Hyperspecialized models

Abstract

The embodiments generally relate to artificial intelligence (AI) model management systems, specifically focusing on a hierarchical framework for optimizing model accuracy, computing-efficiency, and privacy. The embodiments can organize AI models into categories such as generalized, specialized, and hyperspecialized models, and incorporate mechanisms for model selection, execution, training, and management.

Description

Patent T9145-25107W001 HYPERSPECIALIZED MODELS FIELD OF THE INVENTION

[0001] The embodiments of the invention generally relate to artificial intelligence model management systems, and more particularly toward a hierarchical framework for optimizing model accuracy.DISCUSSION OF THE RELATED ART

[0002] The field of artificial intelligence (Al) model management has witnessed significant advancements, paralleling the increasing demand for accuracy, efficiency, and privacy in Al applications. Al systems often deploy models that differ in scope and specialization, including generalized models applicable to a wide array of tasks as well as specialized models. In the context of managing diverse Al models, optimizing the accuracy, cost, and privacy poses substantial challenges.

[0003] Generalized Al models, such as ChatGPT, are used to work effectively in multiple domains. Specialized Al models, such as face recognition models, are built specifically for work in a single domain, but against a wide variety of inputs. An issue with these specialized models is that they can have gaps or can be biased based on the training data available and the algorithms used. They tend to be large, with many input parameters, and can be expensive to execute.

[0004] Another issue is data privacy. AI models need to be trained, and that training data can contain sensitive or personally identifiable information (PII). Since law, policy, and regulation of this are still in the early stages, there will eventually be issues with the use of individual information in Al training. In parallel, the idea of “self-sovereign identity" has emerged, which is the concept that the individual should control their identity data and choose whom to share it with.Patent T9145-25107W001

[0005] Current technologies in Al model management face several limitations.Generalized Al models, while versatile, often lack the precision required for specialized tasks due to their broad training parameters. On the other hand, specialized models, while more precise, can demand extensive resources and data, which may not be available or justified by the task requirements. Moreover, existing systems often do not provide adequate mechanisms to dynamically adapt model architecture based on changing input requirements or privacy regulations, leading to inefficiencies.SUMMARY OF THE INVENTION

[0006] Accordingly, the embodiments of the present invention are directed to hyperspecialized models that substantially obviate one or more problems due to limitations and disadvantages of the related art.

[0007] Additional features and advantages of the invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

[0008] In one aspect, the embodiments relate to systems, devices, methods and instructions for managing Al models, such as those organized hierarchically, to enhance processing efficiency. For example, the embodiments include a plurality of generalized models catering to a wide range of input data, specialized models addressing specific subject classes, and hyperspecialized models focused on individual subjects.

[0009] In an example embodiment, model selection and execution are facilitated through a model selector that utilizes one or more classification algorithms. This functionality determines the suitability of specialized or hyperspecialized models, thereby optimizing the balance between accuracy and computational overhead.Patent T9145-25107W001

[0010] A spectrum manager can be included to determine and maintain the necessary number of models within a hierarchy, thus ensuring the availability of an appropriate model for given input data.

[0011] A model executor and model trainer can execute selected models on input data, and conduct retraining with new data. This feature supports improvement in model performance, leveraging incoming data to refine the Al model hierarchy.

[0012] Yet another object of the embodiments is to address privacy and regulatory requirements through model management capabilities. The system comprises components that allow for the removal of models based on subject requests or privacy considerations. Furthermore, hyperspecialized models may be stored and executed on subject-controlled devices, providing a decentralized approach that enhances privacy.

[0013] In an example embodiment, a spectrum manager operates in an unsupervised manner, autonomously adjusting the count and type of models within the hierarchy. This enhances operational efficiency by ensuring that the model structure aligns with changing data dynamics and processing requirements.

[0014] In another example embodiment, the system’s identity resolver and subject database function to identify subjects based on input claims, maintaining relevant subject information necessary for model mapping and privacy assurance. This enables subject-specific Al processing within the hierarchical framework.

[0015] It is to be understood that both the foregoing general description and the following detailed description are examples and explanatory and are intended to provide further explanation of the invention as claimed.Patent T9145-25107W001 BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention.

[0017] Fig. 1 is a schematic illustrating different classes with training data for each class.

[0018] Fig. 2 is a schematic showing subjects processed by specialized models, with training data for each class.

[0019] Fig. 3 is a diagram depicting an Al model hierarchy transitioning from generalized to hyperspecialized models.

[0020] Fig. 4 is a diagram of an expanded model hierarchy, with more specialized models compared to Fig. 3.

[0021] Fig. 5 is a diagram similar to Fig. 4, showing a structured model hierarchy from generalized to hyperspecialized models.

[0022] Fig. 6 is a flowchart illustrating a method for classifying claims using a selection and evaluation process.

[0023] Fig. 7 illustrates a method for using hyperspecialized models.

[0024] FIG. 8 is a block diagram of a computer system that implements hyperspecialized models.Patent T9145-25107W001 DETAILED DESCRIPTION

[0025] Reference will now be made in detail to the embodiments, examples of which are illustrated in the accompanying drawings. In the following detailed description, numerous specific details are set forth in to provide a thorough understanding of the present invention. However, it will be apparent to one of ordinary skill in the art that the embodiments may be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the embodiments. Wherever possible, like reference numbers will be used for like elements.

[0026] Embodiments of user interfaces and associated methods for using an electronic device are described. In some embodiments, the electronic device can be any of a variety of computing devices, such as a computer, laptop, server, cloud computing device, portable communication device (e.g., a mobile phone or tablet), etc. The user interface may include a touch screen and / or other input / output devices, such as a keyboard or mouse.

[0027] The electronic device may support a variety of applications, such as telephone, text messenger, email, other customer-related applications, etc. The various applications that may be executed on the electronic device may use at least one common physical user-interface device, such as a touch screen. One or more functions of the touch screen as well as corresponding information displayed on the device may be adjusted and / or varied from one application to another and / or within a respective application. In this way, a common physical architecture of the device may support a variety of applications with user interfaces that are intuitive and transparent.Patent T9145-25107W001

[0028] The embodiments of the invention generally relate to artificial intelligence model management systems, specifically focusing on a hierarchical framework for optimizing model accuracy, computing-efficiency, and privacy. The embodiments can organize Al models into categories such as generalized, specialized, and hyperspecialized models, and incorporate mechanisms for model selection, execution, training, and management.

[0029] Provided are computer-based systems, devices, methods, and instructions for collecting subject data and automatically creating and updating a set of increasingly specialized Al models mapped to classes of subjects, and then selecting and using the best models for processing that subject. The set of models includes a generalized model for all subjects, models for classes of subjects, and hyperspecialized models mapped to a single subject. The system includes a method for selecting which candidate models should be used to process a particular subject to optimize accuracy, computing cost, and / or speed based on features of the subject and models in the set. The embodiments also are configured to remove hyperspecialized models and to retrain or remove other models in the set based on privacy regulations.

[0030] Accordingly, the embodiments provide a structured hierarchy of Al models, allowing for the appropriate application of generalized, specialized, and hyperspecialized models according to the task demands. The embodiments optimize the trade-offs between accuracy, computing-efficiency, and privacy, enhancing the adaptability and robustness of Al systems. This includes a dynamic mechanism to efficiently manage the allocation and execution of models, integrating an unsupervised approach to model spectrum management that aligns with data privacy standards and regulatory considerations. Such a framework addresses the inefficiencies and privacy concerns inherent in existing Al model management solutions.Patent T9145-25107W001

[0031] Here, the embodiments create hyperspecialized models tied to, and controllable by, subjects; establish a spectrum of non-individualized specialized models, covering from a “general population” level all the way to a model for a specific class of subjects or individual, controllable by automation; select which model or models are best suited for a given subject, whether general, specialized to a class within the spectrum, or hyperspecialized to the subject; train and improve the model selection process automatically; and automatically manage (e.g., add, remove, adjust) and train models within the spectrum as well as hyper-specialized models.

[0032] In the discussion that follows, some terms and components will be referred to and described. A brief description of the terms and components is provided, but the embodiments are not so limited.

[0033] Terms. Identity - A singular representation of a Subject and their verified Identifiers in the system. In one embodiment, this is a database containing Subjects and Identifiers. Identifier - A feature of the Subject that distinguishes it from other Subjects. Examples include face, name, email address, or passport or identification number. Claim - An assertion that an Identifier belongs to a Subject. General Model - A singular general-purpose model that can be used on all Subjects. Specialized Model - A model that is specialized for a class of Subjects. Hyperspecialized Model - A singular model that is specialized for one Subject. Selection Feature - A feature related to the Subject or of the input that can influence the performance of the Model. Examples for a face recognition embodiment could include skin tone, age, gender, ambient lighting color, time of day, temperature of the environment, camera type, camera orientation, distance from camera, geolocation, etc.

[0034] Components. Requestor - The system (or person) requesting an evaluation of the Subject. Subject - The target of the evaluation; an entity that can be uniquely identified and represented within the system. In a frequently occurring embodiment, this may be a Person. Identity Resolver - Determines the identity of the Subject based on a presented set of Claims, and whether a single unambiguous Identity exists for the Subject, or not. Identity Resolver - Determines the identity of thePatent T9145-25107W001 Subject based on a presented set of Claims, and whether a single unambiguous Identity exists for the Subject, or not. Subject Database - contains Subject data and therefore PII. HSM (Hardware Security Module) Database - contains personalized models and therefore PII. Spectrum Model Database - contains Specialized models that are not attributed to any one Subject. Model Selector - A classification algorithm that selects the best model or models to process the Subject with, given Selection Features. Model Executor - A computing device that executes a selected model on the input data to produce an output. Model Trainer - A process that trains or re-trains models with new input data.

[0035] Fig. 1 is a schematic illustrating different classes with training data for each class.

[0036] As shown, Fig. 1 depicts a schematic representation of training data classification used to train models. The diagram includes multiple groups labeled as ’’Training Data," " Class 1 Training Data," and " Class 2 Training Data," for example. Each group comprises stick figures with distinct geometric head shapes, indicating different classes of data. The first group, labeled " Class 1 Training Data," contains circle and triangle-headed figures. The second, " Class 2 Training Data," includes square-headed figures. The third, labeled simply as " Training Data," is a more general grouping, encompassing figures from the other two classes. A few unclassified figures exist outside these groups, representing potential subjects for classification.

[0037] Fig. 2 is a schematic showing subjects processed by specialized models, with training data for each class.

[0038] Turning to Fig. 2, the figure illustrates the process of classifying various subjects using models trained on specific classes of data. The diagram includes a flow of " Class 1 Subject" and " Class 2 Subject” processed by their respective specialized models, namely " Specialized Model for Class 1" and " Specialized Model for Class 2." Each specialized model is trained on the corresponding class training data depicted in Fig. 1. Additionally, an " Unclassified Subject" is processed by aPatent T9145-25107W001 " Generalized Model" that is trained on the combined training data set, also shown in Fig. 1, highlighting the flexibility of such models to handle diverse unclassified data.

[0039] Fig. 3 is a diagram depicting an Al model hierarchy transitioning from generalized to hyperspecialized models.

[0040] Fig. 3 presents a detailed system of model specialization by demonstrating a progression from a " Generalized Model" through to " Specialized Models" and finally to a " Hyperspecialized Model" for individual subjects. Initially, all subjects are processed using the " Generalized Model." The process then diverges into two models named " Arms - Out" and " Arms - Down" under " Specialized Models," indicating a more refined classification stage based on the arm positions of the subjects. Finally, each subject is further analyzed by a " Hyperspecialized Model,” providing individualized attention per subject for maximum specificity.

[0041] As an example, if the Spectrum Manager determines that two (2) classes of Subjects exist based on the Selection Feature of their arm position, it could create a Model Spectrum as shown below. Subjects with “Arms Out” would be processed by one model, and Subjects with “Arms Down” would be processed by the other.Subjects that could not be fit properly into either class and that do not have a Hyperspecialized Model will be processed by the Generalized Model. Subjects with a Hyperspecialized Model would be processed by it.

[0042] Fig. 4 is a diagram of an expanded model hierarchy, with more specialized models compared to Fig. 3.

[0043] Fig. 4 elaborates further on the specialization process with an added layer of detail. Starting again with a " Generalized Model” encompassing all subjects, the process branches into two distinct classifications within " Specialized Models" labeled " Arms - Out" and " Arms - Down.” Notably, the arm positions are subdivided into four models based on " Head Shape - Triangle" and " Head Shape - Non-Triangle," introducing a more granular differentiation criterion. This refinement leads to the same " Hyperspecialized Model" tier, where individual subjects are processed for optimized classification.Patent T9145-25107W001

[0044] Fig. 5 is a diagram similar to Fig. 4, showing a structured model hierarchy from generalized to hyperspecialized models.

[0045] Fig. 5 maintains a similar format to Fig. 4, showcasing the repetitive structure necessary for specialized classification. The figure follows a pathway from a " Generalized Model" containing all subjects to a segment labeled " Specialized Models," which further refines arms' positions and head shapes. These series of transitions culminate in the " Hyperspecialized Model" segment, sustaining focus on precise classification for each individual subject among the group.

[0046] As another example, if the Spectrum Manager determines that four (4) classes of Subjects exist based on the Selection Features of their arm position and head shape, it could create a Model Spectrum as shown in Fig. 5. Subjects would be processed by one of the four Specialized Models. Subjects that could not be fit properly into these classes and that did not have a Hyperspecialized Model would be processed by the Generalized Model. Subjects with a Hyperspecialized Model would be processed by it.

[0047] As illustrated in Fig. 5 (and Fig. 4), an aspect of the Spectrum Manager is that of determining the “maturity” of the model for a particular class of Subjects. It may be configured to determine that a given class exists but does not (a) have enough training data available in order to be effective, or (b) represents too few Subjects to preserve privacy. For example, the “Triangle Head” models only have 1 Subject and would not be used because they are effectively equivalent to the Hyperspecialized Models for each Subject.

[0048] These examples only illustrate a minimal and human-understandable set of Selection Features, and only illustrate Subject-based features. The Spectrum Manager establishes the Model Spectrum clusters based on hundreds of dimensions given the Selection Features available. The clusters may not be explainable to the operators, in contrast to rules-based and hierarchical methods of creating the Model Spectrum. This is intentional to mitigate opportunities for bias.Patent T9145-25107WO01

[0049] In some configurations, the Spectrum Manager employs a modified density-based spatial clustering algorithm that operates on a multi-dimensional feature space constructed from selection features extracted from training data. Unlike conventional clustering approaches that require predetermined cluster counts, the Spectrum Manager dynamically determines the optimal number of specialized models by iteratively evaluating cluster cohesion metrics. Specifically, the system may calculate a Silhouette coefficient for each potential clustering configuration. The Spectrum Manager may maintain specialized models only when the average Silhouette score across all clusters exceeds a configured threshold, ensuring that each specialized model provides measurably superior performance for its designated subject class compared to the generalized model.

[0050] The Spectrum Manager may further implement a technical solution to the computational challenge of maintaining model diversity while preventing overfitting to small subject populations. When evaluating whether to instantiate a new specialized model for a particular cluster, the system may apply a minimum viable population threshold that accounts for both the dimensionality of the input feature space and the complexity of the model architecture. This ensures that each specialized model has sufficient training examples relative to its complexity, thereby preventing the creation of specialized models that would effectively memorize training data rather than learning generalizable patterns. This technical approach addresses the computational problem of resource allocation in hierarchical model systems by preventing the proliferation of specialized models that consume memory and processing resources without providing accuracy benefits over the generalized model.

[0051] Fig. 6 is a flowchart illustrating a method for classifying claims using a selection and evaluation process.

[0052] Turning to Fig. 6, the figure depicts a flowchart encapsulating the methodological process for classifying claims, beginning with the step of " Establish Claims." Subsequently, the chart demonstrates a selection process for " Select Classifier" followed by " Classify Claim." Two potential outcomes are indicated:Patent T9145-25107WO01 ''Unacceptable Classification," which reverts to the classifier selection stage, and " Acceptable Classification," leading to completion of the classification cycle. This flowchart provides a roadmap for handling claim classifications effectively and ensuring accuracy throughout the procedure.

[0053] Fig. 7 illustrates a method for using hyperspecialized models.

[0054] At 701, identify the Subject and match them to their HSM. Within 710 (1a), a Requestor submits Claims for the Subject - one or more biometric, biographic, or account-based identifiers can be used to identify them. Within 710 (1b), Identity Resolver identifies Subject using Claims to determine their identity as registered in the Subject Database to determine if they already have an HSM established in the HSM Database. If an unambiguous identity was found, propose use of the HSM associated with that identity. If no unambiguous identity was found, create a new identity and HSM associated with that identity and propose its use. If an ambiguous identity was found, continue with no proposed HSM. Within 710 (1c), if an HSM was proposed, evaluate its readiness: If it is “mature”, select it as a Processing Model, and mark it as a Training Model. If it is “not mature”, mark it as a Training Model.

[0055] At 720, determine which Spectrum Model is most appropriate for the Subject given the Selection Features. Within 720 (2a), attempt to select one or more Models from Spectrum Model Database using the Model Selector. Determine Selection Features to be used to select the Processing Mode. Select candidate Models. For each, evaluate its readiness: If it is “mature”, select it as a Processing Model and mark it as a Training Model. If it is “not mature”, mark it as a Training Model. Within 720 (2b), if no mature models were found, use the General Model as the Processing Model.

[0056] At 730, execute the Models (General, Spectrum and / or HSM). Within 730 (3a), with the Model Executor, execute all Processing Models on the Subject using the input data. Within 730 (3b), make results available to the Requestor and to the Model Selection Al. For the Requestor, this could take the form of an API response or updates to an application user interface.Patent T9145-25107WO01

[0057] At 740, retrain the Models. Within 740 (4a), with the Model Trainer, re-train all Training Models using the input data collected from the Subject. Use an “online training" approach to continuously update and refine each Training Model. Within 740 (at 4b), link the Subject in the Subject Database to each Model that was trained.

[0058] At 750, Remove Models. If any Requestor initiated a Subject removal - remove the HSM for the Subject from the HSM Database, and - remove Specialized Models linked to the Subject from the Spectrum Model Database. Based on configuration of the specialization of the Specialized Models to a class of Subjects with too few members, remove any Specialized Models where the Subject was used to train that model. For example, if a Specialized Model was only trained on 3 Subjects, and the configuration was to remove models trained on less than 20 Subjects, that model would be removed.

[0059] At 760, refactor the Spectrum to optimize it. Evaluate the clusters within the Model Spectrum to determine if the addition or removal of a cluster is needed (spectrum refactoring). If the Silhouette Score for each cluster is above a configured threshold, attempt to add a cluster and recompute the new Silhouette Score.

[0060] Various modifications and variations can be made in the hyperspecialized models. One alternative configuration is to remove the Spectrum entirely, and only provide the Generalized Model and the Hyperspecialized Models. This would be appropriate in environments with limited computing power, where the Spectrum Manager would be too impactful on the available computing power or where very specific privacy regulations are in place.

[0061] In another example, an alternative configuration is to store and maintain the Hyperspecialized Model on a communications device controlled by the Subject (e.g., their smartphone) but train and execute it within the system. In this configuration, the Subject uses their device to send their HSM to the system to use in a single processing event. The system processes the input using the provided HSM, conducts the online training of the HSM, transmits the updated HSM to the Subject’s device, then removes the HSM from its own memory and storage. This configurationPatent T9145-25107WO01 increases security, but does not require a high level of computing power on the Subject’s device.

[0062] In another example, an alternative configuration is to entirely store, maintain, train, and execute the Hyperspecialized Model on a communications device controlled by the Subject (e.g., their smartphone) instead of within the system. In this configuration, the system sends the input data to the Subject’s device for processing and receives the output, and the online training of the HSM occurs on the Subject’s device. This configuration increases security, but requires sufficient available computing power on the Subject’s device.

[0063] In yet another example, synthetic data can be used to effectively "precompute" the likely specializations. If the synthetic data is statistically similar to real data, the spectrum of models could be determined without using real data.Additionally, having one of the specialization paths can be configured to identify synthetic versus real data. Here, this could be used could be used for “fake data detection”, such as deepfake or Al video. Also, a synthetic data training model can be used as one of the models that the specialized model creates. There can be instances where there is not enough training data for the system to be able to differentiate or focus in on a certain category. In that instance, the model could create a separate hyper-specialized model that creates synthetic data sets that it has either learned from other sets of training data associated with other categories that generated a hyperspecialized model, or the system could generate its own set of synthetic data that could be useful for training.

[0064] FIG. 8 is a block diagram of a computer system 800 that implements hyperspecialized models.

[0065] As illustrated in FIG. 8, system 800 may include a bus 112 and / or other communication mechanism(s) configured to communicate information between the various components of system 800, such as a processor 122 and a memory 114. In addition, a communication device 120 may enable connectivity between processor 122 and other devices by encoding data to be sent from processor 122 to anotherPatent T9145-25107WO01 device over a network and decoding data received from another system over the network for processor 122.

[0066] For example, communication device 120 may include a network interface card that is configured to provide wireless network communications. A variety of wireless communication techniques may be used including infrared, radio, Bluetooth, Wi-Fi, and / or cellular communications. Alternatively, communication device 120 may be configured to provide wired network connection(s), such as an Ethernet connection.

[0067] Processor 122 may comprise one or more general or specific purpose processors to perform computation and control functions of system 800. Processor 122 may include a single integrated circuit, such as a micro-processing device, or may include multiple integrated circuit devices and / or circuit boards working in cooperation to accomplish the functions of processor 122.

[0068] System 800 may include memory 114 for storing information and instructions for execution by processor 122. Memory 114 may contain various components for retrieving, presenting, modifying, and storing data. For example, memory 114 may store software modules that provide functionality when executed by processor 122. The software modules may include an operating system 115 that provides operating system functionality for system 800. The software modules may further include artificial intelligence, self-learning, and validation modules 116 for hyperspecialized modules configured to concurrently (e.g., simultaneously) execute the functionality described in connection with FIGs. 1-7, as well as other functional modules 118, such as a module configured to automatically create and update a set of increasingly specialized Al models mapped to classes of subjects, and then selecting and using the best models for processing that subject. The Al modules 116 are coupled to bus 112 to provide centralized access to Al functionality.

[0069] Memory 114 may include a variety of computer-readable media that may be accessed by processor 122. For example, memory 114 may include any combination of random access memory (“RAM”), dynamic RAM (“DRAM”), staticPatent T9145-25107WO01 RAM (“SRAM”), read only memory (“ROM”), flash memory, cache memory, and / or any other type of non-transitory or transitory computer-readable medium.

[0070] Processor 122 is further coupled via bus 112 to a display 124, such as a stationary display. A keyboard 126 and a cursor control device 128, such as a computer mouse, are further coupled to communication device 120 to enable a user to interface with system 800.

[0071] Database(s) 117 may store one or more customer related applications.Database 117 may store data in an integrated collection of logically-related records or files. Database 117 may be an operational database, an analytical database, a data warehouse, a distributed database, an end-user database, an external database, a navigational database, an in-memory database, a document-oriented database, a real-time database, a relational database, an object-oriented database, or any other database known in the art.

[0072] In summary, the provided technology offers a scalable and flexible framework for Al model management. By structuring models hierarchically and incorporating privacy-compliant management techniques, the system addresses diverse processing needs while adhering to privacy and regulatory standards.

[0073] It will be apparent to those skilled in the art that various modifications and variations can be made in the hyperspecialized models of the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention cover the modifications and variations of this invention provided they come within the scope of the appended claims and their equivalents.

Claims

Patent T9145-25107WO01 What is claimed is:

1. A system for managing artificial intelligence (Al) models structured in a hierarchical manner, the system comprising:a plurality of generalized models suitable for a broad range of input data; a plurality of specialized models trained for classes of subjects; anda plurality of hyperspecialized models tailored for individual subjects.

2. The system of claim 1, further comprising:a spectrum manager configured to determine and maintain a number of models within the hierarchy;a model selector operable to select suitable models from the hierarchy based on selection features related to input data;a model executor and trainer configured to execute selected models and conduct training with new data; andan identity resolver and a subject database configured to identify subjects based on claims and maintain subject-related information.

3. The system of claim 2, wherein the spectrum manager operates in an unsupervised manner to adjust the number and type of models within the hierarchy.

4. The system of claim 2, wherein the model selector utilizes a classification algorithm to determine the appropriateness of specialized or hyperspecialized models for processing input data.

5. The system of claim 1, further comprising a capability for model management that removes models based on subject requests or privacy regulations.

6. The system of claim 1, wherein the hyperspecialized models are configured to be stored and executed on devices controlled by the subjects.

7. A method for utilizing an artificial intelligence (Al) model hierarchy, the method comprising:identifying subjects from input claims using an identity resolver and a subjectPatent T9145-25107WO01 database;selecting one or more models from a hierarchy comprising generalized, specialized, and hyperspecialized models, based on selection features; and executing the one or more selected models on input data.

8. The method of claim 7, wherein selection of models is performed by a model selector utilizing a classification algorithm to evaluate selection features.

9. The method of claim 7, wherein retraining is performed using new data inputs.

10. The method of claim 7, further comprising removal of hyperspecialized models from the hierarchy when suitable conditions are met, including subject requests or regulatory compliance.

11. The method of claim 7, wherein hyperspecialized models are managed in a decentralized manner, allowing model storage and execution on subject-controlled devices.

12. A non-transitory computer readable storage medium storing one or more programs for utilizing an artificial intelligence (Al) model hierarchy, the one or more programs configured to be executed by a processor, the one or more programs comprising instructions for:identifying subjects from input claims using an identity resolver and a subject database;selecting one or more models from a hierarchy comprising generalized, specialized, and hyperspecialized models, based on selection features; and executing the one or more selected models on input data.

13. The non-transitory computer readable storage medium of claim 12, wherein selection of models is performed by a model selector utilizing a classification algorithm to evaluate selection features.

14. The non-transitory computer readable storage medium of claim 12, wherein retraining is performed using new data inputs.Patent T9145-25107WO01 15. The non-transitory computer readable storage medium of claim 12, further comprising removal of hyperspecialized models from the hierarchy when suitable conditions are met, including subject requests or regulatory compliance.

16. The non-transitory computer readable storage medium of claim 12, wherein hyperspecialized models are managed in a decentralized manner, allowing model storage and execution on subject-controlled devices.

Images