Terminal artificial intelligence (AI) correction method and device

Abstract

The invention discloses a terminal artificial intelligence (AI) correction method and system, and aims to solve the technical bottlenecks of terminal AI in the aspects of transmission efficiency, model iteration and cross-device cooperation in the prior art through efficient integration of feature value conversion and external source feature values. The method comprises two core stages: (1) establishing a same relation characteristic value conversion table between a terminal AI and an external source characteristic value, and realizing characteristic value compatibility through dynamic adjustment of an algorithm patch and a learning data set; and (2) based on a hardware degradation correction and security verification mechanism, ensuring stable application of an external source characteristic value. According to the technical scheme, the operation efficiency, the privacy protection capability and the multi-model cooperation efficiency of the terminal AI are remarkably improved, and the method can be widely applied to the fields of smart home, automatic driving, intelligent robots, industrial Internet of Things and the like.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a correction method for terminal artificial intelligence (AI), which improves the performance, stability, and cross-model and cross-device coordination capabilities of terminal artificial intelligence (AI) through feature value transformation and application of external source feature values. Background Technology

[0002] Currently, the deployment of edge artificial intelligence (AI) faces multiple technical challenges. For example, data transmission bottlenecks are difficult to overcome, model iteration faces significant obstacles, and cross-model and cross-device collaboration is inefficient.

[0003] The existing methods or technical architectures for correcting terminal artificial intelligence (AI) mainly include the following: the concept of AIAgent memory module, research based on neural perception, limb and team coordination patterns, and early neural network research starting with feature values.

[0004] The core functions of an AI Agent's memory module generally include context awareness and understanding, personalized services, task planning and execution, continuous learning and optimization, and knowledge accumulation and reasoning. The implementation of the memory module is primarily based on storage data structures, including vector databases, key-value pair storage, graph databases, natural language processing (such as word embeddings), and knowledge graphs (networks of relationships between entities). Its application scenarios mainly include dialogue systems (contextual and multi-turn dialogues), intelligent recommendation systems, AINIC games, and autonomous driving. The coding practice of the AI ​​Agent memory module presents a series of technical challenges, including privacy and security, data sparsity and noise, and storage and computation costs. The biggest challenge lies in balancing "memory continuity" and "decision timeliness." Just as programmers need to regularly clear code caches to maintain mental agility, AI Agents also need a scientific memory management mechanism to operate continuously and efficiently. The core significance of resetting the AI ​​Agent's memory module to zero lies in information metabolism mechanisms, decision quality assurance, and ethical compliance requirements.

[0005] In summary, current technologies limit the functionality of terminal and mobile AI due to hardware and transmission efficiency. On one hand, open-source and closed-source models typically transmit data from the terminal to the host for computation. Since the transmission environment and speed affect the performance and stability of the AI ​​terminal, they often fail to support terminal AI iteration. On the other hand, terminal AI (such as robotic AI) and cloud AI are currently disconnected. They require different AI architectures. Cloud AI needs to be able to think, correct, and integrate with software suites to complete tasks, while terminal AI often needs to make decisions and adjustments within limited timeframes and with insufficient data. After iteration and updates, communication between the two is like communication between humans and computers, requiring a screen, keyboard, and software suite, which cannot meet the needs of terminal and mobile AI. Furthermore, the operation and interoperability of terminal and host AI also require integration with many external programs or hardware / software modules. Existing AI models, regardless of size or type, require frequent and extensive manual adjustments and corrections, and there is no way to confirm whether they have drifted or malfunctioned. This patent application aims to handle these tasks more efficiently and automatically, achieving better overall performance. Summary of the Invention

[0006] This invention provides a method and apparatus for correcting terminal artificial intelligence (AI). Based on a complete set of learning datasets, algorithm patches, program patches, and a systematic process, it achieves mobile terminal AI correction and error detection in an automated or semi-automated manner. By enabling terminal AI to share feature values ​​with host AI or different terminal AIs, the technical bottlenecks of terminal AI in terms of transmission efficiency, model iteration, and cross-device collaboration in the prior art can be solved.

[0007] To address the aforementioned technical problems, this invention first provides a terminal artificial intelligence (AI) correction method that can utilize external source feature values. This method includes the following steps:

[0008] S1. Receive external source feature values ​​and specified factor relationships transmitted by the host AI, receive learning dataset and algorithm patch, and correct the same relationship feature values ​​of the terminal AI;

[0009] S2. Establish a transformation table or transformation function between the same relation feature values ​​of the terminal AI and the external source feature values ​​through dynamic iteration, and embed the transformation table or transformation function into the neural network;

[0010] S3. Based on the transformation table or transformation function, correct the application mode of the external source feature value;

[0011] S4. Perform hardware degradation correction and eigenvalue clustering verification, wherein the hardware degradation correction may be automated or semi-automated;

[0012] Wherein, the same relational feature value is a feature value that has the same specified factor relationship, and the specified factor relationship includes at least one of the following: statistical correlation between feature values, parameter combination of a specific convolutional layer.

[0013] In one specific embodiment of the present invention, step S1 specifically includes:

[0014] S11. Upload the detection data of the terminal AI to the host AI, supplement the external source feature value field, and add a feature value algorithm module for iteration;

[0015] S12. If the specified factor relationship is not generated in S11, replace the data detection sensor or firmware of the terminal AI, find compatible factor relationships, add special algorithm patches and / or increase the learning dataset, and iterate again.

[0016] S13. Repeat S11 to S12 until the specified factor relationship is generated;

[0017] S14. Input the specified factor relationship and / or the external source feature value data field into the neural network to complete the correction;

[0018] In S11, the host AI provides the learning dataset and performs format conversion by referencing the hardware parameters of similar devices used by the terminal AI; the converted learning dataset is then used for iterative processing. Alternatively,

[0019] The host AI provides the learning dataset and its source, while the terminal AI obtains learning data by detecting similar learning data sources through its hardware and iterates through the data; or,

[0020] The terminal AI obtains similar learning data suitable for its hardware specifications from other external sources and iterates through them;

[0021] Preferably, in S11, the feature value algorithm module has a built-in suggested convolution mode, and the terminal AI iterates according to the suggested convolution mode, which is provided by the host AI or other third-party host.

[0022] In one specific embodiment of the present invention, step S2 specifically includes:

[0023] S21. Compare the specified factor relationship field corrected in S1 with the external feature value data field to form a terminal AI feature value conversion table; and optionally,

[0024] S22. Simplify the feature value mapping relationship through statistical regression or function fitting, and verify that the feature value mapping relationship fit is not lower than a preset safety threshold, thus forming a program patch.

[0025] In one specific embodiment of the present invention, step S3 specifically includes:

[0026] S31. The host AI calculates the feature value and transmits it to the terminal AI;

[0027] S32. The mobile terminal AI determines the application mode of external source feature values. In simple mode, execute S33; in complex mode, execute S34.

[0028] S33. The feature values ​​calculated by the host AI can be used directly or after transformation based on a transformation table or transformation function;

[0029] S34. Use the feature values ​​calculated by the host AI in conjunction with external feature values ​​in a program or a specific learning dataset format.

[0030] In one specific embodiment of the present invention, step S4 specifically includes hardware degradation correction and eigenvalue clustering verification; the hardware degradation correction step includes:

[0031] S41. By comparing the test results of standard products with the factory data, correct hardware parameters or replace faulty components;

[0032] S42. Utilize the simulated scenario dataset provided by the host AI to verify the real-time output accuracy of the terminal sensors;

[0033] The eigenvalue clustering verification step includes:

[0034] S43. Use a clustering algorithm to filter feature value combinations that meet the preset security threshold;

[0035] S44. When clustering fails, readjust the transformation table or transformation function, or supplement the external source feature values ​​to meet the preset clustering conditions.

[0036] In one specific embodiment of the present invention, the method further includes:

[0037] S5. Transmit the transformed same-relationship feature values ​​to other terminal AIs or host AIs in a de-identified format, so that the terminal AIs and host AIs can share the feature values;

[0038] S6. Real-time coordinated computing among multiple terminals is achieved through discrete computing power or emergency coordination allocation.

[0039] According to another aspect of the present invention, the present invention also provides a method for host artificial intelligence (AI) to assist terminal artificial intelligence (AI) in correction using external source feature values, the method comprising: the host AI providing external source feature values ​​or calculating feature values ​​in response to a request from the terminal AI.

[0040] In one specific embodiment of the present invention, providing external source feature values ​​includes:

[0041] In the step of correcting the same relation feature value in the terminal AI, the host AI receives the monitoring data uploaded by the terminal AI, supplements the external source feature value field, adds a feature value algorithm module, and generates a specified factor relationship to be transmitted to the terminal AI.

[0042] When the terminal AI corrects the way external source feature values ​​are used, the external source feature values ​​are provided by the AI.

[0043] During the feature value clustering verification step of the terminal AI, external source feature values ​​are added until the preset clustering conditions are met.

[0044] The calculated feature values ​​include:

[0045] After the terminal AI establishes a feature value conversion table or conversion function, it receives the feature value conversion table or conversion program and uses surplus computing power to assist in the mapping and conversion between the feature values ​​of the terminal AI and the feature values ​​of the host AI.

[0046] In the step of the terminal AI performing hardware degradation correction, the detection data or firmware parameters and specifications transmitted by the terminal AI are received, feature values ​​are calculated and transmitted to the terminal AI for comparison between the detection data and the feature values.

[0047] In one specific embodiment of the present invention, the host AI and the terminal AI jointly calculate the same relational feature value and a specific feature value, and compare them using a feature value conversion table or conversion function to form a multimodal feature value collaborative confirmation.

[0048] In one specific embodiment of the present invention, in the step of multimodal feature value collaborative confirmation, the host AI receives the detection data uploaded by the terminal AI, uses surplus computing power to calculate specific feature values ​​and compares them with external source feature values ​​pre-stored in the host AI, transmits the corrected external source feature values ​​and specific feature values ​​to the terminal AI, and the terminal AI calculates the specific feature values ​​that do not achieve external source feature value matching.

[0049] According to another aspect of the present invention, the present invention also provides a terminal artificial intelligence (AI) correction device, the device comprising the following modules:

[0050] The first correction module is used to correct the same relation feature values ​​through iterative calculation. The first correction module includes a same relation feature value algorithm module and a corresponding external source feature value data field. The algorithm module is used to provide calculation of specified factor relations. When a specified factor relation occurs, the specified factor relation field or the corresponding external source feature value data field is incorporated into the neural network, thus completing the correction.

[0051] The conversion module is used to establish a conversion table or conversion function between the corresponding feature values ​​of the terminal artificial intelligence (AI) and the feature values ​​of the external source;

[0052] The second correction module, based on the conversion table or conversion function established by the conversion module, transmits feature values ​​to the terminal artificial intelligence (AI), provides learning datasets, algorithm patches and program patches, and corrects the way external source feature values ​​are used.

[0053] The security verification module is used to perform hardware degradation correction and feature value clustering verification to ensure the stable transmission and secure application of the external source feature values, and includes a first detection unit and a second detection unit.

[0054] The first detection unit is used to input the data of the standard product detection result into the pre-trained first detection model to obtain the first detection result. The first detection model is used to compare the data of the standard product detection result with the factory data, give instructions to correct hardware parameters or replace faulty components, and use the simulated scenario dataset provided by the host AI to verify the real-time output accuracy of the terminal sensor.

[0055] The second detection unit is used to perform cluster verification on the feature values ​​according to the preset clustering conditions. It uses a clustering algorithm to select feature value combinations that meet the preset safety threshold. When clustering fails, it readjusts the transformation table or transformation function, or supplements the external source feature values ​​to meet the preset clustering conditions.

[0056] In one specific embodiment of the present invention, the terminal artificial intelligence (AI) correction device further includes the following modules:

[0057] The second input module is used to transmit the transformed same-relationship feature values ​​to other terminals or host AI in a desensitized format, so that the terminal AI and the host AI can share the feature values.

[0058] The second algorithm module is used to achieve real-time coordinated computing among multiple terminals through discrete computing power or emergency coordination allocation.

[0059] According to another aspect of the present invention, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the steps of the method of the present invention.

[0060] According to another aspect of the present invention, a computer-readable storage medium is also provided having a computer program stored thereon, which, when executed by one or more processors, implements the steps of the method described in the present invention.

[0061] Technical effect

[0062] This invention provides a systematic method and apparatus for terminal artificial intelligence (AI) calibration, which can replace a large amount of manual calibration, docking, and adjustment work, automating or semi-automating the detection and elimination of problems, improving usability and system security. The terminal AI calibration mode of this invention simultaneously supports terminal model iteration, data privacy requirements, cross-device coordination, and data sharing across large models. It allows terminal AIs and host AIs to save computing power by sharing feature values, and leverages many capabilities of neural networks or large models that are currently not fully realized on terminals. In practical applications, the following technical effects are achieved:

[0063] 1. Saves computing power and transmission costs: This invention provides a mode for storing, transmitting, and sharing large model data transformed from corrected feature values. This mode can be used as big data, or by using corresponding feature values ​​through program patches or algorithm iterations. It can perform distributed or remote computation of corresponding feature values, and share and transmit them after formatting, storing, and compressing the feature values. It has functions such as desensitization or encryption, overcoming hardware and transmission bottlenecks, meeting emergency coordination and supervision needs, saving computing power, and facilitating big data analysis.

[0064] 2. Cross-model dynamic feature value compatibility: The feature value transmission scheme provided by this invention realizes cross-model feature value sharing through a conversion table, supports feature value sharing and coordination between AI models on different terminals. Compared with the current black box of models and privacy infringement of learning data, with program patches, it can meet the needs of supervision and privacy, and reduce data redundancy (saving ≥70% of transmission volume compared with existing technologies).

[0065] 3. External Source Feature Value Security Verification Mechanism: This invention utilizes three types of external source features along with corresponding program and algorithm patches, and leverages different data sources or large datasets to achieve better large model calibration and anchoring effects, and also enables more accurate screening and analysis of large datasets. Furthermore, by integrating hardware degradation detection and cluster verification algorithms, this invention can better avoid illusions or erroneous answers caused by hardware and software limitations or insufficient data quality and quantity. Attached Figure Description

[0066] Figure 1 This is a wireless information transmission system architecture diagram illustrated according to an exemplary embodiment, which schematically shows the interaction process of mobile terminal AI, host AI and shared feature values.

[0067] Figure 2 This is a flowchart illustrating a terminal artificial intelligence (AI) correction method according to an exemplary embodiment.

[0068] Figure 3 This is an exemplary embodiment illustrating the intention of feature value transformation, showing the transformation direction of feature values ​​between the host AI and the terminal AI-B1.

[0069] Figure 4 This is an illustration of the feature value transformation representation according to another exemplary embodiment, showing the transformation direction of feature values ​​between terminal AI-B1 and host AI.

[0070] Figure 5 This is one embodiment of the present invention. Figure 5 A demonstrates the steps taken by the AI ​​in the dialogue terminal to incorporate external source feature values ​​directly into the neural network. Figure 5 B illustrates the process of incorporating external source features into a neural network after correction and security verification.

[0071] Figure 6 This is another embodiment of the present invention. Figure 6 A demonstrates the steps involved in enabling multiple terminal AI systems to incorporate externally sourced feature values ​​directly into neural networks for use. Figure 6 B illustrates the process of incorporating external source features into a neural network after correction and security verification.

[0072] Figure 7 This is a block diagram illustrating an artificial intelligence (AI) correction device for a terminal according to an exemplary embodiment.

[0073] Figure 8 This is a block diagram of a computer device according to an embodiment of the present invention.

[0074] Embodiments of the present invention

[0075] The present invention will now be described in further detail with reference to specific embodiments. The given embodiments are merely illustrative of the invention and not intended to limit its scope. The embodiments provided below can serve as a guide for further improvements by those skilled in the art and do not constitute a limitation on the invention in any way.

[0076] Terminology Explanation

[0077] Host AI (including large models, data transmission modules, or may include big data databases): Host AI is a general term that may refer to cloud, discrete, edge, quantum and server hybrid computing systems, or discrete operations between terminal artificial intelligence (AI). The large model running on it is defined as AI-1 large model.

[0078] Terminal Artificial Intelligence (AI) (including large models, sensor detectors, and data transmission modules): Terminal artificial intelligence (AI) can be a similar hardware terminal B or a different type of hardware C. The large models running on this terminal are defined as AI-B1 large model, AI-B2 large model, ..., AI-C1 large model, AI-C2 large model, etc. These various terminal large models can be lightweight models that are compressed, distilled, or iterated from the AI-1 large model, or neural network systems, large models, or program architectures completely different from AI-1. This invention will focus on terminal AI-B1 for explanation.

[0079] Same-relational features: Features that have the same specified factor relationship, such as statistical correlation under the same convolutional layer, used for cross-model data compatibility.

[0080] Specify factor relationships: relationships between feature values, specific types of convolutional layers, or a combination of both.

[0081] External source features: Feature values ​​provided by the host AI or other terminals to extend the analytical capabilities of the terminal AI.

[0082] Hardware degradation correction: Detect and correct performance deviations of terminal hardware through standard data comparison and simulation analysis.

[0083] Cluster verification: Using machine learning algorithms to classify and detect anomalies in feature value combinations.

[0084] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0085] In the description of the embodiments of the present invention, it should be understood that terms such as "mobile terminal," "mobile end," or "terminal" have the same meaning and, unless otherwise stated, refer to an edge device that provides voice and / or data connectivity to the user, is responsible for communicating with the cloud or host, and processes local data in real time.

[0086] In the description of embodiments disclosed herein, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of the disclosed features, figures, steps, behaviors, components, portions or combinations thereof in this specification, and do not exclude the possibility of the presence of one or more other features, figures, steps, behaviors, components, portions or combinations thereof.

[0087] Unless otherwise stated, " / " means "or". For example, A / B can mean A or B. In this article, "and / or" is merely a way of describing the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A alone, A and B at the same time, and B alone.

[0088] The terms "first," "second," etc., are used only for ease of description to distinguish identical or similar technical features and should not be construed as indicating or implying the relative importance or number of these technical features. Therefore, a feature defined by "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, the term "multiple" means two or more. Detailed Implementation

[0089] Figure 1 This is a schematic diagram of a terminal artificial intelligence (AI) correction method according to an embodiment of the present invention. This correction method can be applied to... Figure 1 The wireless information transmission system shown is illustrated. Those skilled in the art will understand that the wireless information transmission system may also include other conventional network devices, such as core network devices, wireless relay devices, and wireless backhaul devices. Figure 1 Not shown in the image.

[0090] See Figure 1 As shown, this wireless information transmission system includes several terminals and / or mobile devices of the same or different types, as well as a host device. The terminal AIs and the host AI share feature values ​​and transmit and receive information using wireless resources. Furthermore, different terminal AIs can also share feature values ​​(not shown).

[0091] Furthermore, the term "terminal" in this invention broadly refers to electronic devices that integrate artificial intelligence technology and are capable of performing complex tasks, providing intelligent services, and offering interactive experiences. These terminals, through built-in AI algorithms and hardware support, can achieve functions such as speech recognition, image processing, natural language understanding, and predictive analysis, thereby improving user experience and device performance. According to device type, terminal AI can be categorized into smartphones, personal computers, smart wearable devices, smart home devices, smart toys, and in-vehicle information systems, etc. It will be understood by those skilled in the art that the embodiments of this invention do not limit the specific technologies or device forms used in the terminals.

[0092] Furthermore, in this embodiment of the invention, the host AI is used to aggregate datasets from multiple terminals, calculate specified factor relationships, and calculate external source feature values, offering advantages in data throughput and processing speed. The terminal AI is used to establish feature value transformation tables or transformation functions, autonomously iterating according to the convolution mode specified by the host AI. It can also achieve computational power coordination and matching through optimization calculations, enabling rapid response to end-user needs and quickly displaying processed image, video, audio, and text information to users in a low-power, low-cost manner. Therefore, after the commands issued by the end-user are initially processed by the mobile terminal AI, they interact in real-time with the host AI via a wireless network. The host AI's processing results are then fed back to the end-user via the wireless network, improving data processing capabilities and effectively reducing latency.

[0093] In existing technologies, edge AI generally cannot iterate or is prone to drift. Furthermore, it is limited by computing power and transmission speed, failing to effectively utilize the big data analysis advantages of mainframe AI, resulting in wasted computing power and impacting privacy and inter-device coordination needs. Edge devices, such as robots, currently primarily use simple open-source LLMs for requirement analysis and program operation, often requiring manual correction and adjustment. Moreover, mainframe AI often boasts hundreds of thousands of GPUs of computing power, while edge AI, limited by hardware configuration and network speed, often suffers from performance limitations and is more prone to issues like larger illusions due to the scale of its AI models. More complex language and text communication utilizes closed-source models from large companies like GPT, which are limited by transmission speed and compromise data privacy. Larger models like DeepSeek are still limited by hardware and slower algorithms. Significant waste of computing power exists between related systems, and big data cannot be directly shared between models. This black-box nature affects emergency coordination efficiency and post-iteration monitoring.

[0094] This invention provides a terminal artificial intelligence (AI) correction method. The terminal AI receives external source feature values ​​and specified factor relationships transmitted from the host AI. Through learning datasets and algorithm patches, it corrects feature values ​​with the same relationship and establishes a feature value transformation table or transformation function. The specified factor relationship in this invention refers to the relationship between feature values ​​under the same convolutional layer, or the relationship between factors through convolutional combinations. By defining this relationship, suitable convolutional layers can be derived by using some variance, explanatory power extrema, or preset values, while restricting the type of convolutional layer. Feature values ​​with the same specified factor relationship are called identically related feature values. When this specified factor relationship occurs, the corresponding feature value is incorporated into the neural network. Otherwise, the feature value cannot be incorporated into the neural network, but it can be incorporated into big data or used as the basis for communication in large models. Thus, shared feature values ​​can meet the needs of customization and different architectures. Embodiments of this invention utilize shared feature value transmission for neural network learning, which can be used for supervision and data sharing, achieving more efficient cross-platform emergency coordination and significantly saving computing power.

[0095] Figure 2 This is a flowchart illustrating a terminal artificial intelligence (AI) correction method according to an exemplary embodiment. Figure 2 As shown, the terminal artificial intelligence (AI) correction method includes the following steps: In step S1, the terminal AI receives external source feature values ​​and specified factor relationships transmitted from the host AI, receives a learning dataset and algorithm patches, and corrects the same-relation feature values ​​of the terminal AI. In step S2, a transformation table or transformation function between the same-relation feature values ​​of the terminal AI and external source feature values ​​is established through dynamic iteration, and the transformation table or transformation function is embedded into the neural network. In step S3, based on the transformation table or transformation function, feature values ​​are transmitted to the terminal AI or a learning dataset, algorithm patches, and program patches are provided to correct the usage of external source feature values. In step S4, the security of external source feature values ​​is verified by performing hardware degradation correction and feature value clustering verification to ensure the stable transmission and secure application of the external source feature values.

[0096] In this embodiment of the invention, the detection data of the terminal AI is uploaded to the host AI, the external source feature value field is supplemented, a feature value algorithm module is added for iteration, and a specified factor relationship is generated.

[0097] In this embodiment of the invention, if the specified factor relationship is not generated, the data detection sensor or firmware of the terminal AI is replaced, a compatible factor relationship is searched, a special algorithm patch is added and / or the learning dataset is increased, and the iteration is performed again until the specified factor relationship is generated.

[0098] In this embodiment of the invention, the generated specified factor relationship or external feature value data field is input into the neural network to complete the correction of the terminal relationship feature value.

[0099] In one embodiment of this invention, the host AI provides a learning dataset and corresponding learning data hardware parameters, and performs format conversion with reference to similar device hardware parameters of the terminal AI-B1. The conversion can be performed by a program provided by the device manufacturer, directly on behalf of the user, or by estimating the conversion using statistical methods. The converted learning data is combined with the external feature value data field values ​​provided by the host AI, and a related feature value algorithm module is added. This data is then provided to AI-B1 for iteration until a specified factor relationship emerges. Correction is completed when the specified factor relationship field or the corresponding external feature value data field is incorporated into the neural network.

[0100] In one embodiment of this invention, the host AI provides corresponding external feature value fields and a learning dataset and its source. The terminal AI-B1 obtains learning data from a source similar to its hardware detection. The learning data acquired by AI-B1, the external feature value fields provided by the host AI, and the added related feature value algorithm module are provided to AI-B1 for iteration until a specified factor relationship appears. When the specified factor relationship field or the corresponding external feature value field is incorporated into the neural network, the correction of AI-B1 is completed.

[0101] In one embodiment of the present invention, the terminal AI obtains similar learning data suitable for its hardware specifications from other external sources and performs iterations.

[0102] In this embodiment of the invention, the feature value algorithm module has a built-in suggested convolution mode, and the terminal AI iterates according to the suggested convolution mode, which is provided by the host AI or other third-party host.

[0103] In this embodiment of the invention, the method for generating a conversion table or conversion function includes the following steps: In step S21, the specified factor relationship column that has been corrected in S1 is compared with the external feature value data column to form a terminal AI feature value conversion table.

[0104] In one embodiment of the present invention, the method for generating the transformation table or transformation function further includes step S22: simplifying the feature value mapping relationship through statistical regression or function fitting, and verifying that the feature value mapping relationship fit is not lower than a preset safety threshold, thereby forming a program patch. For example, a program patch can be formed by using a function plus statistical variation. When the correspondence between the two is close to a linear relationship in cases of similar devices, convolutional layers, or the existence of compatibility factors, a function can be used instead to speed up the feature value transformation. After regression with a function relationship, the correlation and statistical reliability and validity need to be confirmed (e.g., confirmed by converting to a distribution plot), and then a program can be written to perform feature value transformation.

[0105] After the eigenvalue transformation table is converted, the eigenvalues ​​may differ from those of the same hardware, convolutional layer, or directly calculated eigenvalues, but they still have the same factor relationships. Therefore, different transformation tables are established for different directions of transformation. The eigenvalue transformation table can be directly provided to the host AI, third parties, or internal use.

[0106] Figure 3 This is a conversion table or program built for a specific direction. See also Figure 3In this embodiment, the AI-B1 feature value conversion table or program 1 can be used internally to convert external source feature values, or it can be transmitted to other (mobile) terminals such as the host AI, AI-B2, or AI-C1 to utilize external computing power to convert the external source feature values ​​provided by the host AI. The mapping relationship between AI-B1 and AI-1 feature values ​​is shown below in matrix form:

[0107]

[0108] Figure 4 This is a conversion table or program created for a different direction. See also Figure 4 In this embodiment, the AI-B1 feature value conversion table 2 can be used internally to convert external source feature values, or it can be transmitted to other (mobile) terminals such as the host AI, AI-B2, or AI-C1 to utilize external computing power to convert AI-B1 related feature values, thereby providing external source feature values ​​for general host AI. The feature values ​​after conversion by the conversion table are as follows:

[0109]

[0110] Understandably, even with the same data, the specified factor relationship feature values ​​may differ depending on the convolution method. For a rough comparison, the focus is on whether the factor relationship feature value appears; the accuracy of the specific value is not critical. In this case, one can follow the definition of convolution, such as by comparing the local parts of an unknown pattern with the local parts of a standard pattern, statistically estimating the colors of similar constituent patterns, and selecting the value that best matches the factor relationship. Thus, even after iterations or changes in the convolutional layer, data with the same specified factor relationship can still be shared.

[0111] In this embodiment of the invention, the external source feature value application correction includes the following steps: In step S31, the host AI calculates feature values ​​and transmits them to the mobile terminal AI. There are two ways to use external source feature values: in simple mode, they can be used directly after data transmission conversion; in complex mode, they can be used in conjunction with programs and learning data formats. Therefore, step S32 may be performed, in which the mobile terminal AI determines the external source feature value application mode. In simple mode, step S33 is executed: the feature values ​​calculated by the host AI are used directly or after conversion based on a conversion table or conversion function; in complex mode, step S34 is executed: the host AI provides the learning dataset format and external source feature values, and the (mobile) terminal AI performs learning dataset format conversion according to its initial design, and uses the data after adding or rewriting program patches. Thus, by incorporating good new factor relationships and transmitting feature values, the efficiency of the (mobile) terminal AI can be improved by leveraging the computing power of the host AI or discrete computing power.

[0112] According to the mobile terminal artificial intelligence correction method of the present invention, hardware degradation correction and external source feature value use safety correction are required each time the model is iterated or the hardware is replaced. Correction can also be performed periodically to detect and prevent system failure and help with maintenance judgment.

[0113] In this embodiment of the invention, the external source feature value security verification includes a hardware degradation correction step: In step S41, by comparing the standard product test results with the factory data, hardware parameters are corrected or faulty components are replaced. In step S42, the real-time output accuracy of the terminal sensor is verified using a simulated scenario dataset provided by the host AI. The correction can be automated or semi-automated.

[0114] For example, the CCD hardware manufacturer of the mobile terminal AI-B1 provides a standard light source to cover the lens, displaying different colors and shapes at different times. The host AI receives the data, calculates feature values, and transmits them to the mobile terminal AI-B1. The mobile terminal AI-B1 compares these with factory parameters and historical feature values, and corrects or replaces hardware parameters and components based on the comparison results.

[0115] For example, the mobile terminal AI-B1 provides firmware parameters and specifications, while the host AI simulates image data of a certain route, calculates feature values, and transmits them to the mobile terminal AI-B1. The mobile terminal AI-B1 compares these features during operation to detect the degree of degradation of the output and input devices, determining whether hardware, detectors, or firmware correction are necessary. If the feature values ​​are inconsistent, the hardware parameters are corrected, or faulty components are detected and replaced.

[0116] In this embodiment of the invention, the security verification of external source feature values ​​further includes a feature value clustering verification step: In step S43, a clustering algorithm is used to filter feature value combinations that meet the preset security threshold. In step S44, when clustering fails, the transformation table or transformation function is readjusted, or the external source feature values ​​are supplemented to meet the preset clustering conditions.

[0117] Furthermore, the security correction for external source feature values ​​requires two sets of learning datasets, one containing correct judgments and the other containing incorrect judgments (outputs). These learning datasets must be data from (mobile) terminals or simulated data, along with the corresponding external feature values. The specific correction process is as follows:

[0118] (1) First, confirm whether the combination of external source feature values, transformed feature values, and feature values ​​with the same relation to the terminal can achieve clustering in the two sets of learning datasets. If it can be achieved, the correction is complete. It should be noted that if there are no external feature values ​​with the same relation, then it is only necessary to confirm the external source feature values ​​themselves.

[0119] (2) If perfect clustering is not possible, further screening of combinations of other terminal feature values ​​and terminal-related feature values ​​is required to see if they can cluster in both sets of training datasets. Then, the combination of external source feature values ​​and other terminal feature values ​​is tested to determine if they can cluster in both sets of training datasets. Similarly, if no external feature values ​​with the same relationship exist, only the external source feature values ​​themselves need to be confirmed. If the above tests can achieve clustering, further confirmation is needed to see if the transformed feature values ​​can also be perfectly clustered. If not, the feature value transformation table is corrected until clustering is confirmed, at which point the correction is complete.

[0120] (3) If the correction cannot be completed, it is necessary to add or replace the hardware, add external source feature values, and iterate again.

[0121] In one embodiment of the present invention, taking voice recognition function as an example, when certain commands need to be issued by specific personnel, the construction of the learning dataset can be carried out as follows: simultaneously covering the images and voices of different people, and attaching corresponding common feature values ​​(or similar feature values), and considering the common situation of missing external data, integrating them into a comparative positive and negative learning dataset.

[0122] Furthermore, learning data can also incorporate elements such as external habits and regulatory rules. For example, in traffic scenarios, for the rule "when the right-turn traffic light is on, straight-going vehicles must yield to right-turning vehicles," the learning data can include different scenarios where vehicles yield or straight-going vehicles fail to yield, resulting in collisions, as well as different examples of vehicles using turn signals and not using turn signals.

[0123] Figure 5 This is one embodiment of the present invention. Figure 5 A demonstrates the steps taken by the AI ​​in the dialogue terminal to incorporate external source feature values ​​directly into the neural network. Figure 5 B illustrates the process of incorporating external source features into a neural network after correction and security verification.

[0124] In today's rapidly globalizing and aging society, daily communication scenarios are becoming increasingly complex and diverse. On the one hand, there are significant differences in communication styles among people from different regions, cultural backgrounds, and age groups. For example, when communicating with the elderly, their language habits, speaking speed, and acceptance of new vocabulary differ from those of younger people. In cross-cultural and ethnic communication, differences in language systems and cultural customs can easily lead to semantic misunderstandings and communication barriers. On the other hand, traditional dialogue AI technology mainly relies on the analysis of written language, neglecting the key semantics, emotions, and subtexts contained in non-verbal information such as facial expressions, body language, and tone of voice. This results in low accuracy in understanding complex interpersonal communication scenarios, making it difficult to achieve precise and effective communication and interaction. Furthermore, existing terminal AI, limited by its own hardware performance and algorithm complexity, often falls short when faced with massive and complex multimodal communication data collection, in-depth analysis, and efficient computation and iteration tasks, failing to meet users' needs for high-quality dialogue interaction.

[0125] See Figure 5 A. This embodiment relates to intelligent dialogue interaction technology in the field of artificial intelligence. The method includes the following steps:

[0126] (1) Collaborative confirmation of multimodal eigenvalues:

[0127] Regularly scheduled close collaborations between AI-1 and AI-B1 focus on two key dimensions: facial expressions and body language, and grammatical and logical coherence. During this process, both jointly calculate and confirm correlation feature values ​​and physical feature values, which quantify key characteristics of non-linguistic and linguistic structural information from different perspectives. Simultaneously, AI-B1 correlation feature values ​​are fully utilized and compared with correlation feature value transformation tables or transformation models for verification.

[0128] For example, in a cross-border business negotiation scenario, AI-B1 uses a camera to capture the opponent's slightly furrowed brows and forward-leaning facial expressions and body movements. Combining this with grammatical and logical rationality analysis within the current context, it calculates the corresponding related and physical feature values. Then, using a related feature value conversion table, it accurately interprets whether the other party may have doubts or different opinions about the current negotiation terms, thus providing a strong basis for adjusting subsequent dialogue strategies.

[0129] (2) Iterative optimization and updates of mobile terminal AI:

[0130] On the one hand, AI-B1 can autonomously iterate according to a preset cycle, continuously optimizing its model parameters and algorithm structure to adapt to increasingly complex and ever-changing dialogue scenarios. On the other hand, AI-B1 continuously accumulates various types of dialogue-related data, including the communication habits of different user groups, high-frequency vocabulary, and typical facial expressions and body language examples. With the assistance of AI-1, convolutional iterative correction technology is used to deeply mine the accumulated data and optimize the model.

[0131] For example, when it is discovered that elderly users in a certain region frequently use specific dialect words and gestures when asking about healthcare, AI-B1, with the assistance of AI-1, can quickly adjust its understanding and response strategies for this type of multimodal information through convolutional iterative correction. Furthermore, to further accelerate AI-B1's evolution, external program updates or internal model iteration patches can be easily downloaded from the original model vendor, ensuring it always remains at the forefront of intelligence.

[0132] (3) Customized dialogue understanding driven by hybrid analytics:

[0133] In the initial stage of a conversation or in specific complex situations, such as in cross-cultural tourism consultation scenarios, when a tourist has just started asking a question and the language information is not yet complete, AI-B1 quickly uploads relevant data by combining their rich facial expressions, body movements and unique tone of voice. AI-B1 then uses its powerful multimodal analysis capabilities to deeply analyze more than 70% of the semantics, emotions or subtexts conveyed by facial expressions, body movements and tone of voice, as well as grammatical and logical feature values.

[0134] Subsequently, when the program detects inconsistencies or confusion between the AI-1 analysis results and the AI-B1's initial local assessment, it will immediately conduct a secondary confirmation. For example, if AI-B1 determines from the tourist's initial statement that they might want to learn about local cuisine, but AI-1 discovers through multimodal analysis that the tourist is actually more interested in local traditional architecture, the program will automatically trigger a reconfirmation mechanism.

[0135] (4) Generate a customized neural network analysis to confirm the actual meaning:

[0136] At the same time, based on the massive amount of dialogue data accumulated by (mobile) terminals in the past, combined with the neural network after the aforementioned feature value iterative correction, a customized neural network analysis for actual meaning confirmation can be generated for each user.

[0137] For example, for a user who frequently inquires about technology products, AI-B1 can quickly and accurately understand the user's newly raised questions based on previously accumulated data, and provide professional answers that meet the user's needs, achieving a highly personalized dialogue and interaction experience.

[0138] The methods employed in this invention enable dialogue AI to more accurately understand user intent, effectively reducing communication misunderstandings caused by semantic ambiguity and cultural differences, and significantly improving communication comprehension accuracy. It can quickly adapt to new communication scenarios, language habits, and changes in user needs, continuously optimizing its performance to ensure the provision of up-to-date, high-quality services to users. Furthermore, this invention employs a hybrid analysis method combined with accumulated historical data to tailor dialogue understanding and response strategies for each user, meeting diverse and personalized dialogue needs and enhancing user experience.

[0139] In one embodiment of this invention, external source feature values ​​can be used in conjunction with a program or a specific dataset format. In this case, the AI ​​host needs to provide a specific format learning dataset associated with the feature values, as well as a program patch. If necessary, it can also provide corresponding data format specifications and user manuals. The terminal AI developer then writes a matching program patch or dataset conversion program based on this content.

[0140] At the neural network level, the correction also requires two steps: hardware degradation correction and external feature value security correction. Only after these corrections are completed can the stability and accuracy of the system be ensured.

[0141] In one embodiment of the present invention, a method for accurately locating and confirming the route of a moving or flying object based on feature value comparison is provided in the field of artificial intelligence and navigation positioning technology. The method includes the following steps:

[0142] (1) Data acquisition stage:

[0143] Before commencing any movement or flight operation, the terminal AI-B1 (either a vehicle or drone) sends a request to AI-1, demanding that the host AI provide a dataset of feature values ​​organized in a specific format related to the planned route. This dataset contains rich and structured information, providing crucial foundational data for subsequent positioning and route confirmation.

[0144] (2) Calculation and comparison of feature values ​​during the movement:

[0145] When the terminal is in motion, AI-B1 uses a specified calculation method to accurately calculate specific feature values ​​and compares the feature values ​​to locate and confirm the route.

[0146] Taking an aircraft as an example, its bottom camera dynamically determines the number of grid segments based on its altitude. Then, the terminal computer calculates the characteristic values ​​of specified factor relationships, and subsequently compares these calculated characteristic values ​​with the corresponding map connection characteristic values ​​provided by the AI ​​host. Through this meticulous comparison process, the relative position of the aircraft on the map can be accurately located, achieving high-precision positioning.

[0147] For a car, its cameras precisely divide and segment the surrounding landscape according to a preset program, acquiring specific landscape feature values. These feature values ​​are then compared one by one with the surrounding landscape feature values ​​connected by the route provided by the AI ​​host. In this way, the car's real-time position and direction of travel can be accurately determined.

[0148] Furthermore, the terminal must undergo regular hardware degradation calibration (such as lens light source detection) in order to correct hardware parameters in a timely manner, replace faulty components, and ensure data accuracy.

[0149] (3) Positioning and direction of travel output:

[0150] Through the rigorous feature value calculation and comparison process described above, both aircraft and automobiles can accurately output their own positioning information and direction of travel, providing a reliable basis for subsequent navigation decisions and ensuring the safety of travel or flight and the efficiency of real-time information communication.

[0151] This invention, through its innovative feature value organization and comparison method, significantly improves the positioning accuracy of moving or flying objects, effectively overcoming the limitations of traditional positioning technologies. It is adaptable to different types of moving or flying vehicles, such as aircraft and automobiles, possessing broad applicability and providing a unified and efficient positioning solution for diverse transportation and aviation scenarios.

[0152] In one embodiment of the present invention, a method for improving the positioning accuracy and computational efficiency of moving objects based on a priority calculation strategy is provided in the field of intelligent navigation and artificial intelligence collaborative computing. The method includes the following steps:

[0153] (1) Initial stage of data interaction:

[0154] When AI-B1 is in motion, it has the ability to upload its approximate location and real-time images to the host AI in real time. This allows the host AI to quickly obtain preliminary location information of the moving object and intuitive image data of the surrounding environment, laying the foundation for subsequent accurate calculations.

[0155] (2) Host AI calculation and positioning:

[0156] After receiving the data from AI-B1, the host AI immediately initiates a specific segmentation program to finely divide the acquired image data. Subsequently, it uses its powerful computing capabilities to calculate relevant feature values ​​for each segmented region.

[0157] These calculated feature values ​​are rigorously compared with the surrounding landscape feature values ​​pre-stored within the host AI. Through this high-precision comparison, the host AI is able to accurately locate the precise position of AI-B1 in the current environment.

[0158] After localization is completed, the host AI further filters the feature values ​​in the image that match the surrounding features, and transmits these feature values ​​and precise localization information to AI-B1.

[0159] (3) AI-B1 priority calculus strategy:

[0160] After receiving data from the host AI, AI-B1 initiates a priority calculation mechanism, prioritizing the concentration of computing power to calculate all feature values ​​of blocks that have not yet received external feature values.

[0161] The advantage of this strategy is that it makes full use of external computing power (i.e. the computing power of the host AI) to handle the relatively complex and critical localization and feature value calculation work, while focusing its own computing power on filling the remaining local computing power needs. This truly makes the most of computing power and greatly improves the overall computing efficiency and positioning accuracy.

[0162] This embodiment significantly improves the positioning accuracy of moving objects through collaborative computation between the host AI and AI-B1, combined with precise feature value comparison, effectively avoiding positioning errors caused by insufficient computing power or a single algorithm in traditional navigation systems. Utilizing an innovative priority computation strategy, the computing resources of the host AI and AI-B1 are rationally allocated, fully leveraging the advantages of external computing power and ensuring efficient use of local computing power, reducing unnecessary computing waste and improving the overall system computational efficiency. This embodiment is applicable to various travel scenarios, flexibly handling complex urban roads, remote suburban environments, and high-speed traffic routes, providing stable and accurate navigation services for moving objects.

[0163] After completing the aforementioned correction work, subsequent operations can be carried out based on the corresponding feature values ​​of the host AI. On the one hand, the results learned from the corresponding learning dataset can be directly transmitted and used as a program patch; on the other hand, the corresponding learning dataset can be used in conjunction with operational context data, and even transformed to simulate the operational context data of other devices, to continuously promote evolution and iteration, thereby achieving the desired effect.

[0164] Furthermore, the corresponding learning dataset of this invention can be a specific set of cases specified by regulatory agencies, or it can incorporate elements such as regulations and customs. Regardless of the form, it must include both positive and negative cases to allow for verification using external feature values ​​and a security correction process.

[0165] In this embodiment of the invention, the mobile terminal artificial intelligence correction method supports cross-model and cross-device feature value sharing. The method further includes the following steps: In step S5, the transformed same-relation feature values ​​are transmitted to other mobile terminals or host AI in a desensitized format, so that the mobile terminal AI and the host AI share feature values; In step S6, the discrete calculation is coordinated by a program to achieve shared computing power, or the transmission mode is flexibly switched, and emergency coordination and allocation are carried out to achieve real-time coordinated calculation between multiple terminals to meet the needs of different scenarios.

[0166] The method of this invention has particularly prominent advantages in the field of autonomous driving. In autonomous driving systems, changes in the state of vehicles ahead require rapid information transmission and processing to ensure safe driving. Current autonomous driving systems typically rely on real-time data transmission and computing capabilities, but in complex traffic environments (such as rainy days or sudden road surface changes), control errors, signal loss, or delays may occur due to interrupted information transmission or insufficient computing resources. Therefore, ensuring efficient communication of information between different devices and fully utilizing computing resources to improve system response speed and reliability are urgent technical challenges that need to be addressed.

[0167] Figure 6 This is one embodiment of the present invention. Figure 6 A demonstrates the steps involved in enabling multiple terminal AI systems to incorporate externally sourced feature values ​​directly into neural networks for use. Figure 6 B illustrates a schematic diagram of how external source feature values, after correction and security verification, are incorporated into a neural network. This implementation provides an intelligent driving assistance decision-making method based on internal model construction and multi-terminal collaboration in the fields of artificial intelligence and intelligent transportation.

[0168] See Figure 6 A, the method includes the following steps:

[0169] (1) Internal model construction and iterative correction:

[0170] In this embodiment, the naming convention for the various autonomous vehicle terminals can be, for example, AI-B1, AI-B2, AI-C1, AI-C2, and so on. AI-B1 and AI-C1 are each equipped with unique technical modules, namely, simulating causal datasets and algorithm modules based on their respective contradictory or missing logic. For example, AI-B1 simulates the motion logic and behavioral changes of the vehicle in front, while AI-C1 simulates the response logic of the vehicle behind, including actions such as braking and deceleration, and performs an iterative correction process at the terminal.

[0171] In automotive driving scenarios, authoritative and targeted data can be provided by regulatory or judicial authorities to optimize models and make them more closely match actual driving needs.

[0172] By utilizing these datasets and modules, we can not only continuously optimize our own models, but also use them as a basis to train the tactical coordination mode between AI-B1 and AI-B2 or AI-C1, thereby enhancing the collaborative response capabilities of multiple terminals in complex scenarios.

[0173] (2) Multi-terminal data interaction and intelligent decision-making:

[0174] In real-world driving scenarios, taking the typical driving behavior of "slowing down when the vehicle in front slows down" as an example, the learning dataset contains such information, and the external feature values ​​correspond to and match it. For instance, the state of the vehicle in front (such as speed and acceleration) serves as an external feature value, processed collaboratively by AI-B1 and AI-C1. The state of the vehicle in the middle (such as distance and speed difference) serves as an external feature value, used to assist the vehicle in front in making decisions. The state of the vehicle behind (such as hazard lights and braking status) serves as an external feature value, used to optimize the reaction of the vehicle behind. When encountering extreme emergencies such as "it's raining and the road ahead suddenly collapses," despite obstructed visibility, AI-B1, AI-B2, and AI-C1 can quickly share information by exchanging data.

[0175] Specifically, the external feature values ​​output by each terminal (such as the external feature values ​​output by AI-B1, the external feature values ​​transmitted by AI-B2, and the external feature values ​​transmitted by AI-C1) rapidly activate the decision-making mechanism of the new neural network through information interaction, based on the original neural network. Under the condition of limited computing resources (such as the single-core CPU inside the vehicle), by sharing external feature values, the existing computing power is maximized, so that the vehicles behind, the vehicles further behind, and even the vehicles further back can know the situation in time and quickly take evasive actions such as braking, effectively avoiding traffic accidents.

[0176] This invention, through the construction of an internal model based on contradictory or missing logic, can accurately simulate the causal relationships of various road conditions, providing a scientific basis for driving decisions and greatly improving the accuracy of decision-making in response to emergencies. Iterative correction and tactical coordination mode training across multiple terminals enable efficient collaboration between different terminals, ensuring the timeliness and accuracy of information transmission under complex road conditions and improving the overall traffic flow efficiency.

[0177] The method of this invention can also be applied to scenarios involving the autonomous and continuous evolution of robotic soccer teams. In robotic soccer matches, teams need to respond to various match scenarios and opponent strategies through teamwork and real-time decision-making. Current robotic soccer systems typically rely on centrally-controlled systems or highly manual intervention processes, lacking autonomous learning capabilities. When facing new environments, new opponents, or unexpected situations, they struggle to quickly adjust strategies and achieve effective collaboration, thus limiting the adaptability and flexibility of robotic soccer systems.

[0178] In another embodiment of the present invention, a method is provided that enables a robot soccer team to achieve autonomous and continuous evolution in the field of artificial intelligence and robot competition. The method includes the following steps:

[0179] (1) Iterative learning based on the learning dataset:

[0180] A meticulously constructed learning dataset, along with a vast amount of past match data, was provided to the machine soccer team for iterative learning. This dataset encompasses a wealth of content, including soccer tactical theory, classic match cases, and player technique specifications, providing the machine players with a comprehensive knowledge base.

[0181] Meanwhile, by deeply analyzing past match data, the system can accurately pinpoint problems and optimization directions by examining the robot players' decision-making, action execution, and teamwork performance in different scenarios. For example, when facing a dense defense, the system can identify successful breakthroughs from historical match data, summarize key elements such as tactical application, player positioning, and passing timing, and then guide current training and optimization.

[0182] Based on these abundant data resources, the robot player periodically optimizes and improves itself according to the set iterative rules. Each iteration integrates newly learned knowledge into its decision-making model, continuously adapting to the increasingly complex and ever-changing competition environment.

[0183] (2) Sharing feature values ​​improves computing power and communication efficiency:

[0184] To address the bottlenecks in computing power and communication difficulties, this embodiment introduces a shared feature value technology. In the robot soccer team, the robot players achieve efficient utilization of computing power by sharing a unified feature value system.

[0185] For example, when multiple robotic players simultaneously analyze the situation in a certain area of ​​the field, previously each player had to independently calculate a large amount of repetitive data, consuming a significant amount of computing power. Now, each player's motion data (such as speed and acceleration) can be used as external feature value 1 to optimize that player's action response. Interaction data between players (such as passing distance and receiving timing) can be used as external feature value 2 to optimize team collaboration strategies.

[0186] By sharing feature values, they only need to calculate the differentiated parts, greatly saving overall computing power and allowing more computing power to be allocated to key decision-making stages. For example, the allocation ratio of computing resources can be dynamically adjusted according to the competition situation, increasing the computing power of AI-B1 at critical moments when rapid decision-making is required. In scenarios requiring large-scale data processing (such as data analysis and model updates), AI-C1 or other AI modules are used first.

[0187] In terms of communication efficiency, shared feature values ​​act as an efficient "language." Based on these feature values, robot players can quickly transmit key information, such as the ball's position, teammates' runs, and the opponent's defensive posture, avoiding the lengthy transmission and interpretation of complex information. This achieves near real-time information sharing, greatly improving the team's tacit understanding and comparable to the efficient communication model demonstrated by professional athletes in emergency coordination training.

[0188] This invention, through an adaptive learning and iterative approach, enables the robotic soccer team to quickly adjust its strategies and achieve efficient collaboration in different matches. Simultaneously, by fully utilizing optimized allocation of computing resources, the overall performance and flexibility of the system are further enhanced, allowing the robotic soccer team to demonstrate greater autonomy and competitiveness in matches.

[0189] Based on the same concept, embodiments of the present invention also provide a mobile terminal artificial intelligence correction device.

[0190] It is understood that the mobile terminal artificial intelligence correction device provided in this embodiment of the invention includes hardware structures and / or software modules corresponding to each function in order to achieve the above-mentioned functions. In conjunction with the units and algorithm steps of the various examples disclosed in this embodiment, this invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the technical solutions of this disclosure.

[0191] Figure 7 This is a block diagram illustrating an artificial intelligence correction device for a mobile terminal according to an exemplary embodiment. (Refer to...) Figure 7 The communication processing device 100 includes a first correction module 101, a conversion module 201, a second correction module 102, and a security verification module 301.

[0192] The first correction module 101 is used to correct the same relation feature values ​​through iterative calculation. The first correction module includes a same relation feature value algorithm module and a corresponding external source feature value data field. The algorithm module is used to provide calculation of specified factor relations. When a specified factor relation occurs, the specified factor relation field or the corresponding external source feature value data field is incorporated into the neural network, thereby completing the correction.

[0193] The conversion module 201 is used to establish a conversion table or conversion function between the same relation feature values ​​of the mobile terminal AI and the external source feature values.

[0194] The second correction module 102, based on the conversion table or conversion function established by the conversion module, transmits feature values ​​to the mobile terminal AI, provides learning datasets, algorithm patches and program patches, and corrects the way external source feature values ​​are used.

[0195] The security verification module 301 is used to perform hardware degradation correction and feature value clustering verification to ensure the stable transmission and secure application of the external source feature values, and includes a first detection unit and a second detection unit.

[0196] The first detection unit is used to input the data of the standard product detection result into the pre-trained first detection model to obtain the first detection result. The first detection model is used to compare the data of the standard product detection result with the factory data, give instructions to correct hardware parameters or replace faulty components, and use the simulated scenario dataset provided by the host AI to verify the real-time output accuracy of the terminal sensor.

[0197] The second detection unit is used to perform cluster verification on the feature values ​​according to the preset clustering conditions. It uses a clustering algorithm to select feature value combinations that meet the preset safety threshold. When clustering fails, it readjusts the transformation table or transformation function, or supplements the external source feature values ​​to meet the preset clustering conditions.

[0198] In one embodiment of the present invention, the terminal artificial intelligence correction device further includes the following modules:

[0199] The second input module is used to transmit the transformed same relation feature values ​​to other terminals or host AI in a desensitized format, so that the mobile terminal and the host AI can share the feature values.

[0200] The second algorithm module is used to achieve real-time coordinated computation among multiple terminals through discrete computing power or emergency coordination allocation, enabling cross-model and cross-device feature value sharing.

[0201] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments of the method, and will not be elaborated upon here.

[0202] According to another aspect of the invention, the invention also provides a computer device including a memory and a processor.

[0203] Figure 8 This is a block diagram illustrating a computer device 400 according to an embodiment of the present invention. The computer device 400 may be a mobile phone, computer, digital broadcasting terminal, messaging device, game console, tablet device, medical device, wearable device, fitness equipment, personal digital assistant, etc.

[0204] Reference Figure 8 The computer device 400 may include one or more of the following components: processor 401, memory 402. It is understood that the computer device may also include common components such as power components, multimedia components, audio components, input / output (I / O) interfaces, sensor components, and communication components, etc. Figure 8 Not shown in the image.

[0205] Processor 401 typically controls the overall operation of computer device 400, such as operations associated with display, telephone calls, data communication, camera operation, and recording operations. Processor 401 may include one or more processors 402 to execute instructions to perform all or part of the steps of the methods described above. Furthermore, processor 401 may include one or more modules to facilitate interaction between processor 401 and other components. For example, processor 401 may include a multimedia module to facilitate interaction between multimedia components and processor 401.

[0206] Memory 402 is configured to store various types of data to support the operation of computer device 400, such as the steps of implementing the methods described in this invention when executing a computer program. Examples of such data include instructions for any application or method operating on computer device 400, contact data, phone book data, messages, pictures, videos, etc. Memory 402 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0207] According to another aspect of the present invention, a computer-readable storage medium is also provided, such as a memory 402 including instructions, on which a computer program is stored, which, when executed by one or more processors 401 of a computer device 400, implements the steps of the method described in the present invention. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0208] The present invention has been described in detail above. For those skilled in the art, the invention can be implemented in a wide range of ways with equivalent parameters, scenarios, and conditions without departing from its spirit and scope, and without requiring unnecessary experimentation. Although specific embodiments have been given, it should be understood that this disclosure is not limited to the specific implementations disclosed, and further improvements can be made to the invention. In summary, according to the principles of the invention, this application is intended to include any changes, uses, or improvements to the invention, including changes made using conventional techniques known in the art that depart from the scope disclosed herein. Some basic features can be applied within the scope of the following appended claims.

Claims

1. A terminal artificial intelligence (AI) correction method that can utilize external source feature values, characterized in that, Includes the following steps: S1. Receive external source feature values ​​and specified factor relationships transmitted by the host AI, receive learning dataset and algorithm patch, and correct the same relationship feature values ​​of the terminal AI; S2. Establish a transformation table or transformation function between the same relation feature values ​​of the terminal AI and the external source feature values ​​through dynamic iteration, and embed the transformation table or transformation function into the neural network; S3. Based on the transformation table or transformation function, correct the application mode of the external source feature value; S4. Perform hardware degradation correction and eigenvalue clustering verification, wherein the hardware degradation correction may be automated or semi-automated; Wherein, the same relational feature value is a feature value that has the same specified factor relationship, and the specified factor relationship includes at least one of the following: statistical correlation between feature values, parameter combination of a specific convolutional layer.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S11. Upload the detection data of the terminal AI to the host AI, supplement the external source feature value field, and add a feature value algorithm module for iteration; S12. If the specified factor relationship is not generated in S11, replace the data detection sensor or firmware of the terminal AI, find compatible factor relationships, add special algorithm patches and / or increase the learning dataset, and iterate again. S13. Repeat S11 to S12 until the specified factor relationship is generated; S14. Input the specified factor relationship and / or the external source feature value data field into the neural network to complete the correction; In S11, the host AI provides the learning dataset and performs format conversion by referencing the hardware parameters of similar devices used by the terminal AI; the converted learning dataset is then used for iterative processing. Alternatively, The host AI provides the learning dataset and its source, while the terminal AI obtains learning data by detecting similar learning data sources through its hardware and iterates through the data; or, The terminal AI obtains similar learning data suitable for its hardware specifications from other external sources and iterates through them; Preferably, in S11, the feature value algorithm module has a built-in suggested convolution mode, and the terminal AI iterates according to the suggested convolution mode, which is provided by the host AI or other third-party host.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: S21. Compare the specified factor relationship field corrected in S1 with the external feature value data field to form a terminal AI feature value conversion table; and optionally, S22. Simplify the feature value mapping relationship through statistical regression or function fitting, and verify that the feature value mapping relationship fit is not lower than a preset safety threshold, thus forming a program patch.

4. The method according to claim 1, characterized in that, Step S3 specifically includes: S31. The host AI calculates the feature value and transmits it to the terminal AI; S32. The mobile terminal AI determines the application mode of external source feature values. In simple mode, execute S33; in complex mode, execute S34. S33. The feature values ​​calculated by the host AI can be used directly or after transformation based on a transformation table or transformation function; S34. Use the feature values ​​calculated by the host AI in conjunction with external feature values ​​in a program or a specific learning dataset format.

5. The method according to claim 1, characterized in that, Step S4 specifically includes hardware degradation correction and eigenvalue clustering verification; The hardware degradation correction steps include: S41. By comparing the test results of standard products with the factory data, correct hardware parameters or replace faulty components; S42. Utilize the simulated scenario dataset provided by the host AI to verify the real-time output accuracy of the terminal sensors; The eigenvalue clustering verification step includes: S43. Use a clustering algorithm to filter feature value combinations that meet the preset security threshold; S44. When clustering fails, readjust the transformation table or transformation function, or supplement the external source feature values ​​to meet the preset clustering conditions.

6. The method according to claim 1, characterized in that, The method further includes: S5. Transmit the transformed same-relationship feature values ​​to other terminal AIs or host AIs in a de-identified format, so that the terminal AIs and host AIs can share the feature values; S6. Real-time coordinated computing among multiple terminals is achieved through discrete computing power or emergency coordination allocation.

7. A method for host artificial intelligence (AI) to assist terminal artificial intelligence (AI) correction using external source feature values, characterized in that, include: In response to a request from the terminal AI, provide external source feature values ​​or computed feature values.

8. The method according to claim 7, characterized in that: The provision of external source feature values ​​includes: In the step of correcting the same relation feature value in the terminal AI, the host AI receives the monitoring data uploaded by the terminal AI, supplements the external source feature value field, adds a feature value algorithm module, and generates a specified factor relationship to be transmitted to the terminal AI. When the terminal AI corrects the way external source feature values ​​are used, the external source feature values ​​are provided by the AI. During the feature value clustering verification step of the terminal AI, external source feature values ​​are added until the preset clustering conditions are met. The calculated feature values ​​include: After the terminal AI establishes a feature value conversion table or conversion function, it receives the feature value conversion table or conversion program and uses surplus computing power to assist in the mapping and conversion between the feature values ​​of the terminal AI and the feature values ​​of the host AI. In the step of the terminal AI performing hardware degradation correction, the detection data or firmware parameters and specifications transmitted by the terminal AI are received, feature values ​​are calculated and transmitted to the terminal AI for comparison between the detection data and the feature values.

9. The method according to claim 7, characterized in that: The host AI and the terminal AI jointly calculate the same relational feature values ​​and specific feature values, and compare them using a feature value transformation table or transformation function to form a multimodal feature value collaborative confirmation.

10. The method according to claim 9, characterized in that: In the multimodal feature value collaborative verification step, the host AI receives the detection data uploaded by the terminal AI, uses surplus computing power to calculate specific feature values ​​and compares them with external source feature values ​​pre-stored within the host AI, transmits the corrected external source feature values ​​and specific feature values ​​to the terminal AI, and the terminal AI calculates the specific feature values ​​that do not achieve external source feature value matching.

11. A terminal artificial intelligence (AI) correction device, characterized in that, Includes the following modules: The first correction module is used to correct the same relation feature values ​​through iterative calculation. The first correction module includes a same relation feature value algorithm module and a corresponding external source feature value data field. The algorithm module is used to provide calculation of specified factor relations. When a specified factor relation occurs, the specified factor relation field or the corresponding external source feature value data field is incorporated into the neural network, thus completing the correction. The conversion module is used to establish a conversion table or conversion function between the same relation feature values ​​of the terminal AI and the external source feature values; The second correction module, based on the conversion table or conversion function established by the conversion module, transmits feature values ​​to the terminal AI, provides learning datasets, algorithm patches and program patches, and corrects the way external source feature values ​​are used. The security verification module is used to perform hardware degradation correction and feature value clustering verification to ensure the stable transmission and secure application of the external source feature values, and includes a first detection unit and a second detection unit. The first detection unit is used to input the data of the standard product detection result into the pre-trained first detection model to obtain the first detection result. The first detection model is used to compare the data of the standard product detection result with the factory data, give instructions to correct hardware parameters or replace faulty components, and use the simulated scenario dataset provided by the host AI to verify the real-time output accuracy of the terminal sensor. The second detection unit is used to perform cluster verification on the feature values ​​according to the preset clustering conditions. It uses a clustering algorithm to select feature value combinations that meet the preset safety threshold. When clustering fails, it readjusts the transformation table or transformation function, or supplements the external source feature values ​​to meet the preset clustering conditions.

12. The terminal artificial intelligence (AI) correction device according to claim 11, characterized in that, It also includes the following modules: The second input module is used to transmit the transformed same relation feature values ​​to other terminals or host AI in a desensitized format, so that the terminal and host AI can share the feature values. The second algorithm module is used to achieve real-time coordinated computing among multiple terminals through discrete computing power or emergency coordination allocation.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program that can run on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6 or 7-10.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by one or more processors, it implements the steps of the method according to any one of claims 1 to 6 or 7-10.

Images