ABSTRACT
Abstract
An advanced model management platform for optimizing and securing generative artificial intelligence systems such as large language models (LLMs) and diffusion models. The platform incorporates various techniques to address the limitations of current generative AI systems, such as hallucination, lack of validation, security vulnerabilities, and inadequate model management. The system employs reinforcement learning algorithms for model optimization, retrieval augmented generation (RAG) for hallucination mitigation, domain-specific validation against expert knowledge, model distillation and similarity scoring for security, adversarial training for robustness, and attention mechanism search and model blending for advanced management and neuro symbolic AI routine combinations. By integrating these techniques, the platform significantly improves the performance, reliability, and security of generative AI across a wide range of tasks and domains leveraging the best elements of symbolic and connectionist techniques alongside automated planning and modeling simulation.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
Priority is claimed in the application data sheet to the following patents or patent applications, each of which is expressly incorporated herein by reference in its entirety:
Ser. No. 18/668,137 Ser. No. 18/656,612 63/551,328
BACKGROUND OF THE INVENTION
Field of the Art
The present invention is in the field of large-scale cloud computing, and more particularly to distributed computing platforms, leveraging graph-based specifications for data flows, transport and storage, and logic locality in data handling to enable artificial intelligence enhanced decision-making and automation and planning including those employing large language models, neuro-symbolic AI, and associated services across heterogeneous cloud, managed data center, edge, and wearable/mobile devices.
Discussion of the State of the Art
Current artificial intelligence (AI) systems, including large language models (LLMs) and generative AI (Gen AI), have limitations in their utility stemming from their lack of symbolic reasoning capabilities and integration of symbolic knowledge and limited capabilities for automated planning alongside the limited explainability of connectionist modeling approaches. There is a need for AI system architectures that can bridge symbolic and non-symbolic, i.e. connectionist, representations to enable more advanced, contextual reasoning while also considering critical factors like security, traceability, and collaborative model development and legal, privacy, and data owner restrictions around the locality of data across transport, storage and compute stages of data flow specified processes at both logical and physical levels.
Large language models and other connectionist AI techniques have demonstrated remarkable capabilities in various natural language processing, classification and inference tasks across text, sound, video, and even in chemical, molecule, mathematical and other specialized areas. However, they suffer from limitations such as hallucination, lack of validation, security vulnerabilities, and inadequate model management and suitability monitoring that negatively impact trust and applicability to mission critical utilization.
What is needed is an advanced model management platform for optimizing and securing large language models, neuro-symbolic AI systems, and aiding in efficient use of RAGs, transfer and federated learning, structured expert judgment, and fine-tuning capabilities in production environments.
SUMMARY OF THE INVENTION
Accordingly, the inventor has conceived and reduced to practice, an advanced model management platform for optimizing and securing artificial intelligence enhanced processes including those employing large language models and neuro symbolic models in real-world processes and production environments. The platform incorporates various techniques to address the limitations of current AI systems, including generative AI applications, such as hallucination, lack of validation, security vulnerabilities, and inadequate model management and fitness evaluation or challenges and explainability. The system employs combinations of random search, stochastic search and reinforcement learning algorithms, embedding and retrieval augmented generation (RAG) and knowledge graph enhancements for prompt or input improvements or prompt security/constraint satisfaction or output validation or use limitation, domain-specific knowledge corpus validation against expert or group of expert knowledge, transfer or federated learning enhancements or logic migration across mobile or edge or cloud resources, probabilistic consideration of distributed computing operational challenges, model distillation and output similarity scoring for security based on acceptable output checks and distance or consensus logic, adversarial training for robustness, and model type or attention mechanism search and model blending or consensus for advanced management or output weighting or downstream logic or planning inputs. By integrating these techniques, the platform significantly improves the performance, reliability, and security of generative AI systems across a wide range of reasoning tasks and domains. System may also take collections of specialized models, e.g. text, audio, image, physics, chemical, and combine such elements into spatial intelligence representations for scene or scenario-level planning of interest in many practical personal, business and robotics applications in domestic, commercial and industrial environments.
According to a preferred embodiment, a computing system for optimizing and securing generative AI models employing an advanced model management platform is disclosed, the computing system comprising: one or more hardware processors configured for: employing reinforcement learning algorithms to optimize a model's settings based on task-specific reward functions; using retrieval augmented generation (RAG) to retrieve relevant information from internal and external knowledge sources combined into a curated knowledge corpus and conditioning the model's output on retrieved facts or beliefs; validating model performance against human experts or crowds and authoritative databases or rule sets using cross-validation and domain-specific question-answering; employing model distillation, differential privacy, similarity scoring, and nearness models to create and secure the model and detect infringement; using adversarial training and input perturbation to test the robustness the model against poisoning attacks and adversarial examples; and incorporating attention mechanism search, model blending, and RAG to optimize the performance and reliability of the model for specific tasks.
According to another preferred embodiment, a computer-implemented method executed on an advanced model management platform for optimizing and securing generative AI models is disclosed, the computer-implemented method comprising: employing reinforcement learning algorithms to optimize a model's settings based on task-specific reward functions; using retrieval augmented generation (RAG) to retrieve relevant information from internal and external knowledge sources combined into a curated knowledge corpus and conditioning the model's output on retrieved facts or beliefs; validating model performance against human experts or crowds and authoritative databases or rule sets using cross-validation and domain-specific question-answering; employing model distillation, differential privacy, similarity scoring, and nearness models to create and secure the model and detect infringement; using adversarial training and input perturbation to test the robustness the model against poisoning attacks and adversarial examples; and incorporating attention mechanism search, model blending, and RAG to optimize the performance and reliability of the model for specific tasks.
According to another preferred embodiment, a system for optimizing and securing generative AI models employing an advanced model management platform is disclosed, comprising one or more computers with executable instructions that, when executed, cause the system to: employ reinforcement learning algorithms to optimize a model's settings based on task-specific reward functions; use retrieval augmented generation (RAG) to retrieve relevant information from internal and external knowledge sources combined into a curated knowledge corpus and conditioning the model's output on retrieved facts or beliefs; validate model performance against human experts or crowds and authoritative databases or rule sets using cross-validation and domain-specific question-answering; employing model distillation, differential privacy, similarity scoring, and nearness models to create and secure the model and detect infringement; use adversarial training and input perturbation to test the robustness the model against poisoning attacks and adversarial examples; and incorporate attention mechanism search, model blending, and RAG to optimize the performance and reliability of the model for specific tasks.
According to another preferred embodiment, non-transitory, computer-readable storage media having computer-executable instructions embodied thereon that, when executed by one or more processors of a computing system employing an advanced model management platform for optimizing and securing generative AI models, cause the computing system to: employ reinforcement learning algorithms to optimize a model's settings based on task-specific reward functions; use retrieval augmented generation (RAG) to retrieve relevant information from internal and external knowledge sources combined into a curated knowledge corpus and conditioning the model's output on retrieved facts or beliefs; validate model performance against human experts or crowds and authoritative databases or rule sets using cross-validation and domain-specific question-answering; employing model distillation, differential privacy, similarity scoring, and nearness models to create and secure the model and detect infringement; use adversarial training and input perturbation to test the robustness the model against poisoning attacks and adversarial examples; and incorporate attention mechanism search, model blending, and RAG to optimize the performance and reliability of the model for specific tasks.
According to an aspect of an embodiment, the reinforcement learning algorithms comprise Proximal Policy Optimization or Asynchronous Advantage Actor-Critic algorithms.
According to an aspect of an embodiment, the model's output on retrieved facts further comprises entity linking and knowledge graph, vector database, or vectorized knowledge graph hybrid integration for enhancing model outputs or refining knowledge corpora.
According to an aspect of an embodiment, validating model performance further comprises comparing the model's responses to those provided by certified experts or reliable sources in specific domains.
According to an aspect of an embodiment, the platform uses cosine similarity or Euclidean distance in hyper dimensional vector space to detect model infringement, replication, influence, or theft.
According to an aspect of an embodiment, the adversarial training incorporates malicious examples into the training data to make the model more resilient against manipulated predictions.
According to an aspect of an embodiment, optimizing the performance and reliability of the model comprises searching through different model types, attention architectures and RAG configurations, model consensus or blending mechanisms to find the most effective setup for a given task and then to determine additional viable configurations to address operational resilience concerns for alternate, contingent or emergency data and logic configurations under ranges of operational scenarios and parameters, such as specific scenarios of interest declared by users, or scenarios of interest identified by the system.
According to an aspect of an embodiment, model blending comprises selecting from or blending outputs from multiple models or authoritative knowledge bases, each trained on, or obtained from, defined resource collections or with different retrieval strategies or hyperparameters.
BRIEF DESCRIPTION OF THE DRAWING FIGURES
FIG. 1 A is a block diagram illustrating an exemplary system architecture for a distributed, advanced model management platform for optimizing and securing AI models, according to an embodiment.
FIG. 1 B is a block diagram illustrating an exemplary system architecture for a distributed generative artificial intelligence reasoning and action platform, according to an embodiment.
FIG. 2 is a block diagram illustrating an exemplary aspect of a distributed generative AI reasoning and action platform incorporating various additional contextual data.
FIG. 3 is a diagram illustrating incorporating symbolic reasoning in support of LLM-based generative AI, according to an aspect of a neuro-symbolic generative AI reasoning and action platform.
FIG. 4 is a block diagram illustrating an exemplary architecture for a neuro-symbolic generative AI reasoning and action platform configured for federated learning at a plurality of edge devices, according to an embodiment.
FIG. 5 is a block diagram illustrating an exemplary architecture for a neuro-symbolic generative AI reasoning and action platform configured to utilize a midserver to act as a computing intermediary between a plurality of edge devices and the platform.
FIG. 6 is a block diagram illustrating an exemplary mobile device configured for experience curation using embedded capabilities and functionality provided by a neuro-symbolic generative AI reasoning and action platform, according to an embodiment.
FIG. 7 is a block diagram illustrating an exemplary aspect of a distributed generative artificial intelligence reasoning and action platform, a curation computing system.
FIG. 8 is a block diagram illustrating an exemplary aspect of a distributed generative artificial intelligence reasoning and action platform, a marketplace computing system.
FIG. 9 is a block diagram illustrating a simple example of a distributed computational graph representation for providing neuro-symbolic artificial intelligence capabilities, including generative AI, according to an aspect.
FIG. 10 is a block diagram illustrating an exemplary aspect of an embodiment of a distributed computational graph computing system utilizing an advanced cyber decision platform (ACDP) for external network reconnaissance and contextual data collection.
FIG. 11 is a block diagram illustrating another exemplary aspect of an embodiment of a distributed computational graph computing systems utilizing an advanced cyber decision platform.
FIG. 12 is a diagram of an exemplary architecture for a system for rapid predictive analysis of very large data sets using an actor-driven distributed computational graph, according to one aspect.
<figref idrefs="DRA
CROSS-REFERENCE TO RELATED APPLICATIONS
Priority is claimed in the application data sheet to the following patents or patent applications, each of which is expressly incorporated herein by reference in its entirety:
Ser. No. 18/668,137 Ser. No. 18/656,612 63/551,328
BACKGROUND OF THE INVENTION
Field of the Art
The present invention is in the field of large-scale cloud computing, and more particularly to distributed computing platforms, leveraging graph-based specifications for data flows, transport and storage, and logic locality in data handling to enable artificial intelligence enhanced decision-making and automation and planning including those employing large language models, neuro-symbolic AI, and associated services across heterogeneous cloud, managed data center, edge, and wearable/mobile devices.
Discussion of the State of the Art
Current artificial intelligence (AI) systems, including large language models (LLMs) and generative AI (Gen AI), have limitations in their utility stemming from their lack of symbolic reasoning capabilities and integration of symbolic knowledge and limited capabilities for automated planning alongside the limited explainability of connectionist modeling approaches. There is a need for AI system architectures that can bridge symbolic and non-symbolic, i.e. connectionist, representations to enable more advanced, contextual reasoning while also considering critical factors like security, traceability, and collaborative model development and legal, privacy, and data owner restrictions around the locality of data across transport, storage and compute stages of data flow specified processes at both logical and physical levels.
Large language models and other connectionist AI techniques have demonstrated remarkable capabilities in various natural language processing, classification and inference tasks across text, sound, video, and even in chemical, molecule, mathematical and other specialized areas. However, they suffer from limitations such as hallucination, lack of validation, security vulnerabilities, and inadequate model management and suitability monitoring that negatively impact trust and applicability to mission critical utilization.
What is needed is an advanced model management platform for optimizing and securing large language models, neuro-symbolic AI systems, and aiding in efficient use of RAGs, transfer and federated learning, structured expert judgment, and fine-tuning capabilities in production environments.
SUMMARY OF THE INVENTION
Accordingly, the inventor has conceived and reduced to practice, an advanced model management platform for optimizing and securing artificial intelligence enhanced processes including those employing large language models and neuro symbolic models in real-world processes and production environments. The platform incorporates various techniques to address the limitations of current AI systems, including generative AI applications, such as hallucination, lack of validation, security vulnerabilities, and inadequate model management and fitness evaluation or challenges and explainability. The system employs combinations of random search, stochastic search and reinforcement learning algorithms, embedding and retrieval augmented generation (RAG) and knowledge graph enhancements for prompt or input improvements or prompt security/constraint satisfaction or output validation or use limitation, domain-specific knowledge corpus validation against expert or group of expert knowledge, transfer or federated learning enhancements or logic migration across mobile or edge or cloud resources, probabilistic consideration of distributed computing operational challenges, model distillation and output similarity scoring for security based on acceptable output checks and distance or consensus logic, adversarial training for robustness, and model type or attention mechanism search and model blending or consensus for advanced management or output weighting or downstream logic or planning inputs. By integrating these techniques, the platform significantly improves the performance, reliability, and security of generative AI systems across a wide range of reasoning tasks and domains. System may also take collections of specialized models, e.g. text, audio, image, physics, chemical, and combine such elements into spatial intelligence representations for scene or scenario-level planning of interest in many practical personal, business and robotics applications in domestic, commercial and industrial environments.
According to a preferred embodiment, a computing system for optimizing and securing generative AI models employing an advanced model management platform is disclosed, the computing system comprising: one or more hardware processors configured for: employing reinforcement learning algorithms to optimize a model's settings based on task-specific reward functions; using retrieval augmented generation (RAG) to retrieve relevant information from internal and external knowledge sources combined into a curated knowledge corpus and conditioning the model's output on retrieved facts or beliefs; validating model performance against human experts or crowds and authoritative databases or rule sets using cross-validation and domain-specific question-answering; employing model distillation, differential privacy, similarity scoring, and nearness models to create and secure the model and detect infringement; using adversarial training and input perturbation to test the robustness the model against poisoning attacks and adversarial examples; and incorporating attention mechanism search, model blending, and RAG to optimize the performance and reliability of the model for specific tasks.
According to another preferred embodiment, a computer-implemented method executed on an advanced model management platform for optimizing and securing generative AI models is disclosed, the computer-implemented method comprising: employing reinforcement learning algorithms to optimize a model's settings based on task-specific reward functions; using retrieval augmented generation (RAG) to retrieve relevant information from internal and external knowledge sources combined into a curated knowledge corpus and conditioning the model's output on retrieved facts or beliefs; validating model performance against human experts or crowds and authoritative databases or rule sets using cross-validation and domain-specific question-answering; employing model distillation, differential privacy, similarity scoring, and nearness models to create and secure the model and detect infringement; using adversarial training and input perturbation to test the robustness the model against poisoning attacks and adversarial examples; and incorporating attention mechanism search, model blending, and RAG to optimize the performance and reliability of the model for specific tasks.
According to another preferred embodiment, a system for optimizing and securing generative AI models employing an advanced model management platform is disclosed, comprising one or more computers with executable instructions that, when executed, cause the system to: employ reinforcement learning algorithms to optimize a model's settings based on task-specific reward functions; use retrieval augmented generation (RAG) to retrieve relevant information from internal and external knowledge sources combined into a curated knowledge corpus and conditioning the model's output on retrieved facts or beliefs; validate model performance against human experts or crowds and authoritative databases or rule sets using cross-validation and domain-specific question-answering; employing model distillation, differential privacy, similarity scoring, and nearness models to create and secure the model and detect infringement; use adversarial training and input perturbation to test the robustness the model against poisoning attacks and adversarial examples; and incorporate attention mechanism search, model blending, and RAG to optimize the performance and reliability of the model for specific tasks.
According to another preferred embodiment, non-transitory, computer-readable storage media having computer-executable instructions embodied thereon that, when executed by one or more processors of a computing system employing an advanced model management platform for optimizing and securing generative AI models, cause the computing system to: employ reinforcement learning algorithms to optimize a model's settings based on task-specific reward functions; use retrieval augmented generation (RAG) to retrieve relevant information from internal and external knowledge sources combined into a curated knowledge corpus and conditioning the model's output on retrieved facts or beliefs; validate model performance against human experts or crowds and authoritative databases or rule sets using cross-validation and domain-specific question-answering; employing model distillation, differential privacy, similarity scoring, and nearness models to create and secure the model and detect infringement; use adversarial training and input perturbation to test the robustness the model against poisoning attacks and adversarial examples; and incorporate attention mechanism search, model blending, and RAG to optimize the performance and reliability of the model for specific tasks.
According to an aspect of an embodiment, the reinforcement learning algorithms comprise Proximal Policy Optimization or Asynchronous Advantage Actor-Critic algorithms.
According to an aspect of an embodiment, the model's output on retrieved facts further comprises entity linking and knowledge graph, vector database, or vectorized knowledge graph hybrid integration for enhancing model outputs or refining knowledge corpora.
According to an aspect of an embodiment, validating model performance further comprises comparing the model's responses to those provided by certified experts or reliable sources in specific domains.
According to an aspect of an embodiment, the platform uses cosine similarity or Euclidean distance in hyper dimensional vector space to detect model infringement, replication, influence, or theft.
According to an aspect of an embodiment, the adversarial training incorporates malicious examples into the training data to make the model more resilient against manipulated predictions.
According to an aspect of an embodiment, optimizing the performance and reliability of the model comprises searching through different model types, attention architectures and RAG configurations, model consensus or blending mechanisms to find the most effective setup for a given task and then to determine additional viable configurations to address operational resilience concerns for alternate, contingent or emergency data and logic configurations under ranges of operational scenarios and parameters, such as specific scenarios of interest declared by users, or scenarios of interest identified by the system.
According to an aspect of an embodiment, model blending comprises selecting from or blending outputs from multiple models or authoritative knowledge bases, each trained on, or obtained from, defined resource collections or with different retrieval strategies or hyperparameters.
BRIEF DESCRIPTION OF THE DRAWING FIGURES
FIG. 1 A is a block diagram illustrating an exemplary system architecture for a distributed, advanced model management platform for optimizing and securing AI models, according to an embodiment.
FIG. 1 B is a block diagram illustrating an exemplary system architecture for a distributed generative artificial intelligence reasoning and action platform, according to an embodiment.
FIG. 2 is a block diagram illustrating an exemplary aspect of a distributed generative AI reasoning and action platform incorporating various additional contextual data.
FIG. 3 is a diagram illustrating incorporating symbolic reasoning in support of LLM-based generative AI, according to an aspect of a neuro-symbolic generative AI reasoning and action platform.
FIG. 4 is a block diagram illustrating an exemplary architecture for a neuro-symbolic generative AI reasoning and action platform configured for federated learning at a plurality of edge devices, according to an embodiment.
FIG. 5 is a block diagram illustrating an exemplary architecture for a neuro-symbolic generative AI reasoning and action platform configured to utilize a midserver to act as a computing intermediary between a plurality of edge devices and the platform.
FIG. 6 is a block diagram illustrating an exemplary mobile device configured for experience curation using embedded capabilities and functionality provided by a neuro-symbolic generative AI reasoning and action platform, according to an embodiment.
FIG. 7 is a block diagram illustrating an exemplary aspect of a distributed generative artificial intelligence reasoning and action platform, a curation computing system.
FIG. 8 is a block diagram illustrating an exemplary aspect of a distributed generative artificial intelligence reasoning and action platform, a marketplace computing system.
FIG. 9 is a block diagram illustrating a simple example of a distributed computational graph representation for providing neuro-symbolic artificial intelligence capabilities, including generative AI, according to an aspect.
FIG. 10 is a block diagram illustrating an exemplary aspect of an embodiment of a distributed computational graph computing system utilizing an advanced cyber decision platform (ACDP) for external network reconnaissance and contextual data collection.
FIG. 11 is a block diagram illustrating another exemplary aspect of an embodiment of a distributed computational graph computing systems utilizing an advanced cyber decision platform.
FIG. 12 is a diagram of an exemplary architecture for a system for rapid predictive analysis of very large data sets using an actor-driven distributed computational graph, according to one aspect.
FIG. 13 is a diagram of an exemplary architecture for a system for rapid predictive analysis of very large data sets using an actor-driven distributed computational graph, according to one aspect.
FIG. 14 is a diagram of an exemplary architecture for a system for rapid predictive analysis of very large data sets using an actor-driven distributed computational graph, according to one aspect.
FIG. 15 is a block diagram of an architecture for a transformation pipeline within a system for predictive analysis of very large data sets using a distributed computational graph computing system which can coordinate primary, alternate, contingent and emergency configurations (i.e., resilient logic and data persistent specifications relying on different cyber physical graph states or dependencies).
FIG. 16 is a process flow diagram of a method for predictive analysis of very large data sets using the distributed computational graph
FIG. 17 is a process flow diagram of a method for an aspect of modeling the transformation pipeline module as a directed graph using graph theory.
FIG. 18 is a flow diagram illustrating an exemplary method for providing experience curation, according to an aspect of an embodiment.
FIG. 19 is a flow diagram illustrating an exemplary method for providing experience curation with using rich contextual data, according to an aspect of an embodiment.
FIG. 20 is a flow diagram illustrating an exemplary method for using a distributed computation graph system for creating structured representations or knowledge graphs from various data sources, and setting up a pipeline for continuous processing and monitoring of that data, according to an embodiment.
FIG. 21 is a block diagram illustrating an exemplary system architecture for a distributed, composite symbolic and non-symbolic AI platform for advanced reasoning, according to an embodiment.
FIG. 22 is a block diagram illustrating an exemplary model architecture of the Transformer, consisting of an Encoder and a Decoder.
FIG. 23 is a block diagram illustrating an exemplary basic embedding layer generation process, according to an embodiment.
FIG. 24 is a flow diagram illustrating an exemplary method for routing processing based on certainty threshold and/or challenge-based verification, according to an embodiment.
FIG. 25 is a flow diagram illustrating an exemplary method for retrieving relevant contextual data from a knowledge graph database and enriching vector embeddings with the contextual data, according to an embodiment.
FIG. 26 is a flow diagram illustrating an exemplary method for applying expressive weighting schemes to model combinations, according to an embodiment.
FIG. 27 is a flow diagram illustrating an exemplary method for using feedback loops considering security, licensing, provenance, and collaborative development, according to an embodiment.
FIG. 28 is a flow diagram illustrating an exemplary method for multi-modal alignment for consistent representations across data types, according to an embodiment.
FIG. 29 is a flow diagram illustrating an exemplary method for hyperparameter optimization using information-theoretic guidance, according to an embodiment.
FIG. 30 is a flow diagram illustrating an exemplary method for linking embeddings to knowledge graphs, according to an embodiment.
FIG. 31 is a flow diagram illustrating an exemplary method for advanced reasoning using a composite artificial intelligence platform, according to an embodiment.
FIG. 32 is a flow diagram illustrating an exemplary method for fitness for purpose optimization of a generative AI model, according to an embodiment.
FIG. 33 is a flow diagram illustrating an exemplary method for hallucination mitigation of a generative AI model, according to an embodiment.
FIG. 34 is a flow diagram illustrating an exemplary method for model validation of a generative AI model, according to an embodiment.
FIG. 35 is a flow diagram illustrating an exemplary method for model training and sandboxing of a generative AI model, according to an embodiment.
FIG. 36 is a flow diagram illustrating an exemplary method for providing security and defense of a generative AI model, according to an embodiment.
FIG. 37 is a flow diagram illustrating an exemplary method for model management of a generative AI model, according to an embodiment.
FIG. 38 illustrates an exemplary computing environment on which an embodiment described herein may be implemented.
DETAILED DESCRIPTION OF THE INVENTION
The inventor has conceived, and reduced to practice, an advanced model management platform for optimizing and securing artificial intelligence enhanced processes including those employing large language models and neuro symbolic models. The platform incorporates various techniques to address the limitations of current AI systems, including generative AI applications, such as hallucination, lack of validation, security vulnerabilities, and inadequate model management and fitness evaluation or challenges. The system employs semantic meaning, stochastic search, reinforcement learning algorithms, retrieval augmented generation (RAG) or knowledge graph enhancement for prompt or input improvements or prompt security/constraint satisfaction or output validation or use limitation, domain-specific knowledge corpus validation against expert or group of expert knowledge, transfer or federated learning enhancements or logic migration across mobile or edge or cloud resources, probabilistic consideration of distributed computing operational challenges, model distillation and output similarity scoring for security based on acceptable output checks and distance or consensus logic, adversarial training for robustness, and model type or attention mechanism search and model blending or consensus for advanced management or output weighting or downstream logic or planning inputs or modeling simulation systems. By integrating these techniques, the platform significantly improves the performance, reliability, and security of artificial intelligence enhanced reasoning and generative AI systems across a wide range of tasks and domains.
The platform emphasizes the importance of considering various types of semantics, including, but not limited to, symbolic, distributional, compositional distributional, and information-theoretic compositional distributional semantics. This consideration allows the platform to capture and represent meaning at different levels of abstraction and compositionality, enabling more comprehensive and nuanced understanding of the input data. By explicitly addressing these different types of semantics, the platform can leverage the strengths of each approach and combine them in a unified framework.
Symbolic semantics relies on explicit, structured representations of meaning using symbols and logical expressions. In this approach, the meaning of a concept is defined by its relationships to other concepts in a symbolic knowledge base, often represented using ontologies or logic-based formalisms. Symbolic semantics enables precise and interpretable reasoning, as the meaning is explicitly encoded in the symbols and their relationships. However, symbolic semantics can be brittle and struggle with handling ambiguity, context-dependence, and the open-ended nature of language and historical reliance on expert codified data corpora.
Distributional semantics is based on the idea that the meaning of a word or concept can be inferred from its distribution across a large corpus of text data. This approach represents words as dense vectors (embeddings) in a high-dimensional space, where the proximity between vectors reflects their semantic similarity. Distributional semantics is driven by the statistical co-occurrence patterns of words in the data, capturing the idea that words with similar meanings tend to appear in similar contexts. While distributional semantics can capture rich semantic relationships and handle ambiguity, it lacks the explicit structure and interpretability of symbolic semantics.
Compositional distributional semantics aims to combine the strengths of distributional semantics with the compositionality of language, allowing for the construction of meaning from smaller units. In this approach, the meaning of a phrase or sentence is computed by composing the distributional representations (embeddings) of its constituent words or subphrases. Compositional distributional semantics enables the generation of embeddings for novel or unseen phrases, based on the compositionality principle that the meaning of a complex expression is determined by the meanings of its parts and their mode of combination. Various compositional models have been proposed, such as additive models, multiplicative models, and neural network-based models (e.g., recursive neural networks, transformers).
Information-theoretic compositional distributional semantics incorporates principles from information theory to quantify and optimize the information content and transmission in compositional distributional models. This approach aims to capture the mutual information between the components of a compositional representation, ensuring that the composed meaning preserves the relevant information from the individual constituents. Information-theoretic measures, such as entropy, mutual information, and cross-entropy, are used to guide the learning and composition process, promoting representations that are informative, compact, and generalizable. By grounding compositional distributional semantics in information theory, this approach seeks to improve the interpretability, robustness, and efficiency of the resulting semantic representations.
The distinction and integration of these different types of semantics are important for the platform's goal of achieving advanced reasoning and understanding in AI systems. By considering symbolic semantics, the invention can leverage the structured and interpretable aspects of meaning representation. Distributional semantics allows for capturing the statistical patterns and relationships in large-scale text data. Compositional distributional semantics enables the construction of meaning from smaller units, while information-theoretic principles guide the optimization of the compositional process.
By combining these different semantic approaches, the platform aims to create a more comprehensive and expressive semantic representation that can handle the complexities of language and reasoning. The integration of symbolic and distributional semantics, along with compositional and information-theoretic principles, allows for a richer and more robust understanding of the input data, leading to improved performance in various AI tasks such as natural language understanding, knowledge representation, and inference.
The platform can employ reinforcement learning algorithms, such as Proximal Policy Optimization (PPO), Asynchronous Advantage Actor-Critic (A3C), or Soft Action-Critic (SAC) to optimize model settings like temperature and token length based on task-specific reward functions.
Soft Actor-Critic is an off-policy reinforcement learning algorithm that combines the benefits of both policy-based and value-based methods. It is designed to maximize both the expected reward and the entropy of the policy, which encourages exploration and helps prevent the agent from getting stuck in suboptimal solutions. SAC aims to maximize the entropy of the policy alongside the expected return. By adding an entropy term to the objective function, the agent is encouraged to explore more diverse actions, leading to better exploration and improved sample efficiency. SAC is an off-policy algorithm, meaning it can learn from data collected by any policy, not just the current policy being optimized. This allows for more efficient use of collected experience and enables the use of replay buffers to store and reuse past experiences. SAC employs an actor-critic architecture, where the actor (policy) is responsible for selecting actions, and the critic (value function) estimates the expected return for state-action pairs. The critic guides the actor towards better actions by providing a learned value estimate. SAC learns a stochastic policy, which means the agent outputs a probability distribution over actions rather than a deterministic action. This allows for better exploration and can handle continuous action spaces effectively. SAC introduces a temperature parameter that balances the trade-off between exploration and exploitation. The algorithm automatically adjusts the temperature during training to maintain a desired level of entropy, eliminating the need for manual tuning.
The platform may use retrieval augmented generation (RAG) to retrieve relevant information from external knowledge sources and condition the model's prompts and/or output on retrieved data and/or facts. This ensures that the generated text is more grounded in reality and less likely to contain hallucinated information.
The platform can validate model performance against human experts (or in some cases expert models) and authoritative databases (e.g., relational, NoSQL, graph, knowledge graph, vector, document, hybrid vectorized knowledge graph) using techniques like cross-validation and domain-specific question-answering. This helps assess the model's reliability and performance in specific domains.
The platform may employ model distillation and differential privacy to create smaller, more secure models that retain the performance of larger models. Similarity scoring and nearness models are used to detect model infringement and theft.
The platform can use adversarial training and input perturbation to test the robustness of models against poisoning attacks and other adversarial examples. This helps improve the model's resilience and stability.
The platform may incorporate attention mechanism/model type search and model blending and/or consensus to optimize the performance and reliability of models for specific tasks. RAGs, knowledge graph verification, and/or composite/authoritative vectorized knowledge graphs may be integrated into the model architecture to retrieve and incorporate relevant information from external knowledge sources. Such a process may also be used in single stages of evaluation such as a safety or security measure on model inputs or on RAG or input augmentation or refinement, and again on model output comparison, blending, model-based, or rule-based or simulation-based checks.
Hyperparameter optimization explores ideal embedding generation techniques, training datasets, model architectures and other factors, guided by information theoretic metrics.
Symbolic knowledge extraction from embeddings enables linkage of the learned representations to structured ontologies, allowing neural and symbolic knowledge to be bridged for reasoning.
According to an embodiment, the disclosed invention is capable of probabilistic modeling of compute, transport, and storage asset availability, performance, reliability and security to address practical operational challenges which may cause objective function maximization routines to encourage system to have some level of data or logic duplication in support of system resilience. This stems from practical challenges in distributed computing made famous in the fallacies of distributed computing originally published at Sun Microsystems: The network is reliable; Latency is zero Bandwidth is infinite; The network is secure; Topology doesn't change; There is one administrator; Transport cost is zero; The network is homogeneous. Acknowledging such limitations in probabilistic modeling of distributed computing, especially in AI enhanced processes with logic and data federation across wearable, mobile, edge, and cloud resource pools and the potential for inter-device gossip (e.g. phone to watch) for data and compute tasks necessitates consideration of viable data flow/logic processes in a distributed computational graph specified process across primary, alternate, contingent, and emergency configurations depending on enabling technology asset availability and performance.
One or more different aspects may be described in the present application. Further, for one or more of the aspects described herein, numerous alternative arrangements may be described; it should be appreciated that these are presented for illustrative purposes only and are not limiting of the aspects contained herein or the claims presented herein in any way. One or more of the arrangements may be widely applicable to numerous aspects, as may be readily apparent from the disclosure. In general, arrangements are described in sufficient detail to enable those skilled in the art to practice one or more of the aspects, and it should be appreciated that other arrangements may be utilized and that structural, logical, software, electrical and other changes may be made without departing from the scope of the particular aspects. Particular features of one or more of the aspects described herein may be described with reference to one or more particular aspects or figures that form a part of the present disclosure, and in which are shown, by way of illustration, specific arrangements of one or more of the aspects. It should be appreciated, however, that such features are not limited to usage in the one or more particular aspects or figures with reference to which they are described. The present disclosure is neither a literal description of all arrangements of one or more of the aspects nor a listing of features of one or more of the aspects that must be present in all arrangements.
Headings of sections provided in this patent application and the title of this patent application are for convenience only, and are not to be taken as limiting the disclosure in any way.
Devices that are in communication with each other need not be in continuous communication with each other, unless expressly specified otherwise. In addition, devices that are in communication with each other may communicate directly or indirectly through one or more communication means or intermediaries, logical or physical.
A description of an aspect with several components in communication with each other does not imply that all such components are required. To the contrary, a variety of optional components may be described to illustrate a wide variety of possible aspects and in order to more fully illustrate one or more aspects. Similarly, although process steps, method steps, algorithms or the like may be described in a sequential order, such processes, methods and algorithms may generally be configured to work in alternate orders, unless specifically stated to the contrary. In other words, any sequence or order of steps that may be described in this patent application does not, in and of itself, indicate a requirement that the steps be performed in that order. The steps of described processes may be performed in any order practical. Further, some steps may be performed simultaneously despite being described or implied as occurring non-simultaneously (e.g., because one step is described after the other step). Moreover, the illustration of a process by its depiction in a drawing does not imply that the illustrated process is exclusive of other variations and modifications thereto, does not imply that the illustrated process or any of its steps are necessary to one or more of the aspects, and does not imply that the illustrated process is preferred. Also, steps are generally described once per aspect, but this does not mean they must occur once, or that they may only occur once each time a process, method, or algorithm is carried out or executed. Some steps may be omitted in some aspects or some occurrences, or some steps may be executed more than once in a given aspect or occurrence.
When a single device or article is described herein, it will be readily apparent that more than one device or article may be used in place of a single device or article. Similarly, where more than one device or article is described herein, it will be readily apparent that a single device or article may be used in place of the more than one device or article.
The functionality or the features of a device may be alternatively embodied by one or more other devices that are not explicitly described as having such functionality or features. Thus, other aspects need not include the device itself.
Techniques and mechanisms described or referenced herein will sometimes be described in singular form for clarity. However, it should be appreciated that particular aspects may include multiple iterations of a technique or multiple instantiations of a mechanism unless noted otherwise. Process descriptions or blocks in figures should be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process. Alternate implementations are included within the scope of various aspects in which, for example, functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those having ordinary skill in the art.
Definitions
As used herein, âexplainabilityâ (also referred to as âinterpretabilityâ) is the concept that a machine learning model and its output can be explained in a way that âmakes senseâ to a human being at an acceptable level.
As used herein, âgraphâ is a representation of information and relationships, where each primary unit of information makes up a ânodeâ or âvertexâ of the graph and the relationship between two nodes makes up an edge of the graph. Nodes can be further qualified by the connection of one or more descriptors or âpropertiesâ to that node. For example, given the node âJames R,â name information for a person, qualifying properties might be â183 cm tall,â âDOB 08/13/1965â and âspeaks Englishâ. Similar to the use of properties to further describe the information in a node, a relationship between two nodes that forms an edge can be qualified using a âlabelâ. Thus, given a second node âThomas G,â an edge between âJames Râ and âThomas Gâ that indicates that the two people know each other might be labeled âknows.â When graph theory notation (Graph=(Vertices, Edges)) is applied this situation, the set of nodes are used as one parameter of the ordered pair, V and the set of 2 element edge endpoints are used as the second parameter of the ordered pair, E. When the order of the edge endpoints within the pairs of E is not significant, for example, the edge James R, Thomas G is equivalent to Thomas G, James R, the graph is designated as âundirected.â Under circumstances when a relationship flows from one node to another in one direction, for example James R is âtallerâ than Thomas G, the order of the endpoints is significant. Graphs with such edges are designated as âdirected.â In the distributed computational graph system, transformations within transformation pipeline are represented as directed graph with each transformation comprising a node and the output messages between transformations comprising edges. Distributed computational graph stipulates the potential use of non-linear transformation pipelines which are programmatically linearized. Such linearization can result in exponential or impractical growth of resource consumption. The most sensible approach to overcome this possibility is to introduce new transformation pipelines or elements of pipelines just as they are needed, creating only those that are ready to compute. Such method results in transformation graphs which are highly variable in size and node, edge composition as the system processes data streams and persists elements of the process or of the involved data at various timescales and localities. Those familiar with the art will realize that transformation graph may assume many shapes and sizes with a vast topology of edge relationships and node types. It is also important to note that the resource topologies available at a given execution time for a given pipeline may be highly dynamic due to changes in available node or edge types or topologies (e.g. different servers, data centers, devices, network links, etc.) being available, and this is even more so when legal, regulatory, privacy and security considerations are included in a DCG pipeline specification or recipe in the DSL. Since the system can have a range of parameters (e.g. authorized to do transformation x at compute locations of a, b, or c) the JIT, JIC, JIP elements can leverage system state information (about both the processing system and the observed system of interest) and planning or modeling modules to compute at least one parameter set (e.g. execution of pipeline may say based on current conditions use compute location b) at execution time. This may also be done at the highest level or delegated to lower level resources when considering the spectrum from centralized cloud clusters (i.e. higher) to extreme edge (e.g. a wearable, or phone or laptop). The examples given were chosen for illustrative purposes only and represent a small number of the simplest of possibilities. These examples should not be taken to define the possible graphs expected as part of operation of the invention and may include additional kinds of optimizations or resilience considerations (e.g. in cislunar systems) where uncertainty or variance in assets or viable topologies is particularly problematic compared to earthbound assets.
As used herein, âtransformationâ is a function performed on zero or more streams of input data which results in a single stream of output which may or may not then be used as input for another transformation. Transformations may comprise any combination of machine, human or machine-human interactions Transformations need not change data that enters them, one example of this type of transformation would be a storage transformation which would receive input and then act as a queue for that data for subsequent transformations. As implied above, a specific transformation may generate output data in the absence of input data. A time stamp serves as an example. In the invention, transformations are placed into pipelines such that the output of one transformation may serve as an input for another. These pipelines can consist of two or more transformations with the number of transformations limited only by the resources of the system. Historically, transformation pipelines have been linear with each transformation in the pipeline receiving input from one antecedent and providing output to one subsequent with no branching or iteration. Other pipeline configurations are possible. The invention is designed to permit several of these configurations including, but not limited to: linear, afferent branch, efferent branch and cyclical.
A âpipeline,â as used herein and interchangeably referred to as a âdata pipelineâ or a âprocessing pipeline,â refers to a set of data streaming activities and batch activities. Streaming and batch activities can be connected indiscriminately within a pipeline and compute, transport or storage (including temporary in-memory persistence such as Kafka topics) may be optionally inferred/suggested by the system or may be expressly defined in the pipeline domain specific language.
The system may leverage a Distributed Computational Graph (DCG) framework to instantiate pipelines in a cyber-physical graph specific manner. Given a set of available or expected technology and communication assets, the system determines the applicable DCG topology that is contextualized and optimized for the resource pool and tasks at hand over finite time horizons. This allows the DCG to efficiently map and distribute the pipeline activities across heterogeneous compute resources that may span mobile, edge, cloud, and other environments in historical, current, and probabilistic future states to include resources such as but not limited to physical networking, logical networking, hypervisor, physical host, virtualized host or container, application, session, or trace when considering log data, error and administrative data, performance monitoring data, and other forms of observability and security information made known to the system.
The DCG maintains an adaptive resource allocation strategy, dynamically activating and deactivating subgraphs of the pipeline based on factors such as system objectives, resource availability, performance requirements, and cost constraints leveraging historical and current estimated system states and data as well as various forward estimates, models, simulations or projections, be they obtained through simulation modeling, machine learning, statistical approaches, or artificial intelligence techniques, generative or otherwise. At any point in time, only a limited number of DCG subgraphs may be active, with the inactive portions persisted for later use from in-memory, disk, database, or archival storage with various times and costs associated with such placement for rehydration as needed. DCGs also need not be executed solely within a given resource pool as the DLS supports evaluation and exchange with DCG execution processes at cloud, edge, content delivery network (CDN), host, mobile, wearable and other resource types. The system may also leverage emerging network protocols such as Disruption Tolerant Networking (DTN) for space and cislunar communications where more extensive intermediate data persistence and transmission nodes and capabilities may be needed for interplanetary Internet services.
Events flow through the streaming activity actors in the active DCG subgraphs in a reactive way. At the junction of a streaming activity to batch activity, there exists a StreamBatchProtocol data object. This object is responsible for determining when and if the batch process is run, based on one or more of three possibilities: regular timing interval, every N events, a certain data size or chunk, or optionally an internal (e.g. APM or trace or resource based trigger) or external trigger (e.g. from another user, pipeline, or exogenous service). The events or data about the events are held in a queue (e.g. Kafka or Redis) or similar until processing. In cases where event types are extensive and similar it is possible to engage in further compression of data for transmission and analysis processes with techniques such as vectorization with an associated legendâsuch as in a case where 100 static event types with many words might be reduced to 100 integersâone corresponding to each unique event and a one-time ledger transmission. System may engage in exchange of such ledgers or vector equivalents of them between system components to further increase performance across composite objective function evaluation of cumulative network compute, storage and transport costs across finite time horizons.
Each batch activity may contain a âsourceâ data context (this may be a streaming context if the upstream activities are streaming), and a âdestinationâ data context (which is passed to the next activity). Streaming activities may sometimes have an optional âdestinationâ streaming data context (optional meaning: caching/persistence of events vs. ephemeral).
The system may also comprise a database containing all data pipelines as templates, recipes, or as run at execution time to enable post-hoc reconstruction or re-evaluation with a modified topology of the resources (e.g., compute, transport or storage), transformations, or data involved. This allows the DCG to dynamically re-plan and optimize the pipeline deployment based on changes in the underlying cyber-physical resource graph and its historical, current, or future state estimates.
Conceptual Architecture
FIG. 1 A is a block diagram illustrating an exemplary system architecture for a distributed, advanced model management platform for optimizing and securing AI models 170 , according to an embodiment. According to the embodiment, the platform 170 aims to enable vast automation of modeling/analysis workflows by exploring large potential parameter combinations, model optimization, and underlying resource uses/configurations to support. Platform 170 can be configured for extracting and curating knowledge into structured ontologies to complement neuro-symbolic AI capabilities.
Furthermore, platform 170 may incorporate a hierarchical architecture that enables the migration and replication of logic and data across different layers, such as peer-to-peer communication between distributed nodes or upward propagation to cloud-based services. This functionality allows for efficient dissemination and synchronization of model updates, knowledge bases, and operational data throughout the system. In some implementations, the platform can employ gossip protocols to facilitate peer-to-peer communication, enabling decentralized sharing of information and updates among distributed nodes. This approach ensures robust and scalable propagation of data and logic across the network, even in the presence of intermittent connectivity or node fa
CLAIMS
Claims ( 23 )
1 . A computing system for optimizing generative AI models, the computing system comprising:
one or more hardware processors configured for:
optimizing an AI model's performance by selecting one or more settings for the AI model using reinforcement learning algorithms based on a task-specific reward function that measures the AI model's performance on a specified task;
optimizing the AI model's content by using retrieval augmented generation (RAG) to retrieve information for the AI model from a curated knowledge corpus based on contextual matching with the input query and conditioning the AI model's prompts or outputs based on content from the retrieved information;
validating the AI model's performance against one or more authorities using cross-validation, wherein the one or more authorities includes human experts, crowdsourcing, authoritative databases, rule sets, or expert judgment models that perform synthesis of multiple responses using consensus mechanisms;
optimizing the AI model's robustness using adversarial training;
optimizing the AI model's stability using input perturbation; and
optimizing the AI model's reliability for the specified task using one or more techniques measured against a fitness function, wherein the techniques include model type search, attention mechanism search, model blending with weighted consensus, expert synthesis, RAG, knowledge graph verification, or composite vectorized knowledge graphs, and wherein optimization is performed through a distributed computational graph that automatically parallelizes processing across heterogeneous computing resources.
2 . The computing system of claim 1 , wherein optimizing the AI model's performance includes one or more reinforcement learning algorithms comprising Proximal Policy Optimization or Asynchronous Advantage Actor-Critic algorithms.
3 . (canceled)
4 . The computing system of claim 1 , wherein validating the AI model's performance comprises comparing the AI model's responses to responses provided by the one or more authorities.
5 . (canceled)
6 . The computing system of claim 1 , wherein the adversarial training incorporates malicious examples into training data to make the AI model more resilient against manipulated predictions.
7 . (canceled)
8 . The computing system of claim 1 , wherein model blending comprises selecting from or blending outputs from multiple models or authoritative knowledge bases, each trained on, or obtained from, defined resource collections or with different retrieval strategies or hyperparameters.
9 . A computer-implemented method for optimizing generative AI models, the computer-implemented method comprising:
optimizing an AI model's performance by selecting one or more settings for the AI model using reinforcement learning algorithms based on a task-specific reward function that measures the AI model's performance on a specified task; optimizing the AI model's content by using retrieval augmented generation (RAG) to retrieve information for the AI model from a curated knowledge corpus based on contextual matching with the input query and conditioning the AI model's prompts or outputs based on content from the retrieved information; validating the AI model's performance against one or more authorities using cross-validation, wherein the one or more authorities includes human experts, crowdsourcing, authoritative databases, or rule sets, or expert judgment models that perform synthesis of multiple responses using consensus mechanisms; optimizing the AI model's robustness using adversarial training; optimizing the AI model's stability using and input perturbation; and optimizing the AI model's reliability for the specified task using one or more techniques, measured against a fitness function, wherein the techniques include model type search, attention mechanisms search, model blending with weighted consensus, expert synthesis, RAG, knowledge graph verification, or composite vectorized knowledge graphs, and wherein the optimization is performed through a distributed computational graph that automatically parallelizes processing across heterogeneous computing resources.
10 . The computer-implemented method of claim 9 , wherein optimizing the AI model's performance include one or more reinforcement learning algorithms comprising Proximal Policy Optimization or Asynchronous Advantage Actor-Critic algorithms.
11 . (canceled)
12 . The computer-implemented method of claim 9 , wherein validating the AI model's performance comprises comparing the AI model's responses to responses provided by the one or more authorities.
13 . (canceled)
14 . The computer-implemented method of claim 9 , wherein the adversarial training incorporates malicious examples into training data to make the AI model more resilient against manipulated predictions.
15 . (canceled)
16 . The computer-implemented method of claim 9 , wherein model blending comprises selecting from or blending outputs from multiple models or authoritative knowledge bases, each trained on, or obtained from, defined resource collections or with different retrieval strategies or hyperparameters.
17 - 24 . (canceled)
25 . Non-transitory, computer-readable storage media having computer-executable instructions embodied thereon that, when executed by one or more processors of a computing system for optimizing generative AI models, cause the computing system to:
optimize an AI model's performance by selecting one or more settings for the AI model using reinforcement learning algorithms based on a task-specific reward function that measures the AI model's performance on a specified task; optimize the AI model's content by using retrieval augmented generation (RAG) to retrieve information for the AI model from a curated knowledge corpus based on contextual matching with the input query and conditioning the AI model's prompts or outputs based on content from the retrieved information; validate the AI model's performance against one or more authorities using cross-validation, wherein the one or more authorities includes human experts, crowdsourcing, authoritative databases, or rule sets, or expert judgment models that perform synthesis of multiple responses using consensus mechanisms; optimize the AI model's robustness using adversarial training; optimize the AI model's stability using input perturbation; and optimize the AI model's reliability for the specified task using one or more techniques measured against a fitness function, wherein the techniques include model type search, attention mechanism search, model blending with weighted consensus, expert synthesis, RAG, knowledge graph verification, or composite vectorized knowledge graphs, and wherein optimization is performed through a distributed computational graph that automatically parallelizes processing across heterogeneous computing resources.
26 . The non-transitory, computer-readable storage media of claim 25 , wherein optimizing the AI model's performance includes one or more reinforcement learning algorithms comprising Proximal Policy Optimization or Asynchronous Advantage Actor-Critic algorithms.
27 . (canceled)
28 . The non-transitory, computer-readable storage media of claim 25 , wherein validating the AI model's performance comprises comparing the AI model's responses to responses provided by the one or more authorities.
29 . (canceled)
30 . The non-transitory, computer-readable storage media of claim 25 , wherein the adversarial training incorporates malicious examples into training data to make the AI model more resilient against manipulated predictions.
US18/731,339
2024-02-08
2024-06-02
Advanced model management platform for optimizing and securing ai systems including large language models
Pending
US20250259075A1
( en )
Priority Applications (1)
Application Number
Priority Date
Filing Date
Title
US18/731,339
US20250259075A1
( en )
2024-02-08
2024-06-02
Advanced model management platform for optimizing and securing ai systems including large language models
Applications Claiming Priority (4)
Application Number
Priority Date
Filing Date
Title
US202463551328P
2024-02-08
2024-02-08
US18/656,612
US20250259047A1
( en )
2024-02-08
2024-05-07
Computing platform for neuro-symbolic artificial intelligence applications
US18/668,137
US12536213B2
( en )
2024-02-08
2024-05-18
Composite symbolic and non-symbolic artificial intelligence system for advanced reasoning and automation
US18/731,339
US20250259075A1
( en )
2024-02-08
2024-06-02
Advanced model management platform for optimizing and securing ai systems including large language models
Related Parent Applications (1)
Application Number
Title
Priority Date
Filing Date
US18/668,137
Continuation-In-Part
US12536213B2
( en )
2015-10-28
2024-05-18
Composite symbolic and non-symbolic artificial intelligence system for advanced reasoning and automation
Publications (1)
Publication Number
Publication Date
US20250259075A1
true
US20250259075A1 ( en )
2025-08-14
Family
ID=96661196
Family Applications (1)
Application Number
Title
Priority Date
Filing Date
US18/731,339
Pending
US20250259075A1
( en )
2024-02-08
2024-06-02
Advanced model management platform for optimizing and securing ai systems including large language models
Country Status (1)
Country
Link
US
( 1 )
US20250259075A1
( en )
Cited By (31)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
CN120744423A
( en )
*
2025-08-28
2025-10-03
æå·éä¹å®éªå®¤ç§ææéå ¬å¸
Task processing model evaluation method, role playing model evaluation method and task processing method
CN120764697A
( en )
*
2025-09-08
2025-10-10
å½ç½æ¹åççµåæéå ¬å¸
Power large model performance optimization method and system based on declarative self-prompt
CN120892543A
( en )
*
2025-10-09
2025-11-04
浪潮éç¨è½¯ä»¶æéå ¬å¸
A method, system, and device for optimizing the process of a conversational business intelligence platform.
CN120910224A
( en )
*
2025-10-09
2025-11-07
åå°æ»¨å·¥ä¸å¤§å¦(æ·±å³)(åå°æ»¨å·¥ä¸å¤§å¦æ·±å³ç§æåæ°ç 究é¢)
Security event association analysis and question-answering method, system and medium
CN120929611A
( en )
*
2025-10-14
2025-11-11
常å·å¸¸ä¾çµå设计é¢æéå ¬å¸
Electric power design knowledge base construction method integrating multi-mode data and RAG technology
CN120975358A
( en )
*
2025-10-19
2025-11-18
æé½ä¿¡æ¯å·¥ç¨å¤§å¦
A combinatorial optimization solution system and method based on a general attack and defense framework
CN120975681A
( en )
*
2025-10-21
2025-11-18
å±±ä¸ç计ç®ä¸å¿ï¼å½å®¶è¶ 级计ç®æµåä¸å¿ï¼
A Customs Risk Early Warning Method and System Based on Cross-Modal Retrieval Enhancement
CN120975268A
( en )
*
2025-10-23
2025-11-18
æ¦æ±å·¥ç¨å¤§å¦
Training methods based on dynamically adjusted reward mechanisms
CN120994845A
( en )
*
2025-10-23
2025-11-21
æ¹åæºå¦ä¼è¯¾ä¿¡æ¯ç§ææéå ¬å¸
A Method and System for Intelligent PDF Retrieval and Generation Based on RAG
CN121034657A
( en )
*
2025-08-15
2025-11-28
ä¸å½ç¾ç é¢é²æ§å¶ä¸å¿æ ¢æ§éä¼ ææ§ç¾ç é¢é²æ§å¶ä¸å¿
A knowledge graph-based AI recommendation method and system for preventing hallucinations in chronic disease management
CN121071162A
( en )
*
2025-10-31
2025-12-05
é¿åä¿¡æ¯è¡ä»½æéå ¬å¸
Bank intelligent operation knowledge base implementation method and system based on large language model and RAG
US20250384063A1
( en )
*
2024-06-18
2025-12-18
International Business Machines Corporation
Securing retrieval augmented generation
CN121170273A
( en )
*
2025-11-20
2025-12-19
ä¸å½éå¡è¡ä»½æéå ¬å¸
Multi-mode large model training method and system
CN121258528A
( en )
*
2025-12-01
2026-01-02
éåºèèæ¶è´¹éèæéå ¬å¸
Activity strategy generation method and device, storage medium and electronic equipment
CN121301536A
( en )
*
2025-12-15
2026-01-09
ä¸æµ·ç¾æ°æºè½ç§ææéå ¬å¸
Hybrid AI Adaptive Optimization System and Method Based on Agent Context Engineering
US12524416B1
( en )
*
2024-07-09
2026-01-13
Palo Alto Networks, Inc.
Query response system implementing a retrieval-augment generation architecture
JP7799364B1
( en )
*
2025-09-30
2026-01-15
ãã£ã³ã¼ãã£ã³ã°ã¨ãã¯ã¹æ ªå¼ä¼ç¤¾
System, method, and program for generating a response to a user's voice
CN121388199A
( en )
*
2025-11-14
2026-01-23
å½å®¶åºç¡å°çä¿¡æ¯ä¸å¿
Intelligent classification and retrieval method for mapping files
US20260059074A1
( en )
*
2024-08-21
2026-02-26
Matthew P. Hars
Computer-implemented methods and systems for virtual legal proceedings
CN121578652A
( en )
*
2026-01-23
2026-02-27
æ¸ ååå·è½æºäºèç½ç 究é¢
An energy management method and system based on sparse federated reinforcement learning
US12566923B1
( en )
*
2025-07-11
2026-03-03
Geotab Inc.
Systems and methods for training and using custom vector embedding models
CN121660106A
( en )
*
2026-02-06
2026-03-13
é½é²å·¥ä¸å¤§å¦(å±±ä¸çç§å¦é¢)
Cross-domain large model method and system based on federal RAG
CN121681720A
( en )
*
2026-02-10
2026-03-17
æµæ±åé»ç§ææéå ¬å¸
Method and system for generating large language model diagram query language based on reinforcement learning
CN121706996A
( en )
*
2026-02-24
2026-03-20
å京æç¨æ¶ä»£ç§ææéå ¬å¸
Large model reasoning path optimization method and system based on dynamic entropy sensing
US20260088970A1
( en )
*
2025-07-25
2026-03-26
Joshua A Wright
Hybrid AI Failover and Secure Field-Data Capture System with Shadow-Mode Validation and Training-to-Inference Reassignment
CN121743346A
( en )
*
2026-02-26
2026-03-27
æ±è¥¿ç§æå¸è大å¦
Automatic questionnaire filling method based on cooperation of progressive divide-and-congress and multi-question fusion
US12591612B1
( en )
*
2025-02-21
2026-03-31
Palo Alto Networks, Inc.
Multimodal web page content classification with knowledge graph-based explainability
CN121835820A
( en )
*
2026-03-13
2026-04-10
çå°æµ·é¢è½¯ä»¶è¡ä»½æéå ¬å¸
Large model self-evolution training method for electric drawing understanding
US20260105078A1
( en )
*
2024-10-16
2026-04-16
Dell Products L.P.
Recommendation Generation with User Values
US20260113310A1
( en )
*
2024-10-22
2026-04-23
Capital One Services, Llc
Systems and methods for improved data processing of secured datasets across secured computing networks while maintaining encryption of secured data
US20260119554A1
( en )
*
2024-10-24
2026-04-30
Nomura Research Institute, Ltd.
Security countermeasure support system
Citations (1)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20260024014A1
( en )
*
2022-09-30
2026-01-22
Intel Corporation
Improving accuracy of machine learning operations by compensating for lower precision with scale shifting
2024
2024-06-02
US
US18/731,339
patent/US20250259075A1/en
active
Pending
Patent Citations (1)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20260024014A1
( en )
*
2022-09-30
2026-01-22
Intel Corporation
Improving accuracy of machine learning operations by compensating for lower precision with scale shifting
Non-Patent Citations (7)
* Cited by examiner, â Cited by third party
Title
Amazon. What Is RAG? 2023 (Year: 2023)
*
Beecham et al. Using an expert panel to validate a requirements process improvement model. The Journal of Systems and Software 76 (2005) 251â275 (Year: 2005)
*
Brownlee, Jason. Blending Ensemble Machine Learning With Python. Ensemble Learning. April 27.i. 2021 (Year: 2021)
*
Kariyappa et al. PROTECTING DNNS FROM THEFT USING AN ENSEMBLE OF DIVERSE MODELS. ICLR 2021 (Year: 2021)
*
Lange et al. Privacy at Risk: Exploiting Similarities in Health Data for Identity Inference. 2023 (Year: 2023)
*
Sridhar et al. Towards Alternative Techniques for Improving Adversarial Robustness: Analysis of Adversarial Training at a Spectrum of Perturbations. 2022 (Year: 2022)
*
Zhou et al. Large Language Model as a Policy Teacher for Training Reinforcement Learning Agents. Jan 2024 (Year: 2024)
*
Cited By (31)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20250384063A1
( en )
*
2024-06-18
2025-12-18
International Business Machines Corporation
Securing retrieval augmented generation
US12524416B1
( en )
*
2024-07-09
2026-01-13
Palo Alto Networks, Inc.
Query response system implementing a retrieval-augment generation architecture
US20260059074A1
( en )
*
2024-08-21
2026-02-26
Matthew P. Hars
Computer-implemented methods and systems for virtual legal proceedings
US20260105078A1
( en )
*
2024-10-16
2026-04-16
Dell Products L.P.
Recommendation Generation with User Values
US20260113310A1
( en )
*
2024-10-22
2026-04-23
Capital One Services, Llc
Systems and methods for improved data processing of secured datasets across secured computing networks while maintaining encryption of secured data
US20260119554A1
( en )
*
2024-10-24
2026-04-30
Nomura Research Institute, Ltd.
Security countermeasure support system
US12591612B1
( en )
*
2025-02-21
2026-03-31
Palo Alto Networks, Inc.
Multimodal web page content classification with knowledge graph-based explainability
US12566923B1
( en )
*
2025-07-11
2026-03-03
Geotab Inc.
Systems and methods for training and using custom vector embedding models
US20260088970A1
( en )
*
2025-07-25
2026-03-26
Joshua A Wright
Hybrid AI Failover and Secure Field-Data Capture System with Shadow-Mode Validation and Training-to-Inference Reassignment
CN121034657A
( en )
*
2025-08-15
2025-11-28
ä¸å½ç¾ç é¢é²æ§å¶ä¸å¿æ ¢æ§éä¼ ææ§ç¾ç é¢é²æ§å¶ä¸å¿
A knowledge graph-based AI recommendation method and system for preventing hallucinations in chronic disease management
CN120744423A
( en )
*
2025-08-28
2025-10-03
æå·éä¹å®éªå®¤ç§ææéå ¬å¸
Task processing model evaluation method, role playing model evaluation method and task processing method
CN120764697A
( en )
*
2025-09-08
2025-10-10
å½ç½æ¹åççµåæéå ¬å¸
Power large model performance optimization method and system based on declarative self-prompt
JP7799364B1
( en )
*
2025-09-30
2026-01-15
ãã£ã³ã¼ãã£ã³ã°ã¨ãã¯ã¹æ ªå¼ä¼ç¤¾
System, method, and program for generating a response to a user's voice
CN120910224A
( en )
*
2025-10-09
2025-11-07
åå°æ»¨å·¥ä¸å¤§å¦(æ·±å³)(åå°æ»¨å·¥ä¸å¤§å¦æ·±å³ç§æåæ°ç 究é¢)
Security event association analysis and question-answering method, system and medium
CN120892543A
( en )
*
2025-10-09
2025-11-04
浪潮éç¨è½¯ä»¶æéå ¬å¸
A method, system, and device for optimizing the process of a conversational business intelligence platform.
CN120929611A
( en )
*
2025-10-14
2025-11-11
常å·å¸¸ä¾çµå设计é¢æéå ¬å¸
Electric power design knowledge base construction method integrating multi-mode data and RAG technology
CN120975358A
( en )
*
2025-10-19
2025-11-18
æé½ä¿¡æ¯å·¥ç¨å¤§å¦
A combinatorial optimization solution system and method based on a general attack and defense framework
CN120975681A
( en )
*
2025-10-21
2025-11-18
å±±ä¸ç计ç®ä¸å¿ï¼å½å®¶è¶ 级计ç®æµåä¸å¿ï¼
A Customs Risk Early Warning Method and System Based on Cross-Modal Retrieval Enhancement
CN120994845A
( en )
*
2025-10-23
2025-11-21
æ¹åæºå¦ä¼è¯¾ä¿¡æ¯ç§ææéå ¬å¸
A Method and System for Intelligent PDF Retrieval and Generation Based on RAG
CN120975268A
( en )
*
2025-10-23
2025-11-18
æ¦æ±å·¥ç¨å¤§å¦
Training methods based on dynamically adjusted reward mechanisms
CN121071162A
( en )
*
2025-10-31
2025-12-05
é¿åä¿¡æ¯è¡ä»½æéå ¬å¸
Bank intelligent operation knowledge base implementation method and system based on large language model and RAG
CN121388199A
( en )
*
2025-11-14
2026-01-23
å½å®¶åºç¡å°çä¿¡æ¯ä¸å¿
Intelligent classification and retrieval method for mapping files
CN121170273A
( en )
*
2025-11-20
2025-12-19
ä¸å½éå¡è¡ä»½æéå ¬å¸
Multi-mode large model training method and system
CN121258528A
( en )
*
2025-12-01
2026-01-02
éåºèèæ¶è´¹éèæéå ¬å¸
Activity strategy generation method and device, storage medium and electronic equipment
CN121301536A
( en )
*
2025-12-15
2026-01-09
ä¸æµ·ç¾æ°æºè½ç§ææéå ¬å¸
Hybrid AI Adaptive Optimization System and Method Based on Agent Context Engineering
CN121578652A
( en )
*
2026-01-23
2026-02-27
æ¸ ååå·è½æºäºèç½ç 究é¢
An energy management method and system based on sparse federated reinforcement learning
CN121660106A
( en )
*
2026-02-06
2026-03-13
é½é²å·¥ä¸å¤§å¦(å±±ä¸çç§å¦é¢)
Cross-domain large model method and system based on federal RAG
CN121681720A
( en )
*
2026-02-10
2026-03-17
æµæ±åé»ç§ææéå ¬å¸
Method and system for generating large language model diagram query language based on reinforcement learning
CN121706996A
( en )
*
2026-02-24
2026-03-20
å京æç¨æ¶ä»£ç§ææéå ¬å¸
Large model reasoning path optimization method and system based on dynamic entropy sensing
CN121743346A
( en )
*
2026-02-26
2026-03-27
æ±è¥¿ç§æå¸è大å¦
Automatic questionnaire filling method based on cooperation of progressive divide-and-congress and multi-question fusion
CN121835820A
( en )
*
2026-03-13
2026-04-10
çå°æµ·é¢è½¯ä»¶è¡ä»½æéå ¬å¸
Large model self-evolution training method for electric drawing understanding
Similar Documents
Publication
Publication Date
Title
US12602375B2
( en )
2026-04-14
Composite symbolic and non-symbolic artificial intelligence system for advanced reasoning and semantic search
US12536213B2
( en )
2026-01-27
Composite symbolic and non-symbolic artificial intelligence system for advanced reasoning and automation
US20250259144A1
( en )
2025-08-14
Platform for integration of machine learning models utilizing marketplaces and crowd and expert judgment and knowledge corpora
US20250156898A1
( en )
2025-05-15
System and method for ai based tailored advertising content generation
US20250258708A1
( en )
2025-08-14
Federated distributed graph-based computing platform with hardware management
US20250258685A1
( en )
2025-08-14
System and method for ai based dynamic user experience curation
US20250045304A1
( en )
2025-02-06
Method and system for interpreting inputted information
WO2025160388A1
( en )
2025-07-31
Artificial intelligence driven systems of systems for converged technology stacks
CN109101217B
( en )
2023-09-29
Methods and systems for purposeful computing
Das et al.
2018
Hands-On Automated Machine Learning: A beginner's guide to building automated machine learning systems using AutoML and Python
US20250139378A1
( en )
2025-05-01
Event-driven personalized recommendation systems
US20250259047A1
( en )
2025-08-14
Computing platform for neuro-symbolic artificial intelligence applications
Bass et al.
2025
Engineering AI systems: architecture and DevOps essentials
US20250259032A1
( en )
2025-08-14
Federated distributed graph-based computing platform
US20260010774A1
( en )
2026-01-08
System and Method for Persistent Cognitive Machine on Neuromorphic Platform
US20260010730A1
( en )
2026-01-08
Latent Cognitive Manifolds with Lensing Potentials
Xu
2022
AI theory and applications in the financial industry
Jangid et al.
2023
The AI Renaissance: Innovations, Ethics, and the Future of Intelligent Systems
Patil et al.
2024
Trustworthy artificial intelligence in industry and society
Haviv et al.
2023
Implementing MLOps in the enterprise
US20250307749A1
( en )
2025-10-02
Electronic document obligation monitoring
Mendelevitch et al.
2016
Practical Data Science with Hadoop and Spark: Designing and Building Effective Analytics at Scale
US20260105438A1
( en )
2026-04-16
Artificial intelligence driven systems of systems for converged technology stacks
Subramanian
2025
Applied Machine Learning for Data Science Practitioners
Cooper
2018
Data science from scratch: The# 1 data science guide for everything a data scientist needs to know: Python, linear algebra, statistics, coding, applications, neural networks, and decision trees
Legal Events
Date
Code
Title
Description
2024-07-15
AS
Assignment
Owner name : QOMPLX LLC, VIRGINIA
Free format text : ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:CRABTREE, JASON;KELLEY, RICHARD;HOPPER, JASON;AND OTHERS;REEL/FRAME:067992/0429
Effective date : 20240417
2025-07-08
STPP
Information on status: patent application and granting procedure in general
Free format text : NON FINAL ACTION MAILED
2025-10-09
STPP
Information on status: patent application and granting procedure in general
Free format text : RESPONSE TO NON-FINAL OFFICE ACTION ENTERED AND FORWARDED TO EXAMINER
Free format text : FINAL REJECTION COUNTED, NOT YET MAILED
2025-10-19
STPP
Information on status: patent application and granting procedure in general
Free format text : RESPONSE TO NON-FINAL OFFICE ACTION ENTERED AND FORWARDED TO EXAMINER
2026-02-05
STPP
Information on status: patent application and granting procedure in general
Free format text : FINAL REJECTION COUNTED, NOT YET MAILED
2026-02-19
STPP
Information on status: patent application and granting procedure in general
Free format text : FINAL REJECTION MAILED