and training sample set are not shared between different smart terminals.

[0050] The personality control interface inputs personality options or personality parameter matrix, instantiates the generalized model in the local personality model module of the smart terminal, and generates the personality (differentiation characteristics) of the smart terminal.

[0051] The personalized services and functions provided by this personalized smart terminal are divided into two types, one is the private personality unique to the smart terminal, and the other is the universal personality. The personalized services and functions provided by the universal personality are also commonly provided in other smart terminals, but the universal personality has personalized differences between different smart terminals. The universal personality is realized by the local personalized model of the smart terminal. The personality of the local personalized model is formed by local instantiation of the downloaded generalized model. The instantiation is completed by setting the weight parameter matrix of the model. In the local smart terminal, the personality parameter matrix can be determined by setting the personality options, and then the dynamic weight parameter matrix can be determined to generate a personalized local personalized model. The private personality is realized by the private personalized model of the smart terminal. Its function is not available in the local personalized model, and it generates the unique personality of the smart terminal. Generally speaking, its personality cannot be mirrored and used to build other smart terminal models. The training and reasoning of the private personalized model are private, and its corresponding local data is not shared outside the smart terminal.

[0052] As a second aspect of the present invention, the present invention provides a method for upgrading a personalized intelligent terminal artificial intelligence (AI) model architecture, comprising:

[0053] (1) When the smart terminal model is upgraded OTA, the personalized parameters of the local personalized model remain unchanged, so that the smart terminal maintains the original personalized features. The local personalized model prototype is a generalized model trained by the backend server. When the generalized model is trained, the personality control interface still uses the original personality parameter matrix, the structure of the personality options, and the parameter settings (protocol definition). The upgraded model only changes the model structure and weight parameter matrix. After the upgrade, the local instantiation uses the original personality parameter matrix or the original personality option settings, and the upgraded local personalized model module will maintain the original personalized features.

[0054] (2) When the smart terminal model is upgraded via OTA, the backend server provides a virtual interface module for the original private personalized model to run. The private personalized model is a private model of the smart terminal. After the smart terminal model is upgraded via OTA, the front and back modules of the private personalized model module change. If the private personalized model runs directly, the smart terminal will not work properly. During the model OTA upgrade, the backend server provides the front and back virtual interface modules for the original private personalized model module to run. The front and back virtual interface modules are transient modules used to support backward compatibility for smart terminal upgrades. Because of the existence of the virtual interface module, an environment is created for the original private personalized model to run. The private personalized model module can still operate normally after the smart terminal model is upgraded via OTA.

[0055] Through the above-mentioned upgrade steps, the personalized smart terminal can work immediately after the OTA upgrade of the back-end server and maintain the personalized features of the original smart terminal.

[0056] In combination with a possible implementation of the second aspect, when the smart terminal model is upgraded OTA, the personality parameter matrix increases, and the newly added items are the recommended values ​​or default values ​​of the upgraded model provided by the back-end server. When the smart terminal model is upgraded OTA, the generalized model used for the local personalized model changes significantly, and the interface protocol of its personality control interface changes, such as the increase of the personality parameter matrix and personality options. After the smart terminal model is upgraded OTA, in order to ensure that the local personalized model runs as much as possible according to the personality before the model upgrade, one processing method is to keep the original corresponding items in the personality parameter matrix or personality options unchanged, and the newly added items in the personality parameter matrix or personality options are the recommended values ​​or default values ​​of the upgraded model provided by the back-end server, so that the smart terminal can maintain the original personalized characteristics as much as possible.

[0057] In combination with a possible implementation of the second aspect, after the smart terminal model is OTA upgraded, the personalized smart terminal AI model is locally trained and optimized, including:

[0058] (1) The backend server provides a virtual interface module for the original private personalized model to run during the OTA upgrade of the personalized smart terminal model. In the subsequent work of the smart terminal, the private personalized model module and the virtual interface module in the smart terminal are trained locally together, gradually integrated, and the virtual interface module is gradually digested. In this way, the private personalized model will eventually complete its own training and upgrade work for this smart terminal model OTA upgrade. Because the training of the new private personalized model is based on the original private personalized model, the implementation of the private personalized model can adopt an algorithm with time effect such as the EWC algorithm. During training, most of the neural network structure in the model is retained, and only a part needs to be covered. In this way, the training of the private personalized model can be completed quickly. After training and optimization, the private personalized model accurately realizes the unique personality of the smart terminal.

[0059] (2) The local personalized model module is trained locally during the operation of the smart terminal. Through the fine-tuning of the personalized options and the personalized parameter matrix, the fine-tuning of the weight parameter matrix is ​​further optimized, making the personalized realization of the smart terminal more accurate. This fine-tuning method only involves the personalized parameter matrix and the dynamic weight parameter matrix, with a small amount of calculation, which improves the efficiency of model training.

[0060] As a third aspect of the present invention, the present invention provides a training method for a generalized model used in a local personalized model. The generalized model is trained on a backend server (cloud side, edge side), and the training process is as follows, including: (1) aggregating the training sample sets of multiple users; (2) performing model training to obtain a generalized model, and generating a weight parameter matrix (including weight parameters and bias parameters), including a fixed weight parameter matrix and a dynamic weight parameter matrix; (3) defining and selecting personalized features, extracting personality parameters, and generating a personality parameter matrix; (4) selecting typical features from the personalized features and generating personalized options. The personality options, personality parameter matrix, and weight parameter matrix are linked to form a personality option-personality parameter matrix-weight parameter matrix mapping.

[0061] Traditional AI models use fixed weight parameters, but this method is difficult to describe the distribution of sample multi-user data sets, and it is insufficient to characterize the differences between different types of user groups and cannot adapt to all user samples. The dynamic weight parameter method provides a reasonable solution to the above problems. In addition to the fixed weight parameters generated by common factors in the AI ​​model, dynamic weight parameters are also generated according to the personality differences of different user groups in the sample. In this way, the model has a stronger expression ability in the personality of different sample types. A set of dynamic weight parameter matrices is generated for the different types of user groups in the sample. And the definition and extraction of different types of user groups generate a personality parameter matrix. During reasoning, according to the distinction of user groups, the corresponding personality parameter matrix is ​​selected, and the dynamic weight parameter matrix is ​​determined to obtain a model that accurately expresses the user's personality.

[0062] Embodiment 1

[0063] The present invention provides a personalized intelligent terminal artificial intelligence (AI) model architecture, such as Figure 1 ,include:

[0064] Global model module A 101 , global model module C 104 , global model module B 106 , local personalized model module 102 , private personalized model module 103 , and personalized control interface 105 .

[0065] The global model module is a common model module shared among different smart terminals, which realizes common functions shared by different smart terminals.

[0066] The local personalized model module 102 is a universal personality realization module for smart terminals, which forms the personality of smart terminals by locally instantiating the generalized model downloaded from the backend server;

[0067] The private personalized model module 103 is a private personality realization module of the smart terminal, and the personality it realizes is the unique personality of the smart terminal. This module is a private personalized model module of the smart terminal, and its model training is completed locally, and the training sample set is also obtained locally. The model and training sample set are not shared between different smart terminals.

[0068] The personality control interface 105 inputs personality options or personality parameter matrices, instantiates the generalized model in the local personality model module of the smart terminal, and generates personality differentiation characteristics of the smart terminal.

[0069] Embodiment 2

[0070] The smart terminal model is downloaded or upgraded from the backend server. The system architecture of the smart terminal and the backend server is as follows: Figure 2 As shown, it includes: multiple intelligent terminals 2011, 2012, ..., 201n, a deep machine learning engine 202, a local personalized model builder 203, a global model builder 204, a private personalized interface model builder 205, a local personalized model 206, a global model 207, a private personalized interface model 208, and a network 209. Among them, the deep machine learning engine 202, the local personalized model builder 203, the global model builder 204, the local personalized model 206, and the global model 207 are located in the backend server.

[0071] The backend server calls the deep machine learning engine 202 for learning through the local personalized model builder 203, the global model builder 204, and the private personalized interface model builder 205, and generates a local personalized model 206, a global model 207, and a private personalized interface model 208. The generated local personalized model 206 is a generalized model with a personalized control interface, which can be instantiated at the smart terminal. The global model 207 is a shared universal model used at the smart terminal, and the examples are global model A, global model B, and global model C. The local personalized model and the global model (A, B, C) are downloaded or upgraded to the smart terminal 2011, 2012, ..., 201n through the network 209, forming the local personalized model 20121, 20122, ..., 2012n of the smart terminal (1-n), and the global model (A, B, C) 20131, 20132, ..., 2013n. The private personalized models 20111, 20112, ..., 2011n of the smart terminal (1-n) are private personality implementation modules of the smart terminal, and the personality implemented is the unique personality of the smart terminal. This module is a private personalized model module of the smart terminal, and its model training is completed locally, and the training sample set is also obtained locally. The model and the training sample set are not shared between different smart terminals, so the private personalized model in the smart terminal does not need to be generated and downloaded by the back-end server. However, the back-end server needs to download the private personalized interface model 208 to the smart terminal to build a virtual interface module for the normal operation and upgrade of the private personalized interface model.

[0072] Embodiment 3

[0073] The local personalized model is implemented by instantiating the generalized model downloaded from the backend server at the smart terminal, such as Figure 3-1 , including: local personalized model module 301, personalized parameter matrix generation module 302, dynamic weight matrix generation module 303. Its local instantiation process is as follows: determine personalized option settings; generate personalized parameter matrix through personalized parameter matrix generation module 302; generate dynamic weight parameter matrix through dynamic weight matrix generation module 303; dynamic weight parameter matrix, combined with fixed weight parameter matrix, generates personalized local personalized model.

[0074] The generalized model is trained on the cloud side and edge side of the backend server. Since smart terminals have their own personalities, the training samples in the backend server are also aggregates of the personalized data of smart terminals. For the generalized model, since the sample data implements the same function, part of the model's weights are fixed; but the sample data comes from users with different personalities, so the model can use part of the model's weights to be dynamically distributed to fit users with different personalities, that is, the trained model contains a fixed weight parameter matrix and a dynamic weight parameter matrix. Figure 3-2, which is a flowchart of the training method of the generalization model. The specific steps of the method are as follows, including:

[0075] S301, aggregate the training sample sets of multiple users. Assume that the sample set targeted by the local personalized function of user 1 is K1, the sample set targeted by the local personalized function of user 2 is K2, ..., the sample set targeted by the local personalized function of user n is Kn, and the aggregated sample set is {K1, K2, ..., Kn};

[0076] S302, perform model training, obtain a generalized model, generate a weight parameter matrix containing weight parameters and bias parameters, including a fixed weight parameter matrix and a dynamic weight parameter matrix. Local personalized model module decomposition, assuming composition see Figure 3-3 , the local personalized model A3011, the local personalized model B3012, and the local personalized model C3013, the corresponding fixed weight parameter matrix is ​​decomposed into a fixed weight parameter matrix ShPM1 and a fixed weight parameter matrix ShPM2, the model training is completed, and the fixed weight parameter matrix ShPM1, the fixed weight parameter matrix ShPM2 and the dynamic weight parameter matrix SpPM are obtained;

[0077] S303, define and select the personalized features, extract the personalized parameters, and generate a personalized parameter matrix. Because the personalized features of the local personalized model, even with the same input, its output also presents personalized distribution characteristics, and its distribution has a certain range, and within this range, it also presents distributions of different densities. According to the input-output relationship of the local personalized module, its personalized distribution characteristics are obtained. First, the distribution features are ranged and selected, and the number of selected features is the number of parameter items; then, for a certain parameter, the number of parameter values ​​is determined according to the parameter distribution range and parameter resolution; finally, the parameter values ​​are determined to generate a personalized parameter matrix. The size of the personality parameter matrix is ​​related to the degree of personality difference of the smart terminal. The larger the personality parameter matrix, the more personalized the smart terminal; the determination of the size of the personality parameter matrix is ​​related to the personality complexity of the AI ​​model, and the size is indicated as [number of parameter items, number of parameter values]. The selection of personality parameter items should be orthogonal. For example, the personality parameter matrix is ​​[A1, A2, A3, ..., Am], [B1, B2, B3, ..., Bm], [C1, C2, C3, ..., Cn], [D1, D2, D3, ..., Dn], where A, B, C, D are personality parameter items, and the number of parameter items is 4. The number of parameter item values ​​of A and B parameters is m, and the number of parameter items of C and D parameters is n. The values ​​of A, B, C, D parameters are orthogonal, that is, A, B, C, D personality parameters represent independent personality characteristics;

[0078] S304, select typical features from the personalized features to generate personalized options. The size of the personality parameter matrix is ​​[number of parameter items, number of parameter values], which reflects the personality summary of the localized model. However, there are some typical personalities in these personality summaries. The typical personalities are determined as personality options, and the corresponding personality parameters are selected and classified from the personality parameter matrix.

[0079] Comparison Figure 3-1 , Figure 3-3 As described above, there is a corresponding relationship between personality options, personality parameter matrix and weight parameter matrix. Personality options, personality parameter matrix and weight parameter matrix are linked to form a personality option-personality parameter matrix-weight parameter matrix mapping. The mapping relationship is Figure 3-1 The individual parameter matrix conversion module 302 and the dynamic weight matrix conversion module 303 are embodied in the algorithm.

[0080] The generalized model is trained on the cloud side and edge side of the backend server. First, the local sample sets of each user are aggregated S301, and then the model is trained to generate a fixed weight parameter matrix and a dynamic weight parameter matrix containing weight parameters and bias parameters S302. At the same time, the aggregated personalized data undergoes feature extraction, such as using unsupervised machine learning or semi-supervised machine learning to analyze the algorithm's personality behavior patterns and divide similar features into different personality groups. Then, through feature distribution analysis and personality definition, a personality parameter matrix is ​​formed S303. The main clustering points in the personality parameter matrix form typical personality options S304. The personality options, personality parameter matrix, and weight parameter matrix are linked to form a personality option-personality parameter matrix-weight parameter matrix mapping to complete the model's personality control interface design.

[0081] Embodiment 4

[0082] When the smart terminal model is upgraded OTA, such as the global model or local personalized model, the personalized model needs to be reset or trained to ensure the normal operation of the smart terminal. In particular, the operation of the private personalized model may not work properly due to changes in the previous and next modules of the model. Figure 4-1 , which is a schematic diagram of a personalized smart terminal artificial intelligence (AI) model architecture upgrade method, including: global model modules 401, 404, 408, a local personalized model module 402, a private personalized model module 403, a personalized control interface 405, and virtual interface modules 406, 407.

[0083] The personalized smart terminal performs OTA model upgrades through the backend server, in which the global model is upgraded to an AI model trained by the server and can be used directly; the generalized model corresponding to the local personalized model module 402 still uses the original personalized parameter matrix, the structure of the personalized options, and the parameter setting protocol definition during training. The upgraded model only changes at the model structure and weight parameter matrix level. Therefore, as long as the local personalized model module 402 continues to use the original personalized parameter matrix or the original personalized option settings, the upgraded local personalized model module 402 will maintain the original personalized features.

[0084] However, after the smart terminal model is OTA upgraded, the processing of the private personalized model module 403 is more complicated. Since the front and back modules of the private personalized model module 403 have changed, if the private personalized model is directly run, the smart terminal will not work properly. One processing method is that the back-end server provides the front and back virtual interface modules 406 and 407 for the original private personalized model module 403 to run during the upgrade process of the global model and personalized model of the smart terminal. The front and back virtual interface modules 406 and 407 are transient modules used to support backward compatibility for smart terminal upgrades. The functions implemented by this module are as follows: Figure 4-2 , providing the upgraded local personalized model interface including interface protocol, global model B interface including interface protocol, global model C interface including interface protocol and private personalized model interface including interface protocol before upgrade. Because of the existence of virtual interface modules 406 and 407, the private personalized model module 403 can still operate normally after the smart terminal model OTA upgrade. In the subsequent work of the smart terminal, the private personalized model module 403 and multiple virtual interface modules 406 and 407 in the smart terminal are trained locally together, gradually merged, and the virtual interface module is gradually digested, so that the private personalized model module 403 finally completes its own training and upgrade work for this smart terminal model OTA upgrade. Because the training of the new private personalized model is based on the original private personalized model, the implementation of the private personalized model can adopt an algorithm with time effect such as the EWC algorithm. Most of the neural network structure in the model is retained during training, and only a part needs to be covered. In this way, the training of the private personalized model can be completed quickly. After training and optimization, the private personalized model accurately realizes the unique personality of the smart terminal. In addition, the local personalized model module 402 is trained locally during the operation of the smart terminal, and the fine-tuning of the personalized options and the personalized parameter matrix is ​​further optimized, so that the personalized realization of the smart terminal is more accurate. The personalized parameter matrix or personalized options of the smart terminal can be uploaded and backed up to the back-end server through the network 209 to prevent the smart terminal from malfunctioning and the loss of personalized parameter data.

[0085] Embodiment 5

[0086] The OTA upgrade of the smart terminal model causes a significant change in the generalized model used for the local personalized model, and the interface protocol of the personality control interface 405 changes, such as the increase in personality parameter matrix and personality options. Since the upgraded local personalized model module 402 is quite different from the local personalized model module before the upgrade, in order to ensure that the local personalized model module 402 can maintain the personality before the upgrade as much as possible, a processing method is provided, see Figure 5 , the steps include:

[0087] S501, the original corresponding items in the personality parameter matrix or personality options remain unchanged, and the newly added items in the personality parameter matrix or personality options take the recommended values ​​or default values ​​of the upgrade model provided by the backend server, so that the smart terminal maintains the original personalized features as much as possible;

[0088] S502, and then the newly added personality options and the newly added personality parameter matrix parameters can be trained through local training, and the original personality options and personality parameter matrix parameters can be fine-tuned, so as to achieve a more accurate smart terminal personality.

[0089] Embodiment 6

[0090] The basic function of intelligent robots is to perceive, analyze, and identify multimodal sensor input signals, and then perform prediction, path planning, language analysis, reasoning / judgment, and decision-making, and finally output control instructions to the control module. In addition, intelligent robots also need to have personality, smoother human-computer interaction, and new skill learning. This personalized intelligent terminal artificial intelligence (AI) model architecture is an application in the field of intelligent robots, such as Figure 6 A personalized smart terminal artificial intelligence (AI) model architecture that can be used for intelligent robots includes: a perception module 601, a local personalized reasoning module 602, a private personalized reasoning module 603, a global reasoning module 604, a control module 608, a personalized control interface 605, and virtual interface modules 606 and 607.

[0091] The perception module 601 is implemented using a global model to realize the perception, analysis, and recognition of external and internal multimodal sensor information in the intelligent robot. The internal and external sensor information includes visual sensor input, auditory sensor input, tactile sensor input, azimuth sensor input, laser radar signal input, etc.

[0092] The global reasoning module 604 is implemented using a global model to realize functions such as prediction, path planning, language analysis, reasoning / judgment, and decision-making in the intelligent robot.

[0093] The local personalized reasoning module 602 is implemented using a local personalized model to realize personalized prediction, path planning, language analysis, reasoning / judgment, decision-making and other functions in the intelligent robot.

[0094] The private personalized reasoning module 603 is implemented using a private personalized model to realize the unique functions and private decision-making of the intelligent robot.

[0095] The control module 608 is implemented using a global model to realize the control system and actuators of the intelligent robot, such as audio and video output, motor, and robotic arm control.

[0096] The personality control interface 605 inputs personality options or personality parameter matrices, instantiates the generalized model in the local personality reasoning module 402 of the intelligent robot, and generates a personality-differentiated intelligent robot model.

[0097] The virtual interface modules 606 and 607 provide backward compatibility support for the OTA upgrade of the intelligent robot model. After the model OTA upgrade, they provide a virtual compatible operating environment for the private personalized reasoning module 603 to ensure the normal operation of the private personalized reasoning module 603 and subsequent training and upgrades.

[0098] The above implementation modes of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A personalized intelligent terminal artificial intelligence model architecture, characterized in that: The model architecture includes: Global model module, local personalized model module, private personalized model module, personalized control interface; The global model module is a common model module shared among different smart terminals, realizing common functions shared among different smart terminals; The local personalized model module is a module for realizing the universal personality of the intelligent terminal, and forms the personality of the intelligent terminal by locally instantiating the generalized model downloaded from the back-end server; The private personalized model module is a module for realizing the private personality of the smart terminal, and the realized personality is the unique personality of the smart terminal; the private personalized model training is completed locally, and the training sample set is also obtained locally, and the model and the training sample set are not shared between different smart terminals; The personality control interface inputs personality options or personality parameter matrix, instantiates the generalized model in the local personality model module of the smart terminal, and generates the personality, ie, differentiated characteristics, of the smart terminal.

2. The model architecture according to claim 1, characterized in that: The personality control interface provides personality option input or optional personality parameter matrix input; The user selects the personalized parameter matrix by inputting personalized options, and then sets the weight parameter matrix of the local personalized model; Or directly set the weight parameter matrix of the local personalized model through the input of the personality parameter matrix.

3. The model architecture according to claim 1, characterized in that After the user sets the personality parameter matrix through the personality control interface, the smart terminal optimizes the personality parameters through local training and back propagation to obtain the personality of the smart terminal; Training is limited to fine-tuning of the personality parameter matrix.

4. The model architecture according to claim 1, characterized in that By adding a personalized control interface to the AI ​​model, the training of the personalized model can be deployed on the back-end server including the cloud side and the edge side.

5. The model architecture according to claim 4, characterized in that: The personalized model is the generalized model used in the local personalized model. The training process is as follows: Aggregate the training sample sets of multiple users; Perform model training to obtain a generalized model and generate a weight parameter matrix (including weight parameters and bias parameters), including a fixed weight parameter matrix and a dynamic weight parameter matrix; Define and select personalized features, extract personalized parameters, and generate a personalized parameter matrix; Select typical features from the personalized features to generate personalized options; The personality options, personality parameter matrix and weight parameter matrix are linked to form a personality option-personality parameter matrix-weight parameter matrix mapping.

6. The model architecture according to claim 1, characterized in that The size of the personality parameter matrix and the degree of personality difference of the intelligent terminal are reflected in that the larger the personality parameter matrix, the more personalized the intelligent terminal; The determination of the size of the personality parameter matrix is ​​related to the personality complexity of the AI ​​model, and the size is represented by [number of parameter items, number of parameter values]; the personality parameter item values ​​are orthogonal.

7. The model framework according to claim 1, characterized in that The personalized parameter matrix or personalized options that characterize the personalization of the smart terminal are uploaded and backed up to the back-end server through the network to prevent the smart terminal from malfunctioning and the loss of personalized parameter data.

8. A personalized intelligent terminal artificial intelligence model upgrade method, the method is used to upgrade any one of the architectures in claims 1 to 7, characterized in that: Methods include: When the smart terminal model is upgraded via OTA, the backend server downloads the generalized model of the local personalized model generated by the backend server training, the global model, and the private personalized interface model used for upgrading the private personalized model of the smart terminal to the smart terminal through the network to complete the upgrade of the smart terminal; When the smart terminal model is upgraded OTA, the personalized parameter values ​​of the local personalized model remain unchanged, so that the smart terminal maintains the original personalized features; Through local training, fine-tune the setting parameters of the personality options and the personality parameters in the personality parameter matrix.

9. The method according to claim 8, characterized in that When the smart terminal model is upgraded OTA, the generalized model used for the local personalized model undergoes significant changes, and the interface protocol of the personalized control interface changes; After the smart terminal model is OTA upgraded, the original corresponding items in the personality parameter matrix or personality options remain unchanged, and the new items in the personality parameter matrix or personality options take the recommended values ​​or default values ​​of the upgraded model provided by the backend server; Through local training, the newly added personality options and personality parameter matrix parameters are trained, and the original personality options and personality parameter matrix parameters are fine-tuned.

10. The method according to claim 8, characterized in that When the smart terminal model is upgraded OTA, the backend server provides a virtual interface module for the original private personalized model to run; In the operation of smart terminals, the private personalized model and the virtual interface module are trained together and gradually integrated. The virtual interface module is gradually digested by the private personalized model, and the private personalized model finally completes its own upgrade of the smart terminal model OTA upgrade.

Images