Systems and Methods for Dynamic Neural Network Enhancement and Adaptive Edge … — Veea Inc. (US20250292091A1)

ABSTRACT

Abstract

Systems and methods for adaptive edge computing using artificial intelligence (AI) include monitoring real-time accuracy of a neural network by using a feedback loop configured to detect changes in inference accuracy and dynamically adjusting the structure of the neural network by adding or removing hidden layers based on monitored error rates and predetermined computational constraints. A Kalman gain computation determines neural network weight adjustments based on monitored error rates. Weight matrices undergo incremental updates derived from these adjustments. Incremental weight adjustments remain stored in memory to enable low-bandwidth model updates. The neural network stores inference results and refined weights in an inference result database. Pre-trained models periodically receive incremental updates based on stored adjustments. Predictive holistic inference logic (PHIL) applied to stored inference results improves the accuracy of the inference results.

Description

RELATED APPLICATIONS

This application claims the benefit of priority to U.S. Provisional Application No. 63/565,792 entitled “Artificial Intelligence Edge Middleware” filed Mar. 15, 2024, the entire contents of which are hereby incorporated by reference for all purposes.

BACKGROUND

The growing complexity of distributed computing systems has necessitated advancements in methodologies and architectures for managing computational resources, data processing, and communication networks. Edge computing has emerged as a decentralized approach to augment traditional cloud-based systems by bringing computational resources closer to data sources. This paradigm seeks to address challenges such as reducing latency, improving bandwidth efficiency, and enhancing real-time decision-making capabilities by offloading tasks to devices or nodes at the network edge.

AI techniques, particularly those utilizing neural networks, have demonstrated significant utility in various domains, including data analysis, pattern recognition, and automated decision-making. Neural networks, such as CNNs and DNNs, consist of multiple layers of interconnected processing nodes, each performing transformations on input data to generate meaningful outputs. These networks rely on training processes to adjust weight parameters and biases, allowing the models to adapt to complex datasets and generate accurate predictions or inferences.

Dynamic environments, such as edge computing networks, introduce unique challenges for the deployment and operation of AI models. Constraints on computational resources, such as processing power, memory, and energy, necessitate efficient optimization techniques to achieve consistent performance. In addition, the dynamic nature of network conditions, including varying workloads, changes in topology, and fluctuating resource availability, requires adaptive methodologies to maintain system efficiency.

Enhancement techniques, such as those incorporating feedback-based algorithms, are being increasingly used to refine the performance of AI systems in real-time. Kalman filters, for example, provide a robust framework for estimating and adjusting parameters in systems with noise or uncertainty. These filters have been applied in various computational contexts to improve the precision of predictive models by iteratively updating parameters based on incoming data.

The integration of neural networks with advanced enhancement techniques and feedback mechanisms has potential applications in managing distributed computing environments. These systems may dynamically adapt to changing conditions, improve resource allocation, and improve decision-making processes. As technologies continue to advance, addressing the complexities of distributed and resource-constrained environments remains an area of ongoing development.

SUMMARY

The various aspects include methods for adaptive edge computing using artificial intelligence (AI), which may include monitoring real-time model accuracy of a neural network by using a feedback loop configured to detect changes in inference accuracy, dynamically adjusting a structure of the neural network by adding or removing hidden layers based on monitored error rates and predetermined computational constraints, computing a Kalman gain based on monitored error rates, and determining weight adjustments for the neural network based on the computed Kalman gain, incrementally updating weight matrices of the neural network based on the determined weight adjustments, storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates, storing inference results generated by the neural network and refined weights in an inference result database, periodically updating pre-trained models associated with the neural network based on the stored incremental weight adjustments, and applying predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy.

In some aspects, monitoring real-time model accuracy by using the feedback loop includes tracking model performance metrics by repeatedly analyzing inference accuracy in response to changes in real-time network conditions experienced by the neural network. In some aspects, the methods may include comparing predicted outputs generated by the neural network to actual network behavior to identify deviations between predicted outputs and actual network behavior, and triggering recalibration of the neural network in response to determining that an identified deviation exceeds a predefined accuracy threshold value. In some aspects, the methods may include dynamically adjusting the predefined accuracy threshold value based on measured variability in network conditions to reduce the frequency of recalibrations of the neural network.

In some aspects, dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints may include analyzing neuron activation patterns of at least one hidden layer of the neural network to determine whether the at least one hidden layer contributes to inference accuracy, and improving computational efficiency of the neural network by removing the at least one hidden layer from the structure of the neural network in response to determining that the at least one hidden layer does not substantially contribute to inference accuracy.

In some aspects, analyzing neuron activation patterns of the at least one hidden layer may include analyzing neuron activation levels to determine whether the at least one hidden layer consistently exhibits low activation during inference operations, and removing the at least one hidden layer from the neural network structure in response to determining that the at least one hidden layer consistently exhibits low activation. In some aspects, dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints may include implementing a self-pruning mechanism configured to identify and remove neurons from hidden layers exhibiting low activation gradients.

In some aspects, dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints may include replacing an activation function used by neurons within hidden layers with an alternative activation function selected from a rectified linear unit (ReLU) or a Leaky ReLU to increase learning efficiency. In some aspects, computing the Kalman gain based on monitored error rates and determining weight adjustments based on the computed Kalman gain may include computing the Kalman gain by integrating weighted error estimates derived from real-time inference feedback, and refining weight adjustments of the neural network based on the computed Kalman gain to improve convergence rates during inference operations. In some aspects, storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates further may include applying federated learning techniques to distribute incremental weight differentials instead of distributing full neural network model updates.

In some aspects, updating pre-trained models associated with the neural network periodically based on stored incremental weight adjustments may include incorporating newly observed network traffic patterns and associated device behavior into an existing architecture of the pre-trained models, and executing asynchronous updates of the pre-trained models to maintain uninterrupted real-time inference operations. In some aspects, applying predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy may include integrating multiple inference sources associated with the neural network to refine model predictions through a consensus-based approach.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings, which are incorporated herein and constitute part of this specification, illustrate exemplary embodiments of the claims, and together with the general description given and the detailed description, serve to explain the features herein.

FIG. 1 is a component block diagram illustrating an artificial neuron model in accordance with some embodiments.

FIG. 2 is a component block diagram illustrating a multi-layer artificial neural network in accordance with some embodiments.

FIG. 3 is a component block diagram illustrating an adaptive neural network model incorporating a Kalman filter in accordance with some embodiments.

FIGS. 4 A- 4 H are component block diagrams illustrating a complex adaptive neural network model that uses multiple Kalman filter sub-algorithms in accordance with some embodiments.

FIG. 5 is a component block diagram illustrating the key factors involved in edge compute node optimization, in accordance with some embodiments.

FIG. 6 is a process flow diagram illustrating a method for network configuration using neural networks and Kalman filters in accordance with some embodiments.

FIG. 7 is a process flow diagram illustrating a method for dynamic network management in a computing device in accordance with some embodiments.

FIG. 8 is a process flow diagram illustrating a method for managing an edge computing network in accordance with some embodiments.

FIG. 9 is a process flow diagram illustrating a method for enhancing a deep neural network using a Kalman filter and PHIL for dynamic learning and adaptation in accordance with some embodiments.

FIGS. 10 A- 10 F are process flow diagrams illustrating a method 1000 for AI-driven adaptive edge computing in accordance with some embodiments.

FIG. 11 is a process flow diagram illustrating a method 1100 for dynamically configuring an edge computing network in accordance with some embodiments.

FIG. 12 is a process flow diagram illustrating a method 1200 for enhancing edge computing systems using neural networks and dynamic decision-making in accordance with some embodiments.

FIG. 13 A is a component block diagram that illustrates an example edge computing system suitable for implementing the various embodiments.

FIG. 13 B is a component block diagram that illustrates an edge device that interconnected modules designed for specialized and general-purpose computations in distributed edge computing systems.

FIG. 14 is a component diagram of server suitable for implementing some embodiments.

DETAILED DESCRIPTION

Various embodiments will be described in detail with reference to the accompanying drawings. Wherever possible, the same reference numbers will be used throughout the drawings to refer to the same or like parts. References made to particular examples and implementations are for illustrative purposes and are not intended to limit the scope of the claims.

In overview, the various embodiments include methods of integrating and using advanced artificial intelligence (AI) technologies and techniques that facilitate proactive and real-time decision-making to enhance, augment, or optimize (herein collectively “enhance”) edge computing infrastructures for a wide range of applications. Some embodiments may include a highly adaptive and modular AI system capable of self-tuning, self-optimization, rapid modification or reconfiguration, and/or continual learning. The embodiments may overcome various technical challenges associated with managing resource-constrained edge nodes in a dynamic environment by using AI techniques and technologies for adaptive, real-time decision-making. In some embodiments, the methods may include dynamically adjusting the size and complexity of a neural network in real-time by adding or removing layers based on a confidence level (or performance value, etc.) to refine the output, enhance processing efficiency, and/or adapt to resource constraints. In some embodiments, the methods may include using a Kalman filter to adjust or fine-tune the weights of individual nodes within the neural network layers for real-time optimization and better alignment with desired outputs. In some embodiments, one or more weights may be set to zero to remove one or more hidden layers from the neural network.

Further aspects may include a computing device (edge computing device) having a processor configured with processor-executable instructions to perform various operations corresponding to the methods discussed above. Further aspects may include a computing device (edge computing device) having various means for performing functions corresponding to the method operations discussed above. Further aspects may include a non-transitory processor-readable storage medium having stored thereon processor-executable instructions configured to cause a processor to perform various operations corresponding to the method operations discussed above.

The term “neural network” may be used herein to refer to an interconnected group of processing nodes, or neuron models, that operate as a computational framework to perform functions of a computing device or produce inference results. Each processing node may receive input data, perform defined operations to produce an output referred to as activation, and transmit the activation to subsequent nodes. A neural network may include weights associated with the connections between nodes, defining the relationships between inputs and outputs. Adjustments to these weights during a training process may allow the neural network to refine its performance. Training may involve processing tasks with known outcomes, comparing activations to expected outputs, and adjusting weights based on comparisons. After training concludes, the neural network may perform inference using the adjusted weights to process new tasks.

<div id="p-0034" num="00

RELATED APPLICATIONS

This application claims the benefit of priority to U.S. Provisional Application No. 63/565,792 entitled “Artificial Intelligence Edge Middleware” filed Mar. 15, 2024, the entire contents of which are hereby incorporated by reference for all purposes.

BACKGROUND

The growing complexity of distributed computing systems has necessitated advancements in methodologies and architectures for managing computational resources, data processing, and communication networks. Edge computing has emerged as a decentralized approach to augment traditional cloud-based systems by bringing computational resources closer to data sources. This paradigm seeks to address challenges such as reducing latency, improving bandwidth efficiency, and enhancing real-time decision-making capabilities by offloading tasks to devices or nodes at the network edge.

AI techniques, particularly those utilizing neural networks, have demonstrated significant utility in various domains, including data analysis, pattern recognition, and automated decision-making. Neural networks, such as CNNs and DNNs, consist of multiple layers of interconnected processing nodes, each performing transformations on input data to generate meaningful outputs. These networks rely on training processes to adjust weight parameters and biases, allowing the models to adapt to complex datasets and generate accurate predictions or inferences.

Dynamic environments, such as edge computing networks, introduce unique challenges for the deployment and operation of AI models. Constraints on computational resources, such as processing power, memory, and energy, necessitate efficient optimization techniques to achieve consistent performance. In addition, the dynamic nature of network conditions, including varying workloads, changes in topology, and fluctuating resource availability, requires adaptive methodologies to maintain system efficiency.

Enhancement techniques, such as those incorporating feedback-based algorithms, are being increasingly used to refine the performance of AI systems in real-time. Kalman filters, for example, provide a robust framework for estimating and adjusting parameters in systems with noise or uncertainty. These filters have been applied in various computational contexts to improve the precision of predictive models by iteratively updating parameters based on incoming data.

The integration of neural networks with advanced enhancement techniques and feedback mechanisms has potential applications in managing distributed computing environments. These systems may dynamically adapt to changing conditions, improve resource allocation, and improve decision-making processes. As technologies continue to advance, addressing the complexities of distributed and resource-constrained environments remains an area of ongoing development.

SUMMARY

The various aspects include methods for adaptive edge computing using artificial intelligence (AI), which may include monitoring real-time model accuracy of a neural network by using a feedback loop configured to detect changes in inference accuracy, dynamically adjusting a structure of the neural network by adding or removing hidden layers based on monitored error rates and predetermined computational constraints, computing a Kalman gain based on monitored error rates, and determining weight adjustments for the neural network based on the computed Kalman gain, incrementally updating weight matrices of the neural network based on the determined weight adjustments, storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates, storing inference results generated by the neural network and refined weights in an inference result database, periodically updating pre-trained models associated with the neural network based on the stored incremental weight adjustments, and applying predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy.

In some aspects, monitoring real-time model accuracy by using the feedback loop includes tracking model performance metrics by repeatedly analyzing inference accuracy in response to changes in real-time network conditions experienced by the neural network. In some aspects, the methods may include comparing predicted outputs generated by the neural network to actual network behavior to identify deviations between predicted outputs and actual network behavior, and triggering recalibration of the neural network in response to determining that an identified deviation exceeds a predefined accuracy threshold value. In some aspects, the methods may include dynamically adjusting the predefined accuracy threshold value based on measured variability in network conditions to reduce the frequency of recalibrations of the neural network.

In some aspects, dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints may include analyzing neuron activation patterns of at least one hidden layer of the neural network to determine whether the at least one hidden layer contributes to inference accuracy, and improving computational efficiency of the neural network by removing the at least one hidden layer from the structure of the neural network in response to determining that the at least one hidden layer does not substantially contribute to inference accuracy.

In some aspects, analyzing neuron activation patterns of the at least one hidden layer may include analyzing neuron activation levels to determine whether the at least one hidden layer consistently exhibits low activation during inference operations, and removing the at least one hidden layer from the neural network structure in response to determining that the at least one hidden layer consistently exhibits low activation. In some aspects, dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints may include implementing a self-pruning mechanism configured to identify and remove neurons from hidden layers exhibiting low activation gradients.

In some aspects, dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints may include replacing an activation function used by neurons within hidden layers with an alternative activation function selected from a rectified linear unit (ReLU) or a Leaky ReLU to increase learning efficiency. In some aspects, computing the Kalman gain based on monitored error rates and determining weight adjustments based on the computed Kalman gain may include computing the Kalman gain by integrating weighted error estimates derived from real-time inference feedback, and refining weight adjustments of the neural network based on the computed Kalman gain to improve convergence rates during inference operations. In some aspects, storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates further may include applying federated learning techniques to distribute incremental weight differentials instead of distributing full neural network model updates.

In some aspects, updating pre-trained models associated with the neural network periodically based on stored incremental weight adjustments may include incorporating newly observed network traffic patterns and associated device behavior into an existing architecture of the pre-trained models, and executing asynchronous updates of the pre-trained models to maintain uninterrupted real-time inference operations. In some aspects, applying predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy may include integrating multiple inference sources associated with the neural network to refine model predictions through a consensus-based approach.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings, which are incorporated herein and constitute part of this specification, illustrate exemplary embodiments of the claims, and together with the general description given and the detailed description, serve to explain the features herein.

FIG. 1 is a component block diagram illustrating an artificial neuron model in accordance with some embodiments.

FIG. 2 is a component block diagram illustrating a multi-layer artificial neural network in accordance with some embodiments.

FIG. 3 is a component block diagram illustrating an adaptive neural network model incorporating a Kalman filter in accordance with some embodiments.

FIGS. 4 A- 4 H are component block diagrams illustrating a complex adaptive neural network model that uses multiple Kalman filter sub-algorithms in accordance with some embodiments.

FIG. 5 is a component block diagram illustrating the key factors involved in edge compute node optimization, in accordance with some embodiments.

FIG. 6 is a process flow diagram illustrating a method for network configuration using neural networks and Kalman filters in accordance with some embodiments.

FIG. 7 is a process flow diagram illustrating a method for dynamic network management in a computing device in accordance with some embodiments.

FIG. 8 is a process flow diagram illustrating a method for managing an edge computing network in accordance with some embodiments.

FIG. 9 is a process flow diagram illustrating a method for enhancing a deep neural network using a Kalman filter and PHIL for dynamic learning and adaptation in accordance with some embodiments.

FIGS. 10 A- 10 F are process flow diagrams illustrating a method 1000 for AI-driven adaptive edge computing in accordance with some embodiments.

FIG. 11 is a process flow diagram illustrating a method 1100 for dynamically configuring an edge computing network in accordance with some embodiments.

FIG. 12 is a process flow diagram illustrating a method 1200 for enhancing edge computing systems using neural networks and dynamic decision-making in accordance with some embodiments.

FIG. 13 A is a component block diagram that illustrates an example edge computing system suitable for implementing the various embodiments.

FIG. 13 B is a component block diagram that illustrates an edge device that interconnected modules designed for specialized and general-purpose computations in distributed edge computing systems.

FIG. 14 is a component diagram of server suitable for implementing some embodiments.

DETAILED DESCRIPTION

Various embodiments will be described in detail with reference to the accompanying drawings. Wherever possible, the same reference numbers will be used throughout the drawings to refer to the same or like parts. References made to particular examples and implementations are for illustrative purposes and are not intended to limit the scope of the claims.

In overview, the various embodiments include methods of integrating and using advanced artificial intelligence (AI) technologies and techniques that facilitate proactive and real-time decision-making to enhance, augment, or optimize (herein collectively “enhance”) edge computing infrastructures for a wide range of applications. Some embodiments may include a highly adaptive and modular AI system capable of self-tuning, self-optimization, rapid modification or reconfiguration, and/or continual learning. The embodiments may overcome various technical challenges associated with managing resource-constrained edge nodes in a dynamic environment by using AI techniques and technologies for adaptive, real-time decision-making. In some embodiments, the methods may include dynamically adjusting the size and complexity of a neural network in real-time by adding or removing layers based on a confidence level (or performance value, etc.) to refine the output, enhance processing efficiency, and/or adapt to resource constraints. In some embodiments, the methods may include using a Kalman filter to adjust or fine-tune the weights of individual nodes within the neural network layers for real-time optimization and better alignment with desired outputs. In some embodiments, one or more weights may be set to zero to remove one or more hidden layers from the neural network.

Further aspects may include a computing device (edge computing device) having a processor configured with processor-executable instructions to perform various operations corresponding to the methods discussed above. Further aspects may include a computing device (edge computing device) having various means for performing functions corresponding to the method operations discussed above. Further aspects may include a non-transitory processor-readable storage medium having stored thereon processor-executable instructions configured to cause a processor to perform various operations corresponding to the method operations discussed above.

The term “neural network” may be used herein to refer to an interconnected group of processing nodes, or neuron models, that operate as a computational framework to perform functions of a computing device or produce inference results. Each processing node may receive input data, perform defined operations to produce an output referred to as activation, and transmit the activation to subsequent nodes. A neural network may include weights associated with the connections between nodes, defining the relationships between inputs and outputs. Adjustments to these weights during a training process may allow the neural network to refine its performance. Training may involve processing tasks with known outcomes, comparing activations to expected outputs, and adjusting weights based on comparisons. After training concludes, the neural network may perform inference using the adjusted weights to process new tasks.

The term “inference” may be used herein to refer to a process performed at runtime by a neural network to generate output values, also referred to as activations or inference results. Inference may involve traversing the network along a forward computational path, where input data passes through processing nodes to produce one or more output values.

The term “deep neural network” (DNN) may be used herein to refer to a neural network including multiple layers organized hierarchically. In a DNN, activations from earlier layers may serve as inputs for subsequent layers, creating a distributed computational chain across a population of processing nodes. DNNs may include activation functions such as rectified linear units (ReLU) applied between layers to introduce non-linear transformations. The input layer may accept raw data, the output layer may produce results, and intermediate layers may perform feature extraction or data transformations.

The term “layer” in a neural network may be used herein to refer to a set of interconnected processing nodes that receive inputs from preceding layers and provide outputs to subsequent layers. Each layer may have multiple inputs and outputs. Although embodiments may describe layers with single inputs or outputs for simplicity, the described operations may apply to multiple inputs and outputs.

Each layer in a neural network may have multiple inputs and thus multiple previous or preceding layers. Said another way, multiple layers may feed into a single layer. For ease of reference, some of the embodiments are described with reference to a single input or single preceding layer. However, it should be understood that the operations disclosed and described in this application may be applied to each of multiple inputs to a layer and multiple preceding layers.

The term “recurrent neural network” (RNN) may be used herein to refer to a neural network designed for processing sequence data. Unlike feedforward networks, an RNN may include cycles or loops in its architecture, allowing it to retain information about prior inputs. This ability may help the RNN process tasks where the order or context of input data affects the outcome.

The term “convolutional neural network” (CNN) may be used herein to refer to a deep neural network employing convolution operations to process input data. A CNN may include layers that use filters to detect patterns in spatial data, such as images or videos. Each layer may extract increasingly complex features by applying convolution operations and pooling techniques. The network may generate inference results such as classifications, object detections, or feature maps.

The term “network configuration” may be used herein to refer to the setup of network devices and software to manage data flow and communication. This process may include configuring routing protocols, assigning node roles, allocating bandwidth, and implementing security policies. Network configuration may apply to both physical devices and software components to meet operational demands.

The term “processing system” may be used herein to refer to one or more processors, including multi-core processors, organized to perform computational functions. A processing system may reside within a computing device or as part of a system-on-chip and may execute software to perform operations described herein.

The term “computing device” may be used herein to refer to systems equipped with processors and communication capabilities, such as personal computers, edge devices, smartphones, IoT devices, or servers. A computing device may execute software applications and exchange data with other systems.

The term “edge device” may be used herein to refer to a computing system deployed at the network edge with capabilities for processing data locally and communicating with consumer devices or network components. Examples may include access points, routers, gateways, and devices supporting edge computing techniques.

The term “container” may be used herein to refer to a software component supporting virtualization and resource abstraction. A container may isolate computing resources and operate independently from underlying hardware while executing software applications.

The terms “component,” “system,” and the like may be used herein to refer to a computer-related entity, such as hardware, firmware, a combination of hardware and software, software, or software during its execution. These entities may be configured to carry out specific operations or functionalities. For example, a component may encompass a process operating on a processor, a processor itself, an object, an executable, a thread of execution, a program, or a computing device. As an illustrative example, both an application running on a computing device and the computing device itself could be termed a component. One or more components may be situated within a process and/or thread of execution and/or may be localized on a single processor or core or distributed across multiple processors or cores. In addition, these components may execute from various non-transitory computer-readable media having various instructions and/or data structures stored thereon. Components may communicate by way of local and/or remote processes, function or procedure calls, electronic signals, data packets, memory read/writes, and other known computer, processor, and/or process-related communication methodologies.

The term “system on chip” (SOC) may be used herein to refer to a single integrated circuit (IC) chip that contains multiple resources and/or processors integrated on a single substrate. A single SOC may contain circuitry for digital, analog, mixed-signal, and radio-frequency functions. A single SOC may also include any number of general purpose and/or specialized processors (digital signal processors, modem processors, video processors, etc.), memory blocks (e.g., ROM, RAM, Flash, etc.), and resources (e.g., timers, voltage regulators, oscillators, etc.). SOCs may also include software for controlling the integrated resources and processors, as well as for controlling peripheral devices.

The term “system in a package” (SIP) may be used herein to refer to a single module or package that contains multiple resources, computational units, cores and/or processors on two or more IC chips, substrates, or SOCs. For example, a SIP may include a single substrate on which multiple IC chips or semiconductor dies are stacked in a vertical configuration. Similarly, the SIP may include one or more multi-chip modules (MCMs) on which multiple ICs or semiconductor dies are packaged into a unifying substrate. A SIP may also include multiple independent SOCs coupled together via high speed communication circuitry and packaged in close proximity, such as on a single backplane, single motherboard or in a single wireless device. The proximity of the SOCs facilitates high speed communications and the sharing of memory and resources.

The term “multicore processor” may be used herein to refer to a single integrated circuit (IC) chip or chip package that contains two or more independent processing cores (e.g., CPU core, IP core, GPU core, etc.) configured to read and execute program instructions. A SOC may include multiple multicore processors, and each processor in a SOC may be referred to as a core. The term “multiprocessor” may be used herein to refer to a system or device that includes two or more processing units configured to read and execute program instructions.

The term “computing mesh” may be used herein to refer to distributed computing frameworks connecting resources across multiple nodes. A computing mesh may dynamically allocate resources and execute applications across interconnected components.

The term “edge computing” may be used herein to refer to computational techniques that offload processing tasks to devices or servers at the edge of a network. Edge computing may reduce latency and optimize resource use for applications requiring real-time data analysis.

An edge computing system may integrate remote cloud servers and proximate edge devices to form a collaborative computing environment that may enhance the performance, latency, and energy consumption characteristics of user computing devices. Traditional cloud servers may provide substantial computational and storage capacities but may introduce propagation delays when transmitting large volumes of data due to their remote geographical deployment. In contrast, edge devices may reduce these propagation delays by operating closer to the user but may have more limited computational and storage capacities. The processing and storage limitations of an edge device may be offset by its lower access times and latency characteristics when compared to a cloud server. An edge computing system may intelligently and dynamically balance tradeoffs between performance and latency by distributing computational tasks across cloud and edge resources.

In some embodiments, the edge computing system may balance tradeoffs such as performance, latency, and power consumption by implementing a computing mesh, an application mesh, or a connectivity mesh. The edge computing system may include multiple edge devices connected through wired or wireless communication links, forming a computing mesh that enables resource sharing among edge devices. When multiple edge devices in a computing mesh are served by the same cloud server, the system may allocate cloud computational resources dynamically based on the workload, local computation capacities, and performance requirements of each edge device.

Edge computing may allow applications to execute at the network&#39;s edge through computing platforms that interact with user devices and sensors. This may reduce latency, increase throughput, improve contextual and location awareness, decrease backhaul data transmission, and enhance security and isolation.

Edge computing nodes may form the structural foundation of edge computing by providing localized processing and storage capabilities. These nodes may function as computing platforms with middleware components that manage distributed processing and resource allocation. By shifting computing tasks from the cloud or upstream services to edge computing nodes, edge computing may provide context-aware data processing and reduce reliance on centralized infrastructure.

Edge computing nodes may operate independently or as part of a distributed ecosystem interconnected through wired or wireless communication links. This interconnected structure may support coordinated resource sharing and adaptive workload distribution.

Edge computing nodes may function as standalone devices or as components of a distributed ecosystem. These nodes may interact by sharing backhaul connections, coordinating resource usage, or running applications accessed by other nodes.

By executing applications and functions at the network&#39;s edge rather than in a remote data center, edge computing nodes may reduce latency and improve processing efficiency. However, operating at the network&#39;s edge may introduce resource constraints, including limitations in processing capacity, memory, storage, power, and backhaul connectivity.

A fixed configuration of an edge computing node with stable applications and connectivity may provide consistent performance. However, edge computing environments may be dynamic, requiring configuration adjustments that may be performed manually rather than through automated orchestration.

Elastic edge computing environments may involve continuous changes, including node additions and removals, backhaul modifications, upstream network resource reallocations, node resource adjustments, and inter-node resource distribution. AI integrated into edge computing nodes may support adaptive management, addressing configuration and application delivery challenges in such environments.

Some embodiments may include a processing system with one or more processors configured to manage an edge computing environment using AI. The processing system may receive real-time network data (e.g., resource utilization metrics, network topology information, connectivity status, etc.) from a plurality of edge computing nodes, detect a change in the edge computing environment (e.g., a node addition, a node removal, a backhaul modification, an upstream resource reallocation, etc.), classify a current network state based on the detected change and a set of predefined performance criteria, generate a resource allocation strategy for enhancing workload distribution across the plurality of edge computing nodes, and transmit configuration updates based on the resource allocation strategy.

In some embodiments, the processing system may generate the resource allocation strategy by determining a workload assignment for each edge computing node based on the resource utilization metrics and connectivity status and selecting a backhaul resource allocation for each edge computing node based on a latency threshold and bandwidth availability. The processing system may monitor the post-adjustment performance of the edge computing environment and refine a machine learning model based on the observed performance. Refining the machine learning model may include adjusting at least one model parameter to improve future resource allocation strategies.

Edge computing nodes may be classified based on their operational roles. A node group may represent a collection of nodes, and a master node group may include nodes with backhaul connectivity. Similarly, applications may be categorized into application groups, and resource availability may be classified based on internal resources, local network resources, upstream network resources, or off-network resources such as the internet. Additional classifications may be established based on network requirements. Each classification may include attributes for identification and management. For example, a backhaul group may distinguish between wired and wireless connections and whether a connection supports inter-node communication.

In some embodiments, the processing system may dynamically allocate network resources in an edge computing environment. The processing system may receive operational data (e.g., network traffic patterns, device attributes, resource availability information) from a plurality of edge computing nodes, classify each edge computing node into a corresponding node group based on the operational data, classify each application executed on the plurality of edge computing nodes into an application group based on execution priority and resource requirements, assign resource availability classifications to each edge computing node (e.g., internal resources, local network resources, upstream network resources, off-network resources), and allocate computing resources based on the assigned classifications. The processing system may dynamically update at least one of the classifications based on changes in operational data.

AI using neural networks may rely on weights and biases to transform input data into meaningful outputs. Weights may define the influence of input features and biases may adjust the decision boundary of the network. The process of enhancing these parameters may be central to neural network training. During training, optimization/enhancement algorithms such as stochastic gradient descent may iteratively update weights and biases based on loss function gradients to refine the network&#39;s predictive accuracy. Once trained, the neural network may perform inference by applying learned parameters to new input data to generate outputs based on patterns identified during training. The accuracy of inference may depend on the quality and distribution of training data, the network&#39;s architecture, and the choice of optimization/enhancement techniques.

Neural network learning may extend beyond the initial training phase. As new data and problem domains emerge, continuous learning mechanisms such as online learning, transfer learning, and incremental learning may refine model parameters while preserving previously acquired knowledge. These adaptive techniques may allow neural networks to update their responses dynamically to improve performance in evolving environments.

Weights may include synaptic weights and biases, in which weights determine input influence, and biases adjust neuron activation thresholds.

The Kalman filter may provide a framework for enhancing weights and biases during initial training and continuous learning by leveraging its covariance matrix. These solutions may allow neural networks to refine parameter estimates while adapting to dynamic conditions. Kalman filters may adjust weights and biases independently or in conjunction with other AI architectures. By integrating the Kalman filter, lower-confidence data may be excluded from weight updates, reducing gradient skew and improving convergence efficiency. This filtering process may accelerate the optimization of network parameters by prioritizing higher-confidence data.

In some embodiments, the processing system may be configured to use a Kalman filter to enhance neural network weight updates. The processing system may receive real-time input data (e.g., feature values and corresponding confidence scores), predict an updated weight estimate for at least one neural network parameter using a prior weight estimate and a covariance matrix associated with the Kalman filter, compute a Kalman gain based on the covariance matrix and an error measurement associated with the real-time input data, filter the real-time input data by discarding input data having a confidence score below a predefined threshold, adjust the neural network parameter based on the Kalman gain and the filtered real-time input data, and store the adjusted neural network parameter in memory for subsequent inference operations.

Incremental weight updates may refine neural network performance without requiring full model retraining. Instead of replacing an entire weight set, updates may apply selectively to modified weights, reducing bandwidth consumption and processing overhead. Incremental weight updates may allow seamless integration into an active neural network without requiring power cycles or task interruptions.

In some embodiments, the processing system may perform incremental weight updates by identifying modified weights (e.g., weights differing from a previously stored weight set), determining an update window, applying the modified weights during the update window, and storing the applied weights in memory. The processing system may initiate the next incremental update cycle based on detected changes in network conditions.

AI-based adaptive responses may allow real-time network optimization without requiring predefined policy rules. Artificial neural networks may model and control complex nonlinear systems, such as edge computing networks, by mapping intricate relationships between inputs and outputs.

AI may determine a best-fit configuration rather than an optimal configuration depending on adjustment urgency. A best-fit approach may allow rapid adaptation to new tasks or network configurations, and continuous learning may refine the model toward an optimal configuration.

AI at the network&#39;s edge may extend beyond traditional machine learning by addressing large-scale, dynamic network configurations. Unlike conventional approaches, AI may adapt to evolving network topologies influenced by device locality rather than predefined patterns. Dynamic edge networks may exhibit mobility rather than remaining stationary, requiring continuous self-aware adjustments. The variability and time constraints associated with these adjustments may exceed the capacity of predefined policy rules, necessitating real-time adaptation.

FIG. 1 illustrates an example of a neuron 100 , which is a fundamental element of the neural network. The illustrated neuron has multiple inputs, labeled X 1 through Xn, that are each associated with a respective weight, w 1 through wn. These inputs and weights converge at a summation junction 102 that aggregates the weighted inputs. The resultant sum is then passed to an activation function 104 , denoted as f(a), which processes the input to generate an output 106 . The activation function 104 serves as a nonlinear transformation to determine the neuron&#39;s output in response to its aggregated input. The output 106 may represent the neuron&#39;s contribution to the next layer in a neural network or the final output in a single-layer perceptron model.

FIG. 2 illustrates the architecture of a multi-layer artificial neural network, including an input layer, hidden layers, and an output layer. Data flows sequentially through the layers via weighted connections to allow the network to process and transform input data into actionable outputs. Each node in FIG. 2 may represent an individual neuron (e.g., neuron 100 illustrated in FIG. 1 ). The input layer may receive input data, the hidden layers may process this data using nonlinear transformations, and the output layer may produce the final output. The weights and biases of the connections between nodes may be set during a training process to reduce or minimize the error between actual outputs and desired outputs.

In particular, FIG. 2 illustrates an example artificial neural network (ANN) 200 model may include an input layer, one or more hidden layers, and an output layer. In the example, illustrated in FIG. 2 , the system includes the ANN 200 , input data 202 , a pre-processor 204 , an input layer 206 , interconnections 208 linking neurons between layers, hidden layers 212 , 216 , and 220 , interconnections 222 connecting neurons within the ANN, an output layer 224 , a post-processor 226 , and output data 228 .

The ANN 200 may process input data 202 through interconnected layers, including the input layer 206 , hidden layers 212 , 216 , and 220 , and the output layer 224 . The network may utilize weighted connections between neurons to perform tasks such as pattern recognition, data classification, or inference. Training the ANN may involve adjusting these weights to optimize accuracy and reduce errors.

Input data 202 may represent raw information provided to the ANN for processing. This data may include numerical, categorical, or multidimensional values, such as sensor readings, image pixels, or network traffic statistics. The input data may serve as the initial information for analysis and may be passed to the pre-processor 204 .

The pre-processor 204 may standardize raw input data into a format suitable for the ANN. Preprocessing may involve operations such as normalization, feature extraction, dimensionality reduction, or encoding of categorical data. These operations may improve consistency in data representation and enhance the ANN&#39;s ability to identify patterns.

The input layer 206 may include neurons that receive preprocessed data. Each neuron in this layer may correspond to a feature of the input data and may pass information to subsequent layers. The input layer may act as an interface between raw data and the computational processes of the hidden layers.

Interconnections 208 may link neurons in the input layer to those in subsequent layers, representing weighted pathways that determine how each input feature contributes to the ANN&#39;s output. These weights may be adjusted during training to optimize performance.

Hidden layers 212 , 216 , and 220 may hierarchically process the input data. Each hidden layer may apply nonlinear transformations to extract meaningful abstractions, utilizing activation functions such as ReLU, sigmoid, or tanh. These layers may progressively refine the data to uncover higher-level patterns.

The first hidden layer 212 may process raw features received from the input layer and generate intermediate representations. The second hidden layer 216 may refine these representations by emphasizing patterns relevant to the task. The final hidden layer 220 may prepare the data for classification or inference in the output layer.

Interconnections 222 between neurons in adjacent layers may transmit weighted signals. These connections represent the contribution of one neuron&#39;s activation to another, and their weights may be iteratively optimized during training to improve predictive accuracy.

The output layer 224 may generate the ANN&#39;s final result. Each neuron in the output layer may correspond to a specific outcome, prediction, or decision. For instance, in classification tasks, the output layer may assign probabilities to various categories.

The post-processor 226 may refine the raw outputs of the ANN to produce actionable insights. Operations performed by the post-processor may include scaling probabilities, applying thresholds, or mapping outputs to decision labels. These refinements may prepare results for downstream applications.

Output data 228 may represent the final product of the ANN&#39;s processing. This data may include actionable insights, such as classifications, predictions, or control signals, as refined by the post-processor.

A Multilayer ANN (e.g., ANN 200 illustrated in FIG. 2 , etc.) may be particularly well suited for addressing problems in application domains involving recognition, classification, and vector quantization. Edge computing may represent a key application domain for such implementations.

The ANN for an edge computing node may execute within a neural processing unit (NPU), which may be implemented as a hardware-based platform or as a software-only platform referred to as NPU-lite.

An NPU may represent a dedicated processor or processing unit integrated within a larger SoC and specifically designed to execute ANN operations. Unlike general-purpose CPUs and GPUs, NPUs may be optimized for parallel computing, enabling efficient data processing for neural networks. By offloading ANN functions to the NPU, computational loads on the CPU and GPU may be reduced, enhancing overall system performance.

The NPU-lite may execute as software running on a local edge computing node, using programming languages such as Python or other suitable languages. Alternatively, the NPU-lite may interact with an NPU residing on another edge computing node or within a cloud-based AI system.

CNNs, a type of DNN, may also be utilized within edge computing nodes for tasks such as classification, application of rules, heuristics, analysis of prior cases, and generating expectations.

Deep neural networks, including CNNs, may address challenges associated with exponentially large parameter spaces by incorporating multiple hidden layers. These hidden layers enable hierarchical feature processing, allowing the network to manage complex input data effectively.

The hidden layers in CNNs may include convolution layers, pooling layers, and fully connected layers. Convolution and pooling layers may serve primarily as feature extractors, analyzing data derived from images, applications, or other functions. Pooling layers may also reduce the dimensionality of feature maps while preserving key attributes. Fully connected layers, typically positioned at the final stage of the network, may perform classification tasks.

CNNs may exhibit two defining characteristics: weight sharing and local connectivity. Weight sharing involves using identical weights across nodes within a layer, enhancing computational efficiency. Local connectivity ensures that each node processes input data from a limited subset of the input vector, enabling the network to focus on localized features and produce outputs related to specific input regions.

FIG. 3 illustrates an adaptive neural network model 300 configured to process multiple inputs, labeled X 1 through Xn, each associated with an adaptive weight (e.g., w 1 through wn) and an overall weight of w 0 . These weighted inputs are aggregated at a summation junction, represented by the summation symbol (Σ). The resultant sum, denoted as ‘a,’ is passed through a quantizer activation function 302 , represented as f(a), which transforms the input into a quantized output 106 . This output may correspond to the network&#39;s decision, classification, or other inference result.

The model 300 incorporates a Kalman filter algorithm component 304 as a feedback mechanism. The Kalman filter 304 receives both the actual output from the activation function 302 and the desired output. By comparing these outputs, the Kalman filter dynamically adjusts the weights (e.g., w 0 through wn) to reduce the difference between the actual and desired outputs. This feedback process enhances the network&#39;s precision by iteratively refining the weights, allowing the system to improve its accuracy and efficiency over time.

Model 300 also has the ability to reduce training time as well as provide selective adjustments to specific nodes or layers. For example the weights w 0 through wn can be preset to initiate training, provide a trained model, or provide updates to the existing weights.

The weights can be set so they do not change allowing for trained model to be deployed in multiple diverse locations all providing similar responses based on inputs. The weights can be updated either on a whole model, individual layer or layers or on selective nodes. The weights can also be dynamically updated in real time based on the feedback, recursive feedback process either on a system, layers or nodes. Through multiple methods of updating node and layer weights the AI network can be optimized for the desired objective.

The proposed weight adjustments derived my PHIL may not be implemented or partially implemented or fully implemented based on the desired objective. The updated weights from PHIL can be provided to an new or existing model for enhanced training. The output of PHIL can be a flat file, a matrix, or any format representing the weighted values for each layer and or node allowing for an optimized approach to training models.

The weights and or covariant matrix used by PHIL can be updated via an external source for providing a revised model. The updates could be for the entire model, individual layer or layers, or for a specific node, nodes or set of nodes depending on what updates are needed. Though performing selective updates a revised model can be efficiently updated in real-time or near real time providing a method of improving AI models that allowed for a federated learning process.

FIG. 3 exemplifies the functionality of a single neuron, similar to those present in a CNN. Each input undergoes a weighting process before aggregation at the summation junction. The Kalman filter algorithm determines these weights by analyzing expected versus actual outputs. Inputs with low confidence or poor data quality may be assigned reduced weights or even excluded to reduce or minimize their impact on the output. This may balance noise in the decision-making process while retaining sufficient information for robust inferences.

The Kalman filter algorithm may allow continuous learning within the model. By iteratively adjusting the weights in real time, the neural network may adapt to changing conditions and incorporate new data into its processing pipeline.

The model 300 may also operate as part of a fuzzy logic system. Prior to quantization, the signal represented by ‘a’ may exist as a non-quantized value, which may vary continuously within a specified range, such as 0 to 1 or −1 to 1. These non-quantized values may be interpreted as fuzzy logic states, which may inform more complex decision-making processes. Such values may also serve as inputs for additional neurons to allow for the representation of multiple states before rendering a final quantized output suitable for external actions.

In the context of deeper neural architectures, models with three or more hidden layers may be classified as DNNs. FIG. 2 depicts a basic example of a DNN architecture in which layered relationships between neurons enable the network to process data hierarchically. During training, these relationships may be reinforced through backpropagation. Following training, the network applies the learned relationships to predict outputs based on new inputs.

DNNs may operate in various configurations, including centralized, decentralized, and hybrid models. A centralized DNN may involve training in a cloud environment using diverse data sources. A decentralized DNN may allo

CLAIMS

Claims ( 36 )

What is claimed is:

1 . A computing device, comprising:

a processor configured to:

monitor real-time model accuracy of a neural network by using a feedback loop configured to detect changes in inference accuracy;

dynamically adjust a structure of the neural network by adding or removing hidden layers based on monitored error rates and predetermined computational constraints;

compute a Kalman gain based on monitored error rates, and determining weight adjustments for the neural network based on the computed Kalman gain;

incrementally update weight matrices of the neural network based on the determined weight adjustments;

store incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates;

store inference results generated by the neural network and refined weights in an inference result database;

periodically update pre-trained models associated with the neural network based on the stored incremental weight adjustments; and

apply predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy.

2 . The computing device of claim 1 , wherein the processor is configured to monitor real-time model accuracy by using the feedback loop by tracking model performance metrics by repeatedly analyzing inference accuracy in response to changes in real-time network conditions experienced by the neural network.

3 . The computing device of claim 2 , wherein the processor is further configured to:

compare predicted outputs generated by the neural network to actual network behavior to identify deviations between predicted outputs and actual network behavior; and trigger recalibration of the neural network in response to determining that an identified deviation exceeds a predefined accuracy threshold value.

4 . The computing device of claim 3 , wherein the processor is further configured to dynamically adjust the predefined accuracy threshold value based on measured variability in network conditions to reduce the frequency of recalibrations of the neural network.

5 . The computing device of claim 1 , wherein the processor is configured to dynamically adjust the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints by:

analyzing neuron activation patterns of at least one hidden layer of the neural network to determine whether the at least one hidden layer contributes to inference accuracy; and improving computational efficiency of the neural network by removing the at least one hidden layer from the structure of the neural network in response to determining that the at least one hidden layer does not substantially contribute to inference accuracy.

6 . The computing device of claim 5 , wherein the processor is further configured to:

analyze neuron activation patterns of the at least one hidden layer comprises analyzing neuron activation levels to determine whether the at least one hidden layer consistently exhibits low activation during inference operations; and remove the at least one hidden layer comprises removing the at least one hidden layer from the neural network structure in response to determining that the at least one hidden layer consistently exhibits low activation.

7 . The computing device of claim 1 , wherein the processor is configured to dynamically adjust the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints by implementing a self-pruning mechanism configured to identify and remove neurons from hidden layers exhibiting low activation gradients.

8 . The computing device of claim 1 , wherein the processor is configured to dynamically adjust the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints by replacing an activation function used by neurons within hidden layers with an alternative activation function selected from a rectified linear unit (ReLU) or a Leaky ReLU to increase learning efficiency.

9 . The computing device of claim 1 , wherein the processor is configured to compute the Kalman gain based on monitored error rates and determining weight adjustments based on the computed Kalman gain by:

computing the Kalman gain by integrating weighted error estimates derived from real-time inference feedback; and refining weight adjustments of the neural network based on the computed Kalman gain to improve convergence rates during inference operations.

10 . The computing device of claim 1 , wherein the processor is configured to storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates further by applying federated learning techniques to distribute incremental weight differentials instead of distributing full neural network model updates.

11 . The computing device of claim 1 , wherein the processor is configured to updating pre-trained models associated with the neural network periodically based on stored incremental weight adjustments by:

incorporating newly observed network traffic patterns and associated device behavior into an existing architecture of the pre-trained models; and executing asynchronous updates of the pre-trained models to maintain uninterrupted real-time inference operations.

12 . The computing device of claim 1 , wherein the processor is configured to apply predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy by integrating multiple inference sources associated with the neural network to refine model predictions through a consensus-based approach.

13 . A method for adaptive edge computing using artificial intelligence (AI), the method comprising:

monitoring real-time model accuracy of a neural network by using a feedback loop configured to detect changes in inference accuracy; dynamically adjusting a structure of the neural network by adding or removing hidden layers based on monitored error rates and predetermined computational constraints; computing a Kalman gain based on monitored error rates, and determining weight adjustments for the neural network based on the computed Kalman gain; incrementally updating weight matrices of the neural network based on the determined weight adjustments; storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates; storing inference results generated by the neural network and refined weights in an inference result database; periodically updating pre-trained models associated with the neural network based on the stored incremental weight adjustments; and applying predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy.

14 . The method of claim 13 , wherein monitoring real-time model accuracy by using the feedback loop comprises tracking model performance metrics by repeatedly analyzing inference accuracy in response to changes in real-time network conditions experienced by the neural network.

15 . The method of claim 14 , further comprising:

comparing predicted outputs generated by the neural network to actual network behavior to identify deviations between predicted outputs and actual network behavior; and triggering recalibration of the neural network in response to determining that an identified deviation exceeds a predefined accuracy threshold value.

16 . The method of claim 15 , further comprising dynamically adjusting the predefined accuracy threshold value based on measured variability in network conditions to reduce the frequency of recalibrations of the neural network.

17 . The method of claim 13 , wherein dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints comprises:

analyzing neuron activation patterns of at least one hidden layer of the neural network to determine whether the at least one hidden layer contributes to inference accuracy; and improving computational efficiency of the neural network by removing the at least one hidden layer from the structure of the neural network in response to determining that the at least one hidden layer does not substantially contribute to inference accuracy.

18 . The method of claim 17 , wherein:

analyzing neuron activation patterns of the at least one hidden layer comprises analyzing neuron activation levels to determine whether the at least one hidden layer consistently exhibits low activation during inference operations; and removing the at least one hidden layer comprises removing the at least one hidden layer from the neural network structure in response to determining that the at least one hidden layer consistently exhibits low activation.

19 . The method of claim 13 , wherein dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints comprises implementing a self-pruning mechanism configured to identify and remove neurons from hidden layers exhibiting low activation gradients.

20 . The method of claim 13 , wherein dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints comprises replacing an activation function used by neurons within hidden layers with an alternative activation function selected from a rectified linear unit (ReLU) or a Leaky ReLU to increase learning efficiency.

21 . The method of claim 13 , wherein computing the Kalman gain based on monitored error rates and determining weight adjustments based on the computed Kalman gain comprises:

computing the Kalman gain by integrating weighted error estimates derived from real-time inference feedback; and refining weight adjustments of the neural network based on the computed Kalman gain to improve convergence rates during inference operations.

22 . The method of claim 13 , wherein storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates further comprises applying federated learning techniques to distribute incremental weight differentials instead of distributing full neural network model updates.

23 . The method of claim 13 , wherein updating pre-trained models associated with the neural network periodically based on stored incremental weight adjustments comprises:

incorporating newly observed network traffic patterns and associated device behavior into an existing architecture of the pre-trained models; and executing asynchronous updates of the pre-trained models to maintain uninterrupted real-time inference operations.

24 . The method of claim 13 , wherein applying predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy comprises integrating multiple inference sources associated with the neural network to refine model predictions through a consensus-based approach.

25 . A non-transitory processor-readable storage medium having stored thereon processor-executable instructions configured to cause a processing system in a computing device to perform operations for adaptive edge computing using artificial intelligence (AI), the method comprising:

monitoring real-time model accuracy of a neural network by using a feedback loop configured to detect changes in inference accuracy; dynamically adjusting a structure of the neural network by adding or removing hidden layers based on monitored error rates and predetermined computational constraints; computing a Kalman gain based on monitored error rates, and determining weight adjustments for the neural network based on the computed Kalman gain; incrementally updating weight matrices of the neural network based on the determined weight adjustments; storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates; storing inference results generated by the neural network and refined weights in an inference result database; periodically updating pre-trained models associated with the neural network based on the stored incremental weight adjustments; and applying predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy.

26 . The non-transitory processor-readable storage medium of claim 25 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that monitoring real-time model accuracy by using the feedback loop comprises tracking model performance metrics by repeatedly analyzing inference accuracy in response to changes in real-time network conditions experienced by the neural network.

27 . The non-transitory processor-readable storage medium of claim 26 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations further comprising:

comparing predicted outputs generated by the neural network to actual network behavior to identify deviations between predicted outputs and actual network behavior; and triggering recalibration of the neural network in response to determining that an identified deviation exceeds a predefined accuracy threshold value.

28 . The non-transitory processor-readable storage medium of claim 27 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations further comprising dynamically adjusting the predefined accuracy threshold value based on measured variability in network conditions to reduce the frequency of recalibrations of the neural network.

29 . The non-transitory processor-readable storage medium of claim 25 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints comprises:

analyzing neuron activation patterns of at least one hidden layer of the neural network to determine whether the at least one hidden layer contributes to inference accuracy; and improving computational efficiency of the neural network by removing the at least one hidden layer from the structure of the neural network in response to determining that the at least one hidden layer does not substantially contribute to inference accuracy.

30 . The non-transitory processor-readable storage medium of claim 29 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that:

analyzing neuron activation patterns of the at least one hidden layer comprises analyzing neuron activation levels to determine whether the at least one hidden layer consistently exhibits low activation during inference operations; and removing the at least one hidden layer comprises removing the at least one hidden layer from the neural network structure in response to determining that the at least one hidden layer consistently exhibits low activation.

31 . The non-transitory processor-readable storage medium of claim 25 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints comprises implementing a self-pruning mechanism configured to identify and remove neurons from hidden layers exhibiting low activation gradients.

32 . The non-transitory processor-readable storage medium of claim 25 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that dynamically adjusting the structure of the neural network by adding or removing hidden layers based on the monitored error rates and predetermined computational constraints comprises replacing an activation function used by neurons within hidden layers with an alternative activation function selected from a rectified linear unit (ReLU) or a Leaky ReLU to increase learning efficiency.

33 . The non-transitory processor-readable storage medium of claim 25 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that computing the Kalman gain based on monitored error rates and determining weight adjustments based on the computed Kalman gain comprises:

computing the Kalman gain by integrating weighted error estimates derived from real-time inference feedback; and refining weight adjustments of the neural network based on the computed Kalman gain to improve convergence rates during inference operations.

34 . The non-transitory processor-readable storage medium of claim 25 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that storing incremental weight adjustments of the neural network in memory for performing low-bandwidth model updates further comprises applying federated learning techniques to distribute incremental weight differentials instead of distributing full neural network model updates.

35 . The non-transitory processor-readable storage medium of claim 1 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that updating pre-trained models associated with the neural network periodically based on stored incremental weight adjustments comprises:

incorporating newly observed network traffic patterns and associated device behavior into an existing architecture of the pre-trained models; and executing asynchronous updates of the pre-trained models to maintain uninterrupted real-time inference operations.

36 . The non-transitory processor-readable storage medium of claim 25 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that applying predictive holistic inference logic (PHIL) to the inference results stored in the inference result database to produce enhanced inference accuracy comprises integrating multiple inference sources associated with the neural network to refine model predictions through a consensus-based approach.

US19/075,637

2024-03-15

2025-03-10

Systems and Methods for Dynamic Neural Network Enhancement and Adaptive Edge Computing

Pending

US20250292091A1

( en )

Priority Applications (1)

Application Number

Priority Date

Filing Date

Title

US19/075,637

US20250292091A1

( en )

2024-03-15

2025-03-10

Systems and Methods for Dynamic Neural Network Enhancement and Adaptive Edge Computing

Applications Claiming Priority (2)

Application Number

Priority Date

Filing Date

Title

US202463565792P

2024-03-15

2024-03-15

US19/075,637

US20250292091A1

( en )

2024-03-15

2025-03-10

Systems and Methods for Dynamic Neural Network Enhancement and Adaptive Edge Computing

Publications (1)

Publication Number

Publication Date

US20250292091A1

true

US20250292091A1 ( en )

2025-09-18

Family

ID=97028843

Family Applications (1)

Application Number

Title

Priority Date

Filing Date

US19/075,637

Pending

US20250292091A1

( en )

2024-03-15

2025-03-10

Systems and Methods for Dynamic Neural Network Enhancement and Adaptive Edge Computing

Country Status (1)

Country

Link

US

( 1 )

US20250292091A1

( en )

Cited By (18)

* Cited by examiner, † Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

CN120949858A

( en )

*

2025-10-16

2025-11-14

南通弘铭机械科技有限公司

A device temperature regulation method integrating long short-term memory networks

CN121009401A

( en )

*

2025-10-28

2025-11-25

迈拓仪表股份有限公司

A self-optimizing ultrasonic flow monitoring method based on cloud-centric neural networks

CN121212230A

( en )

*

2025-12-01

2025-12-26

零八一电子集团有限公司

Model Adaptive Compression Method for Edge Inference and Rack-Mounted AI Edge Inference Terminal

CN121234163A

( en )

*

2025-12-02

2025-12-30

大连理工大学宁波研究院

Artificial Intelligence-Based Methods and Systems for Watershed Flood Forecasting

CN121333063A

( en )

*

2025-11-11

2026-01-13

湖南工学院

An Adaptive Voltage and Current Regulation Control Method for Switching Power Supplies Based on AI Feedforward Compensation

CN121334421A

( en )

*

2025-12-12

2026-01-13

南昌首页科技股份有限公司

A method and system for spatiotemporal multiplexing of network bandwidth based on traffic prediction

CN121456753A

( en )

*

2025-11-03

2026-02-03

中国地质大学（武汉）

Time sequence anomaly detection method based on multi-window clustering symbolization

CN121479219A

( en )

*

2026-01-09

2026-02-06

国家能源费县发电有限公司

A Method and System for Online Early Warning of the Health Status of Thermal Storage Tanks Based on LSTM

CN121523073A

( en )

*

2026-01-19

2026-02-13

奥特莱博(苏州)实验室科技有限公司

Artificial Intelligence-Based Intelligent Control System for Motion Positioning in Microscopic Imaging

US20260067164A1

( en )

*

2024-08-28

2026-03-05

The United States Of America, As Represented By The Secretary Of The Navy

Hierarchical Management Policies for Data Queues

CN121660113A

( en )

*

2025-12-16

2026-03-13

南宁超森科技有限公司

Lightweight Model Edge Deployment Optimization Method Based on Incremental Learning

CN121751257A

( en )

*

2026-02-28

2026-03-27

南京航空航天大学

Network slice resource allocation method based on Tsallis-MAAC

CN121808710A

( en )

*

2026-03-10

2026-04-07

众数(厦门)信息科技有限公司

Method for realizing real-time reasoning acceleration of multi-mode large model

CN121809596A

( en )

*

2026-03-09

2026-04-07

南京师范大学

Logits knowledge distillation method and system based on fuzzy self-adaptive PID temperature control

US20260100879A1

( en )

*

2024-09-19

2026-04-09

Dell Products L.P.

Trackable node provisioning for multi-cloud on link-local network

US20260105458A1

( en )

*

2024-10-16

2026-04-16

Notion Labs, Inc.

Unified contact database system with customizable multi-source data integration

CN121906586A

( en )

*

2026-01-09

2026-04-21

新能瑞驰（北京）能源技术有限公司

Urban distributed energy storage configuration method, system and platform

CN122114687A

( en )

*

2026-04-29

2026-05-29

贵州云端辉鸿科技有限公司

Semi-real-time prediction model updating method based on dual-model parallel verification

2025

2025-03-10

US

US19/075,637

patent/US20250292091A1/en

active

Pending

Cited By (19)

* Cited by examiner, † Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

US12647321B2

( en )

*

2024-08-28

2026-06-02

United States Of America As Represented By The Secretary Of The Navy

Hierarchical management policies for data queues

US20260067164A1

( en )

*

2024-08-28

2026-03-05

The United States Of America, As Represented By The Secretary Of The Navy

Hierarchical Management Policies for Data Queues

US20260100879A1

( en )

*

2024-09-19

2026-04-09

Dell Products L.P.

Trackable node provisioning for multi-cloud on link-local network

US20260105458A1

( en )

*

2024-10-16

2026-04-16

Notion Labs, Inc.

Unified contact database system with customizable multi-source data integration

CN120949858A

( en )

*

2025-10-16

2025-11-14

南通弘铭机械科技有限公司

A device temperature regulation method integrating long short-term memory networks

CN121009401A

( en )

*

2025-10-28

2025-11-25

迈拓仪表股份有限公司

A self-optimizing ultrasonic flow monitoring method based on cloud-centric neural networks

CN121456753A

( en )

*

2025-11-03

2026-02-03

中国地质大学（武汉）

Time sequence anomaly detection method based on multi-window clustering symbolization

CN121333063A

( en )

*

2025-11-11

2026-01-13

湖南工学院

An Adaptive Voltage and Current Regulation Control Method for Switching Power Supplies Based on AI Feedforward Compensation

CN121212230A

( en )

*

2025-12-01

2025-12-26

零八一电子集团有限公司

Model Adaptive Compression Method for Edge Inference and Rack-Mounted AI Edge Inference Terminal

CN121234163A

( en )

*

2025-12-02

2025-12-30

大连理工大学宁波研究院

Artificial Intelligence-Based Methods and Systems for Watershed Flood Forecasting

CN121334421A

( en )

*

2025-12-12

2026-01-13

南昌首页科技股份有限公司

A method and system for spatiotemporal multiplexing of network bandwidth based on traffic prediction

CN121660113A

( en )

*

2025-12-16

2026-03-13

南宁超森科技有限公司

Lightweight Model Edge Deployment Optimization Method Based on Incremental Learning

CN121479219A

( en )

*

2026-01-09

2026-02-06

国家能源费县发电有限公司

A Method and System for Online Early Warning of the Health Status of Thermal Storage Tanks Based on LSTM

CN121906586A

( en )

*

2026-01-09

2026-04-21

新能瑞驰（北京）能源技术有限公司

Urban distributed energy storage configuration method, system and platform

CN121523073A

( en )

*

2026-01-19

2026-02-13

奥特莱博(苏州)实验室科技有限公司

Artificial Intelligence-Based Intelligent Control System for Motion Positioning in Microscopic Imaging

CN121751257A

( en )

*

2026-02-28

2026-03-27

南京航空航天大学

Network slice resource allocation method based on Tsallis-MAAC

CN121809596A

( en )

*

2026-03-09

2026-04-07

南京师范大学

Logits knowledge distillation method and system based on fuzzy self-adaptive PID temperature control

CN121808710A

( en )

*

2026-03-10

2026-04-07

众数(厦门)信息科技有限公司

Method for realizing real-time reasoning acceleration of multi-mode large model

CN122114687A

( en )

*

2026-04-29

2026-05-29

贵州云端辉鸿科技有限公司

Semi-real-time prediction model updating method based on dual-model parallel verification

Similar Documents

Publication

Publication Date

Title

Liang et al.

2024

Communication-efficient large-scale distributed deep learning: A comprehensive survey

US10262019B1

( en )

2019-04-16

Distributed management optimization for IoT deployments

US12566635B2

( en )

2026-03-03

Systems and methods for dynamic allocation of compute resources via a machine learning-informed feedback sequence

Gopinathan

2025

AI-powered Kubernetes orchestration for complex cloud-native workloads

Bulla et al.

2021

Anomaly detection in industrial IoT applications using deep learning approach

US12619410B2

( en )

2026-05-05

Systems and methods for automated mesh service-based deployment intelligence

Wang et al.

2025

Federated anomaly detection for multi-tenant cloud platforms with personalized modeling

Qiu et al.

2025

Spatiotemporal traffic prediction in distributed backend systems via graph neural networks

Asif et al.

2025

R-IBN: A reinforcement learning-based intent-driven framework for end-to-end service orchestration and optimization

Kumar et al.

2025

A hybrid neural network and cooperative PSO model for dynamic cloud workloads prediction

Guo et al.

2022

Hierarchical design space exploration for distributed CNN inference at the edge

Sadykova

2024

Artificial intelligence for automated data workflow optimization in cloud-based big data systems

Hassan et al.

2025

Intelligent load balancing in cloud computing with reinforcement learning

US20250133491A1

( en )

2025-04-24

Closed Loop Machine Learning Based Power Optimization Techniques

US11809915B1

( en )

2023-11-07

Parallel processing techniques for expediting reconciliation for a hierarchy of forecasts on a computer system

Mu et al.

2019

BGElasor: elastic-scaling framework for distributed streaming processing with deep neural network

Bukhari et al.

2024

Network Traffic Prediction: Using AI to Predict and Manage Traffic in High-Demand IT Networks

Mahmoudi et al.

2026

A novel state-space ODE framework with tensor decomposition for resource allocation in fog-DSDN networks

Muhamad et al.

2024

Governing Resource Failures through Reinforcement Learning Scheduling in Fog/Edge Computing: A Review

Farahi et al.

2026

DeepWK-MSTC: a novel approach for adaptive controller placement in software-defined networks via deep learning

Ahmadpanah et al.

2026

Towards Decentralized Hybrid Container Elasticity With Fault‐Tolerant Orchestration and Predictive, Risk‐Aware Thresholding

Narsimhulu et al.

2026

A hybrid RL–GA–LSTM–AE framework for energy-aware and SLA-driven task scheduling in cloud computing environments

Mishra et al.

2025

Supervised Machine Learning for Load Balancing in Fog Environments

US20260044768A1

( en )

2026-02-12

System for resource allocation in a hybrid distributed computational environment

Makinde

2025

Machine Learning-Aided Auto-Scaling Strategies in Serverless Big Data Architectures

Legal Events

Date

Code

Title

Description

2025-03-10

AS

Assignment

Owner name : VEEA INC., NEW YORK

Free format text : ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:SMITH, CLINT;REEL/FRAME:070462/0035

Effective date : 20250310

2025-03-27

STPP

Information on status: patent application and granting procedure in general

Free format text : DOCKETED NEW CASE - READY FOR EXAMINATION

Related documents

• Channel-Adaptive Edge AI: Maximizing Inference Throughput by Adapting Computational Complexity to Channel States#820377
• Adaptive Depth and Expert Refinement for Efficient Speech Enhancement#1037507
• Edge GPU Aware Multiple AI Model Pipeline for Accelerated MRI Reconstruction and Analysis#1129120
• Brain-inspired AI for Edge Intelligence: a systematic review#2719
• Does per-frame early exit pay? A compute-matched study of dynamic depth for on-device speech enhancement#1077794