NO349371B1 - Transparent Evolutionary Intelligence System - Google Patents

Description

[0001] 349371

[0003] 1

[0006] TRANSPARENT EVOLUTIONARY INTELLIGENCE SYSTEM

[0008] FIELD OF INVENTION

[0010] Aspects of the present disclosure relate to an artificial intelligence system. Specifically, but not exclusively, aspects of the present disclosure are directed to pruning and bridging 5 mechanisms of artificial intelligence systems. Specially, but not exclusively, aspects of the present disclosure are directed to a transparent evolutionary intelligence system using recursive feedback, pruning, and bridging mechanisms.

[0012] BACKGROUND

[0014] Artificial intelligence (AI) systems have seen significant advancements in recent years, 10 particularly with the development of neural networks for processing complex data. Traditional recurrent neural networks (RNNs) have been enhanced by attention-based encoder and decoder mechanisms, as described in publication number US10452978B2. These systems transduce input sequences into output sequences, enabling parallelisation that significantly reduces training and inference times while optimising computational resource usage. The 15 attention mechanism within these systems improves the network's ability to learn dependencies between distant positions, thereby increasing accuracy in tasks such as machine translation, speech recognition, natural language processing, medical diagnosis, and image processing.

[0016] Despite these advancements, attention-based systems face significant limitations. The 20 reliance on attention mechanisms results in high computational costs, particularly for long sequences, as the computation scales quadratically with the sequence length. Furthermore, these systems struggle to adapt to varying levels of input complexity across domains, leading to inefficiencies and reduced accuracy combined with increased computational inefficiency in complex tasks. Additionally, static recursive feedback mechanisms prevents systems from 25 refining outputs across layers, while the lack of efficient pruning techniques results in resource wastage during low activity periods or offline reconfiguration.

[0018] [0004] AI systems often struggle to adapt to new contexts and real-time processing requirements, which are important for applications such as autonomous driving, real-time language translation, and adaptive user interfaces. Current systems frequently experience 30 latency issues, misinterpretation of context, and inefficiencies when handling real-time data, undermining their effectiveness in time-critical applications. A related challenge is the ability of AI systems to generalise across multiple domains. Cross-domain learning, which involves transferring knowledge from one domain to another, enhances the system's versatility and robustness. However, current systems lack sophisticated mechanisms to address domain-35 specific variations, context dependencies, and the integration of diverse data sources. This results in inconsistent performance across domains and difficulties in achieving seamless knowledge transfer. Additionally, although intermediate and final outputs of neural networks are often stored for reference, these are rarely used by the system during operation, which

[0019] 349371

[0021] 2

[0023] limits the ability to evaluate patterns, detect anomalies, or identify opportunities for improvement, thereby restricting the potential for continuous optimisation.

[0025] Mechanisms for integrating outputs from multiple neural network modules or domains are similarly deficient. Current fusion techniques often fail to account for dynamic 5 interdependencies between domains. For example, when outputs from disparate modules must be combined for tasks such as multi-modal data analysis or cross-lingual translation, existing methods lack the flexibility to adjust modular weights based on real-time context or historical performance metrics. This leads to suboptimal synergy and reduced system accuracy in complex tasks requiring modular interactions. Bridging transformations, which map 10 outputs between different modules, traditionally rely on fixed mappings that cannot adapt to variations in domain characteristics or input dynamics. This rigidity is especially problematic for applications requiring dynamic alignment of representations, such as image-to-text conversion or multi-domain knowledge synthesis. Without adaptive refinement based on context or emotive signals, these systems struggle to achieve accurate and efficient integration 15 of diverse data.

[0027] Moreover, AI systems lack robust mechanisms for scaling context and emotive weighting parameters across domains. While emotional and contextual cues can enhance relevance and accuracy, current systems apply uniform weighting strategies that fail to account for evolving user needs or domain-specific requirements. This limitation impacts performance, 20 particularly in tasks involving human-computer interaction where sensitivity to context is critical. Predictive mechanisms for anticipating and pre-adjusting system parameters based on future inputs are either absent or rudimentary in current systems. Memory logs, though potentially valuable, are rarely utilised for simulating future scenarios or pre-adjusting parameters. This inability to prepare for novel inputs or dynamic changes in the operational 25 environment results in delays, inefficiencies, and a lack of adaptability, further constraining system reliability and performance.

[0030] US 2024/386015 A1 DISCLOSES A SYSTEM FOR OPTIMISING SEARCH PERFORMANCE AND EFFICIENCY

[0031] THROUGH INDEXING TECHNIQUES, DISTRIBUTED COMPUTING, AND CONTINUOUS LEARNING. WITH A

[0032] MODULAR ARCHITECTURE AND SCALABLE INFRASTRUCTURE, THE SEMANTIC SEARCH SYSTEM ENABLES

[0033] 30 USERS TO RETRIEVE RELEVANT, MEANINGFUL, AND CONTEXT-SPECIFIC INFORMATION FROM VAST

[0034] AMOUNTS OF STRUCTURED AND UNSTRUCTURED DATA. EP 4288906 A1 DISCLOSES METHOD FOR

[0035] PRUNING WEIGHTS OF AN ARTIFICIAL NEURAL NETWORK BASED ON A LEARNED THRESHOLD INCLUDES

[0036] DESIGNATING A GROUP OF PRE-TRAINED WEIGHTS OF AN ARTIFICIAL NEURAL NETWORK TO BE

[0037] EVALUATED FOR PRUNING. THE METHOD ALSO INCLUDES DETERMINING A NORM OF THE GROUP OF PRE-

[0038] 35 TRAINED WEIGHTS, AND PERFORMING A PROCESS BASED ON THE NORM TO DETERMINE WHETHER TO

[0039] PRUNE THE ENTIRE GROUP OF PRE-TRAINED WEIGHTS. WO 2023/159760 A1 DISCLOSES A METHOD AND

[0040] APPARATUS FOR PRUNING CONVOLUTIONAL NEURAL NETWORK MODELS INVOLVING CALCULATING

[0041] FILTER SIMILARITY AND PRUNING IMPORTANCE INDICES TO IDENTIFY AND REMOVE REDUNDANT FILTERS

[0042] BASED ON A PRESET PRUNING RATE. SUMMARY OF INVENTION

[0044] 40 [0007] According to an aspect of the present disclosure, there is provided a method of processing data using an artificial intelligence system. This system comprises several processing modules, each specific to a domain and containing a hierarchical neural network.

[0045] 349371

[0047] 3

[0049] The method involves receiving input data and processing it through the hierarchical neural network. The network has multiple layers that extract increasingly complex features, with each layer transforming its output based on the previous layers. The depth and progression of these layers can be adjusted according to the complexity of the domain. The method also comprises 5 refining the layer outputs using recursive feedback, which is applied to each layer's output.

[0050] During periods of low activity or offline reconfiguration, neural connections are pruned by calculating a pruning threshold for each layer and deactivating low-contributing weights. The layer outputs are modified by integrating transformed outputs from other processing modules through weighted bridging transformations. After generating outputs from domain-specific 10 modules in parallel, these outputs are weighed using relevance scores. If the system comprises context signals and emotive weighting, the bridging and modular weighting parameters are adjusted accordingly. The method integrates the layer outputs, recursive feedback, domainspecific modules, and bridging transformations into a probabilistic output function. A prediction is then selected based on the highest probable output of this function. Intermediate and final 15 outputs are stored in a memory log for purposes such as offline replay, hierarchical reconfiguration, layer tuning, or layer pruning. Finally, the prediction is outputted for one or more downstream actions.

[0052] According to an additional aspect of the present disclosure, the method further comprises generating sub-lane outputs for each domain of the input data by subdividing 20 domains into sub-lanes. Bridging transformations are applied to these sub-lane outputs. If the artificial intelligence system comprises emotive weighting, the bridging outputs are scaled using emotive weighting. These bridging outputs are then summed into a synergy vector, which is integrated into the probabilistic output function.

[0054] According to an additional aspect of the present disclosure, the artificial intelligence 25 system comprises emotive weighting, and this emotive weighting for scaling bridging outputs is dynamically adjusted based on historical false-positive rates or synergistic contribution patterns.

[0056] According to an additional aspect of the present disclosure, the method further comprises logging the impact of applied emotive weighting on domain outputs or performance 30 metrics to generate feedback. Future emotive weighting is adjusted, with a supervisory module evaluating performance feedback and overriding scaling parameters based on historical falsepositive rates or synergistic contribution patterns.

[0058] According to an additional aspect of the present disclosure, the method further comprises calculating the synergy vector by summing outputs from the domain-specific 35 modules or sub-lanes and scaling the outputs based on domain-specific emotive weighting.

[0059] The global state is recursively updated using a recursion factor. The persistence of prior context is evaluated by monitoring the norm of the global state and adjusting the recursion factor during offline processes or periods of reduced activity.

[0061] [0012] According to an additional aspect of the present disclosure, the method further 40 comprises encoding sequential data from multiple time steps into the memory log. Each time step is associated with a step-wide record that comprises time, domain input, emotive weighting synergy vector, global state, and one or more downstream actions. An offline replay of the memory log is conducted during reduced activity phases, where synergy patterns are

[0062] 349371

[0065] 4

[0067] analysed across consecutive steps to identify multi-step anomalies. Emotive weighting and bridging transformations are adjusted based on cumulative signals or repeated patterns. A novelty value is evaluated during the offline replay, and recommendations are generated for creating new domain sub-lanes, adjusting emotive baselines, or refining bridging transforms if 5 the novelty exceeds a threshold. Recursive feedback, pruning thresholds, or domain-lane hierarchy are refined during subsequent system operations based on the offline replay.

[0069] According to an additional aspect of the present disclosure, the method further comprises integrating emotive weighting, synergy vectors, and global states across multiple domains to detect cross-domain patterns using an alignment mechanism for refining bridging 10 transformations. During the offline replay, anomalous sub-lanes are identified by analysing synergy deviations. Novelty values are re-evaluated during the offline replay based on sub-lane contributions and cross-domain interactions. Sub-lanes are expanded or merged when repeated patterns across time steps satisfy preset conditions. A future domain input is simulated based on the memory log, and the predicted future domain input is used to pre-15 adjust emotive weighting and bridging parameters. Modular relevance scores are enhanced during system operation by incorporating historical novelty trends based on normalized novelty values over a defined historical window.

[0071] According to an additional aspect of the present disclosure, the method further comprises analysing the memory log to identify patterns. A new sub-lane is created if the 20 patterns are not captured by existing sub-lanes, or sub-lanes within a domain are unified if the patterns indicate redundancy by combining bridging transforms.

[0073] According to an additional aspect of the present disclosure, the method further comprises analysing the memory log to identify overactive emotive weighting during system operation, where synergy vectors are excessively impacted by the emotive weighting. The 25 overactive emotive weighting is restricted by providing a minimum or maximum value.

[0075] According to an additional aspect of the present disclosure, the method further comprises analysing the memory log to detect cause-effect relationships across domain lanes. Detected cause-effect relationships are stored as symbolic rules in a cause-effect repository. These relationships are used to anticipate and modify synergy vectors, and a symbolic 30 explanation is generated for modifying the synergy vectors for explainability.

[0077] According to an additional aspect of the present disclosure, the method further comprises using the cause-effect repository during meta-updates. Proactive adjustments are applied if a confidence value associated with a cause-effect relationship is greater than a confidence threshold. These proactive adjustments include adding a pre-emptive emotive 35 weighting offset for a domain-lane or scaling a bridging transformation matrix.

[0079] According to an additional aspect of the present disclosure, each domain sub-lane is evaluated for pruning based on usage metrics. Sub-lanes are pruned if both metrics fall below respective thresholds.

[0081] [0019] According to an additional aspect of the present disclosure, the method further 40 comprises, during periods of offline reconfiguration or low activity, replaying the memory log to analyse multi-step patterns of synergy and emotive signals. Repeated anomalies or patterns

[0082] 349371

[0085] 5

[0087] missed by single-step evaluation are identified. Recommendations are generated for structural adjustments to bridging transformation matrices or emotive weightings.

[0089] Beneficially, the methods enable enhanced adaptability and efficiency in processing complex, domain-specific data through hierarchical feature extraction, dynamic layer 5 adjustments, recursive feedback, efficient pruning mechanisms, and the integration of emotive weighting and context signals. This results in improved accuracy and performance in varied applications, addressing the limitations of traditional attention-based neural networks. For example, the described system and method significantly reduces computational inefficiency in processing large datasets for medical imaging analysis and, when applied to cancer detection, 10 achieves a substantial reduction in computational requirements while maintaining diagnostic accuracy of over 99%, addressing the technical challenge of real-time medical analysis.

[0091] The skilled person will understand that any above-described apparatus, process, system, and method is not limited to hierarchical abstraction and may be applied to alternative contexts and usage scenarios. For example, he described system could be adapted for flat 15 feature representations in simpler domains, graph-based data structures for relational insights, or modular integration for domain-specific applications across industries such as healthcare diagnostics, control system risk assessment, or energy management.

[0093] BRIEF DESCRIPTION OF DRAWINGS

[0095] Embodiments of the invention will now be described, by way of example only, and with 20 reference to the accompanying drawings, in which:

[0097] Figure 1 illustrates a system architecture diagram for a transparent evolutionary intelligence system;

[0099] Figure 2 illustrates a flowchart of a method for processing data in a multi-domain artificial intelligence system;

[0101] 25 [0025] Figure 3 illustrates a flowchart of a method for pruning an artificial intelligence system;

[0103] Figure 4 illustrates a flowchart of a method for performing bridging in a multi-domain artificial intelligence system; and

[0105] Figure 5 shows an example computing environment for performing any of the methods described herein.

[0107] 30 [0028] All illustrations of the drawings are for the purpose of describing selected versions of the present invention and are not intended to limit the scope of the present invention.

[0109] [0029] As a preliminary matter, it will readily be understood by one having ordinary skill in the relevant art that the present disclosure has broad utility and application. As should be understood, any embodiment may incorporate only one or a plurality of the above-disclosed 35 aspects of the disclosure and may further incorporate only one or a plurality of the abovedisclosed features. Furthermore, any embodiment discussed and identified as being “preferred” is considered to be part of a best mode contemplated for carrying out the embodiments of the present disclosure.

[0110] 349371

[0112] 6

[0114] Other embodiments also may be discussed for additional illustrative purposes in providing a full and enabling disclosure. Moreover, many embodiments, such as adaptations, variations, modifications, and equivalent arrangements, will be implicitly disclosed by the embodiments described herein and fall within the scope of the present disclosure. Accordingly, 5 while embodiments are described herein in detail in relation to one or more embodiments, it is to be understood that this disclosure is illustrative and exemplary of the present disclosure, and are made merely for the purposes of providing a full and enabling disclosure.

[0116] The detailed disclosure herein of one or more embodiments is not intended, nor is to be construed, to limit the scope of patent protection afforded in any claim of a patent issuing 10 here from, which scope is to be defined by the claims and the equivalents thereof. It is not intended that the scope of patent protection be defined by reading into any claim a limitation found herein that does not explicitly appear in the claim itself. Additionally, it is important to note that each term used herein refers to that which an ordinary artisan would understand such term to mean based on the contextual use of such term herein. To the extent that the meaning 15 of a term used herein—as understood by the ordinary artisan based on the contextual use of such term—differs in any way from any particular dictionary definition of such term, it is intended that the meaning of the term as understood by the ordinary artisan should prevail. Furthermore, it is important to note that, as used herein, “a” and “an” each generally denotes “at least one,” but does not exclude a plurality unless the contextual use dictates otherwise.

[0117] 20 When used herein to join a list of items, “or” denotes “at least one of the items,” but does not exclude a plurality of items of the list. Finally, when used herein to join a list of items, “and” denotes “all of the items of the list.”

[0119] The following detailed description refers to the accompanying drawings. Wherever possible, the same reference numbers are used in the drawings and the following description 25 to refer to the same or similar elements. While many embodiments of the disclosure may be described, modifications, adaptations, and other implementations are possible. For example, substitutions, additions, or modifications may be made to the elements illustrated in the drawings, and the methods described herein may be modified by substituting, reordering, or adding stages to the disclosed methods. Accordingly, the following detailed description does 30 not limit the disclosure. Instead, the proper scope of the disclosure is defined by the appended claims.

[0121] The present disclosure contains headers. It should be understood that these headers are used as references and are not to be construed as limiting upon the subjected matter disclosed under the header. Other technical advantages may become readily apparent to one 35 of ordinary skill in the art after review of the following figures and description. It should be understood at the outset that, although exemplary embodiments are illustrated in the figures and described below, the principles of the present disclosure may be implemented using any number of techniques, whether currently known or not. The present disclosure should in no way be limited to the exemplary implementations and techniques illustrated in the drawings 40 and described below.

[0123] [0034] Unless otherwise indicated, the drawings are intended to be read together with the specification, and are to be considered a portion of the entire written description of this invention. As used in the following description, the terms “horizontal”, “vertical”, “left”, “right”,

[0124] 349371

[0126] 7

[0128] “up”, “down” and the like, as well as adjectival and adverbial derivatives thereof (e.g., “horizontally”, “rightwardly”, “upwardly”, “radially”, etc.), simply refer to the orientation of the illustrated structure as the particular drawing figure faces the reader. Similarly, the terms “inwardly,” “outwardly” and “radially” generally refer to the orientation of a surface relative to 5 its axis of elongation, or axis of rotation, as appropriate.

[0130] The present disclosure comprises many aspects and features. Moreover, while many aspects and features relate to, and are described in the context of a multi-domain artificial intelligence system, embodiments of the present disclosure are not limited to use only in this context. In the context of the present invention any systems, methods or processes disclosed 10 herein comprise an at least one processing unit whereby said at least one processing unit performs the process of the present invention.

[0132] DETAILED DESCRIPTION

[0134] Embodiments of the present disclosure will now be described with reference to the attached figures. It is to be noted that the following description is merely used for enabling the 15 skilled person to understand the present disclosure, without any intention to limit the applicability of the present disclosure to other embodiments which could be readily understood and/or envisaged by the reader. In particular, whilst the present disclosure is primarily directed to artificial intelligence systems, the skilled person will appreciate that the apparatus, processes, methods and systems described herein are applicable to computational systems 20 and hardware processing systems more broadly.

[0136] The present disclosure relates to an advanced artificial intelligence (AI) system designed to process complex, domain-specific data with enhanced adaptability and efficiency. This system leverages a plurality of processing modules, each comprising a hierarchical neural network, to achieve superior performance in various applications. The disclosure addresses 25 and overcomes several limitations of traditional attention-based neural networks, providing significant improvements in hierarchical feature extraction, dynamic layer adjustments, recursive feedback, efficient pruning mechanisms, and the integration of emotive weighting and context signals. Existing neural network systems face challenges in adapting weighting parameters dynamically for multi-modal data. The proposed mechanisms provide an adaptive 30 scaling system based on historical and contextual signals, resulting in a significant improvement in cross-domain prediction accuracy.

[0138] [0038] The framework described in the disclosure herein represents a sophisticated multidomain artificial intelligence system, specifically designed to process complex data through layered abstraction and integration mechanisms. The architecture consists of hierarchical 35 layers within domain-specific pathways, referred to as domain-lanes, that progressively refine input data into abstract representations. Each layer performs transformations via matrix operations and non-linear functions, facilitating tasks like noise reduction and feature extraction in lower layers, while higher layers produce advanced feature vectors. These final vectors are utilized by a bridging mechanism, which combines outputs across domains into a unified 40 synergy vector. The synergy vector incorporates emotive weighting to prioritize significant

[0139] 349371

[0141] 8

[0143] domain-lane outputs based on urgency or importance, ensuring a coherent integration of multidomain inputs.

[0145] Key to the framework is the adaptability of its hierarchical abstraction and bridging processes. The system dynamically adjusts its structure based on feedback from a centralised 5 meta-control module termed the Super Prefrontal Cortex (Super PFC). This component evaluates performance metrics, user-defined goals, and logs from the Hippocampus-like Memory (HPC), which stores sequences of prior system states, domain inputs, and actions. Offline processes, resembling a biological sleep phase, replay stored sequences to identify patterns, refine parameters, and prune underutilized pathways. This ensures efficient operation 10 and the creation or adjustment of domain-lanes to address evolving data patterns, enabling the system to adapt to previously unseen scenarios while maintaining optimal processing capacity.

[0147] The integration of causal reasoning into the HPC replay process further enhances the system's capabilities. This reasoning layer identifies cause-and-effect relationships across domain-lanes, formulating symbolic rules that inform real-time adjustments to bridging 15 transforms and emotive weighting. These insights improve the system's ability to anticipate and respond to multi-step data anomalies, reducing false positives and enhancing response accuracy. The recursive state update mechanism ensures that the system retains historical context without requiring extensive recurrent layers, allowing real-time adaptability with minimal computational overhead. By combining biologically inspired principles with advanced 20 machine learning techniques, the framework achieves robust, modular, and interpretable multidomain intelligence, making it highly applicable to scenarios requiring real-time decisionmaking and predictive analytics.

[0149] The AI system comprises multiple processing modules, each equipped with a hierarchical neural network. These networks are structured with a plurality of layers, each 25 configured to extract features of increasing complexity from the input data. The hierarchical nature of the neural networks ensures that each layer transforms its output based on the outputs of previous layers, facilitating dimensional transformations across layers from the input data to a final abstract representation. This hierarchical feature extraction is dynamically adjustable, allowing the system to adapt layer depth and dimensional progression based on 30 the complexity of the specific domain being processed.

[0151] Beneficially, the system incorporates recursive feedback mechanisms. These mechanisms refine the outputs of each layer by applying recursive feedback, enhancing the accuracy and robustness of the feature extraction process. Additionally, the system comprises efficient pruning mechanisms that operate during offline reconfiguration or periods of reduced 35 layer activity. Pruning involves calculating a pruning threshold for each layer and selectively deactivating weights that contribute minimally, thereby optimizing computational resources and maintaining system efficiency.

[0153] [0043] The system further enhances its adaptability using bridging transformations, which modify layer outputs by integrating transformed outputs from other processing modules. This 40 modular integration is dynamically adjusted based on context signals and emotive weighting, allowing the system to fine-tune its processing parameters in real-time. The modular outputs from domain-specific modules are generated in parallel and weighted using relevance scores, ensuring that the most pertinent information is prioritized.

[0154] 349371

[0156] 9

[0158] The integration of layer outputs, recursive feedback, domain-specific modules, and bridging transformations culminates in a probabilistic output function. This function generates predictions based on the highest probable output, ensuring accurate and reliable results. Intermediate and final outputs are stored in a memory log, which can be used for offline replay, 5 hierarchical reconfiguration, layer tuning, or layer pruning. This memory log facilitates continuous improvement and adaptation of the system.

[0160] A beneficial aspect of the disclosure is the integration of emotive weighting and context signals. These elements dynamically adjust bridging and modular weighting parameters, enhancing the system's responsiveness to varying application demands. Emotive weighting is 10 scaled using nonlinear transformations based on domain-specific urgency signals, and its impact on domain outputs or performance metrics is logged for generating feedback. This feedback is used to adjust future emotive weighting, with a supervisory module evaluating performance feedback and overriding scaling parameters based on historical data.

[0162] The system is capable of generating sub-lane outputs for each domain of the input data 15 by subdividing domains into sub-lanes. Bridging transformations are applied to these sub-lane outputs, which are then scaled using emotive weighting and summed into a synergy vector. This synergy vector is integrated into the probabilistic output function, further enhancing the accuracy and relevance of the system's predictions.

[0164] The invention comprises mechanisms for recursively updating a global state using a 20 recursion factor, with the persistence of prior context monitored by evaluating the norm of the global state. During offline processes or periods of reduced activity, the recursion factor is adjusted to maintain optimal performance. Sequential data from multiple time steps is encoded into the memory log, with each time step associated with a comprehensive record of inputs, emotive weighting, global state, and downstream actions.

[0166] 25 [0048] During offline replay, the system analyses synergy patterns across consecutive steps to identify multi-step anomalies. Emotive weighting and bridging transformations are adjusted based on cumulative signals or repeated patterns, with a novelty value evaluated to generate recommendations for creating new domain sub-lanes, adjusting emotive baselines, or refining bridging transforms. Recursive feedback, pruning thresholds, and domain-lane hierarchy are 30 refined based on the offline replay, ensuring continuous system improvement.

[0168] The system integrates emotive weighting, synergy vectors, and global states across multiple domains to detect cross-domain patterns using an alignment mechanism. This mechanism refines bridging transformations and identifies anomalous sub-lanes by analysing synergy deviations during offline replay, ensuring robust and accurate performance across 35 diverse applications.

[0170] In summary, the present disclosure provides a highly adaptable and efficient AI system capable of processing complex, domain-specific data with enhanced accuracy and performance. The integration of hierarchical feature extraction, dynamic layer adjustments, recursive feedback, efficient pruning mechanisms, and emotive weighting and context signals 40 represents a significant advancement over traditional attention-based neural networks.

[0171] Specific aspects of the disclosure are described below in relation to Figures 1 to 4.

[0172] 349371

[0175] 10

[0177] Figure 1 illustrates a system architecture diagram of system 100, which is directed to a modular artificial intelligence framework designed for multi-domain data processing, hierarchical feature extraction, and adaptive decision-making. System 100 integrates domainspecific processing with recursive feedback and probabilistic reasoning to generate 5 predictions for downstream applications.

[0179] System 100 comprises an input interface 102 configured to receive input data; a plurality of processing modules 104, each comprising hierarchical neural networks with a dynamically adjustable number of layers and dimensional transformations tailored to domainspecific complexities; a feedback mechanism 106 that refines layer outputs through recursive 10 feedback; a pruning mechanism 108 for selectively deactivating low-contributing neural connections during offline reconfiguration or periods of reduced activity; a bridging unit 110 that integrates transformed outputs from other processing modules via bridging transformations; a weighting control unit 112 for dynamically adjusting modular and bridging weights based on context signals and emotive weighting; a probabilistic function unit 114 that 15 consolidates the outputs, recursive feedback, domain-specific modules, and bridging transformations into a probabilistic output; selection logic 116 for selecting the prediction with the highest probability; memory units 118 that store intermediate and final outputs for purposes such as offline replay, hierarchical reconfiguration, layer tuning, and pruning; and an output interface 120 configured to output the prediction for downstream actions.

[0181] 20 [0053] System 100 operates by first receiving raw input data via the input interface 102, which structures the data for processing by the domain-specific processing modules 104. Each processing module applies hierarchical feature extraction, transforming the input data across layers of increasing abstraction. The feedback mechanism 106 refines these layers iteratively, while the pruning mechanism 108 optimises neural connections by deactivating those with low 25 contributions. The bridging unit 110 integrates outputs across processing modules to ensure coherent cross-domain synthesis, with the weighting control unit 112 dynamically modifying the importance of specific outputs based on contextual signals or system priorities. The probabilistic function unit 114 combines all processed outputs and transformations into a probabilistic representation, from which the selection logic 116 determines the most probable 30 prediction. The memory units 118 store outputs and intermediate results for further refinement or analysis, and the output interface 120 delivers the prediction for actionable use in downstream systems, enabling adaptive, efficient, and context-aware processing.

[0183] The input interface 102 serves as the entry point for raw input data into system 100 and is responsible for structuring this data for downstream processing. This component is 35 configured to receive data from multiple domains, each with unique characteristics, such as sensor readings, financial metrics, or textual information. By formatting these diverse data types into domain-specific vectors, the input interface ensures that the raw input data can be effectively utilised by the subsequent domain-specific processing modules 104.

[0185] [0055] The input interface 102 handles the allocation of data to appropriate domain-lanes, 40 preserving the distinctiveness of each data source. This modular approach prevents crossdomain conflation at the initial stage and allows for specialised handling of each domain’s data. For example, complex financial signals might require specific preprocessing to extract timeseries features, while textual data might need initial tokenisation or embedding preparation.

[0186] 349371

[0189] 11

[0191] The input interface enables seamless integration of such diverse data into the hierarchical neural networks within the system, ensuring the subsequent layers can focus on progressive feature extraction and transformation without being impeded by data heterogeneity.

[0193] By acting as a bridge between raw external inputs and the highly modular, domain-5 specific architecture of system 100, the input interface 102 supports efficient, domainappropriate preprocessing. This ensures optimal initial conditions for the hierarchical transformation processes that follow, thereby enhancing the system’s ability to extract meaningful features and maintain adaptive, scalable operation across varied data inputs.

[0195] An example of the input interface 102 in use is when system 100 is applied to cancer 10 detection. In this application, the input interface 102 is configured to receive diverse types of medical data, such as imaging scans (e.g., X-rays or MRIs), genomic sequences, and patient clinical records. Each type of data represents a unique domain with distinct characteristics that must be preserved for effective processing. For example, imaging data might be processed as high-dimensional pixel arrays, genomic sequences could be represented as nucleotide 15 patterns or encoded vectors, and clinical records might be parsed into numerical indicators (e.g., blood test results) or categorical features (e.g., family medical history). The input interface 102 organises this data into corresponding domain-lane vectors, ensuring that imaging data is routed to processing modules optimised for spatial feature extraction, genomic data to modules designed for sequence analysis, and clinical records to modules specialised in tabular or 20 categorical data processing. This initial organisation and formatting by the input interface 102 ensures that system 100 is equipped to handle the complex and multimodal nature of cancerrelated data. It enables downstream modules to perform domain-specific feature extraction while maintaining the integrity and relevance of each data type, a foundational step in the accurate prediction and diagnosis of cancer that leverages the system’s full modular and 25 hierarchical processing capabilities.

[0197] The processing modules 104 in system 100 are configured as a plurality of domainspecific hierarchical neural networks, each optimised to handle distinct types of data received from the input interface 102. These modules, labelled individually as 104A, 104B, 104C, and so forth, are specialised for extracting features of increasing complexity from their respective 30 data domains. Each processing module comprises a series of interconnected layers, where each layer applies transformations to its input, producing progressively abstract representations that highlight salient features within the data.

[0199] The hierarchical structure within each module 104 allows for dimensional transformations across layers, adapting the raw input into forms that are increasingly tailored 35 to specific analysis goals. For example, in the cancer detection example, a module processing imaging data (e.g., 104A) may perform convolutional operations to detect patterns indicative of tumours, such as irregular shapes or dense masses. A module handling genomic data (e.g., 104B) might use sequence-based processing layers to identify mutations or gene expression anomalies. Each module dynamically adjusts the depth of its layers and the dimensions of its 40 transformations based on the complexity of the domain it is analysing, enabling a balance between computational efficiency and representational power.

[0201] [0060] The modular outputs generated by each processing module 104 are domain-specific and weighted using relevance scores that reflect the significance of the extracted features in

[0202] 349371

[0204] 12

[0206] the context of the broader system goals. These outputs serve as intermediate results, feeding into the bridging unit 110 for cross-domain integration and further refinement. By structuring processing modules as independent yet interconnected units, system 100 ensures that domain-specific features are robustly extracted and preserved, supporting adaptive, context-5 aware decision-making across complex, multimodal data inputs.

[0208] The processing modules 104 in system 100 are highly versatile, capable of encompassing a wide array of architectures tailored to the specific requirements of each domain. These architectures can include single convolutional neural networks (CNNs) for spatial data like imaging, recurrent neural networks (RNNs) or transformers for sequential data 10 such as genomic sequences or time-series signals, and even ensemble models that combine multiple approaches for enhanced predictive power. For example, in the cancer detection scenario, a processing module 104A may use a CNN with residual connections to extract hierarchical features from imaging data, focusing on tumour localisation and shape analysis. Another module, 104B, might employ a transformer-based model to process long genomic 15 sequences, identifying potential mutations or anomalies in genetic markers. Meanwhile, 104C could implement a gradient boosting machine ensemble to analyse structured clinical records, aggregating tabular and categorical features into actionable insights. Each processing module’s architecture is dynamic and can be reconfigured based on the complexity of its assigned domain, ensuring scalability and adaptability. This breadth of scope allows the 20 modules to function independently and optimise their outputs before integration, enabling system 100 to comprehensively address challenges posed by multimodal and domain-diverse datasets.

[0210] Processing modules 104 preferably comprise sub-lanes, representing specialised subdivisions within each module, designed to handle distinct aspects or subcategories of the 25 domain-specific data. Each processing module may include one or more sub-lanes, with each sub-lane optimised for extracting features or analysing patterns from a specific subset of the domain’s data. This structure enhances modularity and ensures that different facets of the input data are processed independently but cohesively within the overarching hierarchical framework.

[0212] 30 [0063] For example, in the cancer detection scenario, a processing module 104A handling imaging data might include sub-lanes dedicated to different imaging modalities or specific tasks. One sub-lane could process X-ray data to detect anomalies in bone structures, while another focuses on MRI scans to identify soft tissue irregularities. Each sub-lane would operate using a tailored configuration of neural network layers, such as convolutional layers optimised 35 for the spatial features of its respective imaging modality.

[0214] Similarly, in a genomic data processing module 104B, sub-lanes might specialise in analysing different aspects of the genome. One sub-lane could focus on detecting single nucleotide polymorphisms (SNPs), while another identifies structural variations or gene expression profiles. These sub-lanes allow the module to address the diverse complexities 40 within the genomic dataset without conflating or oversimplifying the underlying data.

[0216] [0065] Sub-lanes within a processing module operate in parallel, ensuring that all relevant data subsets are processed simultaneously for efficiency. The outputs from these sub-lanes are aggregated within the module to form a cohesive modular output, which is then forwarded to

[0217] 349371

[0220] 13

[0222] the bridging unit 110 for cross-domain integration. Additionally, sub-lanes can be dynamically adjusted or reconfigured during offline replay, supported by the memory units 118, to align with evolving domain-specific requirements or to address new data patterns.

[0224] The use of sub-lanes in processing modules 104 enables system 100 to achieve a high 5 degree of specialisation and adaptability, ensuring that each domain’s data is processed in a manner that captures its unique characteristics. This modular approach not only improves the accuracy and relevance of the extracted features but also facilitates efficient integration across domains in subsequent stages of the system.

[0226] The feedback mechanism 106 in system 100 is designed to enhance the accuracy and 10 adaptability of the hierarchical neural networks within the processing modules 104 by applying recursive feedback loops. This mechanism ensures that the output of each layer is continuously refined based on both intermediate results and the overall system objectives. The feedback mechanism monitors the performance of each layer in real-time and iteratively adjusts the layer outputs to optimise feature extraction and representation.

[0228] 15 [0068] In operation, the feedback mechanism 106 analyses discrepancies between predicted outputs and target outcomes or system-defined objectives, generating error signals or refinement cues. These signals are propagated backward through the layers of the processing modules, allowing each layer to recalibrate its weights and biases. For example, in the cancer detection example, if a processing module handling imaging data (e.g., 104A) incorrectly 20 identifies a region of interest as non-tumorous, the feedback mechanism would generate a correction signal. This signal propagates through the network, prompting adjustments in convolutional filters to better detect subtle indicators of malignancy in future iterations.

[0230] The recursive nature of feedback allows for real-time and iterative learning, reducing the reliance on static training datasets and enabling the system to adapt dynamically to evolving 25 data patterns. The mechanism can also integrate contextual signals from other components, such as the bridging unit 110 or the weighting control unit 112, to ensure that refinements align with broader system priorities. By continuously fine-tuning layer outputs, the feedback mechanism 106 enhances the precision, reliability, and adaptability of system 100 across a wide range of applications and data modalities.

[0232] 30 [0070] The pruning mechanism 108 in system 100 is a dynamic optimisation component configured to improve efficiency and performance by selectively deactivating low-contributing neural connections within the processing modules 104. This mechanism operates during offline reconfiguration phases or periods of reduced system activity, ensuring that the system remains computationally efficient without compromising its predictive capabilities.

[0234] 35 [0071] Pruning involves the calculation of a pruning threshold for each layer, which is based on metrics such as the contribution of individual weights to the overall network performance, their activity level, or their impact on modular outputs. Connections or weights that fall below this threshold are identified as low-contributing and are selectively deactivated. For example, in the cancer detection application, a processing module 104A analysing imaging data might 40 prune filters in convolutional layers that are found to contribute negligibly to tumour detection, thereby reducing computational overhead while retaining critical feature extraction capabilities.

[0235] 349371

[0238] 14

[0240] The pruning mechanism 108 also leverages insights from the memory units 118, which store historical usage and performance data, to make informed decisions about which connections or layers are underutilised. This data-driven approach ensures that pruning is precise and does not degrade the model's ability to process domain-specific data effectively.

[0241] 5 By continuously refining the architecture in response to performance metrics, the pruning mechanism enables system 100 to maintain high levels of adaptability and scalability, particularly in applications requiring extensive computational resources or real-time responsiveness.

[0243] The bridging unit 110 in system 100 serves as a core integration mechanism, facilitating 10 the synthesis of modular outputs from the domain-specific processing modules 104 into a unified, cross-domain representation. This unit applies bridging transformations to the outputs of each module, ensuring that the diverse features extracted from different domains are coherently combined while preserving their individual significance and relationships.

[0245] The bridging unit 110 uses specialised transformation matrices or functions to align the 15 dimensions and semantics of modular outputs, allowing for seamless integration. For example, in the context of cancer detection, the bridging unit might transform the abstract representations generated by a CNN-based module (104A) processing imaging data, a transformer-based module (104B) analysing genomic sequences, and an ensemble module (104C) handling clinical records into a single synergy vector. This vector represents a holistic 20 view of the patient’s data, capturing spatial, sequential, and structured information in a unified form suitable for downstream analysis.

[0247] Additionally, the bridging unit 110 is dynamic, allowing adjustments to the transformations based on contextual signals and system priorities provided by the weighting control unit 112. This adaptability ensures that the integration process remains robust across 25 varying data distributions and task requirements. By enabling cross-domain collaboration and synergy, the bridging unit 110 enhances the overall predictive and analytical power of system 100, making it capable of addressing complex, multimodal problems with a unified approach.

[0249] The weighting control unit 112 in system 100 is responsible for dynamically adjusting the importance of modular outputs and bridging transformations based on context signals and, 30 if applicable, emotive weighting. This unit ensures that the system can prioritise specific outputs or domains in response to changing conditions or priorities, allowing for adaptive and contextsensitive decision-making.

[0251] [0077] The weighting control unit 112 operates by analysing signals such as the relevance scores of modular outputs, urgency indicators from the input data, or feedback signals from 35 other components like the probabilistic function unit 114. Based on these inputs, it modifies the weights applied to the outputs of processing modules 104 and the bridging transformations performed by the bridging unit 110. For instance, in the cancer detection scenario, if genomic data (processed by 104B) indicates a significant mutation, the weighting control unit may increase the weight of the corresponding modular output to ensure it has a greater influence 40 on the final prediction, while slightly reducing the weights of less critical domains like clinical records.

[0252] 349371

[0254] 15

[0256] The unit is also capable of incorporating emotive weighting, which amplifies or diminishes the importance of domain-specific outputs based on predefined urgency or significance levels. For example, if an anomaly in imaging data suggests a potentially lifethreatening condition, the weighting control unit may amplify its contribution to the final 5 decision, ensuring rapid and accurate response. By dynamically balancing the influence of modular outputs and bridging transformations, the weighting control unit 112 optimises the integration process, ensuring that system 100 remains responsive, efficient, and aligned with its overall objectives.

[0258] Emotive weighting in the system described herein is an optional but highly beneficial 10 mechanism that allows the system to dynamically prioritise domain-specific outputs or modular transformations based on contextual urgency or importance. This weighting mechanism is designed to amplify or diminish the influence of certain outputs in response to real-time signals, such as anomalies or critical thresholds, enabling the system to adapt its decision-making focus effectively.

[0260] 15 [0080] The emotive weighting process works by scaling the outputs of the domain-specific processing modules or their contributions during the bridging transformations. This scaling factor is determined by an emotive signal, which reflects the priority or urgency of the data from a given domain. For instance, in the cancer detection example, if the system detects an anomalous pattern in imaging data indicative of a potential tumour, an emotive signal could 20 elevate the contribution of this data stream. This ensures that the final probabilistic output and subsequent prediction are strongly influenced by the most critical inputs, leading to faster and more accurate responses.

[0262] While emotive weighting is optional and the system and associated units, modules and mechanisms can function without it, its inclusion significantly enhances the system’s capability 25 to handle dynamic, high-stakes environments. By focusing computational and decision-making resources on the most relevant domains or signals, emotive weighting ensures that critical insights are not diluted by less urgent data. This feature is especially advantageous in scenarios where anomalies or priority events require immediate attention, such as healthcare, energy systems, and/or real-time monitoring applications. The modular and adaptive design of the 30 system allows emotive weighting to be seamlessly integrated or excluded based on the requirements of the deployment context.

[0264] [0082] Optionally, the Super Prefrontal Cortex (Super PFC) in system 100 is a centralised meta-control module that dynamically adjusts the system's structure and operations based on real-time feedback and historical data. This sophisticated component is optionally implemented 35 through the integration of the pruning mechanism 108 and the weighting control unit 112, although it is alternatively a separate meta-control module integrated within system 100. The Super PFC evaluates performance metrics, user-defined goals, and logs from the HPC (see below) to ensure the system remains efficient and capable of adapting to new data patterns. By leveraging the pruning mechanism 108, the Super PFC selectively deactivates low-40 contributing neural connections within the processing modules 104 during offline reconfiguration phases or periods of reduced activity. This process optimises the system's computational resources without compromising its predictive capabilities. The weighting control unit 112 dynamically adjusts the importance of modular outputs and bridging

[0265] 349371

[0268] 16

[0270] transformations based on context signals and emotive weighting. This ensures that the system can prioritise specific outputs or domains in response to changing conditions or priorities, allowing for adaptive and context-sensitive decision-making.

[0272] Preferably, the Super PFC continuously monitors the system's performance, assessing 5 how well the hierarchical layers and bridging mechanisms are functioning. Based on feedback and performance metrics, the Super PFC can reconfigure the system's architecture, including adding or removing layers in the processing modules to better handle the complexity of the data. During offline phases, the Super PFC oversees the replay of stored sequences from the HPC to identify patterns, refine parameters, and prune underutilised pathways. This process is 10 akin to a biological sleep phase, where the system optimises its structure for future tasks. By combining the functionalities of the pruning mechanism 108 and the weighting control unit 112, the Super PFC ensures that system 100 remains efficient, adaptive, and capable of handling complex, multi-domain data.

[0274] The probabilistic function unit 114 in system 100 serves as a sophisticated 15 consolidation mechanism that integrates processed outputs from the hierarchical neural networks, recursive feedback, domain-specific modules, and bridging transformations into a coherent probabilistic framework. This unit synthesises the diverse, domain-specific data streams into a unified representation that quantifies the likelihood of various outcomes, providing a robust foundation for the system’s prediction and decision-making processes.

[0275] 20 [0085] The probabilistic function unit 114 leverages the refined and integrated outputs from earlier components, such as the processing modules 104 and the bridging unit 110, combining them with adjustments from the weighting control unit 112 and recursive feedback provided by mechanism 106. It computes a probabilistic output function that assigns confidence scores to each potential outcome based on the cumulative contributions of these inputs. For example, 25 in the context of cancer detection, the unit might integrate imaging data indicating a tumour mass, genomic data suggesting a relevant mutation, and clinical records identifying risk factors. These inputs are combined to produce a probabilistic prediction of whether the condition is malignant, with associated confidence levels.

[0277] The probabilistic function unit 114 is configured to handle the complexity of multi-30 domain data while maintaining interpretability and flexibility. By consolidating all inputs into a probabilistic model, it allows the system to weigh the evidence from each domain proportionally, ensuring that no single data stream dominates unless explicitly prioritised. This approach enables system 100 to provide actionable, context-aware predictions that are both accurate and adaptable to new information. The probabilistic function unit 114 is beneficial for 35 scenarios requiring high-confidence decision-making based on complex, multimodal datasets, ensuring a well-rounded and reliable output for downstream actions.

[0279] [0087] The selection logic 116 in system 100 is a decision-making component that determines the final prediction based on the outputs generated by the probabilistic function unit 114. It identifies the outcome with the highest probability from the probabilistic output function and 40 selects it as the system’s prediction, ensuring a clear and actionable result for downstream applications. Operating at the intersection of statistical inference and deterministic action, the selection logic evaluates the confidence scores assigned to each possible outcome. For example, in the cancer detection scenario, the probabilistic function unit 114 might produce

[0280] 349371

[0283] 17

[0285] predictions such as "benign tumour" with a 30% probability, "malignant tumour" with a 65% probability, and "unknown condition" with a 5% probability. The selection logic would identify "malignant tumour" as the final prediction, as it carries the highest probability.

[0287] The selection logic 116 is configured to operate efficiently even when handling highly 5 complex, multimodal data and integrates seamlessly with the broader system architecture. It ensures that the system outputs a single, definitive prediction, which can then be stored in the memory units 118 for analysis or reconfiguration and delivered via the output interface 120 for downstream actions. By providing a streamlined and deterministic final step in the decisionmaking process, the selection logic 116 enables system 100 to generate precise and actionable 10 insights, important for applications requiring high confidence and reliability in predictions.

[0289] The memory units 118 in system 100 are configured to store intermediate and final outputs generated throughout the processing workflow, serving as a beneficial component for enhancing the system's adaptability, efficiency, and long-term optimisation. These units maintain detailed logs of outputs from the domain-specific processing modules 104, refined 15 representations from the bridging unit 110, and probabilistic predictions from the probabilistic function unit 114. This storage facilitates offline analysis, hierarchical reconfiguration, layer tuning, and layer pruning.

[0291] In practice, the memory units 118 enable system 100 to retain critical information about its operational history, allowing for retrospective evaluation and adjustment. For example, in 20 the cancer detection scenario, the memory units could log the outputs of the imaging module 104A, genomic module 104B, and clinical module 104C, along with the final prediction and its associated probability. During offline replay, the system can analyse these logs to identify patterns, such as recurring anomalies or suboptimal feature extraction in specific modules, and adjust the corresponding layers or connections via the pruning mechanism 108 or the Super 25 PFC.

[0293] Additionally, the memory units 118 support continuity by preserving the global state of the system across time steps. This allows the system to maintain context and refine its predictions based on prior outputs and recursive feedback from the feedback mechanism 106. The stored data also provides a valuable resource for fine-tuning the weighting control unit 112 30 or recalibrating the bridging transformations in unit 110. By enabling iterative learning and optimisation, the memory units 118 ensure that system 100 remains efficient, adaptive, and capable of improving its performance over time.

[0295] [0092] The Hippocampus-like Memory (HPC) in system 100 is preferably implemented through the memory units 118, which store sequences of prior system states, domain inputs, and 35 actions Alternatively, a separate HPC unit is integrated with memory unit 118 and system 100 more broadly. The HPC plays a beneficial optional role in the system's ability to learn from past experiences and improve its performance over time. The memory units 118 are configured to store intermediate and final outputs generated throughout the processing workflow. These units maintain detailed logs of outputs from the domain-specific processing modules 104, 40 refined representations from the bridging unit 110, and probabilistic predictions from the probabilistic function unit 114. By logging detailed sequences of the system's operations, including intermediate and final outputs from various components, the memory units enable the system to retain critical information about its operational history. During offline replay, the

[0296] 349371

[0299] 18

[0301] system can analyse these logs to identify patterns, such as recurring anomalies or suboptimal feature extraction in specific modules. This information is used to refine the system's parameters and improve its predictive accuracy. The memory units ensure that the system retains historical context, which is beneficial for making informed decisions based on prior 5 knowledge. This reduces the need for extensive recurrent layers and allows for real-time adaptability with minimal computational overhead. The HPC supports the integration of causal reasoning into the replay process, allowing the system to formulate symbolic rules that inform real-time adjustments to bridging transforms and emotive weighting. By leveraging the memory units 118, the HPC enhances the system's adaptability and efficiency, enabling system 100 to 10 achieve robust, modular, and interpretable multi-domain intelligence.

[0303] The output interface 120 in system 100 serves as the final component in the architecture, responsible for delivering the system’s predictions and insights to downstream applications or end-users. This interface translates the selected prediction from the selection logic 116 into a format suitable for actionable use, ensuring seamless integration with external 15 systems or workflows. In operation, the output interface 120 takes the definitive prediction, such as the most probable outcome determined by the probabilistic function unit 114 and selected by the selection logic 116, and prepares it for delivery. For example, in the cancer detection scenario, the output interface might present a diagnosis of "malignant tumour" along with supplementary information, such as the confidence score, contributing data sources (e.g., 20 imaging, genomics, clinical records), and recommended next steps like further diagnostic tests or treatment options.

[0305] The output interface is configured to be flexible and adaptable, capable of interfacing with various downstream systems, such as electronic health record systems, alert mechanisms, or visualisation tools. Its ability to present predictions in human-readable formats 25 or structured data streams ensures that the insights generated by system 100 can be effectively utilised by clinicians, decision-makers, or automated systems. By bridging the internal operations of system 100 with external actions, the output interface 120 ensures that the system’s advanced predictive capabilities translate into meaningful, real-world outcomes.

[0307] Optionally, system 100 incorporates one or more interconnected functionalities that 30 contribute to its overall adaptability and robustness. One option is the recursive state update mechanism, which allows the system to maintain a global state across time steps. This mechanism works by partially blending the current state of the system with the outputs generated by the probabilistic function unit 114 and feedback mechanism 106. By doing so, it ensures that system 100 retains contextual awareness of prior operations, enabling it to adapt 35 its processing dynamically without the need for extensive retraining. This functionality is particularly beneficial in scenarios where sequential or time-dependent data patterns influence the predictions.

[0309] [0096] Another option is the integration of offline replay processes, supported by the memory units 118, which allow system 100 to reanalyse historical outputs and refine its internal 40 structures. During these offline phases, the system can perform hierarchical reconfiguration, where it adjusts the depth and dimensionality of layers within processing modules 104 to better align with domain-specific complexities. This process might also involve pruning underutilised

[0310] 349371

[0313] 19

[0315] connections via the pruning mechanism 108 or optimising the modular and bridging weights using insights derived from stored operational data.

[0317] The system also optionally comprises a symbolic reasoning layer, which operates during the offline replay of stored outputs to identify cause-effect relationships within domain-5 specific data. By detecting consistent patterns, such as one domain-lane’s anomaly leading to significant changes in another, the system formulates symbolic rules to inform real-time adjustments in the weighting control unit 112 or the bridging unit 110. These rules allow system 100 to pre-emptively adjust modular outputs or prioritise certain domains, enhancing its ability to respond to critical scenarios with greater precision.

[0319] 10 [0098] The symbolic reasoning layer in system 100 represents an advanced capability for extracting, interpreting, and utilising cause-effect relationships within the data processed by the system. During offline replay phases, supported by the memory units 118, this layer analyses stored outputs and intermediate results to identify patterns where a specific anomaly or event in one domain consistently leads to significant changes in another. By employing 15 techniques inspired by causal inference, the symbolic reasoning layer identifies such dependencies and formulates them as symbolic rules or "if-then" statements, which are stored in a rule repository for use in real-time operations.

[0321] For example, in the cancer detection context, the symbolic reasoning layer might observe across multiple replayed instances that a sharp rise in specific genomic markers 20 (processed by module 104B) frequently precedes an anomaly detected in imaging data (processed by module 104A). The system would then create a rule associating the genomic anomaly with a probable imaging anomaly. Once identified, this rule enables the system to adjust modular weights or bridging transformations dynamically, such that the imaging data is given higher priority in real-time processing when the genomic markers trigger the rule.

[0323] 25 [0100] The symbolic reasoning layer enhances the system’s predictive and adaptive capabilities by enabling proactive adjustments rather than reactive responses. This preemptive approach is facilitated by the weighting control unit 112, which leverages the rules generated by the symbolic reasoning layer to amplify the influence of the affected domains during bridging transformations or probabilistic reasoning. Furthermore, these rules provide 30 interpretability, as they articulate the causal relationships that underpin the system’s adjustments, making system 100 not only more intelligent but also more transparent. By bridging advanced pattern recognition with actionable causal insights, the symbolic reasoning layer significantly enhances the system’s ability to handle complex, multi-domain data environments effectively.

[0325] 35 [0101] Further, the role of emotive weighting as an optional yet significant enhancement to system adaptability. While it has already been outlined above, it interacts closely with the weighting control unit 112, the bridging unit 110, and the probabilistic function unit 114 to amplify or attenuate the influence of domain-specific outputs based on urgency signals. This integration ensures that system 100 is capable of prioritising high-impact data streams 40 dynamically while maintaining a balanced approach to multimodal processing.

[0327] [0102] Alternatively, or preferably additionally, system 100 incorporates mechanisms for tracking and logging its decision-making pathways through the memory units 118. These logs

[0328] 349371

[0331] 20

[0333] include not only intermediate outputs but also the rationale behind predictions, such as the contributions of specific processing modules 104 or the adjustments made by the weighting control unit 112. This feature enhances the system’s transparency and interpretability, making it suitable for applications where traceable decision-making is important, such as healthcare or 5 finance. Together, these interconnected functionalities ensure that system 100 remains a versatile and robust platform capable of addressing diverse and evolving analytical challenges.

[0335] An optimal hardware implementation of system 100 employs a distributed and parallelised architecture tailored to the computational demands of its components, ensuring high throughput, low latency, and scalability. System 100 is optionally deployed on a 10 heterogeneous computing platform that combines GPUs, FPGAs, and ASICs. This configuration supports the diverse tasks of hierarchical neural network processing, real-time feedback, and probabilistic reasoning, providing a robust and efficient hardware solution.

[0337] The input interface 102 is optionally implemented using programmable FPGAs, which preprocess and route data streams from various domains in real-time. FPGAs are configured 15 to normalise, encode, and multiplex imaging, genomic, and clinical data, preparing the inputs for subsequent domain-specific processing modules. The input interface 102 is not limited to FPGA implementation and is also optionally deployed on general-purpose CPUs for flexible software-controlled preprocessing.

[0339] The processing modules 104, comprising hierarchical neural networks, are optionally 20 implemented on GPUs, leveraging their ability to perform matrix computations and parallel operations efficiently. Each module operates as an independent GPU process, allowing simultaneous domain-specific feature extraction. Processing modules 104 are not limited to GPU implementation and are also optionally deployed on ASICs for specialised tasks such as convolutional operations for imaging data or sequence analysis for genomic data. ASICs 25 provide optimal performance with reduced power consumption for domain-specific computations.

[0341] The feedback mechanism 106 and pruning mechanism 108 are optionally implemented using GPUs or FPGAs. GPUs handle gradient computations and recursive adjustments efficiently, while FPGAs are optionally configured for threshold evaluation and selective pruning 30 of low-contributing neural connections. For instance, during offline or low-activity phases, an FPGA implementation evaluates stored usage statistics and deactivates underutilised connections, optimising the system’s computational resources.

[0343] The bridging unit 110 is optionally implemented using FPGAs or Tensor Processing Units (TPUs), which perform the necessary matrix multiplications and dimensional alignments 35 for bridging transformations. For example, an FPGA-based bridging unit integrates the outputs of a CNN processing imaging data with those from a transformer analysing genomic data. The bridging unit 110 is not limited to FPGA or TPU configurations and is also optionally implemented on CPUs for lower-complexity integration tasks.

[0345] [0108] The weighting control unit 112 and the probabilistic function unit 114 are optionally 40 implemented using TPUs, as these units require specialised hardware optimised for neural network operations and probabilistic computations. TPUs adjust modular outputs dynamically and consolidate the processed outputs into probabilistic predictions. These units are not limited

[0346] 349371

[0349] 21

[0351] to TPU implementations and are also optionally executed on GPUs, which provide robust handling of high-dimensional data and adaptive parameter tuning.

[0353] The selection logic 116 and output interface 120 are optionally implemented on general-purpose CPUs, which provide the necessary flexibility for decision-making and 5 downstream interfacing. CPUs efficiently select the most probable outcome and format the output for external applications or user interfaces. The output interface 120 is not limited to CPU implementations and is also optionally deployed on edge computing devices for real-time, localised output delivery.

[0355] The memory units 118 are optionally implemented using high-bandwidth, low-latency 10 storage technologies such as NVMe SSDs or persistent memory modules, or static random access memory (SRAM), dynamic RAM, or other suitable memory modules. These memory units store logs of intermediate outputs, symbolic rules generated by the symbolic reasoning layer, and system state information for offline replay and optimisation. Memory units 118 are not limited to these storage technologies and are also optionally implemented using distributed 15 cloud-based storage solutions to support scalability and remote access.

[0357] An alternative hardware implementation of system 100 focuses on cost-effectiveness and simplicity, optionally leveraging a cloud-based virtual environment with GPUs, CPUs, and scalable storage resources. This implementation optionally runs on high-performance cloud clusters, enabling rapid scaling during periods of high computational demand. System 100 is 20 not limited to cloud-based implementations and is also optionally deployed on edge computing devices, such as NVIDIA Jetson modules or Google Coral TPUs, for decentralised, real-time data processing. Edge devices are particularly advantageous for applications requiring localised analysis, such as remote cancer detection in field environments.

[0359] This alternative implementation, while potentially less powerful than the fully optimised 25 heterogeneous hardware platform, ensures that system 100 remains versatile and accessible.

[0360] By offering hardware configurations tailored to specific operational requirements, system 100 can be deployed effectively in diverse scenarios, from high-performance medical research centres to real-time diagnostic applications in resource-constrained environments.

[0362] System 100 is optionally implemented entirely or partially via software, leveraging 30 existing software frameworks and platforms to achieve flexibility and scalability across various environments. In this implementation, the components of system 100 are designed as modular software entities, executed on general-purpose computing hardware or within virtualised environments such as cloud-based platforms. For example, the input interface 102 and processing modules 104 are optionally implemented using machine learning libraries such as 35 TensorFlow or PyTorch, which enable the creation of hierarchical neural networks and support efficient data preprocessing, feature extraction, and modular processing. The feedback mechanism 106 and pruning mechanism 108 are optionally integrated as software algorithms that adjust neural network parameters or deactivate low-contributing connections based on real-time or offline analysis, using APIs provided by these frameworks.

[0364] 40 [0114] The bridging unit 110, weighting control unit 112, and probabilistic function unit 114 are optionally implemented as custom software modules that perform matrix transformations, weighting adjustments, and probabilistic computations. These functionalities are executed on

[0365] 349371

[0368] 22

[0370] software-optimised hardware accelerators such as GPUs or TPUs but remain compatible with CPU-based execution for environments with limited resources. Selection logic 116 and the output interface 120 are optionally realised as lightweight software modules that perform the selection of high-probability predictions and deliver formatted outputs to downstream systems 5 via APIs or messaging protocols such as REST or gRPC.

[0372] Memory units 118 are optionally implemented as database systems, using software solutions like SQL databases, NoSQL databases (e.g., MongoDB), or cloud-based storage services such as AWS S3 or Google Cloud Storage. These software-configured memory units facilitate the storage and retrieval of logs, system state, and intermediate results, supporting 10 offline replay, hierarchical reconfiguration, and optimisation. By leveraging software, system 100 achieves a high degree of portability, allowing deployment across diverse computational environments, from local servers to cloud-based infrastructures, while retaining its core functionalities and modular architecture.

[0374] The components of system 100 are configured to be highly modular and can be flexibly 15 combined, operate in parallel, or comprise multiple subcomponents arranged in series, parallel, or embedded configurations. This flexibility enhances the adaptability, scalability, and efficiency of the system, accommodating diverse use cases and operational requirements.

[0376] For example, the processing modules 104 optionally comprise multiple sub-modules operating in parallel, each tailored to a specific aspect of a domain. In the cancer detection 20 scenario, one sub-module within 104A might process imaging data focused on tumour shape detection, while another sub-module analyses texture or density features. These sub-modules can feed their outputs to an embedded subcomponent that consolidates their findings before sending the refined result to the bridging unit 110. Alternatively, a single processing module 104 may comprise serially connected sub-layers that progressively refine data, such as a 25 sequence of convolutional layers followed by fully connected layers.

[0378] The feedback mechanism 106 is optionally implemented as a distributed system with parallel feedback loops, each dedicated to different layers or domains within the processing modules. This arrangement allows for independent refinement of modular outputs, enhancing the system's responsiveness and reducing latency. Similarly, the pruning mechanism 108 can 30 be composed of subcomponents dedicated to monitoring specific layers, enabling selective pruning at various levels of granularity.

[0380] The bridging unit 110 optionally comprises multiple transformation pipelines operating in parallel, each specialising in integrating specific subsets of modular outputs. For instance, one pipeline might combine outputs from imaging and genomic modules, while another 35 handles integration between genomic and clinical data. These pipelines can either operate independently or feed into a final consolidation layer that produces the unified synergy vector. Alternatively, the bridging unit itself can be embedded within processing modules to enable intra-module integration before inter-module synthesis.

[0382] [0120] The weighting control unit 112 is optionally decomposed into hierarchical 40 subcomponents, where one level adjusts modular weights within individual processing modules, and another level dynamically tunes bridging weights based on contextual signals or

[0383] 349371

[0385] 23

[0387] emotive weighting. These subcomponents can operate in series to apply finer control over specific data streams or in parallel to manage multiple data domains simultaneously.

[0389] The probabilistic function unit 114 is optionally implemented as a distributed system with parallel probabilistic models, each responsible for synthesising outputs from a subset of 5 processing modules or specific data domains. These parallel models are optionally combined by an overarching probabilistic fusion component that integrates their results into a single output function. Similarly, the selection logic 116 can include embedded subcomponents that rank outcomes for each domain independently before final selection.

[0391] The memory units 118 are optionally implemented as distributed storage systems 10 comprising multiple databases or caches. For example, one database may store intermediate outputs for offline replay, while another logs symbolic rules generated by the symbolic reasoning layer. These memory units can operate in parallel to facilitate high-speed access or in a hierarchical structure where frequently accessed data is cached locally, and less critical data resides in remote storage.

[0393] 15 [0123] The output interface 120 can include embedded components that format predictions for specific downstream applications or parallel components that simultaneously deliver results to multiple systems. For example, one subcomponent may format predictions for humanreadable reports, while another sends structured outputs to an API for integration with automated decision-making systems. This ability to combine, embed, or operate components 20 in series or parallel configurations ensures that system 100 is highly customisable, capable of scaling to meet the demands of any domain or operational context, and optimised for both performance and resource utilisation.

[0395] System 100 operates as a cohesive entity wherein each component interfaces seamlessly through mathematical rigor, supporting adaptability, precision, and domain-specific 25 customisation. The hierarchical neural networks in the plurality of processing modules 104 are dynamically configured through layer transformations described mathematically, with each layer evolving from raw input to final abstract vectors. These transformations, structured by weights and biases, apply domain-specific non-linearities, as detailed in Equations (1)–(3) of the accompanying mathematical description below. These vectors, representing refined 30 domain-specific features, provide critical inputs for bridging transformations performed by the bridging unit 110.

[0397] The bridging unit 110 executes multi-domain synergy through transformations, mapping outputs from processing modules into a unified probabilistic representation . This synergy vector integrates emotive weighting, enabling contextually adaptive responses to 35 critical signals, as described in the equations governing emotive scaling. Probabilistic consolidation within the probabilistic function unit 114 is achieved by summing weighted and scaled outputs, ensuring that predictions incorporate the interplay between probabilistic, recursive, and domain-specific elements, consistent with Equation (3') of the associated bridging framework.

[0399] 40 [0126] Recursive feedback, defined through iterations of the global state, ensures stateful progression while retaining historical context. This recursive state mechanism, mirrored in System 100's memory units 118, enables the system to adaptively reconfigure its operational

[0400] 349371

[0403] 24

[0405] layers during offline phases, optimising processing efficiency and synergy alignment. These reconfigurations align with the pruning criteria established in Equations (1) and (2) of the offline sleep-phase logic, selectively deactivating underperforming layers or components.

[0407] System 100's selection logic 116 identifies the most probable prediction, leveraging the 5 probabilistic distributions and thresholds dynamically defined based on data complexity and time sensitivity. This operation integrates mathematical principles such as dynamic thresholding, ensuring predictions reflect both domain-specific accuracy and cross-domain coherence, consistent with threshold definitions in the probabilistic layer description.

[0409] Through these mathematical principles, System 100 embodies an adaptive, 10 hierarchically structured, and mathematically grounded operational framework. Its design ensures the system's scalability and robustness, facilitating seamless integration of new domains and preserving its predictive and adaptive capabilities in evolving scenarios. These core operations, enhanced by recursive updates and multi-domain synergies, establish System 100 as a robust solution for advanced, context-aware processing.

[0411] 15 [0129] In system 100, each domain-lane optionally has multiple layers to process raw data.

[0412] Conceptually, lower layers handle basic transformations (e.g., noise reduction, feature extraction), while higher layers produce more abstract vectors. Ultimately, these final-layer vectors feed into the multi-domain bridging mechanism of bridging unit 101. Preferably, each layer is a function that transforms an input vector into a higher-level feature representation.

[0413] 20 The process is repeated, stacking transformations until the final layer output is ready for bridging and emotive weighting.

[0415] Suppose each domain-lane has K layers. We refer to them as Layer 1, Layer 2, etc., up to Layer K.

[0417] Equation (1): h_{j,t} {^(i,0)} = x_{j,t} {^(i)}

[0419] 25 [0132] This means the input to the first layer (Layer 1) is the raw domain-lane vector x_{j,t} {^(i)}.

[0421] Equation (2): h_{j,t} {^(i,ℓ)} = f_ℓ( W_{j,i,ℓ} * h_{j,t} {^(i,ℓ-1)} b_{j,i,ℓ} )

[0423] For layer ℓ, the output vector is h_{j,t} {^(i,ℓ)}. W_{j,i,ℓ} is the weight matrix (or transformation) for domain j, sub-lane i, layer ℓ. b_{j,i,ℓ} is an optional bias vector. f_ℓ(...) is a non-linear or linear function. The “*” or multiplication symbol here denotes matrix 30 multiplication.

[0425] Equation (3): z_{j,t} {^(i)} = h_{j,t} {^(i,K)}

[0427] After the top layer K, we call the final output z_{j,t} {^(i)}. This is the abstract representation used by System 100’s bridging transform.

[0429] Dimensional progression is shown with Equation (4) (Layer-by-layer dimension 35 changes):

[0431] dim_{0} = dim_{j,i}, // the raw input size

[0433] dim_{1}, dim_{2}, ... // intermediate layer sizes

[0435] dim_{K} = dim_{j,i} p^rime // final layer size

[0436] 349371

[0439] 25

[0441] For example, if the raw domain-lane input is 8D, the first layer might map 8D -> 16D. Then the second layer might reduce 16D -> 12D, etc., until we end up with a final dimension dim_{j,i} p^rime.

[0443] The process flow in each time step is as follows:

[0445] 5 Raw Input:

[0447] x_{j,t} {^(i)} arrives from domain j, sub-lane i at time t.

[0449] Layer Sequence:

[0451] Successively compute h_{j,t} {^(i,1)}, h_{j,t} {^(i,2)}, …, h_{j,t} {^(i,K)} using Equation (2).

[0452] Final Output for Bridging:

[0454] 10 Define z_{j,t} {^(i)} = h_{j,t} {^(i,K)} (Equation (3)).

[0456] Next Steps:

[0458] The bridging transform uses z_{j,t} {^(i)} for synergy merging and emotive weighting.

[0460] Incremental feature extraction, instead of a single raw to bridging step, using hierarchical layers let system 100 handle complex domain-lane data (e.g. time-series 15 anomalies or sensor waveforms) in smaller transformations, eventually producing a “clean” or “abstract” final vector. There are parallels to biological cognition. In nature, signals pass from lower cortical areas (basic patterns) to higher cortical areas (complex concepts). System 100 mimics this pipeline, but each domain-lane preferably a unique layer structure, specialized for that domain’s needs. If HPC memory logs or user goals indicate a domain-lane is too shallow, 20 extra layers are added. If there are too many layers, layers are pruned to maintain structural integrity, efficiency, and reduce computational expenditure. This combined with offline reconfiguration, e.g. via the Super PFC at unit 108, to keep the architecture efficient.

[0462] When integrating with emotive weighting, the final vector z_{j,t} {^(i)} from hierarchical layers is then scaled by (1 e_{j,t}), e.g. in bridging unit 110. The entire hierarchical feature 25 extraction is then amplified or diminished if that domain-lane is deemed urgent or unimportant.

[0463] Once bridging merges all domain-lane outputs, the global state S_{t+1} partially updates. The hierarchical layers remain the same for step t+1 unless offline reconfiguration changes them, but the synergy context changes. HPC memory, e.g. at memory unit 118, logs each step’s hierarchical final outputs or optionally intermediate layers if needed, depending on complexity.

[0464] 30 In offline replay, it might be detected that a middle layer consistently fails or is underused, leading to potential pruning or retuning. The synergy output, which comprises the effect of these hierarchical layers, is used for gating real-time actions. Any refinement in the hierarchical stack directly improves the decision-making of system 100. If user goals or HPC replay show, e.g., “domain-lane i needs deeper analysis,” the PFC, e.g. pruning mechanism 108, optionally 35 adds an extra layer ℓ+1 or adjust layer sizes, quickly adapting hierarchical depth to reflect domain complexity.

[0466] [0142] Within system 100, hierarchical abstraction ensures each domain-lane is more than a single pass. By splitting the data flow into multiple layers, system 100 obtains an incrementally refined final vector z_{j,t} {^(i)}—ready for bridging synergy. This approach supports: deeper

[0467] 349371

[0470] 26

[0472] domain-specific feature learning, adaptive layer expansions or pruning, better synergy performance when bridging with other domains, alignment with the broader biologically inspired framework (emotive weighting, HPC memory, Basal Ganglia gating where the synergy output, which comprises the effect of all hierarchical layers, is used for gating real-time actions, 5 etc.) of system 100. Hence, hierarchical abstraction is a critical puzzle piece in the architecture of system 100, providing the foundation for domain-lane specialization and bridging synergy that collectively define the cognitive-like multi-domain AI system.

[0474] In system 100, multi-domain bridging is the mechanism by which each domain’s final outputs (from hierarchical layers, if used) are integrated into a single synergy vector. This 10 synergy vector reflects all relevant signals from each domain-lane at a given time step.

[0475] Beneficially, this results in parallel processing of each domain in sub-lanes, each sub-lane generating a specialized output, and enabling the merging of these outputs into a single synergy vector, which system 100 then uses for emotive weighting, recursive state updates, and action selection. This configuration avoids poor concatenation of data, allowing each 15 domain-lane to remain modular and independent while still contributing to a shared synergy representation.

[0477] Equation (5): dim_{j} = sum_{i=1 to L_j}( dim_{j,i} )

[0479] Each domain j of dimension dim_{j} is optionally subdivided into L_j sub-lanes. dim_{j,i} is the dimension of sub-lane i in domain j. Typically, a domain has distinct subcategories and 20 isolation into separate sub-lanes avoids conflating them in one big transform.

[0481] After hierarchical abstraction, each sub-lane produces a final vector z_{j,t} {^(i)}, which has dimension dim_{j,i} p^rime (the sub-lane’s top-layer output size). We define a bridging matrix W_{j,i} that maps z_{j,t} {^(i)} into the synergy dimension d_s.

[0483] Equation (6): b_{j,t} {^(i)} = W_{j,i} * z_{j,t} {^(i)}

[0485] 25 [0148] b_{j,t} {^(i)} is the bridging output for domain j, sub-lane i, at time t. W_{j,i} is of size (d_s x dim_{j,i} p^rime).

[0487] Equation (7): V_t = sum_{j=1 to M} sum_{i=1 to L_j} ( b_{j,t} {^(i)} )

[0489] Once all partial bridging outputs b_{j,t} {^(i)} are obtained, they are summed to form the synergy vector V_t, which is in R (^d_s), e.g. real and of a synergy dimension d_s. Beneficially, 30 each domain-lane’s final vector is multiplied by a bridging transform, producing a partial synergy vector. All partial synergy vectors are then added to make the final synergy vector.

[0491] If emotive weighting is active, Equation (6) is modified.

[0493] Equation (6’): tilde_b_{j,t}^ {(i)} = ( W_{j,i} * z_{j,t} {^(i)} ) * (1 e_{j,t})

[0495] where e_{j,t} is an emotive signal for domain j. Accordingly, Equation (7) becomes a 35 sum of tilde_b_{j,t} {^(i)} instead of b_{j,t} {^(i)}.

[0497] [0154] Instead of handling multiple domain embeddings in isolation, system 100 merges them in a coherent synergy representation each time step. This synergy dimension is tailored to specific applications, e.g. large enough to hold cross-domain features, but not so large as to be unwieldy.

[0498] 349371

[0501] 27

[0503] For example, cancer detection optionally requires three domains, health data, sensor data, and text data, associated with 4, 3, and 128 dimensions respectively in this example. Each domain has 1 sub-lane, L_j=1, for simplicity. The bridging transforms are then:

[0505] W_{health}: (d_s x 4),

[0507] 5 W_{sensor}: (d_s x 3), and

[0509] W_{text}: (d_s x 128).

[0511] At time t:

[0513] 1. The health sub-lane yields a vector z_{health,t} in R 4^.

[0515] 2. The sensor sub-lane yields a vector z_{sensor,t} in R 3^.

[0517] 10 3. The text sub-lane yields a vector z_{text,t} in R 1^28 (e.g. after an LLM embedding).

[0519] Equation (6) is then performed, followed by equation (7), such that:

[0521] V_t = b_{health,t} b_{sensor,t} b_{text,t}.

[0523] If emotive scaling is used, V_t = tilde_b_{health,t} tilde_b_{sensor,t} tilde_b_{text,t}.

[0525] During an offline pass, if HPC memory of memory unit 118 shows that certain sub-lanes 15 rarely yield synergy contribution, pruning occurs by setting bridging transforms to None (or removing them). The Super PFC, a centralised meta-control module that dynamically optimises the architecture of the AI system by pruning underutilised pathways and adjusting modular weights in real time, optionally creates a new bridging transform if HPC logs reveal repeated anomalies not captured by existing sub-lanes. Accordingly, bridging transforms can be added, 20 merged, or removed over time.

[0527] Beneficially, multi-domain reduces parameter bloat, saving energy, increasing speed, and reducing required computational resources. By focusing bridging on final sub-lane outputs, a massive end-to-end model i.e. with everything combined from the start is not needed. Additionally, interpretability and explainability is improved. As bridging logs partial 25 synergy from each domain-lane, it is clearer which domain-lane contributed to the synergy vector each step.

[0529] In system 100, emotive weighting, also referred to as priority weighting, is optionally used to result in urgent or critical signals receiving increased importance during the bridging process. Rather than handling all domain data uniformly, System 100 scales each domain’s 30 sub-lane output by (1 e_{j,t}), where e_{j,t} is the emotive (priority) value for domain j at time t. Beneficially, emotive weighting highlights domains that display unusual or high-impact data (e.g., sensor anomalies, urgent text, series spikes). This approach helps system 100 focus synergy on the data that truly is more impactful for accurate or desired outputs, mimicking how humans give urgent stimuli more cognitive resources.

[0531] 35 [0161] For example, each domain j at time t has an emotive scalar e_{j,t}. Preferably, e_{j,t} > 0 if domain j is “more urgent” than usual. E.g., if a sensor domain sees a spike or a text domain sees alert phrases, e_{j,t} is increased from 0.0 to 0.5 or 1.0, depending on the intended impact. e_{j,t} around 0 is equivalent to a standard weighting, where no emphasis is provided. Further

[0532] 349371

[0535] 28

[0537] optionally, e_{j,t} < 0 if system 100 is configured to de-emphasize a domain-lane, though many use cases keep it ≥ 0 for clarity.

[0539] According to one non-limiting use case, e_{j,t} = max( 0, anomaly_score_{j,t} -threshold_j ). If anomaly_score_{j,t} surpasses threshold_j, emotive weighting increases.

[0540] 5 Additionally, or alternatively, system 100 applies a nonlinear scale (e.g. exponent or logistic).

[0541] For example, tilde_b_{j,t} {^(i)} = b_{j,t} {^(i)} * exp( e_{j,t} ). This approach yields a stronger emphasis if e_{j,t} is high. Any alternative is suitable provided sub-lane outputs are scaled by a function of emotive signals.

[0543] For real-time, including near real-time, usage, during each time step t:

[0545] 10 1. The domain-lanes produce bridging outputs b_{j,t} {^(i)}.

[0547] 2. If domain j at time t is flagged “urgent,” e_{j,t} is > 0.

[0549] 3. The bridging outputs are scaled by (1 e_{j,t}), forming tilde_b_{j,t} {^(i)}.

[0551] 4. The scaled vectors are summed into V_t.

[0553] 5. The synergy vector V_t then partially updates a global state.

[0555] 15 [0164] Preferably, e.g. using memory unit 118, system 100 logs how often each domain-lane’s emotive factor was > 0, and/or how large it was. This is used in pruning, e.g. where a domainlane rarely sees nonzero emotive weighting or synergy, and/or meta-control, where the Super PFC optionally overrides emotive weighting if HPC memory shows too many false alarms from that domain.

[0557] 20 [0165] While standard AI systems sum all signals equally, emotive weighting is a powerful technique to differentiate among domain-lanes based on urgency, akin to how humans quickly prioritize life-threatening stimuli. If multiple domains produce small anomalies, emotive weighting can ensure that a truly big anomaly from one domain-lane is not lost in the synergy sum. If required, emotive weighting can also incorporate negative signals (a domain-lane 25 system 100 will downplay) or apply a more complex functions, e.g.1 alpha * e_{j,t}.

[0559] Optional interaction of emotional weighting with other components of system 100 comprise any of:

[0561] Hierarchy: The domain-lane’s hierarchical abstraction yields a final vector z_{j,t} {^(i)}, which emotive weighting then scales.

[0563] 30 Recursive Updates: The synergy vector V_t (which comprises emotive weighting) updates S_{t+1} = (1 - alpha) * S_t alpha * V_t.

[0565] Hippocampus: HPC memory logs (1 e_{j,t}) per domain-lane each step, so system 100 can see how emotive signals shaped synergy.

[0567] Basal Ganglia gating: The final synergy or synergy norm might reflect emotive weighting 35 strongly, thereby influencing immediate action gating.

[0569] Super PFC: If HPC logs show domain j triggers high emotive weighting too often (causing false alarms), the PFC might reduce baseline emotive offsets for that domain.

[0570] 349371

[0573] 29

[0575] The recursive state updates, comprising computing a synergy vector from domainlanes at each time step and partially merging with the existing global state, allows for system 100 to remember and/or retain context from prior steps without requiring e.g. a large RNN architecture, increasing real-time adaptability.

[0577] 5 [0168] Equation (8):

[0579] S_{t+1} = (1 - alpha) * S_t alpha * V_t

[0581] where S_{t} is the global state at time t, V_{t} is the synergy vector from bridging (optionally including emotive weighting), and alpha (0 < alpha < 1) is a recursion factor controlling how strongly new synergy overrides older context. If alpha is relatively large, e.g.

[0582] 10 close to 1, system 100 heavily weights the new synergy vector and “forgets” the older state more quickly. If alpha is relatively small, e.g. close to 0, system 100 retains more from previous steps, blending new synergy modestly.

[0584] Beneficially, system 100 is not limited to doing a single forward pass and then retrain/update offline. Each step partially updates S_t, incorporating fresh synergy in near real 15 time. For example, if prior time steps indicated a slow-building anomaly, S_t remains partially elevated, ensuring system 100 is primed to detect subsequent related signals. It is preferably integrated with high-level/meta-control systems, e.g. the HPC logs each step at memory unit 118, and/or the Basal Ganglia uses S_t when gating actions at selection logic 116 or output interface 120, and/or the super PFC adjusts alpha over time at weighting control unit 112.

[0585] 20 [0171] For example, optionally, the magnitude of the global state is tracked, e.g.

[0586] Synergy_norm_t = || S_t ||, to gauge how strongly past synergy patterns remain in the system. For example, the Basal Ganglia gating references this norm when determining and/or selecting actions.

[0588] Alternative implementations of Equation (8) comprise S_{t+1} = f( (1 - alpha)*S_t 25 alpha*V_t ), where f(...) is a non-linear function, e.g., ReLU, tanh, or a sigmoid function. Any function is suitable resulting in partial incorporation of new synergy.

[0590] In another example, during an offline routine or periods of low activity, the system 100 optionally adjusts alpha, e.g. if HPC memory shows that S_t is decaying too fast or too slow. The system does not discard S_t across sequences unless explicitly reinitialized. However, in 30 certain scenarios, system 100 performs a partial reset if a new scenario is unrelated to past states.

[0592] [0174] In an AI system, e.g. system 100, comprising multiple domains with hierarchical layers, such as neural networks, a period of low activity refers to a phase where the system exhibits minimal computational engagement or reduced operational intensity. This can manifest in 35 various forms, such as decreased neuron firing rates, lower frequency of data processing events, or diminished inter-layer communication. In neural networks, particularly deep learning models, low activity might be observed when the network is not actively training, when it is processing data that does not stimulate significant changes in weights, or when it is in an idle state awaiting new input. This period can also occur during inference when the input data is 40 either too simple or too familiar, leading to minimal activation across the network layers.

[0593] 349371

[0596] 30

[0598] Detecting a period of low activity in such systems comprises monitoring one or more key indicators. One method is to track the activation levels of neurons across different layers. If the majority of neurons exhibit low activation values, it suggests that the network is not heavily engaged. Another approach is to analyse the gradient magnitudes during training; small 5 gradients indicate that the network is not learning significantly from the current data, which can be a sign of low activity. Additionally, monitoring the computational resource usage, such as CPU/GPU utilization, memory consumption, and data throughput, optionally provides insights into the system's activity levels. Low resource usage typically correlates with periods of low activity. Furthermore, logging and analysing the frequency and intensity of data processing 10 events, such as the number of forward and backward passes per unit time, can help identify when the system is less active. By using or combining these methods, periods of low activity in complex AI systems with hierarchical layers are effectively detect and quantified. For example, after being quantified, a threshold associated with a time interval, e.g.10 seconds, or a metric, e.g. over 50% of neurons exhibiting low activation values, or some combination 15 thereof, a period of low activity is determined.

[0600] In system 100, the HPC is a mechanism by which system 100:

[0602] 1. Captures sequential data over multiple time steps, also referred to as a sequence or episode.

[0604] 2. Replays these episodes in an offline or low-activity sleep-phase, analysing patterns that may not be evident from single-step observations.

[0606] 20 3. Informs bridging transforms, emotive weighting adjustments, or the creation/removal of domain-lane sub-lanes.

[0608] Beneficially, this approach encodes important concepts over time and reactivates them to consolidate memory or refine neural connections. An example step record structure:

[0610] StepRecord_t = {

[0611] 25 'time_index': t,

[0612] 'domain_inputs': X_{j,t} {^(i)}, // hierarchical final outputs

[0613] 'emotive_inputs': e_{j,t}, // emotive signals

[0614] 'synergy_out': V_t, // synergy vector

[0615] 'global_state': S_t, // state before or after update

[0616] 30 'chosen_action': action_t // from Basal Ganglia gating

[0617] }

[0619] A current_sequence accumulates step-by-step records. Once the sequence ends (e.g., 10 steps or completion of an event), the episode is finalised in HPC memory, e.g. stored by memory unit 118. Episode = [ StepRecord_1, StepRecord_2, ..., StepRecord_T ], where T is 35 the number of steps in the sequence, with is preferably tailored for the application of system 100.

[0621] [0179] During offline periods or periods of low activity, e.g. a sleep phase, HPC memory replays one or more, preferably each, stored episode. This is used for any of: identifying multi-step anomalies that may otherwise be missed when analysing a single step, e.g. at processing 40 modules 104 or pruning mechanism 108; correlating how certain emotive weighting or bridging transforms are performed across consecutive steps, e.g. at bridging unit 110 and weighting

[0622] 349371

[0625] 31

[0627] control unit 112; and/or suggesting new bridging transforms if repeated episodes show a pattern not captured by the current sub-lanes, e.g. at pruning mechanism 108, bridging unit 110 and weighting control unit 112. Accordingly, each of the components of system 100 are configured to be integrated into a cohesive artificial intelligence system. For example, each 5 step record of the episode is evaluated to determine synergy patterns over consecutive steps, check cumulative domain-lane usage, and, if there are repeated anomalies, flag anomalies for bridging refinement. System 100 optionally computes a multi-step synergy sum or average emotive weighting across steps to determine if a domain-lane’s signals build up over time.

[0629] During replay, system 100 optionally defines a “novelty” metric.

[0631] 10 [0181] Equation (9), a novelty metric example:

[0633] novelty_e = ( sum of synergy_out magnitudes ) ( sum of emotive_inputs )

[0635] where, if novelty_e is high but system 100 took no strong action or bridging synergy was low, there is a gap. This gap can be plugged by creating a new sub-lane specialised for the relevant pattern, raise the emotive baseline weighting values for the relevant domain-lane, 15 and/or adjust alpha if a multi-step buildup was missed.

[0637] If HPC memory shows that certain sub-lanes rarely contribute synergy or emotive weighting, pruning mechanism 108 prunes them. The HPC replay reveals which domain-lanes or bridging transforms remain underused across multiple episodes. The PFC references HPC data to drive high-level changes, e.g.:

[0639] 20 1. Restructure bridging transforms if repeated HPC episodes identify a new domain pattern.

[0641] 2. Override emotive weighting if HPC memory logs show too many false alarms or missed anomalies.

[0643] Optionally, after computing a synergy vector V_t, i.e. merging multi-domain outputs, 25 and updating a global state S_t, the system 100 picks an immediate action among multiple choices, e.g. akin to a Basal Ganglia mechanism in the brain. Rather than producing only a numeric score, e.g. at probabilistic function unit 114, system 100 additionally makes a discrete decision (e.g., “COLLECT_MORE_DATA” or “RAISE_ALERT”), e.g. at selection logic 116. This step comprises bridging synergy magnitude and emotive signals into a gating score for each 30 action, selecting the highest-scoring action at each time step, and logging the chosen action in the HPC memory, e.g. memory unit 118, for offline replay and analysis as discussed above.

[0645] A set of actions are defined, e.g. { A_1, A_2, …, A_K}, and a gating score for the is calculated. Each action A_k has learned or fixed parameters: w_{k,\text{syn}} , w_{k,\text{emo}} , and a bias b_k.

[0647] 35 [0187] Equation (10), a gating score for action K:

[0649] G_k( S_t, Emo_t ) = w_{k,syn} * || S_t || w_{k,emo} * Emo_t b_k.

[0651] [0189] Emo_t = sum_j(e_{j,t}) is the total emotive signal across domains, || S_t || is optionally a norm (e.g. L2 norm) or some function of S_t summarizing synergy magnitude or important features, and b_k is a bias term that can shift the preference for certain actions.

[0652] 349371

[0655] 32

[0657] Equation (11), action selection:

[0659] A *^ = argmax_{k in [1..K]} ( G_k( S_t, Emo_t ) )

[0661] where action K is picked based on yielding the highest gating score.

[0663] As emotive weighting influences how synergy out V_t is formed, high emotive signals 5 also raise Emo_t, meaning actions that respond to urgency are more likely to be chosen. For example, if emotive signals are large (e.g. urgent domain-lane anomaly), the gating score for “RAISE_ALERT” increases significantly. If synergy is moderate and emotive is low, a more conservative action like “COLLECT_MORE_DATA” will result in a higher gating score

[0664] Preferably, each time an action is chosen, system 100 logs the chosen action in the 10 HPC memory of memory unit 118, e.g. step_record_t['chosen_action'] = action_t. During offline replay, which synergy or emotive conditions led to each chosen action is evaluated by system 100. If HPC replay reveals a repeated scenario that needed a different action, the meta-control of Super PFC optionally adjusts one or more relevant gating parameters (w_{k,\text{syn}}, w_{k,\text{emo}}, b_k) for better alignment.

[0666] 15 [0195] For example, there are three predetermined actions: [DO_NOTHING, RAISE_ALERT, COLLECT_MORE_DATA]. A_1 = DO_NOTHING => parameters (w_{1,syn}, w_{1,emo}, b_1). A_2 = RAISE_ALERT => parameters (w_{2,syn}, w_{2,emo}, b_2). A_3 = COLLECT_MORE_DATA => parameters (w_{3,syn}, w_{3,emo}, b_3).

[0668] For example, at time t:

[0670] 20 1. Synergy norm is computer: synergy_norm_t = || S_t ||.

[0672] 2. Emotive signals are summed across domains: Emo_t = sum_j e_{j,t}.

[0674] 3. Gating scores for each action are calculated.

[0676] Gating for action 2 (RAISE_ALERT) is: G_2( S_t, Emo_t ) = w_{2,syn} * synergy_norm_t w_{2,emo} * Emo_t b_2.

[0678] 25 4. Gating scores G_1, G_2, G_3 are compared, and the highest selected.

[0680] If G_2 (RAISE_ALERT) is highest, selection logic 116 immediately “alerts” the system 100 that an anomaly is found, and HPC logs the step at memory unit 118.

[0682] Continuing the example, during periods of offline or low activity, based on meta-control of the super PFC, weighting control unit 112 raises w_{2,syn} or w_{2,emo} for RAISE_ALERT.

[0683] 30 Alternatively, if too many false alerts for synergy have been raised, w_{2,syn} is reduced.

[0684] Accordingly, the gating parameters adjust over time for improved decisions.

[0686] [0198] The Super PFC, e.g. meta-control module, of system 100 operates at a meta-level, interacting with multiple components, and is configured to: analyse performance metrics, HPC replays, and user-defined goals; modify internal parameters e.g. recursion factor, alpha, 35 bridging transforms, or emotive weighting scales; create or unifying domain sub-lanes if repeated HPC episodes show a pattern not captured or redundantly repeated by existing bridging. Beneficially, by incorporating structural and/or parametric changes during both

[0687] training and during operation, the super PFC enables system 100 to self-tune and improve at a faster rate.

[0689] Input data comprises: performance metrics, e.g. false alarm rate or missed anomaly rate; HPC logs for multi-step episodes comprising synergy usage and, if used, emotive 5 weighting across time; and user goals, e.g. minimise misses, avoid fast positives, or domainspecific priorities. Outputs comprise a parameter update vector or structure, denoted C, specifying changes. For example:

[0691] C = {

[0693] 'alpha_adjustment': DeltaAlpha,

[0695] 10 'emotive_baseline_changes': { domain_j: DeltaE_j, ... },

[0697] 'structure_operations': [ create_new_lane(...), unify_lanes(...), etc. ]

[0699] }

[0701] where alpha_new = alpha_old DeltaAlpha and e_{j,t} b^aseline_new = e_{j,t} b^aseline_old DeltaE_j.

[0703] 15 [0200] If HPC memory reveals, e.g., system 100 underreacted to new synergy, alpha is increased. If it overreacted or forgot valuable past states, alpha is decreased. If domain j triggers too many false alarms, the emotional weighting is reduced. If anomalies at domain j keep getting ignored, emotional weighting is increased by the super PFC. Optionally, memory unit 118 stores emotive weighting as (1 e_{j,t} b^aseline e_{j,t} d^ynamic ) to separate 20 baseline from immediate anomaly signals.

[0705] If HPC memory consistently flags certain patterns or anomalies in domain j that sublane i doesn’t capture, the PFC creates a new sub-lane i_new with a bridging transform, e.g. W_{j,i_new} = random_init or specialized_init(...). The usage for the new lane is then logged. If the HPC indicates two sub-lanes are highly overlapping or redundant, the PFC unifies them by 25 averaging or merging transforms, e.g. W_{j,i_merged} = combine( W_{j,i1}, W_{j,i2} ), and removes the old lanes. This not only keeps system 100 from ballooning in complexity, but results in only necessary computational resources and energy usage for accurate task completion, even as aspects of the task may change, e.g. when used as an AI agent.

[0707] Additionally, emotive or synergy-based tendencies are overridden if HPC episodes or 30 performance logs show consistent misalignment. If a domain-lane is “emotionally” blowing up synergy too often, e.g. the emotive weighting is frequently relatively high, the PFC clamps or reduces the emotive weighting (1 e_{j,t}). For example, (1 e_{j,t})_clamped = min( 1 e_{j,t}, clamp_value ). If HPC replays indicate repeated patterns that synergy fails to capture, the PFC creates a deeper hierarchical stack for that domain-lane. The changes are logged, e.g. in 35 memory unit 118, for the next cycle, iteration, time step, and/or episode.

[0709] [0203] As an addition or alternative to the super PFC, system 100 applies pruning at pruning mechanism 108. During normal operation, system 100 updates a global state each time step, produces a synergy vector, and optionally chooses real-time actions. However, it enters an offline or low activity period sporadically or periodically – a “sleep phase”. In a sleep phase:

[0710] 1. System 100 evaluates usage stats for each domain-lane and bridging transform, removing ones that rarely contribute;

[0712] 2. System 100 replays HPC episodes, identifying multi-step anomalies or patterns that single-step synergy may miss; and/or

[0714] 5 [0206] 3. System 100 invokes Super PFC to reconfigure sub-lanes or adjust emotive weighting/recursion parameters, based on HPC data and performance metrics.

[0716] This is analogous to biological “sleep” or “offline” consolidation, where brains reorder synaptic connections. In system 100, it is a mechanism to prune unhelpful connections, e.g. at pruning mechanism 108, and adopt new structures, e.g. at processing modules 104 using 10 bridging unit 110 and/or weighting control unit 112, if HPC episodes show repeated missed anomalies. Preferably, during the same sleep phase, system 100 replays HPC episodes as described above.

[0718] For one or more, preferably each, domain-lane, system 100 accumulates an invocation count, e.g. how many times a sub-lane was used in bridging, a synergy sum, e.g. the cumulative 15 absolute bridging contribution, and optionally the emotive sum, e.g. the cumulative emotive weighting applied. A synergy ratio and emotive ratio is defined to measure each sub-lane’s average usage.

[0720] Equation (12):

[0722] rho_syn_{j,i} = syn_{j,i} / max(inv_{j,i}, 1),

[0724] 20 rho_emo_{j,i} = emo_{j,i} / max(inv_{j,i}, 1).

[0726] If one or both ratios fall below certain thresholds, the sub-lane is pruned.

[0728] Equation (13):

[0730] If ( rho_syn_{j,i} < Theta_syn ) and ( rho_emo_{j,i} < Theta_emo ) :

[0732] Deactivate bridging transform W_{j,i} or bridging[j,i] = None

[0734] 25 [0212] Beneficially, if a domain-lane’s bridging outputs rarely contribute synergy or emotive weighting, it is removed to avoid overhead.

[0736] Optionally, system 100 is further configured to perform a wave function mechanism for improved adaptability, precision, and scalability. The wave state represents uncertainty as a probability distribution over possible outcomes, e.g. calculated at probability function unit 114.

[0737] 30 An example is provided by Equation (13):

[0739] P(x) = {p1, p2, ..., pn}, where ∑(pi) = 1

[0741] where P(x) is the probability distribution over all outcomes and pi is the probability of outcome i.

[0743] System 100 refines the wave state iteratively using multi-dimensional feedback, e.g. via 35 feedback mechanism 106. Foe example:

[0745] [0217] P_updated(t+1) = P_current(t) η * ∑(ws * ΔFs)

[0746] where P_updated(t+1) is the updated probability distribution at time t+1, P_current(t) is the current probability distribution at time t, η is the feedback learning rate, ws is the weight assigned to feedback source s, and ΔFs is the feedback value from source s.

[0748] Weights for feedback sources are dynamically adjusted based on confidence:

[0750] 5 [0220] ws = Cs / ∑(Cs),

[0752] where Cs is the confidence score of source s.

[0754] The threshold for collapsing the wave state into a particle state adjusts dynamically based on context:

[0756] Threshold_c = Baseline_c α * C_complexity - β * T_urgency.

[0758] 10 [0224] Where:

[0760] Threshold_c: Dynamic confidence threshold.

[0762] Baseline_c: Baseline threshold.

[0764] α: Weight for complexity factor.

[0766] C_complexity: Complexity of the input data.

[0768] 15 β: Weight for urgency factor.

[0770] T_urgency: Time sensitivity of the decision

[0772] When a single probability exceeds the dynamic threshold, the wave state collapses into a definitive particle state:

[0774] x_collapsed = argmax(P_updated), if max(P_updated) > Threshold_c

[0776] 20 [0227] where x_collapsed is the final decision or classification and P_updated is the updated probability distribution calculated above.

[0778] For example, when system 100 is applied for Unidentified Aerial Phenomena (UAP) detection, the wave state initialisation is:

[0780] P(x) = {Meteor: 0.4, Aircraft: 0.35, Anomalous Object: 0.25}.

[0782] 25 [0230] Radar data increases Anomalous Object probability, and hence feedback refinement becomes:

[0784] P_updated = {Meteor: 0.2, Aircraft: 0.3, Anomalous Object: 0.5}.

[0786] Infrared data corroborates Anomalous Object, raising its probability:

[0788] P_updated = {Meteor: 0.1, Aircraft: 0.2, Anomalous Object: 0.7}.

[0790] 30 [0234] Dynamic thresholding with high complexity (e.g. C_complexity = 0.6) and urgency (e.g.

[0791] T_urgency = 0.8):

[0793] [0235] Threshold_c = 0.55 (0.4 * 0.6) - (0.3 * 0.8) = 0.55.

[0794] When assessing particle collapse:

[0796] x_collapsed = "Anomalous Object", as P(Anomalous Object) = 0.7 > Threshold_c.

[0798] Accordingly, system 100 handles uncertainty dynamically, where probability distributions are continuously updated with weighted feedback. System 100 therefore adapts 5 to context, where dynamic thresholds account for complexity and urgency, ensuring optimal decisions. This is scalable across applications, and is applicable to any domain or use-case scenario.

[0800] Preferably, system 100 is configured to integrate a causal symbolic reasoning layer into HPC and Super PFC operations. As described above, the HPC collects multi-step episodes, 10 logging any of synergy, emotive weighting, and chosen actions. The reasoning layer provides a structure for explicit cause-effect or if–then rules. By integrating a causal reasoning pass into HPC replay, system 100: detects persistent triggers where domain-lane X consistently leads to synergy changes in domain-lane Y; formulates these discoveries as symbolic or “if–then” rules, stored in a new cause-effect rule repository; and feeds those rules to the Super PFC, enabling 15 more targeted synergy adjustments or emotive offsets once HPC replay confirms repeated cause-effect patterns.

[0802] An example implementation for each episode e in HPC is provided below:

[0804] For t in [1..T-1]:

[0806] // (1) Identify potential cause: domain-lane X high at time t

[0808] 20 // (2) Identify potential effect: domain-lane Y synergy or emotive weighting high at time t+1

[0809] If condition repeated across multiple episodes:

[0811] increment cause_effect_confidence(X -> Y).

[0813] If domain-lane X is consistently above a threshold at time t, and domain-lane Y synergy or emotive weighting is above a threshold at t+1 in multiple episodes, system 100 infers a 25 possible cause-effect link (X -> Y).

[0815] Each discovered link is stored as a symbolic rule:

[0817] rule_XY = {

[0818] cause_domain_lane: (jX, sub-lane iX, thresholdX),

[0819] effect_domain_lane: (jY, sub-lane iY, thresholdY),

[0820] 30 confidence: c,

[0821] last_updated: sleep_cycle_id

[0822] }

[0824] where confidence c is how strongly HPC replay has confirmed a cause-and-effect link (X->Y).

[0826] 35 [0244] Each offline pass, HPC replay tallies repeated triggers. For example:

[0828] confidence(X->Y) = (number of times effect followed cause) / (number of times cause occurred).

[0829] 349371

[0831] 37

[0833] If that ratio is stable or increasing, it is more likely to be a real cause-effect chain, not mere coincidence. More complex and/or alternative methods of calculating confidence in cause-and-effect links are also suitable.

[0835] One alternative involves using Bayesian inference, which enables probabilistic 5 reasoning about causation. A Bayesian approach updates the belief in a causal relationship between two events based on observed data. In this context, prior probabilities represent the initial belief in the likelihood of causation (e.g., informed by domain knowledge or historical data), while the likelihood term quantifies how well the observed data fits the proposed causal model. Posterior probabilities, computed using Bayes' theorem, provide a refined belief in 10 causation after accounting for evidence. This method is beneficial because it can incorporate uncertainty and prior knowledge, which are particularly relevant in noisy or incomplete datasets.

[0837] Another approach leverages Granger causality, commonly used in time-series analysis. Granger causality does not determine true causation in a philosophical sense but assesses 15 whether one variable provides statistically significant information about the future values of another. The method involves constructing predictive models, typically autoregressive models, for the target variable both with and without the inclusion of the putative cause variable. If the inclusion of the cause variable significantly improves the prediction of the target variable, it is considered to Granger-cause the target. However, this technique assumes temporal 20 precedence and cannot account for confounding variables, so it is best used when temporal ordering is well-established.

[0839] Causal discovery algorithms, such as those based on the PC (Peter and Clark) algorithm or the Fast Causal Inference (FCI) algorithm, offer another sophisticated method for determining causal links. These methods use conditional independence testing to infer causal 25 structures from observational data. The PC algorithm, for example, starts with a fully connected graph and iteratively removes edges by testing for independence between pairs of variables, conditioned on subsets of other variables. The result is a partially directed acyclic graph (DAG) that encodes possible causal relationships. These algorithms often assume that the causal relationships are linear and that there are no unmeasured confounders, although more 30 advanced variants can relax these assumptions.

[0841] Structural causal models (SCMs) provide a more comprehensive framework by explicitly modelling the mechanisms underlying causal relationships. SCMs consist of equations describing how each variable is generated based on its direct causes and a noise term. These models are particularly powerful because they allow the simulation of 35 interventions—what happens if a variable is deliberately manipulated. By comparing outcomes under different interventions, SCMs can robustly distinguish between correlation and causation. Techniques such as do-calculus, e.g. as developed by Judea Pearl, formalise the manipulation of these models to compute causal effects from observational data, even in the presence of confounders, provided sufficient assumptions about the data-generating process.

[0843] 40 [0250] Information-theoretic approaches, such as Transfer Entropy (TE), also provide tools for evaluating causal relationships. Transfer Entropy quantifies the reduction in uncertainty about the future state of one variable given knowledge of the past state of another, over and above the information contained in the variable's own past. Unlike Granger causality, TE does not

[0844] 349371

[0847] 38

[0849] assume linearity and is more robust to non-linear dependencies. However, it typically requires large amounts of data to estimate probability distributions accurately, which can be computationally intensive.

[0851] Optionally, incorporating neural networks into causal inference is used where methods 5 such as causal representation learning aim to extract causally meaningful features from highdimensional data. Variational autoencoders (VAEs) or generative adversarial networks (GANs) can be adapted to learn representations that align with underlying causal structures. For example, disentanglement-focused VAEs are designed to separate causal factors in latent spaces, enabling both prediction and intervention. Additionally, counterfactual reasoning is 10 gaining traction as a method to evaluate causation by answering "what if" questions.

[0852] Counterfactuals involve comparing observed outcomes to hypothetical outcomes that would have occurred under different conditions. In system 100, this optionally requires generating counterfactual data using simulation or generative models, e.g. a dedicated processing module 104. For example, a model predicts how a patient's health trajectory might differ if they had 15 received a different treatment, enabling an evaluation of treatment effectiveness. These counterfactuals can be paired with propensity score matching or inverse probability weighting to account for confounders and ensure fair comparisons between groups.

[0854] After HPC replay sets or updates cause-effect rules, the Super PFC reads them:

[0856] If cause_effect_confidence( X->Y ) > min_confidence:

[0858] 20 // Then adopt a synergy or emotive weighting override for domain-lane Y

[0860] // whenever domain-lane X passes thresholdX

[0862] where min_confidence is a dynamic or predefined threshold.

[0864] Actions performed comprise any of increasing emotive weighting, modifying bridging, or explaining in human-readable format. For example, if HPC logs show domain-lane X leading 25 to domain-lane Y synergy next step, the PFC adds a small “anticipation offset” to emotive weighting for domain-lane Y as soon as X crosses threshold. The PFC also intensifies bridging transforms for domain-lane Y if HPC repeatedly sees Y synergy after X’s spike. System 100 then produces a mini “chain-of-thought,” e.g., “Because HPC replay found X -> Y in 80% of episodes, raises bridging and emotive weighting for Y upon detecting X’s spike”. This is stored 30 at memory unit 118 and/or output at output interface 120 for monitoring of system 100.

[0866] [0255] System 100 integrates causal and symbolic reasoning into HPC replay, resulting in advancement over traditional AI systems, which typically rely on correlation-based anomaly detection or single-step classification. Traditional AI systems often detect anomalies based on patterns observed within a single pass or static data window, lacking the inherent capability to 35 predict how changes in one domain might influence another unless explicitly modelled, such as with specialized RNNs. In contrast, the causal layer of system 100 enables the inference of cause-effect relationships, allowing future surges to be anticipated more effectively. For example, if a sensor-lane spike occurs at time t, System 100 predicts that a synergy or emotive weighting in a related domain-lane will likely surge at time t+1. This capability allows System 40 100 to accelerate detection by a time step or two, which is beneficial in time-sensitive domains

[0867] 349371

[0870] 39

[0872] like health, energy, cybersecurity, and critical sensor monitoring. Even a single-step lead can be the difference between major disruption and timely intervention.

[0874] System 100 detects and reduces missed anomalies. Previously, HPC replay could highlight multi-step anomalies but lacked a direct method to label certain domain-lane changes 5 as causative. Now, with the ability to detect repeated cause-effect links, System 100 proactively adjusts emotive weighting or bridging transforms for related domains, resulting in fewer missed anomalies. This improvement is particularly significant in scenarios where cause-effect relationships drive the main anomalies, such as in health risk assessment or sensor fault detection. Quantitatively, identifying cause-effect relationships can reduce missed anomalies 10 by an additional 10–30% compared to correlation-only HPC logging, depending on the prevalence and strength of the multi-step triggers.

[0876] The incorporation of causal reasoning also helps reduce false alarms through rational overrides. Traditional systems might trigger too many alerts if a domain-lane frequently spikes but the synergy correlation is ambiguous. System 100 optionally determines that a spike in one 15 domain-lane only matters if it is followed by a surge in a related domain-lane. This selective alerting mechanism significantly reduces false positives, as System 100 only raises alarms if HPC replays confirm the subsequent synergy rise. Depending on domain overlap, this can lead to a 10–25% reduction in false alarms, filtering out single-lane spikes that do not cause actual multi-lane synergy surges.

[0878] 20 [0258] The causal and symbolic reasoning capabilities of output interface 120 also enhance interpretability and chain-of-thought, addressing a common limitation of typical AI systems, which often function as black boxes. Even advanced models like LSTMs rarely produce explicit cause-effect statements. System 100 preferably generates mini-statements such as "If domainlane X crosses threshold, domain-lane Y synergy rises next step (80% confidence)." This 25 transparency allows users to understand why System 100 prioritizes certain lanes or synergy signals, fostering trust and enabling more robust auditing. While interpretability may not directly translate to a single performance metric, it significantly impacts adoption and effectiveness in mission-critical fields like energy, healthcare, and cybersecurity.

[0880] For example, System 100 sets symbolic rules when it detects consistent cause-effect 30 patterns, adjusting synergy or emotive weighting accordingly. This capability allows System 100 to reorganize synergy for new cause-effect triggers more quickly, potentially saving multiple offline cycles and accelerating the time-to-optimal by 20–40%.

[0882] Over repeated HPC cycles, cause-effect rules help System 100 refine synergy more selectively, stabilizing bridging transforms and emotive weighting faster. The new cause-effect 35 rules also produce more human-like logic, allowing real-world teams to verify or tune these rules, further boosting real-world efficacy beyond raw numeric improvements. Beneficially, system 100 is more robust, adaptive, and explanatory than current standard AI solutions.

[0884] [0261] Preferably, system 100, e.g. processing modules 104, integrate dual-mode processing. The dual-mode processing mechanism allows for any processing module or the system 100 40 overall to alternative between (1) a decision making mode, optimised for high speed and high accuracy decision making in dynamic environments, and (2) a reflective processing mode, an analysis phase for refining predictions, reducing errors and enhancing long-term learning. This

[0885] 349371

[0888] 40

[0890] allows system 100 to function both reactively and proactively, significantly improving its ability to refine causal models over time.

[0892] For example, the dual-mode processing mechanism comprises:

[0894] a rational processing layer, e.g. Decision_final(t) = Σ(Logical_input * Causal_weight), or 5 processing real-time logical data through weighted causal reasoning;

[0896] an emotional weighting layer, e.g. Urgency_score = Σ(Historical_success_rate * Sentiment_factor), for adjusting decision-making urgency based on contextual factors and/or historical success rates;

[0898] a reflection state, e.g. Refinement_state(t) = Past_decisions (Validated_causal_links * 10 Confidence_factor), for processing prior decisions, preferably in a computational environment without limiting time constraints; and

[0900] a grooming loop, e.g. Learning_adjustment = Learning_rate * (Observed_error -Predicted_error), for error reduction by adjusting predictive models based on feedback to improve long-term accuracy.

[0902] 15 [0263] By dynamically switching between real-time processing and reflective learning, both short-term accuracy and long-term adaptability are improved. While traditional AI systems struggle with frequent recalibration needs, dual-mode processing results in system 100 continuously refining itself, requiring less manual intervention. The combination of logical inference (rational processing) and introspective learning (reflective mode) mirrors human 20 cognition, making system 100 uniquely capable of emotionally adaptive and high-stakes decision-making.

[0904] Figure 2 illustrates a flowchart of method 200, which outlines a process for processing data using an artificial intelligence system comprising a plurality of domain-specific processing modules, each employing hierarchical neural networks. The method demonstrates how input 25 data is processed, refined, integrated, and transformed into actionable predictions through a sequence of adaptive and modular steps.

[0906] [0265] Method 200 comprises steps 202 to 218, beginning with receiving input data (step 202) and processing it through hierarchical neural networks in the domain-specific modules (step 204). These networks dynamically adjust layer depth and dimensional progression to extract 30 features of increasing complexity. Layer outputs are iteratively refined using recursive feedback (step 206), and offline pruning (step 208) optimises the network by deactivating lowcontributing connections. Modular outputs are then modified and integrated using weighted bridging transformations (step 210), followed by the weighting of modular outputs based on relevance scores (step 212). If the system comprises context signals and emotive weighting, 35 these factors are used to adjust the weighting parameters dynamically (step 214). The refined and integrated outputs are consolidated into a probabilistic output function (step 216), and a prediction is selected based on the highest probable output (step 218). Optional steps include storing intermediate and final outputs in memory logs for offline reconfiguration or analysis (step 220) and outputting the prediction for downstream actions (step 222). This sequence 40 ensures that data is processed adaptively and effectively, producing contextually relevant predictions.

[0907] 349371

[0910] 41

[0912] Step 202 of method 200 comprises receiving input data from various domains and preparing it for processing within the artificial intelligence system. This step optionally encompasses capturing raw data from diverse sources and structuring it into representations compatible with the system’s processing capabilities. Depending on the application, these data 5 sources can include imaging devices, genomic sequencing platforms, sensor systems, or databases containing structured and unstructured data. By ensuring that raw input is correctly formatted and routed to appropriate processing modules, this step facilitates the seamless integration of domain-specific data into the broader analytical workflow.

[0914] In a cancer detection scenario, step 202 involves acquiring imaging data, such as X-10 rays or MRIs, as high-dimensional pixel arrays; genomic sequences, which may be represented as encoded nucleotide patterns; and clinical records, formatted as numerical indicators or categorical data points. This diverse input is then organised into domain-specific vectors, ensuring that each type of data is appropriately structured for further analysis by hierarchical neural networks. This preparation ensures the system’s downstream processes operate 15 efficiently and accurately, providing a robust foundation for adaptive data processing.

[0916] Step 204 of method 200 comprises processing the input data through a hierarchical neural network within a domain-specific processing module. This step focuses on progressively extracting features of increasing complexity through a series of layers, where each layer transforms its output based on the results of previous layers. The hierarchical structure enables 20 the system to refine the raw input data into abstract representations that capture domainrelevant patterns and insights.

[0918] The neural network used in this step comprises multiple layers, such as convolutional layers for spatial data like imaging, recurrent layers for sequential data like genomic sequences, or fully connected layers for structured clinical data. Each layer performs dimensional 25 transformations, adapting the raw input into intermediate representations that become increasingly tailored to the specific characteristics of the domain. For example, in a cancer detection application, an imaging module might use convolutional layers to detect low-level features such as edges, followed by deeper layers to identify complex shapes or anomalies indicative of tumours. Similarly, a genomic module might process nucleotide sequences 30 through attention mechanisms to highlight mutations or structural variations.

[0920] The depth and dimensional progression of the hierarchical network are dynamically adjustable in response to the complexity of the domain-specific data. This ensures that simpler data domains are processed efficiently with fewer layers, while more complex data can utilise deeper networks to capture intricate patterns. The ability to adapt the architecture dynamically 35 allows the system to balance computational efficiency with analytical precision, ensuring robust feature extraction across diverse domains. By transforming raw input into high-level abstract representations, step 204 establishes the groundwork for integrating domain-specific insights in subsequent steps.

[0922] [0271] Step 206 of method 200 involves refining the layer outputs of the hierarchical neural 40 network using recursive feedback. This step enhances the system's accuracy and adaptability by iteratively improving the quality of the representations generated at each layer. Recursive feedback ensures that any discrepancies or errors in intermediate outputs are corrected, aligning the feature extraction process with the desired outcomes.

[0923] 349371

[0925] 42

[0927] During this step, feedback signals are propagated backward through the network layers, allowing each layer to recalibrate its weights and biases based on the results of subsequent layers or the system’s broader performance objectives. For example, in a cancer detection scenario, if the neural network fails to correctly identify a tumour-like anomaly in an 5 image, the feedback mechanism generates corrective signals to adjust the convolutional filters in earlier layers. This refinement enables the network to detect subtle features more accurately in future iterations.

[0929] The recursive feedback process operates in real time during active data processing, allowing the system to dynamically adapt to variations in the input data. This refinement step is 10 particularly critical in complex or noisy data environments, where initial outputs may require iterative improvement to achieve high precision. By continuously enhancing layer outputs, step 206 supports the system’s ability to produce robust, domain-specific representations that form the basis for downstream integration and decision-making.

[0931] Step 208 of method 200 comprises pruning neural connections during offline 15 reconfiguration or periods of reduced layer activity. This step is aimed at optimising the hierarchical neural network by selectively deactivating connections that contribute minimally to the overall performance of the system. Pruning ensures that the network maintains computational efficiency without sacrificing predictive accuracy.

[0933] The pruning process begins with calculating a pruning threshold for each layer, which 20 is determined based on metrics such as the contribution of individual weights to the layer’s outputs or the frequency of their activation during processing. Connections or weights that fall below this threshold are identified as low-contributing and are selectively deactivated. For instance, in a cancer detection application, convolutional filters in an imaging module might be pruned if their activation consistently fails to correlate with meaningful features, such as tumour 25 edges or shapes.

[0935] This step leverages data logged during the system's operation, such as intermediate outputs and usage patterns stored in memory units. By analysing these logs, the pruning mechanism identifies underperforming neural connections while preserving those important for extracting critical domain-specific features. Additionally, pruning is performed offline or 30 during periods of low activity, minimising any impact on real-time performance.

[0937] Step 208 not only reduces the computational load of the system but also enhances its adaptability by allowing the network to reallocate resources to more impactful connections. This process is especially valuable in environments where system demands or data characteristics evolve over time, ensuring that the neural networks remain efficient and 35 optimised for their specific tasks.

[0939] Step 210 of method 200 involves modifying the layer outputs of the hierarchical neural networks by integrating transformed outputs from other domain-specific processing modules through weighted bridging transformations. This step ensures that the system synthesises diverse, domain-specific insights into a unified representation, enabling cross-domain 40 collaboration and enhancing the overall predictive capability.

[0941] [0279] The bridging process applies transformation matrices or functions to align the outputs of different modules in terms of dimensions and semantics. These transformations ensure that

[0942] 349371

[0944] 43

[0946] the outputs from distinct domains, such as imaging, genomic, and clinical data, can be coherently combined. For instance, in a cancer detection scenario, the abstract representations from an imaging module that identifies tumour shapes are integrated with outputs from a genomic module detecting relevant mutations. The bridging transformations 5 allow these representations to be harmonised into a single synergy vector that captures both spatial and genetic patterns indicative of cancer.

[0948] Each bridging transformation is weighted to reflect the relevance or significance of the modular outputs being integrated. The weights may be dynamically adjusted based on context signals or relevance scores calculated in earlier steps. By applying these transformations, the 10 system ensures that critical features from different domains are prioritised appropriately during integration.

[0950] Step 210 is beneficial for enabling a holistic analysis of multimodal data, ensuring that the system benefits from the specialised processing capabilities of each domain-specific module while synthesising their outputs into a coherent and actionable format for subsequent 15 probabilistic reasoning and decision-making.

[0952] Steps 208 and 210 work together to enhance the efficiency, precision, and adaptability of the system by optimising the neural network’s structure and enabling effective cross-domain integration, resulting in outcomes that are greater than the sum of their parts. Step 208, which involves pruning neural connections, ensures that the hierarchical neural networks in the 20 processing modules focus only on the most relevant and impactful features of their domainspecific data. This streamlines the processing capabilities, reducing computational overhead while maintaining or even improving the quality of the extracted features.

[0954] Step 210 builds on this optimisation by integrating the refined, domain-specific outputs through weighted bridging transformations. By combining the outputs of multiple, pruned 25 neural networks, the system synthesises a unified representation that captures the most significant aspects of each domain. The weighted bridging process ensures that the critical features retained by each module after pruning are harmonised in a way that highlights their combined importance.

[0956] For example, in a cancer detection scenario, step 208 ensures that the imaging module 30 focuses on important tumour patterns while eliminating redundant or low-contributing filters, and the genomic module emphasises key genetic markers. When these optimised outputs are integrated in step 210, the resulting synergy vector captures complementary insights from imaging and genomic data, providing a more comprehensive and accurate representation of potential malignancies than either domain could achieve alone.

[0958] 35 [0285] This interplay between pruning and bridging not only reduces resource consumption but also ensures that the system maximises the value of the most informative features across all domains. By concentrating computational and analytical efforts on the most impactful data and combining them intelligently, the system achieves a heightened level of performance, making its predictions more reliable and actionable.

[0960] 40 [0286] Step 212 of method 200 involves weighing the modular outputs generated in parallel by the domain-specific processing modules. This step ensures that the contributions of each

[0961] 349371

[0963] 44

[0965] module to the system’s overall prediction are appropriately calibrated based on their relevance to the specific context or task.

[0967] After the modular outputs are generated by the processing modules in step 204 and integrated through bridging transformations in step 210, each output is assigned a relevance 5 score. These scores reflect the importance of the outputs in the context of the data being analysed or the system’s broader objectives. For instance, in a cancer detection scenario, if imaging data reveals a clear tumour-like structure, its modular output may be assigned a higher relevance score compared to other domains, such as genomic data, which may show less significant anomalies.

[0969] 10 [0288] The relevance scores are then used to adjust the weights applied to each modular output, amplifying the influence of highly relevant domains while reducing the impact of less critical ones. This dynamic weighting ensures that the system’s subsequent probabilistic reasoning and decision-making are informed by the most significant insights available. Step 212 also interacts with step 214 if context signals or emotive weighting are included, further 15 refining the weighting process based on real-time conditions or system priorities.

[0971] By calibrating the importance of modular outputs, step 212 ensures that the system maintains a balanced and contextually appropriate approach to multimodal data analysis, supporting accurate and actionable predictions in diverse application scenarios.

[0973] Step 214 of method 200 involves adjusting the bridging and modular weighting 20 parameters based on context signals and, if included, emotive weighting. This step allows the system to dynamically prioritise specific data streams or features depending on real-time conditions, task objectives, or the significance of certain inputs.

[0975] Context signals are derived from the broader operational environment or the characteristics of the input data itself. These signals may indicate urgency, anomalies, or 25 specific patterns that require additional attention. For example, in a cancer detection application, if genomic data reveals a critical mutation associated with a high likelihood of malignancy, the system adjusts the modular weighting parameters to prioritise the outputs of the genomic processing module. Similarly, context signals might influence bridging transformations by increasing the emphasis on integrating imaging and genomic outputs, given 30 their combined diagnostic value in this scenario.

[0977] If the system comprises emotive weighting, this step also incorporates urgency or significance levels to amplify or diminish the influence of certain outputs. For instance, an anomaly detected in imaging data indicative of a potentially life-threatening tumour could trigger an emotive weighting signal that increases its impact on the final prediction. This 35 ensures that the system prioritises critical insights without overloading the analysis with less relevant information.

[0979] [0293] By dynamically adjusting weighting parameters, step 214 enables the system to adapt to changing conditions and focus on the most significant aspects of the data. This ensures that the final outputs remain contextually relevant and optimally prioritised, enhancing the system’s 40 ability to deliver accurate and impactful predictions.

[0980] 349371

[0982] 45

[0984] Step 216 of method 200 involves integrating the layer outputs, recursive feedback, domain-specific modular outputs, and bridging transformations into a probabilistic output function. This step consolidates the diverse and refined data streams from earlier processing steps into a single probabilistic framework, allowing the system to quantify the likelihood of 5 various potential outcomes.

[0986] The integration process begins by synthesising the modular outputs generated in step 204, incorporating refinements from recursive feedback (step 206) and optimisations such as pruning (step 208). These modular outputs, transformed and harmonised through weighted bridging transformations (step 210), are combined into a unified representation. This unified 10 representation serves as the input to the probabilistic function unit, which applies statistical models or neural network-based inference techniques to compute probabilities for each possible prediction. For instance, in a cancer detection application, the system might analyse the combined insights from imaging, genomic, and clinical data to assign probabilities to outcomes such as "benign", "malignant", or "inconclusive".

[0988] 15 [0296] The probabilistic output function provides a confidence score for each potential prediction, reflecting the relative likelihood of each outcome based on the integrated data. This quantitative representation is essential for the subsequent selection step (step 218), where the system identifies the most probable outcome. By unifying diverse data streams into a coherent probabilistic framework, step 216 ensures that the system’s predictions are grounded in robust, 20 multimodal analysis, enabling reliable and context-aware decision-making.

[0990] Step 218 of method 200 involves selecting a prediction based on the highest probable output of the probabilistic output function. This step is the culmination of the processing workflow, where the system consolidates all preceding analysis and refinement into a singular, actionable outcome.

[0992] 25 [0298] The probabilistic output function, generated in step 216, provides confidence scores for various potential predictions, reflecting the likelihood of each outcome based on the combined insights from all domain-specific processing modules, recursive feedback, and bridging transformations. In step 218, the system evaluates these scores and selects the prediction with the highest probability. For instance, in a cancer detection application, the 30 probabilistic output might assign probabilities to different diagnoses, such as "benign tumour" (30%), "malignant tumour" (65%), and "unknown condition" (5%). The system selects "malignant tumour" as the final prediction because it has the highest likelihood.

[0994] This step ensures that the system’s decision-making is deterministic and reliable, translating complex, multimodal analysis into a clear and actionable result. The selection 35 process may also involve additional criteria, such as threshold confidence levels or secondary contextual factors, to ensure robustness and applicability across diverse operational scenarios. By providing a definitive prediction, step 218 enables the system to deliver meaningful insights for downstream actions or further analysis.

[0996] [0300] Step 220 of method 200 is an optional step that involves storing intermediate and final 40 outputs in a memory log for various purposes, including offline replay, hierarchical reconfiguration, layer tuning, or layer pruning. This step provides a mechanism for preserving

[0997] 349371

[1000] 46

[1002] the system’s operational data, enabling analysis, optimisation, and reconfiguration beyond the immediate processing cycle.

[1004] During this step, the system logs data such as the outputs of individual processing modules, refined representations generated through recursive feedback, modular weights, and 5 the probabilistic output function. For example, in a cancer detection scenario, the memory log might store intermediate feature maps from the imaging module, genomic anomaly scores, and the final prediction probabilities. These logs allow the system to retain a detailed record of its decision-making pathway, facilitating retrospective evaluation.

[1006] The stored data is beneficial for offline replay, where past operations are reanalysed to 10 identify patterns or anomalies that can inform future processing. It also supports hierarchical reconfiguration, enabling adjustments to the layer depth or structure of processing modules to align with evolving data complexities. Additionally, the memory log is used in layer tuning and pruning, allowing the system to fine-tune parameters or deactivate low-contributing connections based on historical performance metrics.

[1008] 15 [0303] By maintaining a comprehensive record of intermediate and final outputs, step 220 enhances the system’s adaptability and long-term efficiency. It ensures that the system can learn from past operations, refine its architecture, and continually improve its performance, making it more effective in handling complex and evolving datasets.

[1010] Step 222 of method 200 is an optional step that involves outputting the prediction for 20 one or more downstream actions. This step ensures that the results of the system’s analysis and decision-making are delivered in a format that can be readily used by external systems or end-users for further processing, action, or decision-making.

[1012] In this step, the selected prediction, determined in step 218, is formatted and transmitted to downstream systems through the output interface. The format and delivery 25 mechanism depend on the specific application. For instance, in a cancer detection scenario, the system might output the prediction "malignant tumour" along with associated probabilities, contributing features (e.g., imaging anomalies or genetic markers), and potential recommendations for follow-up actions. This output can be delivered as a structured data stream via an API, a formatted report for medical professionals, or a visual representation for 30 integration into decision-support tools.

[1014] Step 222 allows the system to interface seamlessly with external environments, such as medical record systems, automated diagnostic platforms, or alerting mechanisms. By providing actionable insights in an accessible format, this step ensures that the system’s predictive capabilities translate into meaningful real-world applications. Furthermore, the 35 flexibility of this step enables integration with a wide range of workflows, making the system adaptable to diverse operational contexts.

[1016] [0307] The method optionally comprises generating sub-lane outputs by subdividing domains into sub-lanes, where each sub-lane handles a specific aspect or subset of the domain's data. For instance, an imaging domain may have sub-lanes for detecting edges, texture, and density 40 features. The outputs from these sub-lanes are subjected to bridging transformations, scaled using optional emotive weighting if present, and summed into a synergy vector. This synergy

[1017] 349371

[1019] 47

[1021] vector represents a unified representation of the processed sub-lane outputs and is integrated into the probabilistic output function for downstream decision-making.

[1023] If the system comprises emotive weighting, the method optionally scales the bridging outputs dynamically based on historical false-positive rates or patterns of synergistic 5 contribution. This ensures that the system adjusts its weighting to reduce bias or overemphasis on less reliable sub-lanes, enhancing accuracy and adaptability.

[1025] The method optionally logs the impact of applied emotive weighting on domain outputs or performance metrics to generate feedback. A supervisory module evaluates this feedback, enabling the system to adjust future emotive weighting dynamically. For instance, if historical 10 false-positive rates suggest over-prioritisation of a domain, the supervisory module overrides scaling parameters to balance the contributions from sub-lanes, ensuring improved reliability.

[1027] The method optionally calculates the synergy vector by summing outputs from the domain-specific modules or sub-lanes, scaling these outputs based on domain-specific emotive weighting. A global state is recursively updated using a recursion factor, preserving 15 contextual continuity across processing iterations. The persistence of prior context is evaluated by monitoring the norm of the global state, and the recursion factor is adjusted during offline processes or low-activity phases to optimise context retention.

[1029] The method optionally encodes sequential data from multiple time steps into the memory log, associating each step with inputs, synergy vectors, global states, and downstream 20 actions. During offline replay, the system analyses synergy patterns across steps to identify anomalies and adjust emotive weighting and bridging transformations. Novelty values are evaluated, and if they exceed a threshold, recommendations are generated for creating new sub-lanes, adjusting emotive baselines, or refining bridging transformations. The replay process also refines recursive feedback, pruning thresholds, and domain-lane hierarchies to 25 enhance future operations.

[1031] The method optionally integrates emotive weighting, synergy vectors, and global states across domains using an alignment mechanism to refine bridging transformations. Offline replay identifies anomalous sub-lanes through synergy deviations, evaluates novelty values, and determines whether sub-lanes should be expanded, merged, or pruned. Simulated future 30 domain inputs, based on the memory log, pre-adjust emotive weighting and bridging parameters. Modular relevance scores are enhanced during system operation using historical novelty trends.

[1033] The method optionally analyses the memory log to identify patterns not captured by existing sub-lanes, creating new sub-lanes to address these gaps. Conversely, redundant sub-35 lanes are unified by combining bridging transformations when patterns indicate overlap, ensuring efficient processing without unnecessary duplication.

[1035] [0314] The method optionally detects overactive emotive weighting, where synergy vectors are excessively influenced by emotive adjustments. To mitigate this, the system restricts emotive weighting by enforcing minimum or maximum thresholds, maintaining balance and 40 preventing skewed predictions.

[1036] 349371

[1039] 48

[1041] The method optionally analyses the memory log to detect cause-effect relationships across domain lanes, storing these as symbolic rules in a cause-effect repository. These relationships are used to anticipate and modify synergy vectors and provide symbolic explanations for these modifications, enhancing system transparency and interpretability.

[1042] 5 [0316] The method optionally uses the cause-effect repository during meta-updates, applying proactive adjustments if a cause-effect relationship's confidence value exceeds a threshold. Adjustments include adding pre-emptive emotive weighting offsets for specific domain lanes or scaling bridging transformation matrices to align with anticipated outcomes.

[1044] The method optionally evaluates each domain sub-lane for pruning based on usage 10 metrics, such as contribution to synergy vectors or activation frequency. Sub-lanes are pruned when metrics fall below predefined thresholds, optimising system efficiency without compromising accuracy.

[1046] During offline reconfiguration, the method optionally replays the memory log to analyse multi-step patterns of synergy and emotive signals. Repeated anomalies or overlooked patterns 15 are identified, and recommendations are generated for structural adjustments to bridging transformations, emotive weightings, or sub-lane configurations. Meta-updates, applied via a supervisory module, refine recursion factors, emotive baselines, and system architecture to improve performance.

[1048] The method optionally resets usage statistics, such as invocation counts and synergy 20 or emotive sums, during offline phases. This ensures that subsequent iterations rely on accurate and up-to-date metrics, maintaining system performance and adaptability.

[1050] Method 200 offers significant advancements over existing AI frameworks and processing methods by addressing key limitations in adaptability, efficiency, and integration. Traditional AI models, such as convolutional neural networks (CNNs) and transformers, are 25 often designed as monolithic architectures that perform generalised feature extraction and inference. While these models are highly effective for single-domain tasks, they lack the modularity and specialisation needed for multi-domain or context-sensitive applications. Method 200 overcomes this limitation through the use of domain-specific hierarchical neural networks, where each processing module is tailored to its respective domain and dynamically 30 adjusts its depth and complexity based on the data being processed. This allows method 200 to extract highly relevant and domain-specific features with greater precision, ensuring that each module operates optimally for its input data while minimising resource consumption.

[1052] Another area where method 200 improves upon traditional approaches is its use of recursive feedback (step 206), which provides real-time refinement of neural network outputs.

[1053] 35 Unlike existing models that rely on static pre-trained weights or post-hoc adjustments, method 200 incorporates an iterative feedback mechanism that continuously recalibrates layer outputs during processing. This ensures that the system adapts dynamically to variations or anomalies in the input data, reducing the likelihood of misclassification or underperformance in complex scenarios. For example, in a real-time diagnostic application, recursive feedback enables the 40 system to refine its predictions in response to evolving patterns, such as subtle changes in medical imaging data or shifting genomic indicators, without requiring extensive retraining or manual intervention.

[1054] 349371

[1056] 49

[1058] The incorporation of pruning as an iterative, threshold-based optimisation step (step 208) also represents a significant improvement over conventional techniques. While methods like deep compression or static pruning reduce model size and computational cost after training, they are typically applied as one-time operations and may fail to account for changes 5 in data distributions or task requirements. Method 200 integrates pruning as a dynamic and context-aware process during offline reconfiguration phases, ensuring that low-contributing neural connections are deactivated while preserving the most impactful features. This capability allows the system to remain computationally efficient and scalable, even in environments with evolving data patterns or resource constraints.

[1060] 10 [0323] Method 200 also addresses limitations in cross-domain data integration, which is often a challenge in existing multi-modal AI frameworks. Traditional systems that combine outputs from different data domains tend to rely on rigid architectures or simple concatenation techniques, which can lead to loss of domain-specific nuances or suboptimal synergy between domains. Method 200’s use of weighted bridging transformations (step 210) ensures that 15 outputs from domain-specific modules are harmonised in a coherent manner, with weights dynamically adjusted based on contextual relevance or system priorities. This approach allows the system to synthesise complementary insights from diverse domains, such as combining imaging data with genomic and clinical information in a healthcare application, producing a unified representation that is richer and more actionable than the outputs of individual domains 20 alone.

[1062] The optional inclusion of emotive weighting further enhances the adaptability and decision-making capability of method 200, particularly in high-stakes or context-sensitive scenarios. Unlike traditional AI models, which typically apply uniform importance to all inputs, method 200 can dynamically scale modular and bridging weights based on urgency signals or 25 historical performance feedback. This allows the system to prioritise critical data streams or features, ensuring that predictions reflect the most relevant and impactful aspects of the data. Moreover, the ability to log and analyse the effects of emotive weighting (as described in optional steps) ensures that the system evolves intelligently over time, refining its weighting mechanisms based on empirical performance metrics and reducing the risk of overemphasis 30 on less critical domains.

[1064] By integrating these features into a cohesive framework, method 200 achieves a level of modularity, adaptability, and efficiency that surpasses current AI frameworks. It is particularly well-suited for applications requiring multi-domain analysis, real-time adaptability, and longterm optimisation, offering a robust and scalable solution for complex, dynamic, and high-35 dimensional data environments.

[1066] As a non-limiting example for explanation, there is provided a mathematical implementation of method 200.

[1068] The method 200 comprises:

[1070] step 202, receiving input data (ݔ);

[1072] 40 step 204, processing the input data through the hierarchical neural network, wherein:

[1073] 349371

[1075] 50

[1077] the hierarchical neural network comprises a plurality of layers ݈= 1, 2, … ,ܮ configured to extract features of increasing complexity; and

[1079] each layer transforms its output according to:

[1081] ܪ <(௟) >= (݂ܹ <(௟) >∙ ܪ <(௟ି ଵ) >+ <(ܾ௟)>), wherein:

[1083] 5 ܪ <(௟) >is the output of layer ,݈ ܪ <(௟ି ଵ) >is the input to layer ݈from a preceding layer, ܹ <(௟) >is the weight matrix for layer ,݈ <(ܾ௟) >is the bias vector for layer ,݈ and ݂ is an activation function;

[1085] dimensional transformations occur across layers from raw input (ܪ <(଴) >= ݔ) to final abstract representations (ܪ <(௅)>); and

[1087] 10 the layer depth (ܭ) and dimensional progression (dim<଴>, dim<ଵ>, ⋯ , dim<௄>) are dynamically adjustable based on domain-specific complexity;

[1089] step 206, refining layer outputs (ܪ <(௟)>) using recursive feedback (߂ܪ<୤ୣ ୣୢ ୠୟୡ୩>) and applying the recursive feedback to the output of layer ,݈ computed as:

[1091] 15 ߂ܪ<୤ୣ ୣୢ ୠୟୡ୩ >= ܪ<(௟ି ଵ)>

[1092] <୤ୣ ୣୢ ୠୟୡ୩ >− ܪ <(௟ି ଵ)>, wherein ܪ<(௟ି ଵ)>

[1093] <୤ୣ ୣୢ ୠୟୡ୩ >is feedback-adjusted output from the preceding layer; and

[1095] ܪ <(௟)>୰ୣ ୤୧୬ୣୢ

= ܪ <(௟) >+ ߚ ∙ ߂ܪ<୤ୣ ୣୢ ୠୟୡ୩>, wherein ߚ is a scaling factor for feedback incorporation;

[1097] step 208, pruning neural connections during offline reconfiguration or periods of 20 reduced layer activity, wherein pruning is performed by:

[1099] calculating a pruning threshold ( )߬ for each layer:

[1101] <߬= ߤௐ ߣ ⋅ ߪௐ ;>

[1103] <wherein ߤௐ is the mean average of the weight distribution, ߣ is a scaling >factor for the layer, and ߪ<ௐ >is the standard deviation of the weight 25 distribution; and

[1105] selectively deactivating weights (ܹ ) that are low contributing based on the pruning threshold:

[1107] ܹ <ᇱ |ௐ |ିఛ>= ܹ ⋅ ߪ ቀ ఑ ቁ;

[1109] wherein ܹ <ᇱ >is the pruned weight matrix, ܹ is the original weight matrix, 30 ߪ is a sigmoid function, and ߢ is a sensitivity scaling factor;

[1111] step 210, modifying the layer outputs by integrating transformed outputs from other processing modules using bridging transformations:

[1113] ܪ<(௅)ᇱ>

[1114] <௜ >= ܪ<(௅) )>

[1115] <௜ >+ ߙ<௜௝ >∙ ܹ <௜௝ >∙ ܪ<(௅>

[1116] <௝ >, wherein:

[1118] ܪ<(௅)>

[1119] <௜ >and ܪ<(௅)>

[1120] <௝ >are high-level outputs of modules ݅and ,݆

[1121] 349371

[1124] 51

[1126] ܹ <௜௝ >is the bridging transformation matrix for mapping outputs from one module to another, and

[1128] ߙ<௜௝ >is an intensity coefficient for scaling contributions of the transformed outputs;

[1130] 5 step 212, if the artificial intelligence system comprises context signals and emotive weighting, adjusting the intensity coefficient based on the context signals, such that:

[1132] ߙ<௜௝ >= ߛ(ܧ<௜>,ܧ<௝>) ∙ ൫1 <௝݁,௧>൯;

[1134] wherein ܧ<௜>and ܧ<௝ >are context signals associated with modules ݅and ,݆

[1135] ߛ(∙) is a gating function that maps context signals to scaling coefficients, and

[1136] 10 <௝݁,௧ >represents domain-specific emotive weighting of module ݆at time ݐ;

[1138] step 214, generating modular outputs (ܴ<௜>( <୧୬ܵ ୮୳୲>)) from domain-specific modules in parallel and weighting the modular outputs using relevance scores (ℛ<௜>), wherein:

[1140] <୶ୣ୮ೢ ೔>

[1141] ℛ<௜>= <∑೙>

[1142] <ೕసభ ୶ୣ୮ೢ ೕ >; and

[1144] ݓ<௜ >is the task relevance score of module ݅and ݊ is the total number of modules;

[1145] 15 step 216, integrating the layer outputs, recursive feedback, domain-specific modules, and bridging transformations into a probabilistic output, wherein:

[1147] ܲ(ܿ ∣ ݔ) = softmax(<ଵ ௡>

[1149] <ଵ >ℛ <(௅)ᇱ>

[1150] ௓<∑>௜ୀ ௜௜

∙ ቂܪ<௜ >+ ߂ܪ<୤ୣ ୣୢ ୠୟୡ୩ >+ ܴ<௜>( <୧୬ܵ ୮୳୲>)ቃ ,

[1152] ܲ(ܿ ∣ ݔ) is the probability of class ܿgiven input ݔ, and

[1154] ܼ is a normalisation factor;

[1156] 20 step 218, selecting a prediction (ݕො) based on the highest probability:

[1158] ݕො= argmax<௖∈஼ >ܲ(ܿ ∣ ݔ),

[1160] wherein ܿ is the set of possible classes;

[1162] step 220, storing intermediate and final outputs (ܪ <(௟)>, ߂ܪ<୤ୣ ୣୢ ୠୟୡ୩>, ܴ<௜>( <୧୬ܵ ୮୳୲>)) in a memory log for any of offline replay, hierarchical reconfiguration, layer tuning, or layer 25 pruning; and

[1164] step 222, outputting the prediction (ݕො) for one or more downstream actions.

[1166] The method 200 optionally further comprises:

[1168] generating sub-lane outputs (ݖ<௠ ,௡,௧>) for each domain ܦ<௠ >by subdividing domains into ܰ<௠ >sub-lanes, wherein each sub-laneݏ<௡ >within domainܦ<௠ >produces a final vectorݖ<௠ ,௡,௧ >30 of dimension <௠݀ ,௡>;

[1169] 349371

[1170] 52

[1172] applying bridging transformations to sub-lane outputs, wherein a bridging matrix ܹ<௠ ,௡ >maps each sub-lane output ݖ<௠ ,௡,௧ >into a shared synergy space of dimension <௦݀>, such that:

[1174] <௠ܾ ,௡,௧>= ܹ<௠ ,௡ >∙ ݖ<௠ ,௡,௧>,

[1175] 5 wherein <௠ܾ ,௡,௧ >is the bridging output for domain ܦ<௠ >, sub-lane ݏ<௡>, at time ݐ;

[1176] if the artificial intelligence system comprises emotive weighting, scaling bridging outputs, by calculating:

[1177] ෨<௠>ܾ <,௡,௧>= <௠ܾ ,௡,௧ >∙ ݂൫݁ <௠ ,௧>൯, wherein ݂൫݁ <௠ ,௧>൯is a scaling function;

[1178] else <෨>௠<ܾ >,௡,௧

= <௠ܾ ,௡,௧>;

[1179] 10 summing bridging outputs into a synergy vector ( <௧ܸ>), calculated by:

[1181] <௧ܸ= >∑<ெ>௠ ୀଵ

∑<ே೘>

[1182] <௡ୀଵ >෨<௠>ܾ <,௡,௧ ,>

[1183] wherein ܯ is the total number of domains; and

[1184] integrating the synergy vector <௧ܸ >such that:

[1185] <ଵ >ܲ(ܿ ∣ ݔ) = softmax( <௡>

[1186] ௓<∑>௜ୀଵ

ℛ<௜௜>∙ ൣܸ <௧>+ ߂ܪ<୤ୣ ୣୢ ୠୟୡ୩ >+ ܴ<௜>( <୧୬ܵ ୮୳୲>)൧.

[1187] 15 Optionally, emotive weighting to bridging outputs is dynamically scaled by:

[1188] applying a nonlinear transformation:

[1189] ݂൫݁ <௠ ,௧>൯= 1 <௠݁ ,௧>;

[1190] ݂൫݁ <௠ ,௧>൯= exp ( <௠݁ ,௧>); or

[1191] <0 1 ௠݁ ,௧≤ 0>

[1192] <݂൫݁ >௠ ,௧<൯= ቐ1 >௠݁ ,௧<0 < 1 ௠݁ ,௧≤ max ;>

[1193] <max 1 ௠݁ ,௧> max>

[1194] 20 logging an impact of applied emotive weighting on domain outputs or performance metrics for generating feedback; and

[1195] adjusting future emotive weighting, wherein a supervisory module evaluates performance feedback and overrides scaling parameters <௠݁ ,௧ >for domain ܦ<௠ >based on historical false-positive rates or synergistic contribution patterns.

[1196] 25 The method 200 optionally further comprises:

[1197] calculating the synergy vector <௧ܸ >by summing outputs ܱ<௠ ,௡,௧ >from the domain-specific modules or <௝ܰ >sub-lanes and scaling the outputs based on domain-specific emotive <ே >weighting <௝݁,௡>, such that <௧ܸ>= <∑ே>

[1199] <௝ୀଵ >∑ ೕ

[1201] <௡ୀଵ ௝݁,௡ ௝ܱ,௡,௧ >;

[1202] updating a global state <௧ܵ >recursively using the equation:

[1203] 30 <௧ܵ>= (݂(1 − ߙ) <௧ିܵ ଵ >+ ߙ <௧ܸ>),

[1204] 349371

[1207] 53

[1209] wherein ߙ is a recursion factor (0 < ߙ < 1), wherein ݂ is a non-linear activation function;

[1211] evaluating persistence of prior context by monitoring the norm of the global state <௧ܵ >and adjusting the recursion factor ߙ during offline processes or periods of reduced 5 activity.

[1213] The method 200 optionally further comprises:

[1215] encoding sequential data from multiple time steps into the memory log, wherein each time step is associated with a step-wide record comprising: time ݐ, domain input <(௜)>

[1216] <௝ܺ,௧>, emotive weighting <௝݁,௧>, synergy vector <௧ܸ>, global state <௧ܵ>, and the one or more 10 downstream actions;

[1218] conducting an offline replay of the memory log during reduced activity phases, wherein:

[1220] synergy patterns are analysed across consecutive steps for identifying multistep anomalies; and

[1222] emotive weighting and bridging transformations are adjusted based on 15 cumulative signals or repeated patterns;

[1224] <evaluating a novelty value during the offline replay, computed as:>

[1226] <novelty = ∑| ௧ܸ| ∑ ௝݁,௧, and>

[1228] generating recommendations for creating new domain sub-lanes, adjusting emotive baselines, or refining bridging transforms if novelty exceeds a 20 threshold; and

[1230] refining recursive feedback, pruning thresholds, or domain-lane hierarchy during subsequent system operations based on the offline replay.

[1232] The method 200 optionally further comprises:

[1234] integrating emotive weighting <௝݁,௧>, synergy vectors <௧ܸ >and global states <௧ܵ>across 25 multiple domains to detect cross-domain patterns using an alignment mechanism for refining bridging transformations ܤ<௜,௝>, wherein:

[1236] ܤ <ᇱ>

[1237] <௜,௝ >= ܤ<௜,௝>+ ߙ ∙ diag(∆ <(௜,௝)>

[1238] <௧ܸ >) ∙ <௧ܵ>, and

[1240] ܤ<௜,௝>ᇱ

is the updated bridging matric between domains ݅and ,݆ ߙ is a learning rate for alignment, and ∆ <(௜,௝)>

[1241] <௧ܸ >is synergy deviation between domains;

[1243] 30 during the offline replay, identifying the anomalous sub-lanes <(௜)>

[1244] <௞݈ >by analysing synergy deviations:

[1246] <ብ௏ ഥ ብ೟,೗(೔)ି௏>

[1247] ߜ = <ೖ ೗(೔)>

[1248] <ೖ>

[1249] ௟(೔)

[1250] <ೖ >ఙ<ೇ >,

[1251] ೗<(೔)>

[1252] ೖ

[1253] 349371

[1255] 54

[1257] wherein ത௟ܸ(೔) and ߪ<௏>

[1258] <ೖ ) >are the mean and standard deviation of the synergy ೗<(೔>

[1259] ೖ

[1260] vectors for sub-lane <(௜)>

[1261] <௞݈ >, and

[1263] pruning or refining sub-lane parameters if ߜ௟(೔) > ߠ<ୟ୬୭୫ ୟ୪୷>

[1264] <ೖ >;

[1266] re-evaluating novelty values during the offline replay based on sub-lane contributions

[1267] <5 >ฯܸ <(೔) ௧,௟ೖ ฯand cross-domain interactions, such that:>

[1268] <୬୭୴ ୪୲ୣ୷ adjusted novelty = >,

[1269] ∑<ಿ ೗>

[1270] <ೖసభ >ብ௏

[1272] ೟,೗<(೔)ብ>

[1274] ೖ

[1276] wherein ܰ<௟ >is the total number of sub-lanes in the domain;

[1278] expanding or merging sub-lanes <(௜)>

[1279] <௞݈ >when repeated patterns ܲ௟(೔)

[1280] <ೖ >across time steps satisfy the condition:

[1282] 10 merge if ฯܲ ௟(೔) − ܲ <୰ୣ୥ୣ >and expand if ฯܲ (೔) − ܲ(ೕ)

[1283] <ೖ >௟(ೕ)ฯ< ߠ<୫>

[1284] <ೖ ௟ೖ ௟ >ฯ< ୣߠ <୶୮ୟ୬ୢ>;

[1285] <ೖ>

[1287] simulating a future domain input <(௜)>

[1288] <௝ܺ,௧ାଵ >based on the memory log ℳ, using:

[1290] ෠ <௜)>

[1291] <௝>ܺ<(>

[1292] <,௧ାଵ >= <୮݂୰ୣ ୢ>(ℳ,ݐ),

[1294] wherein <୮݂୰ୣ ୢ >is a trained predictive function; and

[1296] a predicted future domain input <෠(௜)>

[1297] <௝>ܺ<,௧ାଵ >is used to pre-adjust emotive weighting 15 and bridging parameters: ¬-

[1298] <ᇱ >௝݁,௧ାଵ

= <௝݁,௧>+ ߚ ∙ ∆ <(௜)>

[1300] <௝ܺ,௧ାଵ >,

[1302] wherein ߚ is a weighting adjustment rate and ∆ <(௜) ௜) (௜)>

[1303] <௝ܺ,௧ାଵ >= <෠>௝<ܺ(>

[1304] <,௧ାଵ >− <௝ܺ,௧>; and

[1305] enhancing modular relevance scores ℛ<௜>during system operation by incorporating historical novelty trends ℵ<௧>:

[1307] 20 ℛ<௜>ᇱ

= ℛ<௜>+ ߛ ∙ norm(ℵ<௧>),

[1309] wherein ߛ is a scaling factor for novelty influence, and norm(ℵ<௧>) is the normalised novelty over a defined historical window.

[1311] Optionally:

[1313] each domain sub-lane is evaluated for pruning based on usage metrics, defined by 25 synergy ratio (ߩ<ୱ୷୬>) and emotive ratio ( ୣߩ <୫୭>):

[1315] ߩ <೔ ୫ୣ ୭>

[1316] <ୱ୷୬ >= <ୱ୷୬ ೔>

[1317] <୫ ୟ୶(୧୬୴೔,ଵ)>, ୣߩ <୫୭ >= <୫ ୟ୶(୧୬୴೔,ଵ)>,

[1319] where syn<௜ >is the cumulative synergy contribution, emo<௜ >is the cumulative emotive contribution, and inv<௜>is the invocation count of the sub-lane; and

[1320] 349371

[1322] 55

[1324] sub-lanes are pruned if both p and p fall below respective thresholds.

[1326] Optionally, method 200 is used for adjusting one or more parameters of a system in real time to improve performance or resource efficiency. This benefit allows the system to dynamically optimize its operations, ensuring that resources are used efficiently and 5 performance is maximized, which is particularly useful in environments with fluctuating demands or conditions.

[1328] Optionally, method 200 is used for controlling a physical or virtual system by modifying operational parameters to achieve an optimized state. This ensures that the system can adapt to changing conditions and maintain optimal performance, reducing the likelihood of 10 suboptimal operation and improving overall efficiency.

[1330] Optionally, method 200 is used for initiating, modifying, or terminating a workflow process involving data processing, device control, or system operations. This flexibility allows the system to dynamically manage workflows, ensuring that processes are executed efficiently and can be adjusted in response to real-time data and conditions.

[1332] 15 [0330] Optionally, method 200 is used for dynamically reconfiguring input data sources or processing pipelines to enhance subsequent predictions. This capability improves the accuracy and relevance of predictions by ensuring that the most appropriate data and processing methods are used, adapting to new information as it becomes available.

[1334] Optionally, method 200 is used for adjusting feedback parameters in the artificial 20 intelligence system to adapt future predictions to changing conditions. This ensures that the system remains accurate and relevant over time, as it can learn from new data and adjust its predictions accordingly.

[1336] Optionally, method 200 is used for identifying and correcting errors, anomalies, or inconsistencies in a system to maintain operational stability or accuracy. This enhances the 25 reliability and accuracy of the system by ensuring that any issues are promptly identified and addressed, preventing them from affecting overall performance.

[1338] Optionally, the recursive feedback signals are bounded by stability constraints to prevent divergence, ensuring that the magnitude of feedback adjustments remains below a predefined threshold during iterative refinement. This prevents the system from becoming 30 unstable due to excessive feedback adjustments, maintaining consistent and reliable performance.

[1340] Optionally, the pruning threshold for each layer is dynamically computed based on the statistical properties of the weight distribution in the layer, with the threshold being determined using the mean and standard deviation of the weights, and adjusted with a scaling factor to 35 control pruning aggressiveness. This ensures that pruning is performed optimally, maintaining model efficiency without compromising accuracy.

[1342] [0335] Optionally, pruned weights are dynamically restored during periods of high activity in the neural network, with the restoration being performed by incrementally reactivating the pruned weights based on their original values and a predefined restoration rate. This allows 40 the system to adapt to varying workloads, ensuring that it can handle increased demand without sacrificing performance.

[1343] 349371

[1346] 56

[1348] Optionally, method 200 further comprises monitoring the activation sparsity of each layer in the hierarchical neural network, with pruning initiated for a layer if the sparsity, measured as the proportion of non-zero activations to the total number of activations, falls below a predefined threshold. This ensures that the network remains efficient by pruning 5 underutilized layers, reducing computational overhead.

[1350] Optionally, the recursive feedback signals are computed using error gradients to refine predictions, with the error gradients being determined by the difference between the predicted outputs and the expected outputs, scaled by a feedback adjustment rate. This improves the accuracy of predictions by continuously refining the model based on the difference between 10 actual and expected outcomes.

[1352] Optionally, the hierarchical neural network comprises an attention mechanism configured to prioritize features within the input data, with attention weights computed based on the relevance of the features to the task-specific context. This ensures that the most relevant features are given priority, improving the accuracy and efficiency of the model.

[1354] 15 [0339] Optionally, the feedback scaling factor is dynamically adjusted based on the magnitude of the initial prediction error, ensuring that feedback corrections are weighted more heavily for larger errors and less heavily for smaller errors. This allows the system to focus on correcting significant errors, improving overall prediction accuracy.

[1356] Optionally, the integration of outputs from the domain-specific modules is performed 20 using a softmax-weighted mechanism, with the relevance of each module’s contribution dynamically determined based on the context of the input data. This ensures that the most relevant module outputs are given priority, improving the overall performance of the system.

[1358] Optionally, the hierarchical neural network is further configured to optimize its feature representations during training using a contrastive loss function, ensuring that similar inputs 25 are mapped to closer representations while dissimilar inputs are mapped to more distant representations. This improves the model's ability to distinguish between different types of inputs, enhancing its overall accuracy.

[1360] Optionally, method 200 further comprises generating confidence scores for the predictions, with the confidence score determined based on the predicted probability of the 30 output class and an associated uncertainty measure derived from variations in model predictions. This provides a measure of the reliability of the predictions, allowing users to make more informed decisions based on the model's output.

[1362] Optionally, method 200 further comprises dynamically skipping one or more layers of the hierarchical neural network during inference, with the skipping based on the predicted 35 contribution of the layers to the final output for reducing computational overhead. This improves the efficiency of the model by avoiding unnecessary computations, reducing processing time and resource usage.

[1364] [0344] Optionally, the gradients used in the recursive feedback mechanism are clipped to a predefined maximum value during training or refinement for numerical stability and gradient 40 explosion prevention. This ensures that the training process remains stable and prevents the gradients from becoming excessively large, which can disrupt the learning process.

[1365] 349371

[1368] 57

[1370] Optionally, only a subset of the domain-specific modules is activated for a given input, with the activation determined based on the task-specific context or relevance of the input to each module, for optimizing computational efficiency. This ensures that only the most relevant modules are used, reducing computational overhead and improving efficiency.

[1372] 5 [0346] Optionally, a temperature scaling factor is applied to the softmax function during probabilistic output computation, with the scaling factor dynamically adjusted to modify the sensitivity of the probabilistic output to differences in weighted scores. This allows the system to fine-tune the sensitivity of its probabilistic outputs, improving the accuracy of its predictions.

[1374] Optionally, method 200 further comprises processing input data from multiple data 10 modalities, with each data modality processed through a separate feature extraction pipeline prior to integration in the hierarchical neural network, with the integration performed to generate a unified feature representation for task-specific predictions. This ensures that the system can effectively handle and integrate data from multiple sources, improving the accuracy and robustness of its predictions.

[1376] 15 [0348] Optionally, method 200 further comprises compressing the weight matrices of the hierarchical neural network during training or inference using low-rank approximations, with the compression reducing memory requirements and computational complexity while maintaining model accuracy. This improves the efficiency of the model by reducing its memory and computational requirements without sacrificing accuracy.

[1378] 20 [0349] Optionally, the hierarchical neural network comprises one or more auxiliary loss functions applied to intermediate layers during training, with the auxiliary loss functions configured to guide feature extraction at intermediate stages of the network to improve overall task performance. This ensures that the model learns useful features at all stages of the network, improving its overall performance.

[1380] 25 [0350] Optionally, the artificial intelligence system is configured to operate in a resourceconstrained environment, and wherein the pruning mechanism is dynamically adapted to balance energy efficiency and prediction accuracy based on the available computational resources. This ensures that the system can operate efficiently even in environments with limited resources, maintaining a balance between energy use and prediction accuracy.

[1382] 30 [0351] Optionally, method 200 further comprises using the prediction to dynamically adjust hyperparameters of the hierarchical neural network during training, with the adjustment optimizing the training process for improved convergence and reduced training time. This ensures that the model can be trained more efficiently, reducing the time and resources required to achieve optimal performance.

[1384] 35 [0352] Optionally, the system is configured to perform uncertainty quantification for predictions, with the uncertainty determined by generating multiple predictions under stochastic configurations of the hierarchical neural network and computing a measure of variance across the predictions. This provides a measure of the reliability of the predictions, allowing users to make more informed decisions based on the model's output.

[1386] 40 [0353] Optionally, method 200 further comprises augmenting the input data during training with synthetic variations generated by a data augmentation pipeline, with the augmented data

[1387] 349371

[1390] 58

[1392] improving the robustness of the artificial intelligence system to input variations and noise. This ensures that the model can handle a wider range of inputs and is less sensitive to noise, improving its overall robustness.

[1394] Optionally, the pruning mechanism is further configured to identify and deactivate 5 redundant connections between domain-specific modules, with the redundancy evaluated based on module output similarity over multiple predictions. This ensures that the system remains efficient by removing unnecessary connections, reducing computational overhead.

[1396] Optionally, the recursive feedback mechanism incorporates time-series data, and the feedback signals are computed based on historical trends in the input data, enabling the 10 system to capture temporal dependencies in dynamic environments. This improves the system's ability to handle time-dependent data, enhancing its accuracy in dynamic environments.

[1398] Optionally, the hierarchical neural network is configured to implement transfer learning, with the configuration comprising initializing a subset of the layers with pre-trained weights 15 from a source task, and fine-tuning the layers on the task-specific dataset to improve learning efficiency. This ensures that the model can leverage existing knowledge to improve its performance on new tasks, reducing the time and resources required for training.

[1400] Optionally, method 200 further comprises evaluating the relevance scores of domainspecific modules using an attention mechanism, with the attention mechanism prioritizing 20 modules based on their contribution to the predicted outcome in the current context. This ensures that the most relevant modules are given priority, improving the overall performance of the system.

[1402] Optionally, the artificial intelligence system is further configured to generate explanations for its predictions, with the explanations derived from feature importance metrics 25 computed during the hierarchical processing and presented to a user in a human-readable format. This improves the interpretability of the model, allowing users to understand the reasoning behind its predictions.

[1404] Optionally, the system is configured to perform distributed processing of input data, with different subsets of the hierarchical neural network and domain-specific modules 30 executed on separate hardware units, with the outputs synchronized to produce the final prediction. This ensures that the system can handle large-scale data processing efficiently, improving its overall performance.

[1406] Optionally, the system comprises at least two specialized modules, each configured to process data from a distinct domain, with the outputs of the specialized modules integrated 35 through the bridging mechanism to enable cross-domain feature interaction and joint decisionmaking. This ensures that the system can effectively handle and integrate data from multiple domains, improving the accuracy and robustness of its predictions.

[1408] [0361] Optionally, the bridging mechanism applies separate transformation matrices to the outputs of the specialized modules, with each transformation matrix configured to map the 40 output of one module into the representational space of the other module. This ensures that

[1409] 349371

[1412] 59

[1414] the system can effectively integrate data from different domains, improving the accuracy and robustness of its predictions.

[1416] Optionally, the intensity coefficients used in the bridging transformations are dynamically adjusted based on context signals derived from the characteristics of the input 5 data or the activation states of the specialized modules. This ensures that the system can adapt to changing conditions, improving the accuracy and robustness of its predictions.

[1418] Optionally, method 200 further comprises selectively activating only the relevant specialized modules for a given input, with the activation determined based on the context of the input data or its relevance to the task-specific requirements. This ensures that only the 10 most relevant modules are used, reducing computational overhead and improving efficiency.

[1420] Optionally, the bridging mechanism dynamically computes recursive feedback signals across the specialized modules, with the feedback signals used to refine the outputs of the modules for improved cross-domain alignment. This ensures that the system can effectively integrate data from different domains, improving the accuracy and robustness of its 15 predictions.

[1422] Optionally, the bridging mechanism is configured to incorporate context signals derived from environmental parameters, user inputs, or metadata associated with the input data, with the context signals used to gate or scale the bridging transformations. This ensures that the system can adapt to changing conditions, improving the accuracy and robustness of its 20 predictions.

[1424] Optionally, the recursive feedback signals generated within the bridging mechanism are bounded by stability constraints to ensure convergence and prevent numerical instability during iterative refinement. This ensures that the system remains stable and reliable, preventing numerical issues from affecting its performance.

[1426] 25 [0367] Optionally, the bridging mechanism is further configured to optimize the computational load by selectively pruning low-contributing weights in the bridging transformations during lowactivity phases, with the pruning dynamically reversed during high-activity phases. This ensures that the system remains efficient by adjusting its computational load based on activity levels, reducing resource usage during low-activity periods.

[1428] 30 [0368] Optionally, the relevance scores of the specialized modules are evaluated using an attention mechanism, with the attention mechanism prioritizing module outputs based on their contribution to the final prediction in the given context. This ensures that the most relevant module outputs are given priority, improving the overall performance of the system.

[1430] Optionally, the bridging mechanism incorporates historical trends in the input data, 35 enabling the system to dynamically refine the outputs of the specialized modules based on temporal dependencies or recurring patterns. This improves the system's ability to handle time-dependent data, enhancing its accuracy in dynamic environments.

[1432] [0370] Optionally, the outputs of the bridging mechanism are hierarchically incorporated into higher layers of the hierarchical neural network, enabling contextualized integration of cross-40 domain features for task-specific predictions. This ensures that the system can effectively

[1433] 349371

[1435] 60

[1437] integrate data from different domains, improving the accuracy and robustness of its predictions.

[1439] Optionally, the artificial intelligence system is configured to deactivate redundant bridging connections between the specialized modules during periods of low activity, with the 5 redundancy evaluated based on the similarity of module outputs across multiple predictions. This ensures that the system remains efficient by removing unnecessary connections, reducing computational overhead.

[1441] Optionally, method 200 further comprises using the outputs of the bridging mechanism to guide feature extraction in intermediate layers of the hierarchical neural network, with the 10 bridging mechanism influencing representations to improve overall task performance. This ensures that the system learns useful features at all stages of the network, improving its overall performance.

[1443] Optionally, the bridging mechanism incorporates attention weights to prioritize features exchanged between the specialized modules, with the attention weights computed based on 15 the task relevance of the exchanged features. This ensures that the most relevant features are given priority, improving the overall performance of the system.

[1445] Optionally, the artificial intelligence system is configured to generate human-readable explanations for the outputs of the bridging mechanism, with the explanations derived from the feature importance or relevance metrics computed during cross-module integration. This 20 improves the interpretability of the model, allowing users to understand the reasoning behind its predictions.

[1447] Optionally, the bridging mechanism comprises auxiliary loss functions applied to intermediate bridging outputs during training, with the auxiliary loss functions configured to guide the optimization of cross-module transformations for improved generalization. This 25 ensures that the model learns useful features at all stages of the network, improving its overall performance.

[1449] Optionally, method 200 further comprises compressing the transformation matrices of the bridging mechanism using low-rank approximations, with the compression reducing memory and computational requirements while maintaining accuracy in cross-module 30 integration. This improves the efficiency of the model by reducing its memory and computational requirements without sacrificing accuracy.

[1451] Optionally, the bridging mechanism enables distributed processing across hardware units, with the outputs of the specialized modules and bridging transformations synchronized to produce the final integrated prediction. This ensures that the system can handle large-scale 35 data processing, improving its overall performance.

[1453] Figure 3 illustrates a flowchart of method 300, which is a method for optimizing neural network performance through dynamic pruning of weight matrices. This method comprises several steps that systematically reduce computational overhead while maintaining the integrity and performance of the neural network.

[1455] 40 [0379] Method 300 begins with step 302, where a weight matrix is obtained. This weight matrix represents the connections between neurons in a neural network layer. In step 304, a pruning

[1456] 349371

[1458] 61

[1460] threshold is calculated based on the statistical properties of the weight matrix, such as the mean and standard deviation of the weight values. This threshold determines which weights are considered low-impact and can be pruned. Step 306 involves generating a pruned weight matrix by applying the calculated threshold, effectively deactivating weights that fall below the 5 threshold. This pruned weight matrix retains only the most significant connections, reducing the computational load. In step 308, the pruning threshold is adjusted based on runtime activation sparsity, which measures the activity level of the neurons. This dynamic adjustment ensures that the pruning process remains responsive to the current state of the network. Finally, in step 310, the pruned weight matrix is applied to the neural network, optimizing its 10 performance by reducing unnecessary computations.

[1462] The benefits of method 300 are multifaceted. By systematically pruning low-impact weights, the method reduces the number of active connections in the neural network, leading to significant computational savings. This reduction in computational load translates to lower energy consumption, making the neural network more efficient and sustainable. Additionally, 15 the dynamic adjustment of the pruning threshold based on runtime activation sparsity ensures that the network remains adaptable to varying conditions and data inputs. This adaptability enhances the network's ability to maintain high performance even as the nature of the input data changes. Furthermore, the method's ability to selectively prune and restore connections during low-activity phases ensures that the network can optimize its structure without 20 compromising its learning and predictive capabilities. Overall, method 300 provides a robust framework for enhancing the efficiency and performance of neural networks, making it a valuable tool for developing scalable and sustainable AI systems.

[1464] Step 302 of method 300 comprises obtaining a weight matrix, which is a fundamental component in the structure of a neural network. The weight matrix represents the connections 25 between neurons in a specific layer of the network, with each element in the matrix corresponding to the strength of the connection between two neurons. This matrix is beneficial for the network's ability to learn and make predictions, as it encodes the learned parameters that have been adjusted during the training process.

[1466] In this step, the weight matrix is typically extracted from a pre-trained neural network 30 model. This model has undergone a training phase where it has been exposed to a large dataset, allowing it to adjust its weights to minimize the error in its predictions. The weight matrix obtained in step 302 contains these learned parameters, which are essential for the network's performance. The process of obtaining the weight matrix involves accessing the specific layer of the neural network and retrieving the matrix that represents the connections 35 within that layer. This matrix serves as the starting point for the subsequent pruning process, where low-impact weights will be identified and deactivated to optimize computational efficiency.

[1468] Beneficially, by starting with a well-trained weight matrix, the method ensures that the pruning process targets only those weights that are truly redundant or low-impact, preserving 40 the network's overall performance.

[1470] [0384] Step 304 of method 300 comprises calculating a pruning threshold, which is a critical step in the process of optimizing the neural network by identifying and deactivating low-impact weights. This threshold is determined based on the statistical properties of the weight matrix

[1471] 349371

[1473] 62

[1475] obtained in step 302. Specifically, the pruning threshold is calculated by analysing the distribution of the weights within the matrix. This involves computing the mean and standard deviation of the weight values, which provide insights into the overall distribution and variability of the weights.

[1477] 5 [0385] The pruning threshold is dynamically adjusted to ensure that it is appropriate for the specific layer and the overall system activity. This dynamic adjustment takes into account the layer-specific characteristics and the current state of the network. By setting a threshold that is tailored to the specific conditions of the layer, the method ensures that only the weights that have minimal impact on the network's performance are pruned. This approach allows for a 10 more precise and effective pruning process, as it targets the weights that contribute the least to the network's functionality.

[1479] The benefit of calculating a pruning threshold in this manner is that it allows for a more efficient and targeted pruning process. By using statistical properties to determine the threshold, the method ensures that the pruning process is both systematic and adaptive. This 15 results in a more optimized weight matrix, where only the most significant connections are retained. Consequently, the computational load is reduced without compromising the network's performance, leading to improved efficiency and energy savings. This step is beneficial for maintaining the balance between computational efficiency and the accuracy of the neural network.

[1481] 20 [0387] Step 306 of method 300 comprises generating a pruned weight matrix, which is a critical step in optimizing the neural network by reducing computational overhead. This step takes the weight matrix obtained in step 302 and applies the pruning threshold calculated in step 304 to deactivate low-impact weights. The pruning process involves setting the weights that fall below the threshold to zero, effectively removing them from the computational graph.

[1482] 25 This results in a pruned weight matrix that retains only the most significant connections, ensuring that the network's performance is maintained while reducing the number of active weights.

[1484] The process of generating the pruned weight matrix begins by evaluating each weight in the original matrix against the calculated pruning threshold. Weights that are below the 30 threshold are considered low-impact and are set to zero. This selective deactivation of weights ensures that the network focuses its computational resources on the most important connections, thereby improving efficiency. The pruned weight matrix is then stored in a sparse format, which further optimizes memory usage and computational efficiency. This format only records the non-zero weights and their positions, reducing the storage requirements and 35 speeding up subsequent computations.

[1486] [0389] The benefits of generating a pruned weight matrix are significant. By removing lowimpact weights, the method 300 reduces the computational load and energy consumption of the neural network. This makes the network more efficient and sustainable, particularly in resource-constrained environments such as edge devices and IoT sensors. Additionally, the 40 pruned weight matrix helps maintain the network's performance by ensuring that only the most critical connections are retained. This balance between efficiency and performance is beneficial for deploying neural networks in real-world applications, where computational resources and energy consumption are often limited.

[1487] 349371

[1490] 63

[1492] Step 308 of method 300 comprises adjusting the pruning threshold based on runtime activation sparsity. This step ensures that the pruning process remains adaptive and responsive to the current activity levels within the neural network. During runtime, the system continuously monitors the activation sparsity of each layer, which is a measure of how many 5 neurons are actively firing in response to the input data. Activation sparsity is calculated by determining the proportion of non-zero activations in the activation matrix of a given layer.

[1494] Based on the observed activation sparsity, the pruning threshold is dynamically adjusted. If a layer exhibits high activation sparsity, indicating that many neurons are active, the pruning threshold may be lowered to allow more weights to be pruned. Conversely, if a 10 layer shows low activation sparsity, suggesting that fewer neurons are active, the pruning threshold may be raised to preserve more weights and maintain the network's performance. This dynamic adjustment process ensures that the network can adapt to varying levels of activity and data complexity, optimizing computational efficiency without compromising accuracy.

[1496] 15 [0392] Beneficially, this adaptive approach allows the network to maintain an optimal balance between computational load and performance, ensuring that resources are used efficiently. By continuously fine-tuning the pruning threshold, the system can respond to changes in data patterns and workload, enhancing its ability to perform effectively in diverse and dynamic environments.

[1498] 20 [0393] Step 310 of method 300 comprises applying the pruned weight matrix to the neural network. This step is beneficial for realizing the computational and energy efficiency gains achieved through the pruning process. Once the pruned weight matrix has been generated in step 306, it needs to be integrated back into the neural network to replace the original weight matrix. This integration involves updating the network's weights with the pruned matrix, 25 ensuring that only the most significant connections are retained for forward propagation.

[1500] The application of the pruned weight matrix begins by loading the pruned weights into the neural network's architecture. This involves replacing the original weight matrix with the pruned version, which contains fewer active weights. The pruned weight matrix is typically stored in a sparse format, which records only the non-zero weights and their positions. This 30 format optimizes memory usage and speeds up subsequent computations by reducing the number of operations required during forward propagation.

[1502] By applying the pruned weight matrix, the neural network can operate with reduced computational overhead and lower energy consumption. This step ensures that the benefits of the pruning process are fully realized, leading to a more efficient and sustainable neural 35 network. The reduced number of active weights translates to fewer floating-point operations per second (FLOPs), which directly impacts the network's power consumption and processing speed. This optimization is particularly valuable for deployment in resource-constrained environments, such as edge devices and IoT sensors, as well as for large-scale AI systems where efficiency and sustainability are critical.

[1504] 40 [0396] Method 300 is not limited to neural networks and can be applied to a variety of computational systems that utilize weight matrices or similar structures. The core principles of obtaining a weight matrix, calculating a pruning threshold, generating a pruned weight matrix,

[1505] 349371

[1508] 64

[1510] adjusting the pruning threshold based on runtime activation sparsity, and applying the pruned weight matrix are broadly applicable to any system where optimization of computational resources is desired. For example, in decision tree algorithms, the concept of pruning can be adapted to remove branches that contribute minimally to the decision-making process, thereby 5 reducing the complexity and improving the efficiency of the model. Similarly, in support vector machines (SVMs), the method can be used to prune support vectors that have minimal impact on the decision boundary, optimizing the computational load during the classification process.

[1512] Moreover, method 300 can be extended to graph-based algorithms, where the weight matrix represents the edges between nodes. By pruning low-impact edges, the method can 10 simplify the graph, making algorithms like shortest path or community detection more efficient.

[1513] This approach can be particularly beneficial in large-scale graph processing tasks, such as social network analysis or transportation network optimization. In the context of signal processing, method 300 can be applied to filter design, where the weight matrix represents filter coefficients. Pruning low-impact coefficients can lead to more efficient filter 15 implementations, reducing the computational burden in real-time signal processing applications. Overall, the principles of method 300 are versatile and can be adapted to optimize a wide range of computational systems beyond neural networks.

[1515] Optionally, method 300 comprises dynamic weight restoration. Dynamic weight restoration involves monitoring a system's activity levels and identifying periods of increased 20 demand or complexity in the input data. During these high-activity phases, the system evaluates the previously pruned weights to determine if they should be restored. This evaluation is based on the current importance of the weights, which can be influenced by changes in the data patterns or the network's performance requirements. The restoration process is gradual, using a formula that incrementally reactivates the pruned weights, ensuring 25 a smooth transition and preventing abrupt changes that could destabilize the network.

[1517] Optionally, method 300 further comprises the use of sparse tensor representations for storing pruned weights. This technique optimizes memory usage by only recording the nonzero weights and their positions, significantly reducing the storage requirements. Sparse tensor representations are particularly beneficial for large-scale neural networks, where the number 30 of weights can be substantial. By compressing the pruned weights, the system can achieve further computational savings and improve overall efficiency.

[1519] A mathematical interpretation of method 300 is provided below.

[1521] Pruning comprises:

[1523] obtaining a weight matrix ܹ associated with a layer of the machine learning model;

[1524] 35 calculating a pruning threshold ߬for the layer:

[1526] ߬= ߤ<ௐ >+ ߣ ⋅ ߪ<ௐ >, wherein

[1528] ߤ<ௐ >is the mean average of the weight distribution,

[1530] ߣ is a scaling factor for the layer, and

[1531] 349371

[1533] 65

[1535] ߪ<ௐ >is the standard deviation of the weight distribution;

[1537] generating a pruned weight matrix ܹ ′ using a pruning function:

[1539] ܹ <ᇱ |ௐ |ିఛ>= ܹ ⋅ ߪ ቀ ఑ ቁ, wherein

[1541] ߪ is a sigmoid function, and

[1543] 5 ߢ is a sensitivity scaling factor;

[1545] adjusting the pruning threshold ߬based on runtime activation sparsity metrics of the

[1546] machine learning model; and

[1548] applying the pruned weight matrix ܹ ′ in forward propagation during training or

[1549] updating of the machine learning model.

[1551] 10 [0401] Figure 4 illustrates a flowchart of method 400, which is a method for integrating outputs from two computational modules in an artificial intelligence system, where each module comprises a hierarchical neural network. The method is configured to enhance the interaction between the modules to improve the overall system performance.

[1553] Method 400 comprises step 402, where a first output is generated from a first module 15 and a second output from a second module, with each module processing data independently to produce high-level representations. In step 404, the first output is modified by incorporating a transformed representation of the second output using a first bridging transformation, which is weighted by a first intensity coefficient. Concurrently, in step 406, the second output is modified by incorporating a transformed representation of the first output using a second 20 bridging transformation, weighted by a second intensity coefficient. These modifications enable the outputs to be enriched with contextual information from the other module, facilitating a more integrated and comprehensive representation. In step 408, updated states for the first and second modules are generated by applying a recursive state update that incorporates the respective modified outputs. This recursive update ensures that the modules continuously 25 refine their states based on the enriched outputs, leading to more accurate and robust representations. Step 410 involves adjusting the first and second bridging transformations and intensity coefficients based on a gating function derived from additional signals, allowing the system to dynamically adapt the integration process based on real-time data. Finally, in step 412, connections in the bridging transformations are selectively pruned or restored during a 30 low-activity phase to optimize computational efficiency. This selective pruning helps maintain the system's performance while reducing unnecessary computational overhead, making the method both effective and efficient.

[1555] [0403] Step 402 of method 400 comprises generating a first output from a first module and a second output from a second module, where each module processes data independently and 35 outputs a high-level representation, e.g. each module is domain-specific. Each module, similar to a left or right hemisphere in a biological neural network, consists of hierarchical layers that process input data to produce final outputs at the top layer. These outputs are high-level representations that encapsulate the processed information from each module's respective

[1556] 349371

[1558] 66

[1560] hierarchical structure. In this process, each module maintains its own recursive or global state, which updates at every time step. The independent processing in each module ensures that the data is transformed through multiple layers, with each layer potentially extracting different features or patterns from the input data. This hierarchical processing is beneficial for 5 generating rich, high-level representations that can be effectively integrated in subsequent steps. The final outputs represent the culmination of this layered processing, providing comprehensive and abstracted representations of the input data. These outputs serve as the basis for further integration and modification in the following steps of method 400.

[1562] Step 402 is optionally utilised in various artificial intelligence systems where modular 10 processing is advantageous. For example, in a multi-agent system, each agent acts as a module that processes its own set of data independently before sharing its high-level output with other agents. This approach allows for specialized processing within each module, leveraging the strengths of different algorithms or architectures tailored to specific types of data or tasks. By generating high-level representations independently, the system then 15 integrates these outputs to form a more holistic understanding of the data, enhancing overall performance and decision-making capabilities.

[1564] A detailed example of implementing step 402 can be found in an AI system applied towards cancer detection. In this scenario, one module processes imaging data, such as MRI or CT scans, while another module processes genomic data from patient samples. The imaging 20 module independently analyses the visual data through its hierarchical neural network, extracting features related to tumour size, shape, and texture, resulting in a high-level representation of the imaging data. Simultaneously, the genomic module processes the genetic information, identifying mutations, expression levels, and other relevant biomarkers, producing its own high-level representation. These independent outputs, one capturing the visual 25 characteristics of the tumour and the other encapsulating the genetic profile, provide a comprehensive view of the patient's condition. Subsequent steps in method 400 then integrate these high-level representations to enhance diagnostic accuracy and treatment planning, leveraging the strengths of both imaging and genomic data.

[1566] Step 404 of method 400 comprises modifying the first output by incorporating a 30 transformed representation of the second output using a first bridging transformation, wherein the transformed representation is weighted by a first intensity coefficient. Each module, similar to a hemisphere, produces a high-level output that is then transformed and integrated into the other module's output. This transformation is achieved through a bridging weight matrix, which maps features from one module to the other.

[1568] 35 [0407] The first bridging transformation comprises applying a weight matrix to the second module's high-level output, effectively transforming it into a representation that can be integrated with the first module's output. This transformed representation is then weighted by a first intensity coefficient, which controls the influence of the second module's output on the first module's modified output. The intensity coefficient can be adjusted based on various 40 factors, such as the importance of the second module's information or the current state of the system. The resulting modified output of the first module is a combination of its original highlevel representation and the transformed, weighted representation of the second module's output.

[1569] 349371

[1571] 67

[1573] For instance, in the multi-agent system, each agent processes its own data independently and then shares its high-level output with other agents. The bridging transformation allows each agent to incorporate relevant information from other agents, enhancing its own output. This approach enables the system to leverage the strengths of 5 different agents, leading to a more comprehensive and accurate understanding of the data. A detailed example of implementing step 404 can be found in an AI system applied towards cancer detection. In this scenario, the imaging module processes MRI or CT scans to produce a high-level representation of the visual data, while the genomic module processes genetic information to produce a high-level representation of the genomic data. The bridging 10 transformation involves applying a weight matrix to the genomic module's output, transforming it into a representation that can be integrated with the imaging module's output. This transformed representation is then weighted by an intensity coefficient, which controls the influence of the genomic data on the imaging module's modified output. The resulting modified output of the imaging module is a combination of its original high-level representation and the 15 transformed, weighted representation of the genomic data. This integrated output provides a more comprehensive view of the patient's condition, enhancing diagnostic accuracy and treatment planning.

[1575] Step 406 of method 400 comprises modifying the second output by incorporating a transformed representation of the first output using a second bridging transformation, wherein 20 the transformed representation is weighted by a second intensity coefficient. This transformation is achieved through a bridging weight matrix, which maps features from one module to the other, as described in step 404. The second bridging transformation involves applying a weight matrix to the first module's high-level output, effectively transforming it into a representation that can be integrated with the second module's output. This transformed 25 representation is then weighted by a second intensity coefficient, which controls the influence of the first module's output on the second module's modified output. The intensity coefficient can be adjusted based on various factors, such as the importance of the first module's information or the current state of the system. The resulting modified output of the second module is a combination of its original high-level representation and the transformed, weighted 30 representation of the first module's output.

[1577] Step 408 of method 400 comprises generating updated states for the first and second modules by applying a recursive state update that incorporates the respective modified outputs. This step is beneficial for ensuring that the modules continuously refine their states based on the enriched outputs obtained from the previous bridging transformations. In this 35 process, each module updates its recursive or global state by integrating the modified outputs from the other module. The recursive state update mechanism ensures that the modules are not only influenced by their own high-level representations but also by the transformed and weighted representations from the other module. This integration allows the modules to adapt and evolve their internal states, leading to more accurate and robust representations over time.

[1579] 40 [0411] The updated states are generated by applying a function that takes into account the modified outputs and the current states of the modules. This function is optionally a neural submodule or another computational mechanism designed to recast the bridged signals into the module’s new state. The updated states are then used in subsequent processing steps,

[1580] 349371

[1582] 68

[1584] ensuring that the modules continuously learn and adapt based on the integrated information from both modules.

[1586] For example, in the multi-agent system, each agent updates its state based on the integrated outputs from other agents, allowing for more coordinated and adaptive behaviour.

[1587] 5 This approach enables the system to leverage the strengths of different agents, leading to a more comprehensive and accurate understanding of the data. A detailed example of implementing step 408 can be found in an AI system applied towards cancer detection. In this scenario, the imaging module processes MRI or CT scans to produce a high-level representation of the visual data, while the genomic module processes genetic information to 10 produce a high-level representation of the genomic data. After the bridging transformations, the updated states for both modules are generated by integrating the modified outputs. The imaging module updates its state based on the transformed and weighted representation of the genomic data, while the genomic module updates its state based on the transformed and weighted representation of the imaging data. These updated states are then used in 15 subsequent processing steps, enhancing the system's ability to accurately diagnose and plan treatment for cancer patients.

[1589] Step 410 of method 400 comprises adjusting the first and second bridging transformations and intensity coefficients based on a gating function derived from additional signals. This step ensures that the integration process between the two modules remains 20 dynamic and responsive to varying conditions and inputs. The gating function, which is optionally influenced by emotive signals or other contextual data, modulates the intensity coefficients that control the influence of the transformed outputs from one module on the other. These intensity coefficients determine how much weight the transformed representation from one module carries when it is incorporated into the other module's output. By adjusting these 25 coefficients dynamically, the system can fine-tune the level of integration based on real-time data or specific requirements of the task at hand.

[1591] The adjustment process involves evaluating the additional signals, which optionally comprises emotive signals, performance metrics, or other relevant data, and using these signals to modify the bridging transformations and intensity coefficients. This ensures that the 30 system can adapt to changes in the environment or the data it processes, maintaining optimal performance. The gating function acts as a regulatory mechanism, ensuring that the integration between the modules is neither too weak nor too strong, but appropriately balanced to enhance the overall system performance.

[1593] [0415] For instance, in the multi-agent system, the gating function is used to adjust the level of 35 information sharing between agents based on their current performance or the complexity of the task. This dynamic adjustment allows the system to maintain a balance between independent processing and collaborative integration, optimizing overall efficiency and effectiveness. A detailed example of implementing step 410 can be found in an AI system applied towards cancer detection. In this scenario, the gating function can be influenced by the 40 confidence levels of the imaging and genomic modules. If the imaging module has high confidence in its analysis of MRI or CT scans, the intensity coefficient for incorporating genomic data might be reduced, and vice versa. This dynamic adjustment ensures that the system leverages the most reliable data at any given time, enhancing diagnostic accuracy. The gating

[1594] 349371

[1596] 69

[1598] function allows the system to adapt to varying conditions, such as changes in the quality of the input data or the emergence of new biomarkers, ensuring that the integration process remains optimal and responsive to the needs of the task.

[1600] Step 412 of method 400 involves selectively pruning or restoring connections in the 5 bridging transformations during a low-activity phase to optimize computational efficiency, e.g.

[1601] using method 300 and/or pruning mechanism 108 as described above. This step is configured to ensure that the system remains efficient by removing unnecessary or less significant connections, while also allowing for the restoration of important connections that may have been pruned previously. During the low-activity phase, which can be likened to an offline or 10 sleep phase, the system evaluates the significance of the bridging transformations and intensity coefficients. Connections that exhibit low magnitudes or low gradient significance are identified as candidates for pruning. This evaluation can be based on various criteria, such as the magnitude of the weight matrices or the learned coefficients. If a connection is deemed insignificant, it is set to zero, effectively removing it from the computational graph. This pruning 15 process helps reduce the computational load and memory usage, making the system more efficient.

[1603] Optionally, the system also has the capability to restore connections that may have been pruned in previous cycles. If new data or changes in the system's state indicate that a previously pruned connection is now significant, the system can reinstate that connection. This 20 dynamic adjustment ensures that the system remains flexible and adaptive, capable of optimizing its structure based on the current requirements and data.

[1605] For example, in a multi-agent system, the pruning and restoration of connections helps manage the computational resources required for communication and integration between agents. By selectively pruning less significant connections, the system can focus its resources 25 on the most important interactions, enhancing overall performance. A detailed example of implementing step 412 can be found in an AI system applied towards cancer detection. In this scenario, the system processes large amounts of imaging and genomic data, which can be computationally intensive. During the low-activity phase, the system evaluates the significance of the connections between the imaging and genomic modules. Connections that contribute 30 less to the diagnostic accuracy are pruned, reducing the computational load. If new patient data or changes in the system's state indicate that a previously pruned connection is now important, the system restores that connection. This dynamic adjustment ensures that the system remains efficient while maintaining high diagnostic accuracy, ultimately improving the system's ability to detect and treat cancer.

[1607] 35 [0419] Method 400 optionally further comprises several additional components and processes that enhance the bridging mechanism between the two computational modules, analogous to the hemispheres in an AI system. One such component is the use of emotive signals, which provide a gating function for the bridging transformations. These emotive signals can be derived from various sources, such as performance metrics or contextual data, and are used 40 to dynamically adjust the intensity coefficients that control the influence of the transformed outputs from one module on the other. This dynamic adjustment ensures that the integration process remains responsive to real-time data and specific task requirements.

[1608] 349371

[1611] 70

[1613] One method of deriving emotive signals comprises monitoring the performance metrics of each module, such as accuracy, precision, recall, or other relevant indicators. For example, if the system is applied to a diagnostic task, the confidence levels of the predictions made by each module can serve as emotive signals. High confidence levels optionally indicate that the 5 module's output is reliable, leading to an increase in the intensity coefficient for that module's transformed output. Conversely, low confidence levels reduce the intensity coefficient, indicating less reliance on that module's output. Another method comprises contextual data, such as user feedback or environmental conditions. In a user-interactive system, real-time feedback from users can be collected and analysed to generate emotive signals. Positive 10 feedback optionally enhances the intensity coefficients, while negative feedback reduces them.

[1614] Environmental conditions, such as changes in input data quality or external factors affecting the system's performance, can also be used to derive emotive signals. For example, in a surveillance system, variations in lighting or weather conditions might influence the emotive signals, adjusting the intensity coefficients to ensure optimal performance under different 15 scenarios. These methods ensure that the system remains adaptive and responsive to realtime data and contextual factors, enhancing the overall effectiveness of the bridging mechanism.

[1616] Bridging weight matrices transform the high-level outputs and states from one module into a representation that can be integrated with the other module's outputs and states. The 20 weight matrices are learned during the training phase and can be adjusted based on the system's performance. Additionally, method 400 preferably comprises layer-by-layer bridging, where bridging transformations are defined at intermediate layers within the hierarchical structure of each module. This approach allows for deeper synergy between the modules, as information is integrated at multiple levels of abstraction.

[1618] 25 [0422] Another optional feature is partial or time-based bridging. This involves allowing bridging transformations only at certain time steps, which can be triggered by specific conditions such as emotive signals surpassing a predefined threshold. This selective bridging ensures that the system remains efficient by limiting the computational overhead associated with continuous integration. Furthermore, method 400 is preferably integrated with the HPC 30 logs, which allows the bridging mechanism to reference previously stored outputs and states.

[1619] This capability enables the system to bridge not only the current states of the modules but also historical data, providing extended synergy and enhancing the system's ability to learn from past experiences.

[1621] The offline "sleep-phase" pruning or refinement process is another beneficial aspect of 35 the bridging mechanism. During this phase, the system evaluates the significance of the bridging transformations and intensity coefficients based on criteria such as low magnitudes or low gradient significance. Connections that are deemed insignificant are pruned, reducing the computational load and improving efficiency. However, the system also retains the capability to restore previously pruned connections if new data or changes in the system's state 40 indicate their importance. This dynamic adjustment ensures that the system remains flexible and adaptive, capable of optimizing its structure based on current requirements and data.

[1623] [0424] Beneficially, method 400 represents an improvement over known bridging techniques by incorporating dynamic and adaptive mechanisms that enhance the integration of outputs

[1624] 349371

[1627] 71

[1629] from two computational modules. Unlike traditional methods that may rely on static or fixed transformations, method 400 employs a recursive state update process that continuously refines the states of each module based on enriched outputs. This dynamic adjustment is further enhanced by the use of emotive signals, which provide a gating function to modulate 5 the intensity of the bridging transformations. This allows the system to adapt to real-time data and specific task requirements, ensuring optimal performance. Additionally, the method optionally comprises an offline "sleep-phase" pruning process that evaluates the significance of the bridging transformations and intensity coefficients, removing less significant connections to improve computational efficiency. The ability to restore previously pruned connections 10 ensures that the system remains flexible and adaptive. Furthermore, the method supports layer-by-layer bridging, partial or time-based bridging, and integration with a hippocampus-like memory system, providing deeper synergy and extended learning capabilities. These features collectively enable method 400 to achieve richer multi-domain synergy, enhancing the system's performance in tasks that require specialized but cooperative processing streams.

[1631] 15 [0425] An optional mathematical implementation of method 400 is described below.

[1633] Bridging comprises:

[1635] generating a first output from a first module (ܪ<ଵ>(௞)

) and a second output from a second module (ܪ<ଶ>(௞)

), wherein:

[1637] ܪ<ଵ>(௞)

,ܪ<ଶ>(௞)

∈ ℝ<ௗ೓>are high-level representations generated at layer ܭ, the final 20 layer of each module’s hierarchical neural network; and

[1639] the first output and the second output are computer recursively for all layers ݈= <1, 2, … ,ܭ in each module as:>

[1641] ܪ <ଵ)>

[1642] <௜>(௟)

= (݂ܹ<௜>(௟)

∙ ܪ<(௟ି>

[1643] <௜ >+ <(௟)>

[1644] <௜ܾ >), for ݅= 1, 2, wherein:

[1646] layer r݈epresents an intermediate layer of the hierarchical neural

[1647] 25 network, and each layer processes input from the preceding layer ݈− 1

[1649] ; and

[1651] ܪ<௜>(௟)

is the output of layer f݈or module ,݅ ܪ<௜>(௟ି ଵ)

is the input to layer f݈rom

[1652] a preceding layer, ܹ (௟)

[1653] <௜ >∈ ℝ<ௗ೓×ௗ೓ >is the weight matrix for layer ݈ in

[1654] module ,݅ (௟)

[1655] <௜ܾ >∈ ℝ<ௗ೓ >is the bias vector for layer i݈n module ,݅ and ݂ is an

[1656] 30 activation function;

[1658] modifying the first output by incorporating a transformed representation of the second output, such that:

[1660] ܪ<(௞)ᇱ (௞)>

[1661] <ଵ >= ܪ<ଵ >+ ߙ<ଵଶ >∙ ܹ<ଵଶ >∙ ܪ<ଶ>(௞)

, wherein

[1663] ܪ(௞)ᇱ

[1664] <ଵ >∈ ℝ<ௗ೓ >is the modified first output, ܹ<ଵଶ >∈ ℝ<ௗ೓×ௗ೓ >is the bridging 35 transformation matrix that maps the second output to representation space of

[1665] 349371

[1668] 72

[1670] the first module, and ߙ<ଵଶ >∈ ℝ<ௗ೓ >is a first intensity coefficient for controlling the influence of the transformed representation of the second output;

[1671] modifying the second output by incorporating a transformed representation of the first output, such that:

[1673] <௞)ᇱ>

[1674] 5 ܪ<(>

[1675] <ଶ >= ܪ<ଶ>(௞)

+ ߙ<ଶଵ >∙ ܹ <ଶଵ >∙ ܪ<ଵ>(௞)

, wherein

[1677] ܪ(௞)ᇱ

[1678] <ଶ >∈ ℝ<ௗ೓ >is the modified second output, ܹ <ଶଵ >∈ ℝ<ௗ೓×ௗ೓ >is the bridging transformation matrix that maps the first output to representation space of the second module, and ߙ<ଶଵ >∈ ℝ<ௗ೓ >is a second intensity coefficient for controlling the influence of the transformed representation of the second output;

[1680] 10 generating updated states for the first and second modules by applying a recursive state update that incorporates respective modified outputs, such that:

[1682] <(௧ାଵ) (௧) (௧)>

[1683] ଵܵ = <ଵܵ >+ ߚ<ଵଶ >∙ ܹ <௦ଵଶ >∙ <ଶܵ >+ ߛ <(௞)ᇱ>

[1685] <ଵଶ >∙ ܪ<ଵ >,

[1687] <(௧ାଵ) ௧) (௞)ᇱ>

[1688] ଶܵ = <(>

[1689] <ଶܵ >+ ߚ<ଶଵ >∙ ܹ <௧)>

[1691] <௦ଶଵ >∙ <(>

[1692] <ଵܵ >+ ߛ<ଶଵ >∙ ܪ<ଶ >, wherein

[1694] (௧)

[1695] ௜ܵ ∈ ℝ<ௗ೓ >is a recursive state of module ݅at time ݐ, ܹ <௦ଵଶ>,ܹ <௦ଶଵ >∈ ℝ<ௗ೓×ௗ೓ >are 15 state bridging matrices for mapping recursive states between modules,

[1696] ߚ<ଵଶ>,ߚ<ଶଵ >∈ ℝ are intensity coefficients controlling the interaction between recursive states of the modules, and ߛ<ଵଶ>,ߛ<ଶଵ >∈ ℝ are intensity coefficients controlling the contribution of modified hierarchical outputs to the updates states;

[1698] 20 adjusting the first and second bridging transformations and intensity coefficients based on a gating function ߛ(∙), such that:

[1700] ߙ<ଵଶ >= ߛ(ܧ<ଵ>,ܧ<ଶ>), ߚ<ଵଶ >= ߛ(ܧ<ଵ>,ܧ<ଶ>), wherein

[1702] ܧ<ଵ>,ܧ<ଶ >∈ ℝ<ௗ೓ >are additional signals comprising emotive or contextual signals for influencing bridging parameters, and ߛ: ℝ<ௗ೓ >× ℝ<ௗ೓ >→ ℝ adjusts the intensity 25 coefficients based on the additional signals; and

[1704] during a low-activity phase, selectively pruning or restoring bridging using:

[1706] <หௐ หି ఛ >ܹ <ᇱ ೔ೕ>

[1707] <௜௝ >= ܹ <௜௝ >⋅ ߪ ൬ <఑ >൰, wherein

[1709] ܹ <௜௝>ᇱ

is a pruned bridging weight matrix (݆݅ ∈ <{>12, 21,ݏ12,ݏ21<}>), ߪ(ݔ) is a sigmoid function, ߬∈ ℝ is a threshold for pruning determined based on the mean and 30 standard deviation of the bridging weight matrix, and ߢ ∈ ℝ is a sensitivity scaling factor.

[1711] System 100, or any of methods 200, 300 or 400, are applicable to AI systems in a wide range of use cases. The following are non-limiting examples.

[1713] [0427] In the context of patient monitoring for anomaly detection, system 100 and method 200 35 are employed to continuously analyze physiological data collected from patients. The artificial

[1714] 349371

[1717] 73

[1719] intelligence system processes this data to predict potential anomalies, such as irregular heartbeats or abnormal blood pressure levels. Upon identifying such anomalies, the system triggers an automated alert to notify medical staff of critical changes in the patient's condition, ensuring timely intervention and potentially preventing adverse health events. System 100 5 and/or method 200 significantly enhances patient monitoring and anomaly detection by leveraging hierarchical abstraction and multi-domain bridging mechanisms. Traditional AI systems often rely on static thresholds or simple pattern recognition, which can lead to missed anomalies or false alarms. System 100 and/or method 200, however, processes raw physiological data through multiple hierarchical layers, each performing specific 10 transformations such as noise reduction and feature extraction. This results in a refined, abstract representation of the data. The final output vectors are then integrated using a multidomain bridging mechanism, which considers the synergy between different physiological signals. This approach allows for more accurate and timely detection of physiological anomalies. Additionally, the recursive feedback mechanism ensures continuous learning and 15 adaptation, improving the system's ability to detect critical changes in a patient's condition and triggering automated alerts to medical staff more reliably.

[1721] For medical robotics in precision surgery, system 100 and method 200 are utilized to enhance the accuracy of robotic surgical tools. The artificial intelligence system dynamically adjusts the trajectory of these tools based on real-time predictions derived from intraoperative 20 data. This adjustment ensures that the surgical instruments follow the optimal path, minimizing tissue damage and improving the precision of minimally invasive procedures, thereby enhancing patient outcomes and reducing recovery times. In the context of medical robotics and precision surgery, System 100 and/or method 200 offers substantial improvements over existing AI systems by dynamically adjusting the trajectory of robotic surgical tools. Traditional 25 systems may use pre-defined paths or simple real-time adjustments based on immediate feedback, which can lack the precision required for minimally invasive procedures. System 100 and/or method 200 processes intraoperative data through hierarchical layers to extract high-level features and predict the optimal tool trajectory. The multi-domain bridging mechanism integrates these predictions with other relevant data, such as imaging and sensor 30 inputs, to ensure precise adjustments. This dynamic and context-aware approach minimizes tissue damage and enhances surgical outcomes. The recursive feedback mechanism further refines the predictions, adapting to any changes in the surgical environment in real-time.

[1723] In radiotherapy planning, system 100 and method 200 are applied to optimize treatment plans for cancer patients. The artificial intelligence system analyses imaging data to predict the 35 most effective angles and intensities for radiation beams. This prediction helps in generating a treatment plan that targets the tumour while minimizing exposure to surrounding healthy tissues, thereby reducing side effects and improving the efficacy of the radiotherapy. System 100 and/or method 200 improves radiotherapy planning by optimizing the intensity and angle of radiation beams to minimize damage to healthy tissue. Traditional AI systems may use static 40 models or limited real-time adjustments, which can result in suboptimal treatment plans.

[1724] System 100 and/or method 200 processes imaging data through hierarchical layers to generate a detailed, abstract representation of the tumour and surrounding tissues. The multidomain bridging mechanism then integrates these representations to predict the most effective radiation parameters. This approach ensures that the treatment plan is tailored to the specific

[1725] 349371

[1728] 74

[1730] characteristics of the tumour, reducing exposure to healthy tissues. The recursive feedback mechanism allows for continuous refinement of the treatment plan based on real-time data, ensuring optimal outcomes throughout the course of radiotherapy.

[1732] When integrated with wearable health devices, system 100 and method 200 process 5 sensor data to predict critical health events. The artificial intelligence system adjusts device parameters, such as sampling rates and alert thresholds, based on these predictions. This adjustment conserves battery life while ensuring that significant health changes are detected early, allowing for timely medical intervention and continuous health monitoring. For wearable health devices, System 100 and/or method 200 enhances performance by dynamically 10 adjusting device parameters based on real-time predictions. Traditional systems may use fixed sampling rates and alert thresholds, which can lead to inefficient battery usage and delayed detection of critical health events. System 100 and/or method 200 processes sensor data through hierarchical layers to extract meaningful features and predict potential health issues. The multi-domain bridging mechanism integrates these predictions with other relevant data, 15 such as user activity and environmental factors, to adjust device parameters dynamically. This ensures early detection of critical health events while conserving battery life. The recursive feedback mechanism continuously refines the predictions and adjustments, adapting to the user's changing conditions and improving the overall effectiveness of the wearable device.

[1734] In the realm of predictive maintenance for industrial automation, system 100 and 20 method 200 are used to forecast potential failures in machine components. The artificial intelligence system analyzes operational data to identify signs of wear and tear, triggering automated maintenance schedules. This proactive approach minimizes machine downtime, optimizes resource use, and extends the lifespan of industrial equipment. In industrial automation, System 100 and/or method 200 significantly improves predictive maintenance by 25 accurately identifying components likely to fail. Traditional systems may rely on static thresholds or simple trend analysis, which can result in missed failures or unnecessary maintenance. System 100 and/or method 200 processes operational data through hierarchical layers to extract high-level features indicative of wear and tear. The multi-domain bridging mechanism integrates these features with other relevant data, such as environmental 30 conditions and usage patterns, to predict potential failures. This approach ensures timely and accurate identification of components that need maintenance, minimizing downtime and optimizing resource use. The recursive feedback mechanism allows for continuous learning and adaptation, improving the system's predictive accuracy over time.

[1736] For manufacturing process control, system 100 and method 200 dynamically adjust 35 operational parameters such as temperature, pressure, and speed based on real-time predictions. The artificial intelligence system ensures that these adjustments maintain product quality and reduce waste, leading to more efficient and cost-effective manufacturing processes. System 100 and/or method 200 enhances manufacturing process control by dynamically adjusting operational parameters to maintain product quality and reduce waste.

[1737] 40 Traditional systems may use fixed control parameters or simple real-time adjustments, which can be insufficient for complex manufacturing processes. System 100 and/or method 200 processes operational data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as material properties and environmental conditions, to predict the

[1738] 349371

[1741] 75

[1743] optimal control parameters. This dynamic and context-aware approach ensures that the manufacturing process remains within desired quality standards while minimizing waste. The recursive feedback mechanism continuously refines the control parameters based on real-time data, adapting to any changes in the manufacturing environment.

[1745] 5 [0433] In guiding robotic assembly systems, system 100 and method 200 predict the characteristics of components being handled. The artificial intelligence system dynamically adapts the positioning and gripping force of robotic arms, ensuring precise assembly and reducing the risk of damage to components. This adaptability enhances the efficiency and reliability of robotic assembly lines. For robotic assembly systems, System 100 and/or method 10 200 improves performance by dynamically adapting the positioning and gripping force of robotic arms. Traditional systems may use pre-defined paths or simple real-time adjustments, which can lack the precision required for handling delicate components. System 100 and/or method 200 processes data related to the components being handled through hierarchical layers to extract high-level features. The multi-domain bridging mechanism integrates these 15 features with other relevant data, such as component properties and environmental conditions, to predict the optimal positioning and gripping force. This approach ensures precise and reliable assembly, reducing the risk of damage to components. The recursive feedback mechanism continuously refines the predictions and adjustments, adapting to any changes in the assembly environment.

[1747] 20 [0434] For supply chain optimization, system 100 and method 200 analyze logistics data to predict demand and supply fluctuations. The artificial intelligence system adjusts warehouse processing schedules and inventory allocation based on these predictions, reducing energy consumption and transportation costs while ensuring timely delivery of goods. System 100 and/or method 200 enhances supply chain optimization by dynamically adjusting warehouse 25 processing schedules and inventory allocation based on real-time predictions. Traditional systems may use static models or limited real-time adjustments, which can result in inefficiencies and increased costs. System 100 and/or method 200 processes logistics data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as demand 30 forecasts and transportation schedules, to predict the optimal supply chain parameters. This dynamic and context-aware approach ensures efficient energy consumption and transportation costs while maintaining timely delivery of goods. The recursive feedback mechanism continuously refines the supply chain parameters based on real-time data, adapting to any changes in the logistics environment.

[1749] 35 [0435] In renewable energy integration, system 100 and method 200 balance energy supply and demand in grid systems incorporating renewable sources. The artificial intelligence system predicts fluctuations in renewable energy availability and adjusts the operational parameters of energy storage systems or backup generators accordingly. This balancing act ensures a stable and efficient energy supply. In energy management, System 100 and/or method 200 improves 40 renewable energy integration by balancing energy supply and demand in grid systems.

[1750] Traditional systems may use static models or limited real-time adjustments, which can result in inefficiencies and instability. System 100 and/or method 200 processes data related to renewable energy sources through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with

[1751] 349371

[1754] 76

[1756] other relevant data, such as energy consumption patterns and weather forecasts, to predict fluctuations in renewable energy availability. This dynamic and context-aware approach ensures efficient operation of energy storage systems and backup generators, maintaining a stable and efficient energy supply. The recursive feedback mechanism continuously refines 5 the energy management parameters based on real-time data, adapting to any changes in the energy environment.

[1758] For smart grid control, system 100 and method 200 process real-time grid data to predict potential overloads or inefficiencies. The artificial intelligence system controls switching operations to improve energy distribution efficiency and prevent overloading of grid 10 components, ensuring a reliable and resilient power grid. System 100 and/or method 200 enhances smart grid control by dynamically adjusting switching operations to improve energy distribution efficiency and prevent overloading of grid components. Traditional systems may use fixed control parameters or simple real-time adjustments, which can be insufficient for complex grid systems. System 100 and/or method 200 processes real-time grid data through 15 hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as energy consumption patterns and grid topology, to predict potential overloads and inefficiencies. This dynamic and context-aware approach ensures efficient energy distribution and prevents overloading of grid components. The recursive feedback mechanism continuously refines the 20 control parameters based on real-time data, adapting to any changes in the grid environment.

[1760] In optimizing HVAC systems, system 100 and method 200 predict the thermal comfort needs of a building. The artificial intelligence system adjusts operational parameters such as temperature and airflow to reduce energy consumption while maintaining optimal comfort conditions for occupants, leading to more sustainable building management. For HVAC 25 systems, System 100 and/or method 200 improves performance by dynamically adjusting operational parameters to reduce energy consumption while maintaining optimal thermal comfort conditions. Traditional systems may use fixed control parameters or simple real-time adjustments, which can be insufficient for complex building environments. System 100 and/or method 200 processes data related to building conditions through hierarchical layers to 30 generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as occupancy patterns and weather forecasts, to predict the optimal HVAC parameters. This dynamic and context-aware approach ensures efficient energy consumption while maintaining thermal comfort. The recursive feedback mechanism continuously refines the HVAC parameters based on real-time data, 35 adapting to any changes in the building environment.

[1762] [0438] For wind turbine control, system 100 and method 200 optimize the pitch angle and rotational speed of turbines based on wind predictions. The artificial intelligence system ensures that these adjustments maximize energy output while minimizing mechanical stress, enhancing the efficiency and longevity of wind turbines. System 100 and/or method 200 40 enhances wind turbine control by dynamically adjusting the pitch angle and rotational speed to maximize energy output while minimizing mechanical stress. Traditional systems may use fixed control parameters or simple real-time adjustments, which can be insufficient for varying wind conditions. System 100 and/or method 200 processes wind data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism

[1763] 349371

[1766] 77

[1768] integrates these representations with other relevant data, such as turbine properties and environmental conditions, to predict the optimal control parameters. This dynamic and contextaware approach ensures efficient energy output while minimizing mechanical stress. The recursive feedback mechanism continuously refines the control parameters based on real-time 5 data, adapting to any changes in wind conditions.

[1770] In perimeter security for threat identification, system 100 and method 200 analyse data from sensors such as motion detectors and thermal cameras to predict potential security threats. The artificial intelligence system triggers automated responses, including alarms and activation of additional surveillance systems, ensuring robust perimeter security. In perimeter 10 security, System 100 and/or method 200 improves threat identification by dynamically analysing data from sensors such as motion detectors and thermal cameras. Traditional systems may use fixed thresholds or simple pattern recognition, which can result in missed threats or false alarms. System 100 and/or method 200 processes sensor data through hierarchical layers to generate detailed feature representations. The multi-domain bridging 15 mechanism integrates these representations with other relevant data, such as environmental conditions and historical patterns, to predict potential security threats. This dynamic and context-aware approach ensures accurate and timely identification of threats, triggering automated responses such as alarms or activation of additional surveillance systems. The recursive feedback mechanism continuously refines the threat identification parameters based 20 on real-time data, adapting to any changes in the security environment.

[1772] For biometric security systems, system 100 and method 200 process biometric data to enhance identity verification accuracy. The artificial intelligence system predicts potential identity mismatches and adjusts verification parameters, improving the reliability and security of access control systems. For biometric security systems, System 100 and/or method 200 25 enhances identity verification accuracy by dynamically processing biometric data such as facial or iris scans. Traditional systems may use fixed thresholds or simple pattern recognition, which can result in false positives or negatives. System 100 and/or method 200 processes biometric data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as user 30 profiles and environmental conditions, to predict the optimal verification parameters. This dynamic and context-aware approach ensures accurate and reliable identity verification. The recursive feedback mechanism continuously refines the verification parameters based on realtime data, adapting to any changes in the biometric data.

[1774] [0441] In surveillance camera optimization, system 100 and method 200 dynamically adjust 35 camera settings such as resolution, focus, and frame rate based on predictions. The artificial intelligence system ensures that these adjustments improve image quality for identified regions of interest while reducing energy consumption, enhancing the effectiveness of surveillance operations. System 100 and/or method 200 improves surveillance camera optimization by dynamically adjusting camera settings such as resolution, focus, and frame rate based on real-40 time predictions. Traditional systems may use fixed settings or simple real-time adjustments, which can result in suboptimal image quality and increased energy consumption. System 100 and/or method 200 processes surveillance data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as environmental conditions and historical

[1775] 349371

[1778] 78

[1780] patterns, to predict the optimal camera settings. This dynamic and context-aware approach ensures high-quality images for identified regions of interest while reducing energy consumption. The recursive feedback mechanism continuously refines the camera settings based on real-time data, adapting to any changes in the surveillance environment.

[1782] 5 [0442] For network security intrusion detection, system 100 and method 200 analyse network traffic to predict potential cyberattacks. The artificial intelligence system detects anomalies indicative of intrusions and triggers automated countermeasures to isolate affected nodes, preventing system-wide breaches and ensuring network security. In network security, System 100 and/or method 200 improves intrusion detection by dynamically analysing network traffic 10 to predict potential cyberattacks. Traditional systems may use fixed thresholds or simple pattern recognition, which can result in missed intrusions or false alarms. System 100 and/or method 200 processes network traffic data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as historical traffic patterns and known attack 15 signatures, to predict potential intrusions. This dynamic and context-aware approach ensures accurate and timely detection of cyberattacks, triggering automated countermeasures to isolate affected nodes and prevent system-wide breaches. The recursive feedback mechanism continuously refines the intrusion detection parameters based on real-time data, adapting to any changes in network traffic.

[1784] 20 [0443] In air quality monitoring, system 100 and method 200 analyse sensor data to predict pollution hotspots. The artificial intelligence system triggers automated adjustments to ventilation systems or emissions controls, improving air quality in the monitored area and ensuring a healthier environment. For air quality monitoring, System 100 and/or method 200 improves performance by dynamically analysing sensor data to identify pollution hotspots.

[1785] 25 Traditional systems may use fixed thresholds or simple pattern recognition, which can result in missed pollution events or false alarms. System 100 and/or method 200 processes air quality data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as environmental conditions and historical patterns, to predict pollution hotspots. This dynamic 30 and context-aware approach ensures accurate and timely identification of pollution events, triggering automated adjustments to ventilation systems or emissions controls. The recursive feedback mechanism continuously refines the air quality monitoring parameters based on realtime data, adapting to any changes in the monitored area.

[1787] [0444] For wildfire detection, system 100 and method 200 process environmental sensor data 35 to predict early signs of wildfires. The artificial intelligence system triggers alerts and activates firefighting resources in affected regions, enabling rapid response and minimizing the impact of wildfires. System 100 and/or method 200 enhances wildfire detection by dynamically analysing environmental sensor data to predict early signs of wildfires. Traditional systems may use fixed thresholds or simple pattern recognition, which can result in missed wildfires or false 40 alarms. System 100 and/or method 200 processes environmental data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as weather conditions and historical patterns, to predict potential wildfires. This dynamic and context-aware approach ensures accurate and timely detection of wildfires, triggering alerts and activating firefighting

[1788] 349371

[1790] 79

[1792] resources in affected regions. The recursive feedback mechanism continuously refines the wildfire detection parameters based on real-time data, adapting to any changes in the environmental conditions.

[1794] In water management, system 100 and method 200 optimize irrigation systems by 5 predicting water requirements for specific zones. The artificial intelligence system dynamically adjusts flow rates to ensure efficient water distribution, conserving resources and improving agricultural productivity. For water management, System 100 and/or method 200 improves performance by dynamically adjusting water distribution in irrigation systems based on realtime predictions. Traditional systems may use fixed schedules or simple real-time adjustments, 10 which can result in inefficient water use and suboptimal crop yields. System 100 and/or method 200 processes irrigation data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as soil moisture levels and weather forecasts, to predict the optimal water distribution parameters. This dynamic and context-aware approach ensures efficient 15 water use and improved crop yields. The recursive feedback mechanism continuously refines the water management parameters based on real-time data, adapting to any changes in the irrigation environment.

[1796] For urban traffic control, system 100 and method 200 analyse traffic data to predict congestion patterns. The artificial intelligence system optimizes traffic light scheduling to 20 reduce congestion and emissions, improving the flow of vehicles and pedestrians and enhancing urban mobility. System 100 and/or method 200 enhances urban traffic control by dynamically adjusting traffic light scheduling to reduce congestion and emissions while improving the flow of vehicles and pedestrians. Traditional systems may use fixed schedules or simple real-time adjustments, which can be insufficient for complex urban environments.

[1797] 25 System 100 and/or method 200 processes traffic data through hierarchical layers to generate detailed feature representations. The multi-domain bridging mechanism integrates these representations with other relevant data, such as traffic patterns and environmental conditions, to predict the optimal traffic light scheduling. This dynamic and context-aware approach ensures efficient traffic flow and reduced congestion and emissions. The recursive feedback 30 mechanism continuously refines the traffic control parameters based on real-time data, adapting to any changes in the urban environment.

[1799] Figure 5 shows an example computing system. Specifically, Figure 5 shows a block diagram of an embodiment of a computing system 500 according to example embodiments of the present disclosure.

[1801] 35 [0448] Computing system 500 can be configured to perform any of the operations disclosed herein such as, for example, any of the processes or methods described herein. Computing system 500 comprises one or more computing devices 502. Computing device 502 of computing system 500 comprises one or more processors 504 and memory 506. device(s) 502 of computing system 500 comprise one or more processors 504 and memory 506. One 40 or more processors 504 can be any general-purpose processor(s) configured to execute a set of instructions. For example, one or more processors 504 can be one or more general-purpose processors, one or more field programmable gate array (FPGA), and/or one or more application specific integrated circuits (ASIC). In one embodiment, one or more processors

[1802] 349371

[1805] 80

[1807] 504 include one processor. Alternatively, one or more processors 504 include a plurality of processors that are operatively connected. One or more processors 504 are communicatively coupled to memory 506 via address bus 508, control bus 510, and data bus 512. Memory 506 can be a random-access memory (RAM), a read-only memory (ROM), a persistent storage 5 device such as a hard drive, an erasable programmable read-only memory (EPROM), and/or the like. Computing device(s) 502 further comprise input/output (I/O) interface 514 communicatively coupled to address bus 508, control bus 510, and data bus 512.

[1809] Memory 506 can store information that can be accessed by one or more processors 504. For example, memory 506 (e g., one or more non-transitory computer-readable storage 10 mediums, memory devices) can include computer-readable instructions (not shown) that can be executed by one or more processors 504. Alternatively, or additionally, memory 506 can be a combination of information storage and the option to process data in-memory without having to read out the data first, and let the processor 504 perform operations on the memory data. For example, in memory computing is optionally used for CTM implementation, e.g. for 15 “weighting” of clauses, where the weights from a trained model are stored in memory and the weighting is automatically performed when a clause evaluates to 1.

[1811] The computer-readable instructions can be software written in any suitable programming language or can be implemented in hardware. Additionally, or alternatively, the computer-readable instructions can be executed in logically and/or virtually separate threads 20 on one or more processors 504. For example, memory 506 can store instructions (not shown) that when executed by one or more processors 504 cause one or more processors 504 to perform operations such as any of the operations and functions for which computing system 500 is configured, as described herein. In addition, or alternatively, memory 506 can store data (not shown) that can be obtained, received, accessed, written, manipulated, created, and/or 25 stored. The data can include, for instance, the data and/or information described herein in relation to Figures 1 to 6. In some implementations, computing device(s) 502 can obtain from and/or store data in one or more memory device(s) that are remote from the computing system 500.

[1813] Computing system 500 further comprises storage unit 516, network interface 518, input 30 controller 520, and output controller 522. Storage unit 516, network interface 518, input controller 520, and output controller 522 are communicatively coupled to central control unit or computing devices 502 via I/O interface 514.

[1815] Storage unit 516 is a computer readable medium, preferably a non-transitory computer readable medium, comprising one or more programs, the one or more programs comprising 35 instructions which when executed by one or more processors 504 cause computing system 500 to perform the method steps of the present disclosure. Alternatively, storage unit 516 is a transitory computer readable medium. Storage unit 516 can be a persistent storage device such as a hard drive, a cloud storage device, or any other appropriate storage device.

[1817] [0453] Network interface 518 can be a Wi-Fi module, a network interface card, a Bluetooth 40 module, and/or any other suitable wired or wireless communication device. In an embodiment, network interface 518 is configured to connect to a network such as a local area network (LAN), or a wide area network (WAN), the Internet, or an intranet.

[1818] 349371

[1820] 81

[1822] Figure 5 illustrates one example computer system 500 that can be used to implement the present disclosure. Other computing systems can be used as well. Computing tasks discussed herein as being performed at and/or by one or more functional unit(s) can instead be performed remote from the respective system, or vice versa. Such configurations can be 5 implemented without deviating from the scope of the present disclosure. The use of computerbased systems allows for a great variety of possible configurations, combinations, and divisions of tasks and functionality between and among components. Computer-implemented operations can be performed on a single component or across multiple components. Computer-implemented tasks and/or operations can be performed sequentially or in parallel.

[1823] 10 Data and instructions can be stored in a single memory device or across multiple memory devices.

[1825] Regarding the above disclosure, references to items in the singular should be understood to include items in the plural, and vice versa, unless explicitly stated otherwise or clear from the context. Additionally, grammatical conjunctions are intended to express any and 15 all disjunctive and conjunctive combinations of conjoined clauses, sentences, words, and the like, unless otherwise stated or clear from the context. Thus, the term “or” should generally be understood to mean “and/or” and so forth. The use of any and all examples, or exemplary language (“e.g.,” “such as,” “including,” or the like) provided herein, is intended merely to better illuminate the embodiments and does not pose a limitation on the scope of the 20 embodiments or the claims.

[1827] Methods described herein may relate to a computer storage product with a nontransitory computer-readable medium (also can be referred to as a non-transitory processorreadable medium) having instructions or computer code thereon for performing various computer-implemented operations, such that the methods are performed. The computer-25 readable medium (or processor-readable medium) is non-transitory in the sense that it does not include transitory propagating signals per se (e.g., a propagating electromagnetic wave carrying information on a transmission medium such as space or a cable). The media and computer code (also can be referred to as code) may be those designed and constructed for the specific purpose or purposes. Examples of non-transitory computer-readable media 30 include, but are not limited to, magnetic storage media such as hard disks, floppy disks, and magnetic tape, optical storage media such as Compact Disc/Digital Video Discs (CD/DVDs), Compact Disc-Read Only Memories (CD-ROMs), and holographic devices; magneto-optical storage media such as optical disks; carrier wave signal processing modules; and hardware devices that are specially configured to store and execute program code, such as Application-35 Specific Integrated Circuits (ASICs), Programmable Logic Devices (PLDs), Read-Only Memory (ROM) and Random-Access Memory (RAM) devices. Other embodiments described herein relate to a transitory computer program product, which can include, for example, the instructions and/or computer code discussed herein.

[1829] [0457] Some embodiments and/or methods described herein can be performed by software 40 (executed on hardware), hardware, or a combination thereof. Hardware modules include, for example, a general-purpose processor, a field programmable gate array (FPGA), and/or an application specific integrated circuit (ASIC). Software modules (executed on hardware) can be expressed in a variety of software languages (e.g., computer code), including C, C++, Java, Ruby, Visual Basic, Python, and/or other object-oriented, procedural, or other programming

[1830] 349371

[1832] 82

[1834] language and development tools. Examples of computer code include, but are not limited to, micro-code or micro-instructions, machine instructions, such as produced by a compiler, code used to produce a web service, and files containing higher-level instructions that are executed by a computer using an interpreter. For example, embodiments can be implemented using 5 imperative programming languages (e.g., C, Fortran, etc.), functional programming languages (Haskell, Erlang, etc.), logical programming languages (e.g., Prolog), object-oriented programming languages (e.g., Java, C++, etc.) or other suitable programming languages and/or development tools. Additional examples of computer code include, but are not limited to, control signals, encrypted code, and compressed code. Optionally, the embodiments and/or 10 methods described herein are implemented using an operating system such as Robot Operating System (ROS).

[1836] The skilled person will also understand that any use of “or” throughout description herein encompasses use of “or”, “and/or”, and “and”. For example, the term "or" within the discourse is construed to encompass both "and" and "and/or" owing to its inherent inclusivity.

[1837] 15 Within linguistic reasoning, "or" denotes an inclusive disjunction, allowing for the consideration of scenarios wherein either one condition holds true, the other condition holds true, or both conditions hold true concurrently. This interpretation inherently incorporates the conjunction "and", permitting the acknowledgment of scenarios wherein multiple conditions coexist. Additionally, the term "and/or" explicitly acknowledges the possibility of either condition being 20 singularly true or both conditions being true simultaneously, thus aligning with the broader meaning of "or" within the context of this disclosure. Consequently, "or" functions as a flexible connector within the statements of invention, accommodating both exclusive and inclusive interpretations to suit the nuanced requirements of embodiments described herein.

Claims (15)

1. 349371

83

Claims

1. $P HWKRGRISURFHVVLQJGDWDXVLQJDQDUWLÞFLDOLQWHOLJHQFHV\ VWHP FRP SULVLQJDSOXUDOLW\ of processing modules, each module being domain-VSHFLÞFDQGFRP SULVLQJDKLHUDUFKLFDO neural network, the method comprising:

receiving input data;

processing the input data through a hierarchical neural network, wherein:

WKHKLHUDUFKLFDOQHXUDOQHWZRUNFRP SULVHVDSOXUDOLW\ RIOD\ HUVFRQÞJXUHGWR extract features of increasing complexity;

each layer transforms its output according to the output of previous layers, such that dimensional transformations occur across layers from the input GDWDWRDÞQDODEVWUDFWUHSUHVHQWDWLRQVDQG

the layer depth and dimensional progression are dynamically adjustable based on domain-VSHFLÞFFRP SOH[LW\

characterised by:

UHÞQLQJOD\ HURXWSXWVXVLQJUHFXUVLYHIHHGEDFNDQGDSSO\ LQJWKHUHFXUVLYHIHHGEDFNWR the output of each layer;

SUXQLQJQHXUDOFRQQHFWLRQVGXULQJRIIOLQHUHFRQÞJXUDWLRQRUSHULRGVRIUHGXFHGOD\ HU activity, wherein pruning comprises:

calculating a pruning threshold for each layer; and

selectively deactivating weights that are low contributing based on the pruning threshold;

modifying the layer outputs by integrating transformed outputs from other processing modules using weighted bridging transformations;

after generating modular outputs associated with the layer outputs from domain-

VSHFLÞFP RGXOHVLQSDUDOHOZHLJKLQJWKHP RGXODURXWSXWVXVLQJUHOHYDQFHVFRUHV

LIWKHDUWLÞFLDOLQWHOLJHQFHV\ VWHP FRP SULVHVFRQWH[WVLJQDOVDQGHP RWLYHZHLJKWLQJ

adjusting bridging and modular weighting parameters based on the context signals and emotive weighting;

integrating the layer outputs, recursive feedback, domain-VSHFLÞFP RGXOHVDQG bridging transformations into a probabilistic output function;

selecting a prediction based on a highest probable output of the probabilistic output function;

VWRULQJLQWHUPHGLDWHDQGÞQDORXWSXWVin a memory log for any of: offline replay, KLHUDUFKLFDOUHFRQÞJXUDWLRQOD\ HUWXQLQJRUOD\ HUSUXQLQJDQG

outputting the prediction for one or more downstream actions.

2. The method of claim 1, further comprising:

generating sub-lane outputs for each domain of the input data by subdividing domains into sub-lanes;

applying bridging transformations to sub-lane outputs;

LIWKHDUWLÞFLDOLQWHOLJHQFHV\ VWHP FRP SULVHVHP RWLYHZHLJKWLQJVFDOLQJEULGJLQJ

outputs using emotive weighting;

summing bridging outputs into a synergy vector; and

integrating the synergy vector into the probabilistic output function.

3. 7KHP HWKRGRIFODLP ZKHUHLQWKHDUWLÞFLDOLQWHOLJHQFHV\ VWHP FRP SULVHVHP RWLYH weighting, and the emotive weighting for scaling bridging outputs is dynamically scaled based on historical false-positive rates or synergistic contribution patterns.

4. The method of claim 3, further comprising:

349371

84

logging an impact of applied emotive weighting on domain outputs or performance metrics for generating feedback; and

adjusting future emotive weighting, wherein a supervisory module evaluates performance feedback and overrides scaling parameters based on historical falsepositive rates or synergistic contribution patterns.

5. The method of any preceding claim, further comprising:

calculating the synergy vector by summing outputs from the domain-VSHFLÞFP RGXOHV or sub-lanes and scaling the outputs based on domain-VSHFLÞFHP RWLYHZHLJKWLQJ recursively updating a global state using a recursion factor; and

evaluating persistence of prior context by monitoring the norm of the global state and adjusting the recursion factor during offline processes or periods of reduced activity.

6. The method of claim 5, further comprising:

encoding sequential data from multiple time steps into the memory log, wherein each time step is associated with a step-wide record comprising time, domain input, emotive weighting synergy vector, global state, and the one or more downstream actions;

conducting an offline replay of the memory log during reduced activity phases, wherein:

synergy patterns are analysed across consecutive steps for identifying multistep anomalies; and

emotive weighting and bridging transformations are adjusted based on cumulative signals or repeated patterns;

evaluating a novelty value during the offline replay and generating recommendations for creating new domain sub-ODQHVDGMXVWLQJHP RWLYHEDVHOLQHVRUUHÞQLQJEULGJLQJ transforms if novelty exceeds a threshold; and UHÞQLQJUHFXUVLYHIHHGEDFNSUXQLQJWKUHVKROGVRUGRP DLQ-lane hierarchy during subsequent system operations based on the offline replay.

7. The method of claim 6, further comprising:

integrating emotive weighting, synergy vectors and global states across multiple domains to detect cross-GRP DLQSDWWHUQVXVLQJDQDOLJQP HQWP HFKDQLVP IRUUHÞQLQJ bridging transformations;

during the offline replay, identifying anomalous sub-lanes by analysing synergy deviations;

re-evaluating novelty values during the offline replay based on sub-lane contributions and cross-domain interactions;

expanding or merging sub-lanes when repeated patterns across time steps satisfy preset conditions;

simulating a future domain input based on the memory log, wherein a predicted future domain input is used to pre-adjust emotive weighting and bridging parameters; and enhancing modular relevance scores during system operation by incorporating KLVWRULFDOQRYHOW\ WUHQGVEDVHGRQQRUP DOLVHGQRYHOW\ YDOXHVRYHUDGHÞQHGKLVWRULFDO window.

8. The method of any preceding claim, further comprising analysing the memory log to identify patterns, and:

creating a new sub-lane if the patterns are not captured by existing sub-lanes; or unifying sub-lanes within a domain if the patterns indicate redundancy by combining bridging transforms.

349371

85

9. The method of any preceding claim, further comprising:

analysing the memory log to identify an overactive emotive weighting during system operation, wherein synergy vectors are excessively impacted by the emotive weighting; and

restrictive the overactive emotive weighting by providing a minimum or maximum value.

10. The method of any preceding claim, further comprising:

analysing memory log to detect cause-effect relationships across domain lanes; storing detected cause-effect relationships as symbolic rules in a cause-effect repository;

using the detected cause-effect relationships to anticipate and modify synergy vectors; and

generate a symbolic explanation for modifying the synergy vectors for explainability.

11. The method of claim 10, further comprising using the cause-effect repository during meta-updates, wherein:

SURDFWLYHDGMXVWPHQWVDUHDSSOLHGLIDFRQÞGHQFHYDOXHDVVRFLDWHGZLWKDFDXVH-effect UHODWLRQVKLSLVJUHDWHUWKDQDFRQÞGHQFHWKUHVKROGDQG

proactive adjustments comprise adding a pre-emptive emotive weighting offset for a domain-lane or scaling a bridging transformation matrix.

12. The method of any preceding claim, wherein:

each domain sub-lane is evaluated for pruning based on usage metrics; and sub-lanes are pruned if both p and p fall below respective thresholds.

13. The method of any preceding claim, further comprising during periods of offline UHFRQÞJXUDWLRQRUORZDFWLYLW\

replaying the memory log to:

analyse multi-step patterns of synergy and emotive signals;

identify repeated anomalies or patterns missed by single-step evaluation; generate recommendations for structural adjustments to bridging transformation matrixes or emotive weightings;

incorporate meta-updates via a supervisory module, wherein performance metrics are

XVHGWRDGMXVWUHFXUVLRQIDFWRUVHP RWLYHEDVHOLQHVRUVWUXFWXUDOFRQÞJXUDWLRQV

comprising the creation, pruning, or merging of sub-lanes.

14. The method of claim 13, further comprising resetting usage statistics comprising invocation counts, synergy sums, and emotive sums, to ensure accurate metrics for subsequent iterations.

15. A hardware enabled artificial intelligence system configured to perform the method of any of claims 1-14.

349371

86