System and method for establishing new in-network learning heterarchical intelligent collaboration session in communications network
Patent Information
- Application Number
- PCT/CN2025/085409
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-03-27
- Publication Date
- 2026-10-01
Smart Images

Figure CN2025085409_01102026_PF_FP_ABST
Abstract
Description
SYSTEM AND METHOD FOR ESTABLISHING NEW IN-NETWORK LEARNING HETERARCHICAL INTELLIGENT COLLABORATION SESSION IN COMMUNICATIONS NETWORKTECHNICAL FIELD
[0001] The present disclosure relates to In-Network Learning (INL) and Heterarchical Intelligent Collaboration (HiC) in distributed communication networks. Specifically, the present disclosure relates to a system and a method for establishing a new INL HiC session in a communications network.BACKGROUND
[0002] In-network learning (INL) has emerged as a framework for distributed machine learning in communication networks, where multiple nodes collaborate to train neural network models. A subset of INL, heterarchical intelligent collaboration (HiC) , facilitates distributed learning across a dynamically changing network topology. However, when a new learning task arises, establishing an efficient INL HiC session remains challenging due to the lack of standardized mechanisms for leveraging existing learned models. Traditional approaches require significant computational resources and extensive training data, making it difficult to scale learning processes dynamically within a network.
[0003] Current methods for distributed learning often rely on independent task initialization, where each new task requires a complete retraining of the associated models from scratch. While transfer learning has been explored in centralized machine learning paradigms, there has been little to no implementation of transfer learning in the context of HiC-based INL. Specifically, existing techniques do not define efficient signaling protocols for identifying relevant existing sessions, selecting appropriate model parameters, and ensuring compatibility between tasks. The absence of standardized signaling mechanisms results in inefficiencies, such as increased latency, redundant computations, and suboptimal resource allocation.
[0004] Therefore, in light of the foregoing discussion, there exists a need to overcome the aforementioned drawbacks.SUMMARY
[0005] The present disclosure provides a system, a method, and a computer program for establishing a new In-Network Learning (INL) Heterarchical Intelligent Collaboration (HiC) session in a communications network for a new task by leveraging transfer learning using an existing trained model associated with an existing task and an existing INL HiC session. The present disclosure provides a solution to the existing problem of how to efficiently establish a new In-Network Learning (INL) Heterarchical Intelligent Collaboration (HiC) session for a new task by leveraging transfer learning from an existing trained model associated with an active INL HiC session. Traditional approaches to distributed learning in communication networks often require extensive computational resources, prolonged training times, and redundant data processing when initiating new learning tasks. Additionally, the prior art lacks an effective mechanism for determining the similarity between tasks, selecting the most relevant existing session, and efficiently transferring learned parameters, hyperparameters, and communication patterns. These limitations result in inefficient utilization of network resources, increased latency, and suboptimal learning performance.
[0006] An objective of the present disclosure is to provide a solution that overcomes at least partially the problems encountered in the prior art and provides an improved system and an improved method for dynamically establishing a new INL HiC session using knowledge transfer from an existing session. The method optimizes resource utilization by selecting the most relevant active HiC session based on multiple measurements, including task similarity, implementation quality, and node compatibility. It further enhances the efficiency of distributed learning by gathering critical model-related information, such as neural network structures, optimization strategies, and communication paths, and then intelligently distributing this information to nodes in the new INL HiC session. By leveraging a structured taxonomy-based similarity distance measure, Relevance Indicators (RI) , and fine-tuned transition loss functions, the proposed system ensures a seamless adaptation of pre-trained models to new tasks, reducing computational overhead and improving learning accuracy. Additionally, the disclosed solution improves network scalability, minimizes signaling overhead, and enhances fault tolerance, thereby enabling a more autonomous, intelligent, and efficient AI-driven communication network.
[0007] One or more objectives of the present disclosure are achieved by the solutions provided in the enclosed independent claims. Advantageous implementations of the present disclosure are further defined in the dependent claims.
[0008] In one aspect, the present disclosure provides a method which includes requesting information from a plurality of HiC agents in each active INL HiC session for each of a plurality of active INL HiC sessions. The HiC agent is located within a respective node of the communications network. The method further includes receiving responses from HiC agents from each of the plurality of active INL HiC sessions. The method further includes selecting one of the plurality of active INL HiC sessions based on a plurality of measurements. The method further includes gathering information from the selected INL HiC session. The method further includes sending the gathered information to a plurality of nodes in the new INL HiC session, to enable the nodes of the new INL HiC session to perform transfer learning, where the gathered information is related to patterns and parameters related to the responses received from the HiC agents, where the patterns correspond to the active INL HiC session and include information concerning a neural network model, an optimisation method, related hyperparameters and a communication path.
[0009] The method provides the optimization of learning resources through transfer learning, which allows new tasks to inherit patterns and parameters from existing trained models. This eliminates the need to train models from scratch, thereby reducing computational complexity and minimizing power consumption in resource-constrained network nodes, such as user equipment (UE) or edge devices. The ability to leverage knowledge from previously established HiC sessions ensures that the new session benefits from pre-optimized neural network models, hyperparameters, and optimization strategies, resulting in faster convergence rates and improved model accuracy with fewer training iterations. Another key advantage of the method is the dynamic adaptability to network conditions and task similarities through a structured selection process. The method employs a plurality of measurements, including task similarity, implementation quality, and node compatibility, to ensure that the most relevant existing HiC session is selected for transfer learning. This targeted selection mechanism enhances the effectiveness of knowledge transfer, ensuring that the new task is initialized with highly relevant information rather than generic pre-trained models. The incorporation of a taxonomy-based similarity distance measure allows tasks to be categorized based on type, subtype, and properties, enabling a precise and efficient matching of learning models. Furthermore, the use of Relevance Indicators (RI) and data modality considerations ensures that the quality of transferred knowledge is maintained, preventing performance degradation due to poor task alignment.
[0010] Additionally, the proposed method improves network-wide coordination and scalability by efficiently distributing gathered information to multiple nodes within the new INL HiC session. This ensures synchronized knowledge dissemination across network entities, reducing inconsistencies and improving overall system stability. The method also minimizes signaling overhead by structuring information exchange in a heterarchical manner, where the HiC Controller node centrally manages the communication between HiC agents, base stations, and network function entities (AF / NF nodes) . By enabling intelligent task allocation and model adaptation, the method ensures that existing network resources are utilized optimally, leading to lower latency, enhanced bandwidth efficiency, and improved overall learning performance in decentralized AI-driven applications.
[0011] Moreover, the method provides robust fault tolerance and resilience by allowing new tasks to be initialized with backup models and communication paths derived from existing HiC sessions. This ensures that even in cases of partial network failures or changes in node availability, the learning process can continue without significant disruption. The structured approach to gathering and deploying learned knowledge also contributes to modular and scalable AI model deployment, making it well-suited for evolving 5G and 6G communication networks where dynamic, real-time adaptation is critical. By integrating the principles of heterarchical collaboration, the proposed method ensures that learning processes are not solely dependent on a centralized architecture, making AI-driven networks more autonomous, intelligent, and self-optimizing.
[0012] In an implementation form, one of the HiC agents is in a user equipment (UE) device of the communications network.
[0013] Placing a HiC agent in the UE device enables localized processing, reducing reliance on centralized servers and minimizing latency. The reduction in reliance enhances real-time adaptability to network changes, optimizes resource utilization at the edge, and improves overall system efficiency by distributing computational load across network nodes.
[0014] In an implementation form, one of the HiC agents is in a base station (BS) device of the communications network.
[0015] Deploying a HiC agent in the BS device enables efficient coordination between edge devices and the core network, reducing data transmission latency and improving real-time decision-making. The improvement in real-time decision-making enhances network adaptability, optimizes resource allocation, and ensures seamless transfer learning in dynamic network conditions.
[0016] In an implementation form, the transfer learning carried out by the new INL HiC session includes making a copy of the pattern used by the existing INL HiC session.
[0017] Making a copy of the pattern used by the existing INL HiC session allows the new session to quickly inherit optimized neural network architectures, hyperparameters, and communication paths, significantly reducing initialization time and computational overhead. This enhances learning efficiency, ensures seamless adaptation to new tasks, and minimizes the need for redundant model training, leading to improved resource utilization and lower latency in AI-driven communication networks.
[0018] In an implementation form, the transfer learning carried out by the new INL HiC session includes initializing its models and hyperparameters using the parameters used by the existing INL HiC session, and then using a transition loss function to make a transition from the initialized parameters used in the existing task to the parameters associated with the new task.
[0019] Initializing models and hyperparameters using parameters from the existing INL HiC session enables faster convergence, reducing the need for extensive retraining and computational resources. The transition loss function ensures a smooth adaptation from the existing task to the new task, improving accuracy and stability while minimizing performance degradation during the learning transition.
[0020] In an implementation form, the plurality of measurements includes a similarity distance between existing tasks and the new task.
[0021] Using the similarity distance between existing tasks and the new task ensures precise selection of the most relevant INL HiC session for transfer learning, improving knowledge transfer efficiency. This reduces training time, enhances model accuracy, and optimizes resource utilization by leveraging previously learned patterns that closely match the new task requirements.
[0022] In an implementation form, the similarity distance is based on a taxonomy.
[0023] Basing the similarity distance on a taxonomy allows for a structured and systematic comparison of tasks, ensuring accurate selection of the most relevant existing INL HiC session. This enhances transfer learning efficiency by aligning model parameters with task-specific characteristics, leading to improved adaptability, reduced computational overhead, and higher model accuracy.
[0024] In another implementation form, the taxonomy is based on type of task, subtype of task and properties of task.
[0025] Structuring the taxonomy based on task type, subtype, and properties enables a granular and precise classification of tasks, improving the accuracy of task similarity assessment. The accuracy ensures optimal model selection for transfer learning, enhances learning efficiency, and minimizes unnecessary computations, leading to faster convergence and better overall performance.
[0026] In an implementation, a type of task is a classification task.
[0027] Defining the task type as the classification task allows for efficient matching with pre-trained models that specialize in similar learning objectives, improving transfer learning accuracy. This enhances model generalization, reduces training time, and optimizes computational resources by leveraging existing classification frameworks.
[0028] In an implementation form, a subtype of task is image classification.
[0029] Defining image classification as a subtype allows for precise selection of pre-trained models optimized for visual data, improving feature extraction and pattern recognition. The improvement in feature extraction and pattern recognition enhances transfer learning efficiency, reduces training time, and ensures better accuracy by leveraging domain-specific knowledge from similar tasks.
[0030] In an implementation form, for the subtype of image classification, the properties include color or black and white, and 2 dimensional or 3 dimensional.
[0031] Defining image classification properties such as color mode and dimensionality enables accurate matching of pre-trained models with new tasks. The accurate matching improves feature extraction, enhances model adaptation, and reduces computational overhead by utilizing networks optimized for specific image characteristics.
[0032] In an implementation form, the plurality of measurements includes the quality of implementation of an existing task.
[0033] Assessing the quality of an existing task's implementation ensures that only well-optimized models with high accuracy and stability are used for transfer learning. The higher accuracy enhances learning efficiency, reduces computational waste, and improves the reliability of the new INL HiC session.
[0034] In an implementation form, the quality is based on the number of agents supporting the existing task.
[0035] Determining quality based on the number of agents supporting the existing task ensures that transfer learning leverages a well-trained and diverse model. A higher number of agents indicates broader data coverage, improved generalization, and a more robust learning process. The more robust learning enhances model reliability, reduces bias, and increases adaptability to new tasks.
[0036] In an implementation form, the quality is based on the Relevance Indicator (RI) .
[0037] The quality assessment based on the Relevance Indicator (RI) provides the technical advantage of enabling optimized session selection by quantifying the applicability of existing models to new tasks. The RI metric allows the system to automatically identify the most suitable candidate sessions for transfer learning, thereby significantly reducing computational overhead and convergence time when deploying new learning tasks across the network infrastructure.
[0038] In an implementation form, the quality is further based on the quality of views measured by the RI and based on a list of data modalities.
[0039] Basing quality assessment on the quality of views measured by the RI and the list of data modalities provides the technical advantage of enabling more precise transfer learning selection across heterogeneous data environments. The multi-dimensional evaluation ensures optimal model reuse by considering both the relevance and the nature of available data types, thereby significantly improving learning efficiency and reducing resource consumption when implementing new tasks in diverse network conditions.
[0040] In an implementation form, the quality is further based on the quality of the data available or observed at a plurality of children nodes in the communication network.
[0041] Incorporating quality assessment based on the data available or observed at children nodes provides the technical advantage of leveraging distributed insights from edge devices throughout the network hierarchy. This comprehensive evaluation of data quality at multiple network endpoints enables more accurate session selection for transfer learning, resulting in optimized model initialization and faster convergence when deploying new tasks across heterogeneous network environments.
[0042] In an implementation form, the requesting and gathering steps are performed by an HiC Controller (HicC) node.
[0043] Centralizing the requesting and gathering steps within the HiC Controller (HicC) node provides the technical advantage of unified orchestration with minimal signaling overhead. The architectural design enables efficient discovery and selection of optimal transfer learning sources through a single coordination point, eliminating redundant communications and allowing for holistic comparison of candidate sessions across the entire network infrastructure.
[0044] In an implementation form, the HicC controller node communicates with an Application Function / Network Function (AF / NF) node.
[0045] Establishing communication between the HicC controller node and an Application Function / Network Function (AF / NF) node provides the technical advantage of seamless integration with standard network service frameworks. This interconnection enables direct translation of high-level application requirements into optimized learning tasks, creating an efficient bridge between service deployment and the underlying distributed learning infrastructure while maintaining compatibility with existing network management systems.
[0046] In an implementation form, the gathered information includes an HiC identifier of the selected one of the plurality of active INL HiC sessions for a node that is part of the selected session and also which is part of the new session.
[0047] Including the HiC identifier of the selected active INL HiC session in the gathered information provides the technical advantage of enabling direct model inheritance for nodes participating in both sessions. The efficient referencing mechanism eliminates redundant data transfer by allowing dual-session nodes to locally access existing trained models, significantly reducing initialization overhead and bandwidth consumption when deploying transfer learning across the network.
[0048] In an implementation form, a pattern includes the set of initialization information that is derived from the existing task using only the HiC identifier.
[0049] Including a pattern with initialization information derived from the existing task using only the HiC identifier provides the technical advantage of compact and efficient knowledge transfer between sessions. This minimalist reference mechanism enables nodes to reconstruct complete model architectures and parameters from a simple identifier, dramatically reducing signaling overhead while maintaining full transfer learning capabilities across the distributed network environment.
[0050] In another aspect, the present disclosure provides a system comprising means adapted for carrying out all the steps of the method.
[0051] The system achieves all the advantages and technical effects of the method of the present disclosure.
[0052] It is to be appreciated that all the aforementioned implementation forms can be combined.
[0053] It has to be noted that all devices, elements, circuitry, units, and means described in the present application could be implemented in the software or hardware elements or any kind of combination thereof. All steps which are performed by the various entities described in the present application as well as the functionalities described to be performed by the various entities are intended to mean that the respective entity is adapted to or configured to perform the respective steps and functionalities. Even if, in the following description of specific embodiments, a specific functionality or step to be performed by external entities is not reflected in the description of a specific detailed element of that entity which performs that specific step or functionality, it should be clear for a skilled person that these methods and functionalities can be implemented in respective software or hardware elements, or any kind of combination thereof. It will be appreciated that features of the present disclosure are susceptible to being combined in various combinations without departing from the scope of the present disclosure as defined by the appended claims.
[0054] Additional aspects, advantages, features, and objects of the present disclosure would be made apparent from the drawings and the detailed description of the illustrative implementations construed in conjunction with the appended claims that follow.BRIEF DESCRIPTION OF THE DRAWINGS
[0055] The summary above, as well as the following detailed description of illustrative embodiments, is better understood when read in conjunction with the appended drawings. For the purpose of illustrating the present disclosure, exemplary constructions of the disclosure are shown in the drawings. However, the present disclosure is not limited to specific methods and instrumentalities disclosed herein. Moreover, those in the art will understand that the drawings are not to scale. Wherever possible, like elements have been indicated by identical numbers.
[0056] Embodiments of the present disclosure will now be described, by way of example only, with reference to the following diagrams wherein:
[0057] FIG. 1 is a block diagram that depicts a system configured for establishing a new In-Network Learning (INL) Heterarchical Intelligent Collaboration (HiC) session in a communications network for a new task, in accordance with an embodiment of the present disclosure;
[0058] FIG. 2 is a flowchart depicting a method for establishing a new In-Network Learning (INL) Heterarchical Intelligent Collaboration (HiC) session in a communications network for a new task, in accordance with an embodiment of the present disclosure;
[0059] FIG. 3 is an exemplary diagram illustrating a phase for finding a target active HiC session, in accordance with an embodiment of the present disclosure;
[0060] FIG. 4 an exemplary diagram illustrating a phase for gathering information from the target active HiC session, in accordance with an embodiment of the present disclosure; and
[0061] FIG. 5 is a diagram illustrating a phase involving the implementation of the new task on the destination, in accordance with an embodiment of the present disclosure.
[0062] In the accompanying drawings, an underlined number is employed to represent an item over which the underlined number is positioned or an item to which the underlined number is adjacent. A non-underlined number relates to an item identified by a line linking the non-underlined number to the item. When a number is non-underlined and accompanied by an associated arrow, the non-underlined number is used to identify a general item at which the arrow is pointing.DETAILED DESCRIPTION OF EMBODIMENTS
[0063] The following detailed description illustrates embodiments of the present disclosure and ways in which they can be implemented. Although some modes of carrying out the present disclosure have been disclosed, those skilled in the art would recognize that other embodiments for carrying out or practicing the present disclosure are also possible.
[0064] FIG. 1 is a block diagram that depicts a system configured for establishing a new In-Network Learning (INL) Heterarchical Intelligent Collaboration (HiC) session in a communications network for a new task, in accordance with an embodiment of the present disclosure. With reference to FIG. 1, there is shown a block diagram that includes a system 100. The system 100 includes a user equipment (UE) device 102 communicably connected to a base station (BS) device 104 via a communication network 106. The system 100 further includes a fusion centre 108 connected to the BS device 104 and a HiC Controller 110. The HiC Controller 110 is coupled to the BS device 104 and the fusion centre (FC) 108. Further, the HiC Controller 110 is connected to core network function nodes 112.
[0065] The In-Network Learning (INL) enables distributed and collaborative processing by leveraging spatially dispersed clients. The INL approach enhances network resource utilization while preserving data privacy and reducing communication overhead. In the context of communication networks, the INL approach is structured into three distinct levels to handle AI learning tasks efficiently. The first level consists of devices that function as Feature Extractors (FE) , responsible for processing raw data and extracting relevant features before transmitting them further in the network. The devices minimize data transmission requirements by sending only essential information rather than full raw datasets. The second level comprises base stations that act as Local Intermediate Fusioners (LIF) , aggregating and refining the extracted features from multiple devices. The intermediate processing reduces the burden on central processing units and improves response times by performing preliminary inference tasks at the edge. Finally, the third level operates at the core network, serving as the FC 108, where the final aggregation, inference, and decision-making occur. The FC 108 ensures a global understanding of the distributed data, consolidates intermediate results and enhances AI model accuracy through comprehensive learning. By structuring INL into the heterarchical levels, communication networks achieve the balance between computational efficiency, reduced latency, and privacy preservation. The multi-tiered approach allows for adaptive AI learning while ensuring that relevant tasks are executed at the most suitable network nodes.
[0066] The system 100 is a comprehensive goal decomposition and context adaptation framework for distributed intelligent networks, designed to optimize collaboration between network elements using the Heterarchical Intelligent Collaboration (HiC) architecture. The HiC technology (e.g., disclosed in Patent Application Number, PCT / CN2022 / 135723; WO 2024 / 113288 A1) includes multiple nodes engaging in real-time data exchange and distributed decision-making, often under varying network conditions. The present disclosure is an improvement of the HiC technology discussed in Patent Application Number, PCT / CN2022 / 135723; WO 2024 / 113,288 A1.
[0067] HiC is a framework that operates within INL, enabling intelligent coordination, early inference, and efficient decision-making among nodes. While HiC utilizes INL’s three-level structure, it focuses on optimizing collaboration, signaling, and dynamic task allocation across these levels. The HiC is a standard architecture for intelligent collaboration between network elements (NEs) at different layers, including collaboration organization management, protocol interfaces, and interaction processes. The HiC efficiently organizes the collaboration between AI tasks on NEs / terminals in terms of space and time, supports the transmission of diversified intelligent representations between the NEs, and unifies the transmission format. In addition, collaboration is controllable, and collaboration patterns are scalable, and various collaborative learning modes are flexibly supported.
[0068] HiC is an advanced distributed intelligence framework designed for enhancing intelligent tasks across multiple nodes within a communication network. HiC enables the collaborative execution of AI-driven tasks by leveraging a network of heterarchical and self-organizing nodes. The HiC framework enhances the efficiency of AI model training and inference by ensuring seamless coordination among terminals, access network nodes, and core network nodes. The HiC framework consists of two primary types of nodes: the HiC Controller node and the HiC agent node. The HiC Controller node functions as the managing entity responsible for configuring, enhancing, and allocating tasks, while the HiC agent node executes the assigned intelligent collaboration tasks. The HiC framework supports different organizational models for intelligent collaboration. The HiC framework can function in a heterarchical mode, where the HiC Controller 110 manages and configures multiple HiC agent nodes, ensuring efficient execution of distributed tasks. In contrast, the HiC framework can also operate in a self-organizing mode, where HiC agent nodes negotiate among themselves to determine task execution without a central controlling entity. The flexibility allows the HiC framework to be scalable and adaptable, making it suitable for small-scale and large-scale distributed learning environments. In scenarios involving a large number of nodes, a combination of hierarchical and self-organizing modes can be used to balance efficiency and complexity.
[0069] One of the key advantages of HiC framework is its ability to dynamically allocate computational resources across multiple nodes. For example, after completing an intelligent collaboration task, the node can initiate a subsequent task based on the results of the previous one. If the first task involves AI model training, the trained model can be used for further inference or adaptive retraining at another node without needing to restart from an initial state. The approach improves node resource utilization and enhances communication efficiency, as each node can progressively refine its AI models without redundant processing.
[0070] The HiC framework also includes mechanisms for task authorization and configuration. Before the HiC agent node executes an intelligent collaborative task, it must first obtain authorization from the HiC-Controller 110. The authorization ensures that tasks are properly assigned and executed in a coordinated manner. The HiC Controller 110 determines which nodes will participate in a given task, assigns necessary computational resources, and ensures that data samples are sufficiently diverse to meet AI model training requirements. Once the HiC agent nodes are configured, they can execute their assigned tasks and report back to the HiC-Controller 110 for further coordination. In addition, HiC framework incorporates a robust signaling mechanism for monitoring and optimizing ongoing tasks. The nodes participating in an intelligent collaborative task continuously exchange status updates with the HiC Controller 110 or with other HiC-agent nodes, depending on the organizational structure. If the HiC-agent node detects inefficiencies, such as insufficient data diversity or network congestion, it can trigger adaptive reconfiguration of the task execution plan. This allows the HiC framework to dynamically respond to changes in network conditions, computational capacity, and task complexity, ensuring sustained performance across all participating nodes.
[0071] Overall, HiC represents a highly scalable, adaptable, and intelligent framework for distributed AI processing in communication networks. Its combination of hierarchical control and self-organizing capabilities allows for efficient task execution, reduced processing latency, and optimal resource utilization. By dynamically coordinating AI-driven tasks, HiC significantly improves the efficiency of AI applications in telecommunications, edge computing, and autonomous system.
[0072] The system 100, by utilizing the HiC architecture for transfer learning, follows three key phases: finding the target active HiC session, gathering information from the selected active session, and implementing the new task using transfer learning techniques. The structured approach enables more efficient model reuse across sessions, better utilization of pre-trained network resources, improved knowledge transfer between similar tasks, and optimized initialization of new learning processes. Furthermore, the system 100 enhances overall learning efficiency by dynamically selecting the most relevant existing session based on task similarity, implementation quality, and node compatibility metrics. The system 100 effectively manages various forms of transfer learning scenarios, including reusing models across similar classification tasks, leveraging trained parameters for related data modalities, and applying existing optimization strategies to new but similar problems. By integrating the transfer learning mechanisms, the system 100 ensures faster convergence, reduced computational overhead, and more efficient intelligent collaboration across distributed network environments.
[0073] The UE device 102 refers to a mobile or edge computing device that participates in INL transfer learning within the HiC framework. In some implementations, the UE device may be a smartphone, IoT sensor, autonomous vehicle, or any computing-enabled terminal capable of collecting local data, executing learning tasks, and serving as an HiC agent in both existing and new INL HiC sessions. The UE device 102 functions as a distributed HiC agent, assisting in transfer learning by providing information about its current capabilities, data characteristics, relevance indicators, and potential availability for the new task to the HiC Controller. When selected to participate in a new session based on similarity to an existing task, the UE device 102 leverages previously trained models through the HiC identifier and transition parameters provided by the HiC Controller 110 enabling efficient knowledge transfer and faster convergence for the new learning task.
[0074] The BS device 104 refers to a cellular base station that facilitates communication between UE devices and the core network while also serving as a processing and coordination node in the HiC framework for transfer learning. The BS device 104 functions as the HiC agent that can participate in multiple INL HiC sessions simultaneously, making it a valuable node for reusing trained models across similar tasks.
[0075] The communication network 106 includes a medium (e.g., a communication channel) through which the UE device 102 communicates with the BS device 104. The communication network 106 may be wired or wireless. Examples of the communication network 106 may include, but are not limited to, a Local Area Network (LAN) , a wireless personal area network (WPAN) , a Wireless Local Area Network (WLAN) , a wireless wide area network (WWAN) , a cloud network, a Long-Term Evolution (LTE) network, a plain old telephone service (POTS) , a Metropolitan Area Network (MAN) , and / or Internet.
[0076] The FC 108 refers to a centralized or edge-based processing unit that plays a role in data aggregation, model fusion, and higher-level decision-making in the HiC network. The FC 108 processes task results from multiple UEs and BSs, integrates learning insights and refines the global learning model used across the network. It ensures that data from distributed nodes is synthesized, validated, and used effectively for real-time adaptation and optimization in the INL HiC session.
[0077] The HiC Controller 110 refers to the central intelligence node responsible for coordinating transfer learning across multiple INL HiC sessions in the network. The HiC Controller 110 performs the functions of requesting information from HiC agents in active sessions, receiving their capability responses, evaluating sessions based on similarity distances and quality metrics, and selecting the optimal existing session to serve as a foundation for the new task. By interfacing with AF / NF nodes, the HiC Controller 110 receives specifications for new tasks and orchestrates the entire transfer learning process, from discovery to deployment. Once a suitable session is identified, the HiC Controller 110 gathers detailed pattern information, including neural network models, optimization methods, hyperparameters, and communication paths from the selected session, then distributes this information to nodes in the new session along with appropriate transition loss function parameters. In accordance with an embodiment, the requesting, gathering, and sending steps of the transfer learning method are performed by an HiC Controller node.
[0078] The core network function nodes 112 refer to Application Function (AF) and Network Function (NF) nodes within the 5G / 6G service-based architecture that manage network intelligence, policy control, and resource orchestration. The core network function nodes 112 nodes provide key network insights such as congestion levels, mobility patterns, security policies, and service prioritization, enabling the HiC framework to adapt learning strategies accordingly. The core network function nodes 112 also facilitate low-latency coordination between distributed HiC agents and core network services, ensuring efficient execution of in-network learning tasks while maintaining optimal network performance. In accordance with an embodiment, the HiC Controller node communicates with the Application Function / Network Function (AF / NF) node.
[0079] There is provided the system 100, for each of a plurality of active INL HiC sessions, the system 100 is configured to request information from a plurality of HiC agents in each active INL HiC session, for each of a plurality of active INL HiC sessions. The HiC agent is located within a respective node of the communications network. The HiC agent represents an intelligent software component residing within a respective node of the communications network, which may be implemented in various network elements including user equipment (UE) devices, base stations (BS) , or other network functions. The requesting process is orchestrated by the HiC Controller 110, which initiates compatibility queries to all HiC agents within each active session. The compatibility queries are designed to collect comprehensive information about the agents'current operational capabilities, including their processing resources, memory availability, and network connectivity status. Additionally, the queries seek information about the data characteristics available at each node, including modalities, quality, and relevance to the new task under consideration. The system 100 further requests performance metrics of the existing trained models, potential availability for participating in the new task, and estimated accuracy on the new task when using the existing model. The systematic information gathering across all active sessions enables the system 100 to build a comprehensive knowledge base of available resources and trained models, which serves as the foundation for identifying the most suitable existing session to leverage for transfer learning when establishing the new INL HiC session.
[0080] The system 100 is further configured to receive responses from HiC agents from each of the plurality of active INL HiC sessions. Following the transmission of the compatibility queries, the HiC Controller 110 establishes dedicated communication channels with each queried HiC agent to facilitate the secure and efficient reception of response data. These responses comprise comprehensive information packages containing multiple categories of metrics and parameters essential for evaluating session suitability for transfer learning purposes. Specifically, each response includes detailed computational capability metrics of the respective node, such as available processing power, memory resources, and energy status; data characteristic information, including modalities supported, data quality indicators, and volume statistics; current task performance measurements, including accuracy, convergence rates, and resource utilization efficiency; and relevance indicators (RI) quantifying the estimated applicability of the agent's trained model to the new task requirements. Additionally, the responses may include validation test results if the HiC Controller 110 requested preliminary accuracy assessments using sample data from the new task. The system 100 employs robust error handling mechanisms during response collection, including timeout management and retry protocols, to ensure comprehensive data gathering even in challenging network conditions. Each received response is systematically processed, verified for completeness, and stored in a structured database within the HiC Controller 110, enabling subsequent analysis and comparison across all active sessions. The comprehensive response collection mechanism provides the foundation for the sophisticated selection algorithms that will identify the optimal existing session to leverage for transfer learning, thereby maximizing efficiency when establishing the new INL HiC session.
[0081] The system 100 is further configured to select one of the plurality of active INL HiC sessions based on a plurality of measurements. The plurality of measurements refers to a specific set of quantitative metrics used by the system 100 to evaluate and select the optimal active INL HiC session for transfer learning purposes. In accordance with an embodiment, the plurality of measurements includes a similarity distance between existing tasks and the new task. The similarity distance represents a quantitative assessment of how closely related the learning objectives, data characteristics, and computational requirements of an existing task are to those of the new task being established. In accordance with an embodiment, the similarity distance is based on a taxonomy. The taxonomy refers to a hierarchical classification structure that organizes machine learning tasks into categories and subcategories based on their fundamental characteristics. The taxonomy classification framework provides a systematic way to compare tasks by examining their positions within the taxonomic hierarchy.
[0082] In accordance with an embodiment, the taxonomy is based on type of task, subtype of task and properties of task. The type of task refers to the fundamental category of machine learning operation being performed in an INL HiC session. The type of task represents the broadest classification level in the taxonomy and identifies the general purpose of the learning algorithm, such as classification, regression, translation, or other primary machine learning functions. The subtype of task represents a more specific categorization within each type of task, providing greater granularity in task classification. The subtype task is an intermediate taxonomic level and identifies specialized variations of the primary task type. For example, within the classification type, subtypes might include image classification, audio classification, or sentiment classification. The properties of task constitute the most detailed level of the taxonomy, describing specific characteristics and parameters unique to each subtype. The properties of task define particular attributes of the learning task that further differentiate it from other tasks within the same subtype. For example, in image classification, properties may include whether images are color or black-and-white, 2D or 3D, focused on people or objects, or other distinguishing features relevant to the specific learning objective.
[0083] In accordance with an embodiment, a type of task is a classification task. The classification task represents a fundamental type of machine learning operation where the system 100 is trained to categorize input data into predefined classes or categories. The task type involves training models to recognize patterns and features in the input data that distinguish between different classes, enabling the system 100 to assign appropriate labels to new, previously unseen inputs. In some implementations, the classification task may be distributed across multiple network nodes that collectively learn to identify and categorize various phenomena based on their observations.
[0084] In accordance with an embodiment, a subtype of task is image classification. The image classification represents a specific subtype of classification task where the system 100 is trained to categorize visual data into predefined classes based on image content and features. The subtype involves specialized neural network architectures designed to process pixel data, extract meaningful visual features, and recognize patterns that distinguish between different image categories. In distributed INL environments, image classification may leverage multiple network nodes with cameras or image processing capabilities to improve classification accuracy collectively. In accordance with an embodiment, for the subtype of image classification, the properties include color or black and white, and 2 dimensional or 3 dimensional. For the subtype of image classification, defining properties such as color or black and white and 2-dimensional or 3-dimensional provides the technical advantage of enabling more precise transfer learning selection between related visual tasks. This granular property-level classification allows the system 100 to identify the most compatible source models with matching visual processing characteristics, resulting in more effective parameter initialization and significantly faster convergence when establishing new image-processing tasks across the network.
[0085] In accordance with an embodiment, the plurality of measurements includes the quality of implementation of an existing task. The quality of implementation metric evaluates several key performance indicators of the existing session, including accuracy and precision of the trained models, convergence rates achieved during training, stability of the learning process, and overall resource efficiency. Higher quality implementations demonstrate superior learning outcomes while utilizing network resources efficiently, making them more valuable candidates for transfer learning. The quality assessment incorporates multiple sub-measurements, including the number of agents supporting the task, which indicates the breadth of network participation and data diversity; the Relevance Indicator (RI) values, which quantify the relevance and reliability of each node's contribution; the quality of views from different data sources; and the computational capabilities of participating devices. By incorporating quality of implementation as a key measurement, the system 100 ensures that new INL HiC sessions leverage not just similar tasks but specifically those similar tasks that have achieved high-performance levels. The approach prevents the propagation of poorly optimized models and ensures that transfer learning builds upon successful implementations, thereby maximizing the efficiency and effectiveness of the new learning session.
[0086] In accordance with an embodiment, the quality is based on the number of agents supporting the existing task. The quality assessment based on the number of agents supporting the existing task provides insight into the breadth of network participation and data diversity within an active INL HiC session. The quality metric evaluates how extensively the task has been distributed across the network, with a higher agent count typically indicating more robust training through diverse data inputs and perspectives. Sessions with greater agent participation generally offer more comprehensive model training, making them more valuable candidates for transfer learning to new tasks. In accordance with an embodiment, the quality is based on the Relevance Indicator (RI) . The quality assessment based on the RI provides the technical advantage of enabling optimized session selection by quantifying the applicability of existing models to new tasks. The RI metric allows the system to automatically identify the most suitable candidate sessions for transfer learning, thereby significantly reducing computational overhead and convergence time when deploying new learning tasks across the network infrastructure.
[0087] In accordance with an embodiment, the quality is further based on the quality of views measured by the RI and based on a list of data modalities. The quality assessment incorporates the quality of views measured by the RI and is further refined by evaluating a list of data modalities available across the network. This comprehensive evaluation examines how effectively each participating node observes and processes different types of data (such as visual, audio, or sensor readings) relevant to the task, with the RI providing a quantitative measure of each view's contribution to learning outcomes. By considering both the relevance indicators and the diversity of data modalities, the system can identify sessions with rich, multi-perspective observations that provide more robust foundations for transfer learning, thereby enhancing model initialization when establishing new tasks.
[0088] In accordance with an embodiment, the quality is further based on the quality of the data available or observed at a plurality of children nodes in the communication network. The quality of the data available refers to the intrinsic characteristics and reliability of information collected at various network points, encompassing factors such as data freshness, completeness, signal-to-noise ratio, resolution, accuracy, and overall fidelity. Higher-quality data typically contains fewer errors, more relevant information, and better representations of the underlying phenomena being modelled. The plurality of children nodes represents network elements positioned at lower tiers in the architecture, such as end-user devices, sensors, or peripheral equipment that collect raw information directly from the environment. The plurality of children nodes function as data acquisition points distributed throughout the network topology, often serving as the primary sources of input for the learning process.
[0089] The selection process begins with the system 100 requesting information from a plurality of HiC agents located within respective nodes of the communications network. The HiC agents provide responses that include details about task similarity, quality of implementation, and compatibility of the existing session with the new task. The system 100 evaluates the similarity distance between the new task and the existing tasks using a predefined taxonomy that classifies tasks based on type, subtype, and properties. The system 100 then assesses the quality of implementation of the existing task based on factors such as the number of supporting agents, relevance indicators (RI) , and the quality of available data. Additionally, the system 100 evaluates the compatibility of the existing nodes for the new task, considering factors such as the computational capacity of the nodes and the ability to accommodate the new task requirements. After analysing these measurements, the system 100 selects the most suitable existing INL HiC session that maximizes learning efficiency and minimizes resource consumption. The selected session serves as the source for transferring patterns and parameters, ensuring that the new task benefits from prior learning experiences and established network structures.
[0090] The system 100 is further configured to gather information from the selected INL HiC session. Once the system 100 has selected the most suitable INL HiC session based on a plurality of measurements, it initiates a data retrieval process by establishing communication with the HiC Controller node associated with the selected session. The HiC Controller 110 sends a request to the HiC agents located within the nodes participating in the selected session. The request includes specific parameters related to the task requirements, such as the neural network model architecture, optimization method, hyperparameters, and communication paths. The HiC agents within the nodes of the selected session respond by providing detailed information about the model used in the existing session, including layer configurations, weight initializations, activation functions, and learning rates. The gathered information also includes details about the data modalities involved, such as image, audio, or text data, and the specific pre-processing steps applied during the existing task. Additionally, the system 100 collects information on the communication structure used in the selected session, including routing paths, latency, and error handling methods, which are essential for replicating the session's efficiency in the new task. The gathered information is aggregated and processed by the HiC Controller 110, which ensures that the data is organized and aligned with the requirements of the new task. This organized data is then transmitted to the nodes of the new INL HiC session, allowing them to initialize models and parameters based on the successful configurations of the selected session. This process ensures that the new task benefits from the pre-existing training and operational efficiency of the selected session, reducing training time and improving overall learning performance.
[0091] The system 100 is further configured to send the gathered information to a plurality of nodes in the new INL HiC session, to enable the nodes of the new INL HiC session to perform transfer learning. The gathered information is related to patterns and parameters related to the responses received from the HiC agents, where the patterns correspond to the active INL HiC session and include information concerning a neural network model, an optimisation method, related hyperparameters and a communication path. In accordance with an embodiment, the gathered information includes an HiC identifier of the selected one of the plurality of active INL HiC sessions for a node that is part of the selected session and also is part of the new session. The HiC identifier serves as a unique reference code that links the selected session with the new session, enabling the system 100 to maintain consistency and traceability during the transfer process. When the system 100 gathers information from the selected session, the HiC identifier ensures that the collected patterns and parameters are correctly associated with the source session. The HiC identifier allows the system 100 to identify specific nodes that are part of both the selected session and the new session, enabling direct mapping of communication paths and learning models. This ensures that the structural and functional integrity of the learning framework is preserved during the transfer. For instance, if a particular node was responsible for handling a specific type of data or executing a certain layer of the neural network in the selected session, the HiC identifier enables the system 100 to replicate this function within the new session. This direct mapping minimizes the need for reconfiguration and ensures that the nodes in the new session can quickly inherit the operational efficiency of the selected session. The HiC identifier also supports fault tolerance and scalability by allowing the system to dynamically update or replace nodes without disrupting the learning process. By establishing a direct link between the selected and new sessions through the HiC identifier, the system 100 enhances the accuracy and reliability of the transfer learning process, reduces setup time, and improves the overall adaptability of the new session to complex and evolving learning tasks.
[0092] In accordance with an embodiment, a pattern includes the set of initialization information that is derived from the existing task using only the HiC identifier. The HiC identifier acts as a unique reference that links the existing task with the new task, allowing the system 100 to retrieve and apply relevant initialization data without needing to manually configure the new session. The initialization information within the pattern includes the structural details of the neural network model used in the existing task, such as the number of layers, type of layers (e.g., convolutional, fully connected, recurrent) , activation functions, weight initializations, and dropout rates. It also includes hyperparameters such as learning rate, batch size, momentum, and regularization factors, which are critical for efficient model training and convergence. The HiC identifier enables the system 100 to directly access this initialization information from the existing task, eliminating the need for extensive configuration and manual adjustments.
[0093] Furthermore, the pattern includes details about the optimization method used in the existing session, such as stochastic gradient descent (SGD) , Adam, or RMSProp, and the corresponding hyperparameters. It also includes information about the communication path within the existing session, including the routing mechanisms, data exchange protocols, latency management strategies, and load balancing configurations. By using the HiC identifier, the system 100 can accurately replicate these parameters and network configurations in the new session, ensuring that the learning framework is consistent and well-optimized. The structured initialization process reduces the time and computational effort required to set up the new session, improves the accuracy and stability of the learning model, and accelerates the overall learning process. The ability to derive and apply initialization information using the HiC identifier ensures that the new session inherits the efficiency and performance of the existing task, leading to faster convergence, improved generalization, and enhanced adaptability to complex learning environments.
[0094] After gathering the necessary information from the selected INL HiC session, the HiC Controller 110 processes and organizes the data to ensure compatibility with the new task. The gathered information includes detailed patterns and parameters that define the structure and operational framework of the existing session. The patterns correspond to the architecture and configuration of the neural network model used in the selected session, including the number of layers, layer types, activation functions, weight initializations, and learning rates. Additionally, the gathered information includes the optimization method employed in the existing session, such as stochastic gradient descent (SGD) , Adam, or other optimization algorithms, along with the associated hyperparameters, such as learning rate, batch size, and momentum. The system 100 also collects data on the communication path used within the selected session, including the routing of data between nodes, latency management, error correction mechanisms, and load balancing strategies. This information ensures that the new session can replicate the successful communication architecture of the existing session.
[0095] Once the gathered information is processed and verified, the HiC Controller 110 transmits it to the nodes participating in the new INL HiC session. The transmission process is optimized to minimize latency and ensure data integrity. The nodes in the new session receive the information and use it to initialize their models and learning parameters. The initialization process involves setting the neural network structure, loading the pre-trained weights, and configuring the hyperparameters to align with the gathered data. The communication paths between the nodes are also established based on the patterns from the existing session, ensuring that data exchange and synchronization between nodes are efficient and stable. This structured transfer of knowledge allows the new session to bypass the need for extensive model training from scratch, reducing computational overhead and accelerating the learning process. By inheriting the successful patterns and parameters from the selected session, the new INL HiC session can quickly adapt to the new task, achieve faster convergence, and improve overall learning accuracy and performance.
[0096] In accordance with an embodiment, the transfer learning carried out by the new INL HiC session includes making a copy of the pattern used by the existing INL HiC session. In accordance with an embodiment, the transfer learning carried out by the new INL HiC session includes making a copy of the pattern used by the existing INL HiC session to enable efficient model initialization and faster learning convergence. The pattern represents the structural and functional framework of the neural network model used in the existing INL HiC session, including the network architecture, optimization method, hyperparameters, and communication path. When the new INL HiC session is established, the system 100 retrieves the pattern from the selected existing session using the HiC identifier and creates an identical copy of the pattern in the new session. The process allows the new session to bypass the need for building a model from scratch, reducing computational load and improving resource utilization.
[0097] The copied pattern includes detailed information about the neural network architecture, such as the number of layers, layer types (e.g., convolutional, recurrent, fully connected) , weight initializations, and activation functions. It also includes the optimization method applied in the existing session, such as stochastic gradient descent (SGD) , Adam, or RMSProp, along with the associated hyperparameters like learning rate, batch size, momentum, and regularization factors. The communication path defined in the pattern specifies how data is exchanged between nodes within the network, including routing protocols, latency handling mechanisms, and error correction methods. By making an exact copy of this pattern, the new session benefits from the operational efficiency and learning stability already established in the existing session. This ensures that the new session can achieve faster convergence and improved accuracy without the need for extensive retraining.
[0098] Furthermore, copying the pattern enables consistency in the learning framework across different sessions, ensuring that model updates and task-specific adjustments are handled uniformly. The uniform handling also facilitates scalability, as the copied pattern can be adjusted or modified to suit the specific requirements of the new task while retaining the foundational learning structure. The ability to copy a pattern from an existing session enhances the adaptability of the system 100, allowing it to respond to dynamic changes in task requirements and network conditions without compromising learning performance. Overall, this process reduces training time, optimizes resource allocation, and ensures high accuracy and reliability in AI-driven communication networks.
[0099] In accordance with an embodiment, the transfer learning carried out by the new INL HiC session includes initializing its models and hyperparameters using the parameters used by the existing INL HiC session and then using a transition loss function to make a transition from the initialized parameters used in the existing task to the parameters associated with the new task. The initialization process begins when the system 100 selects an existing INL HiC session based on a plurality of measurements, such as task similarity and implementation quality. After the selection, the system 100 retrieves the model architecture, optimization method, hyperparameters, and communication path used in the existing session. The retrieved parameters include the structure of the neural network, such as the number of layers, type of layers (e.g., convolutional, fully connected, recurrent) , weight initializations, and activation functions. The hyperparameters include learning rate, batch size, momentum, and regularization factors, which are essential for controlling the training process and ensuring model stability. These parameters are then used to initialize the model in the new INL HiC session, allowing it to start from a well-trained baseline rather than training from scratch. Once the model and hyperparameters are initialized, the system 100 applies a transition loss function to adapt the model to the new task. The transition loss function acts as a bridge between the existing and new task domains by gradually adjusting the model’s parameters to fit the new task requirements while preserving the foundational learning from the existing session. The transition loss function measures the divergence between the output of the initialized model and the expected output for the new task. During the training process, the transition loss function minimizes this divergence by adjusting the model weights and hyperparameters, allowing the model to adapt to the new task while retaining useful features learned from the existing task. The transition loss function also ensures that the adaptation process is stable, preventing catastrophic forgetting where the model loses previously learned information.
[0100] Additionally, the transition loss function enables a smooth and controlled transition, ensuring that the model retains the generalization ability of the existing session while fine-tuning itself to the specific characteristics of the new task. This approach significantly reduces training time and computational overhead, as the model does not need to learn from scratch but instead leverages the pre-trained knowledge from the existing session. The use of a transition loss function also improves model accuracy and convergence rate by allowing the model to inherit key features and patterns from the existing session while adjusting to the unique requirements of the new task. Overall, the process enhances learning efficiency, reduces resource consumption, and ensures that the new INL HiC session achieves high performance in handling complex and dynamic learning environments.
[0101] FIG. 2 is a flowchart depicting a method of updating an active In-Network Learning (INL) Heterarchical Intelligent Collaboration (HiC) session in a communications network, in accordance with an embodiment of the present disclosure. With reference to FIG. 2, there is shown a flowchart of a method 200 of updating an active In-Network Learning (INL) Heterarchical Intelligent Collaboration (HiC) session in a communications network. The method 200 includes steps 202 to 210.
[0102] At step 202, the method 200 includes for each of a plurality of active INL HiC sessions, requesting information from a plurality of HiC agents in each active INL HiC session, where an HiC agent is located within a respective node of the communications network. The HiC Controller 110 receives a request from an Application Function / Network Function (AF / NF) to establish a new INL HiC session for a specific task. The HiC Controller 110 then systematically identifies all active INL HiC sessions present within the network and transmits compatibility query messages to each HiC agent participating in these sessions. The HiC agents may be deployed in various network elements including user equipment (UE) devices, base stations (BS) , or other network functions. The queries specifically request information regarding the agent's computational capabilities (processing power, memory resources, energy status) , current task characteristics, performance metrics of existing models, data properties available at the node, and estimated relevance to the new task requirements. The comprehensive information-gathering process creates a foundation for subsequent assessment of potential transfer learning candidates.
[0103] At step 204, the method 200 includes receiving responses from HiC agents from each of the plurality of active INL HiC sessions. Following the transmission of compatibility queries, the HiC Controller 110 establishes secure communication channels to receive response data from all queried agents. The responses contain detailed information packages with multiple categories of metrics and parameters. The response data includes quantitative measures of computational resources available at each node, current model performance statistics (accuracy, precision, convergence rates) , data modality information, relevance indicators (RI) that quantify the applicability of the agent's current task to the new requirements, and potential availability for participating in the new session. The HiC Controller 110 implements robust error-handling mechanisms during response collection, including timeout management and retry protocols, to ensure comprehensive data collection even in challenging network conditions. Each received response is verified for completeness, processed, and stored in a structured database within the HiC Controller 110 to enable subsequent comparative analysis.
[0104] At step 206, the method 200 includes selecting one of the plurality of active INL HiC sessions based on a plurality of measurements. The selection process employs sophisticated algorithms to evaluate and rank each active session based on multiple criteria. The first measurement considers the similarity distance between existing tasks and the new task, utilizing a hierarchical taxonomy that categorizes tasks by type (e.g., classification, regression) , subtype (e.g., image classification, audio classification) , and specific properties (e.g., color / black-white, 2D / 3D) . The second measurement evaluates the quality of implementation of the existing task, considering the number of agents supporting the task, their relevance indicators, quality of views based on data modalities, and the quality of data available at children nodes. The third measurement assesses the compatibility of existing nodes with the requirements of the new task, including their capability to accommodate additional processing loads. The HiC controller 110 applies weighted scoring mechanisms to these measurements to identify the optimal existing session that provides the most advantageous foundation for transfer learning to the new task.
[0105] At step 208, the method 200 includes gathering information from the selected INL HiC session. Once the optimal candidate session has been identified, the HiC Controller 110 initiates a more detailed information-gathering process specifically targeting the selected session. This process collects comprehensive model architectures, trained parameters, optimization methodologies, hyperparameter configurations, and communication pathways established within the selected session. For nodes that are present in both the selected session and the intended new session, the gathered information may include an HiC identifier that enables direct access to existing models without requiring full parameter transmission. The gathered information also includes transition loss function parameters that will facilitate efficient adaptation from the existing task to the new task requirements. This detailed information collection provides all necessary components to effectively initialize the new session with pre-trained knowledge, significantly accelerating the learning process.
[0106] At step 210, the method 200 includes sending the gathered information to a plurality of nodes in the new INL HiC session, to enable the nodes of the new INL HiC session to perform transfer learning, where the gathered information is related to patterns and parameters related to the responses received from the HiC agents, where the patterns correspond to the active INL HiC session and include information concerning a neural network model, an optimisation method, related hyperparameters and a communication path. The HiC Controller 110 distributes the collected patterns and parameters to all nodes that will participate in the new session. The information is tailored according to each node's specific role and relationship to the selected session. For nodes that participated in the selected session and will also join the new session, the HiC Controller 110 may provide an HiC Controller 110 identifier that allows direct model access. For nodes that are exclusive to the new session, the Hic Controller 110 provides complete initialization patterns including neural network architectures, pre-trained weights, optimization methods, hyperparameters, and communication protocols. Additionally, all nodes receive transition loss function parameters that guide the fine-tuning process from the existing task to the new task requirements. This distribution of gathered information enables all participating nodes to initialize their learning processes with knowledge transferred from the selected session, thereby significantly reducing training time and computational resources compared to starting with randomly initialized models.
[0107] The steps 202 to 210 are only illustrative, and other alternatives can also be provided where one or more steps are added, one or more steps are removed, or one or more steps are provided in a different sequence without departing from the scope of the claims herein.
[0108] There is provided a computer program comprising instructions that, when executed by a computer system, cause the computer system to implement the method 200. In an example, the instructions are implemented on the computer-readable media, which include, but are not limited to, Electrically Erasable Programmable Read-Only Memory (EEPROM) , Random Access Memory (RAM) , Read-Only Memory (ROM) , Hard Disk Drive (HDD) , Flash memory, a Secure Digital (SD) card, Solid-State Drive (SSD) , a computer-readable storage medium, and / or CPU cache memory.
[0109] FIG. 3 is an exemplary diagram illustrating a phase for finding the target active HiC session, in accordance with an embodiment of the present disclosure. FIG. 3 is explained in conjunction with elements of FIGs. 1 to 2. With reference FIG. 3, there is shown an exemplary diagram 300 illustrating the communication flow for selecting an appropriate active HiC session for transfer learning. The exemplary diagram 300 includes a plurality of active HiC sessions, for example a first active HiC session 302A, a second active HiC session 302B, up to an Nth active HiC session 302N, where each session comprises their respective network components. The first active HiC session 302A includes a node 304A of an HiC Agent in UE device, a node 306A of a HiC Agent in LIF, and a node 308A of a HiC Agent in the FC 108. These components collectively operate according to defined HiC patterns for their specific learning task. The first active session includes a processing and communication unit 312A.
[0110] Similarly, the second active HiC session 302B includes a node 304B of an HiC Agent in UE device, a node 306B of a HiC Agent in LIF, and a node 308B of a HiC Agent in the FC 108 processing and communication unit 312B. Further, the Nth active HiC session 302N includes a node 304N of an HiC Agent in UE device, a node 306N of a HiC Agent in LIF, and a node 308N of a HiC Agent in the FC 108 and a processing and communication unit 312N. The exemplary diagram 300 further includes external control elements, specifically a node 310 representing an Application Function (AF) or Network Function (NF) that initiates requests for new learning tasks.
[0111] In operation, when the node 310 transmits a task request to the HiC Controller 110 via a path 320, the HiC Controller 110 then establishes communication paths to send competitive queries (for example, a path 314A to query the first active HiC session 302A, a path 314B to query the second active HiC session 302B and a path 314N to query the Nth active HiC session 302N) sequentially. The HiC Agents within active HiC each session respond with comprehensive information about their capabilities, current task characteristics, and potential relevance to the new task. For example, the first active HiC session 302A sends query response to HiC Controller 110 via a path 316A, the second active HiC session 302B sends query response to HiC Controller 110 via a path 316B and the Nth active HiC session 302N sends query response to HiC Controller 110 via a path 316N.
[0112] The HiC Controller 110 processes these responses through a plurality of monitoring devices, each monitoring device handles the task of one active HiC session individually, for example, a first monitoring device 318A handles the first active HiC session 302A, a second monitoring device 318B handles the second active HiC session 302B up to an Nth monitoring device 318N handling the Nth active HiC session 302N and evaluates them according to the plurality of measurements. Further, the plurality of monitoring devices also gets status updates of each active HiC session. For example, the status update (sent via a path 322A) of the first active HiC session 302A is monitored by the first monitoring device 318A, the status update (sent via a path 322B) of the second active HiC session 302B is monitored by the second monitoring device 318B and the status update (sent via a path 322N) of the Nth active HiC session 302N is monitored by the Nth monitoring device 318N.
[0113] The evaluation involves assessing task similarity based on taxonomic classification, implementation quality based on performance metrics, and compatibility of nodes with the new task requirements. Based on this comprehensive assessment, the HiC Controller 110 selects the optimal active session that offers the most advantageous foundation for transfer learning, enabling efficient establishment of the new INL HiC session with initialized parameters derived from existing trained models.
[0114] FIG. 4 is an exemplary diagram illustrating a phase for gathering information from the target active HiC session, in accordance with an embodiment of the present disclosure. FIG. 4 is explained in conjunction with elements of FIGs. 1 to 3. With reference FIG. 4, there is shown an exemplary diagram 400. The exemplary diagram 400 depicts the communication flow during the information gathering phase after a suitable active HiC session has been selected for transfer learning. The exemplary diagram 400 illustrates a chosen active HiC session 402 containing multiple HiC Agents deployed across different network elements, specifically an HiC Agent with node 404 in a User Equipment device 102, an HiC Agent with node 406 in a Local Intermediate Fusioner, and an HiC Agent with a node 408 in a Fusion Center. The HiC agents collectively operate according to established HiC patterns through a processing and communication unit 412. The HiC Controller 110 orchestrates the information-gathering process, while maintaining communication with the node 310 of the Application Function / Network Function that initiated the request for the new task.
[0115] The operational flow begins with a transmission line 416A carrying initial communication from the node 310 of AF / NF to the HiC Controller 110, which includes task specifications and confirmation to proceed with the information gathering phase for a selected active HiC session. Following this initial authorization, the HiC Controller 110 establishes a connection with the chosen active HiC session 402 following its selection as the optimal candidate for transfer learning. The HiC Controller 110 then initiates a series of targeted queries to each HiC Agent within the chosen active HiC session 402, requesting detailed model information including neural network architectures, trained parameters, optimization methodologies, hyperparameter configurations, and communication pathways. The series of targeted queries include a first query, a second query and a third query. The first query is sent to the HiC Agent with the node 404 in the UE device 102 via a path 420. The second query is sent to the HiC Agent with the node 406 in a Local Intermediate Fusioner via a path 422. The third query is sent to the HiC Agent with a node 408 in a Fusion Center via a path 424.
[0116] Each agent responds with comprehensive information packages containing their respective model specifications and operational parameters. For example, a first query response is sent from HiC Agent with node 404 in the UE device 102 via a path 426 to the HiC Controller 110. A second query response is sent from the HiC Agent with the node 406 in a Local Intermediate Fusioner via a path 428 to the HiC Controller 110. A third query response is sent from the HiC Agent with the node 408 in a Fusion Center via a path 430 to the HiC Controller 110.
[0117] Throughout the exchange, the HiC Controller 110 performs monitoring via a monitoring device 432 to ensure complete data collection and provides status reports to the node 310 of the AF / NF via a connection pathway 416B. The systematic information-gathering process obtains the detailed implementation specifications that will serve as the foundation for transfer learning, enabling the new INL HiC session to initialize with pre-trained knowledge from the existing session, thereby significantly reducing training time and computational resources compared to starting with randomly initialized models.
[0118] FIG. 5 is a diagram illustrating a phase involving the implementation of the new task on the destination, in accordance with an embodiment of the present disclosure. FIG. 5 is explained in conjunction with elements of FIGs. 1 to 4. With reference to FIG. 5, there is shown an exemplary diagram 500, a new INL HiC session 502 containing multiple HiC Agents deployed across different network elements, specifically a node 504 of an HiC Agent in the UE device 102, a node 506 of an HiC Agent in a Local Intermediate Fusioner, and a node 508 of an HiC Agent in a Fusion Center. The HiC agents will collectively implement the new learning task utilizing transfer learning from the previously selected session. Within the new INL HiC session 502, a processing and communication unit 514 and a handling unit 512 are present.
[0119] The exemplary diagram 500 further includes the HiC Controller 110 that orchestrates the implementation process and external entities, including node 310 of the AF / NF that initiated the request for the new task.
[0120] In operation, the process begins when the HiC Controller 110 establishes connections with the new INL HiC session 502 nodes. The HiC Controller 110 initiates the implementation by transferring learning instructions to each HiC Agent in the new INL HiC session 502. The transmission occurs through dedicated channels with the help of a unit 516 and includes all necessary patterns and parameters for transfer learning. For nodes that participated in the selected session and are also joining the new INL HiC session 502, the HiC Controller 110 may provide an HiC identifier through a connection pathway 518 that allows direct access to existing models. For nodes exclusive to the new INL HiC session 502, the HiC Controller 110 provides complete initialization patterns, including neural network architectures, pre-trained weights, optimization methods, and hyperparameters. Further a start command generated by a unit 522 is sent to the HiC Agents via a path 524 and a response generated by the HiC agents is sent back to the HiC Controller 110 via a path 520. The processing and communication unit 514 and the handling unit 512 help in the generation of the response.
[0121] The HiC Controller monitors the implementation process through monitoring functions with the help of a monitoring device 528 to ensure proper initialization and transition to the new task. A status update is sent to the HiC Controller 110 via path 526. The HiC Agents in the new INL HiC session 502 utilize the received information to initialize their models with parameters from the existing task, and then apply the transition loss functions to fine-tune these parameters specifically for the new task requirements. When the new INL HiC session 502 has to be ended the HiC Controller 110 send an end HiC signal via a path 530 and the new INL HiC session 502 comes to an end.
[0122] Modifications to embodiments of the present disclosure described in the foregoing are possible without departing from the scope of the present disclosure as defined by the accompanying claims. Expressions such as "including" , "comprising" , "incorporating" , "have" , "is" used to describe and claim the present disclosure are intended to be construed in a non-exclusive manner, namely allowing for items, components or elements not explicitly described also to be present. Reference to the singular is also to be construed to relate to the plural. The word "exemplary" is used herein to mean "serving as an example, instance or illustration" . Any embodiment described as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments or to exclude the incorporation of features from other embodiments. The word "optionally" is used herein to mean "is provided in some embodiments and not provided in other embodiments" . It is appreciated that certain features of the present disclosure, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable combination or as suitable in any other described embodiment of the disclosure.
Claims
1.A method (200) of establishing a new In-Network Learning (INL) Heterarchical Intelligent Collaboration (HiC) session in a communications network for a new task by leveraging transfer learning using an existing trained model associated with an existing task and an existing INL HiC session, the method (200) comprising steps of:for each of a plurality of active INL HiC sessions, requesting information from a plurality of HiC agents in each active INL HiC session, where an HiC agent is located within a respective node of the communications network; andreceiving responses from HiC agents from each of the plurality of active INL HiC sessions;selecting one of the plurality of active INL HiC sessions based on a plurality of measurements;gathering information from the selected INL HiC session; andsending the gathered information to a plurality of nodes in the new INL HiC session, to enable the nodes of the new INL HiC session to perform transfer learning, where the gathered information is related to patterns and parameters related to the responses received from the HiC agents, where the patterns correspond to the active INL HiC session and include information concerning a neural network model, an optimisation method, related hyperparameters and a communication path.2.The method (200) of claim 1, wherein one of the HiC agents is in a user equipment (UE) device (102) of the communications network.3.The method (200) of claim 1, wherein one of the HiC agents is in a base station (BS) device (104) of the communications network.4.The method (200) of claim 1, wherein the transfer learning carried out by the new INL HiC session includes making a copy of the pattern used by the existing INL HiC session.5.The method (200) of claim 1, wherein the transfer learning carried out by the new INL HiC session includes initializing its models and hyperparameters using the parameters used by the existing INL HiC session, and then using a transition loss function to make a transition from the initialized parameters used in the existing task to the parameters associated with the new task.6.The method (200) of claim 1, wherein the plurality of measurements includes a similarity distance between existing tasks and the new task.7.The method (200) of claim 6, wherein the similarity distance is based on a taxonomy.8.The method (200) of claim 7, wherein the taxonomy is based on type of task, subtype of task and properties of task.9.The method (200) of claim 8, wherein a type of task is a classification task.10.The method (200) of claim 8, wherein a subtype of task is image classification.11.The method (200) of claim 10, wherein, for the subtype of image classification, the properties include color or black and white, and 2 dimensional or 3 dimensional.12.The method (200) of claim 1, wherein the plurality of measurements includes the quality of implementation of an existing task.13.The method (200) of claim 12, wherein the quality is based on the number of agents supporting the existing task.14.The method (200) of claim 12, wherein the quality is based on the Relevance Indicator (RI) .15.The method (200) of claim 14, wherein the quality is further based on the quality of views measured by the RI and based on a list of data modalities.16.The method (200) of claim 12, wherein the quality is further based on the quality of the data available or observed at a plurality of children nodes in the communication network.17.The method (200) of claim 1, wherein the requesting and gathering steps are performed by an HiC Controller (HicC) node.18.The method (200) of claim 15, wherein the HicC controller node communicates with an Application Function / Network Function (AF / NF) node.19.The method (200) of claim 1, wherein the gathered information includes an HiC identifier of the selected one of the plurality of active INL HiC sessions, for a node which is part of the selected session and also which is part of the new session.20.The method (200) of claim 19, wherein a pattern includes the set of initialization information that is derived from the existing task using only the HiC identifier.21.A system (100) comprising means adapted for carrying out all the steps of the method (200) according to any preceding method claim.22.A computer program comprising instructions for carrying out all the steps of the method according to any preceding method claim, when said computer program is executed on a computer system.