A large model cognitive separation method, system, electronic device and readable storage medium

Abstract

This application relates to a method, system, electronic device, and readable storage medium for large-scale cognitive separation, and pertains to the field of artificial intelligence. The method includes constructing a cognitive separation neurodynamic system, modeling N neurons as N cognitive agents in a non-cooperative game, constructing an influence matrix, solving for the equilibrium parameters of the influence matrix, and outputting a cognitive strategy; updating the current activation state of the cognitive agents and correcting the equilibrium parameters of the influence matrix; monitoring the driving variables of the cognitive agents in real time to maintain the activity of the cognitive separation neurodynamic system; continuously evolving the cognitive separation neurodynamic system in the form of a system of ordinary differential equations; determining whether an external large-scale model needs to inject impulse vectors; if so, after the impulse vector injection, updating the influence matrix parameters and the cognitive strategy based on impulse learning rules. The beneficial effects of this application are: most cognition is completed locally, without consuming large-scale model inference resources; decision-making is guaranteed to converge by an energy function; the system becomes increasingly independent with use; and the dependence on the large model decreases with experience.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence and neural network technology, specifically to a method, system, electronic device, and readable storage medium for large model cognitive separation. Background Technology

[0002] Current mainstream artificial intelligence systems, especially large language models (LLMs), encode cognitive abilities completely in a large-scale parameter matrix. Every step of cognition in the system, such as perceiving the nature of input, judging the type of task, and deciding on the action method, must be completed through forward inference of the model. In continuous data stream processing scenarios that require continuous perception and autonomous judgment, the computational cost of cognition is linearly coupled with the cognitive activity itself, resulting in high cognitive costs.

[0003] The AI ​​Agent frameworks that have emerged in recent years operate under these constraints: the subject of cognition is still the large language model itself, each round of planning and judgment consumes reasoning resources, the system has no continuous internal state independent of the reasoning process, no parameters that continuously evolve across tasks, and no mechanism to change its own structure from experience.

[0004] Therefore, a technical solution is needed that can separate cognitive processes from large language models and entrust them to an independent, continuously operating neurodynamic system. Summary of the Invention

[0005] In view of the above problems, this application is made in order to provide a large-model cognitive separation method that overcomes or at least partially solves the above problems.

[0006] According to one aspect of this application, a method for large-scale cognitive separation is provided, the method comprising: A cognitive separation neurodynamic system is constructed, which models N neurons as N cognitive agents in a non-cooperative game. The N neurons are distributed in the cognitive function layer according to their cognitive functions. An influence matrix is ​​constructed, and the initial parameters of the influence matrix are encoded according to the relationship between the cognitive functions. In response to continuous data stream input, the parameters of the influence matrix are updated in real time. The equilibrium parameters of the influence matrix are solved, and the cognitive strategy is output. The cognitive agent's current activation state is updated based on the state update rules, the equilibrium parameters of the influence matrix are corrected, and the cognitive strategy is updated. The driving variables of the cognitive agent are monitored in real time, and the activity of the cognitive dissociation neurodynamic system is maintained based on the driving variable update rules. The cognitive separation neurodynamic system continues to evolve in the form of a system of ordinary differential equations; Based on the local features of the trajectory of the cognitive separation neurodynamic system, it is determined in real time whether the large model call gate value of the cognitive separation neurodynamic system exceeds a preset threshold. If so, the pulse type is determined, the large model is called to obtain the response vector corresponding to the pulse type, and injected into the cognitive separation neurodynamic system. Calculate the energy function difference of the cognitive separation neurodynamic system before and after pulse injection, determine whether the energy decreases after pulse injection, and if so, update the influence matrix parameters based on the pulse learning rule and update the cognitive strategy.

[0007] Optionally, in the above method, the influence matrix is ​​an N×N real symmetric matrix, constituting an exact potential game; The cognitive functional layers include, but are not limited to: the perception layer, the cognition layer, the evaluation layer, the action layer, the constraint layer, the attention layer, the basic level, the social layer, and the metacognitive layer; The utility function of the cognitive agent i is: , in, Let represent the current activation strength of cognitive agent i, with a value ranging from [0,1]. For the strategy combination of all cognitive agents other than the current cognitive agent i, The element in the i-th row and j-th column of the influence matrix represents the influence coefficient of cognitive agent j on cognitive agent i. The external signal input received by cognitive agent i is provided by a data stream from outside the system. The activation function of the cognitive agent i is the sigmoid function. Let the adjustment cost function of cognitive agent i be defined as the activation function. The integral of the inverse function; The process of solving the equilibrium parameters of the influence matrix includes: solving the spectral radius of the influence matrix in real time, determining whether the spectral radius of the influence matrix is ​​less than the convergence threshold, and if so, determining that the cognitive separation neurodynamic system is in a Nash equilibrium state of the game, solving the local minimum value of the energy function, and obtaining the equilibrium parameters of the influence matrix. The cognitive strategy is a combination of strategies in which all cognitive agents have no unilateral deviation motivation, including the current state vector.

[0008] Optionally, in the above method, updating the current activation state of the cognitive agent based on the state update rule includes: The state update rule is as follows: , in, It is the inertial parameter of cognitive agent i, with a value range of (0,1). It is the basic gain parameter of the cognitive agent i, which controls the sensitivity to the input signal. It is the temperature parameter of the l-th layer at time step t, which decreases according to the exponential annealing rule, and is the initial temperature multiplied by the decay coefficient raised to the power of t. The lower limit is not lower than the minimum temperature value. The pre-activation input aggregates four signal sources, including: , in, Let be the effective fan-in degree of cognitive agent i, and let be the number of non-zero elements in the i-th row of the influence matrix. The fan-in normalization coefficient ensures that cognitive agents with different connection densities have comparable signal amplitudes. This is the inherent bias of cognitive agent i, reflecting the basic activity level of this cognitive process. The value of the k-th emotional component at time step t is calculated by multiplying the emotional readout matrix R by the current state vector s(t). It is the element in the i-th row and k-th column of the emotion feedback matrix, defined as the feedback intensity coefficient multiplied by the transpose of the emotion readout matrix R, where E is the dimension of the emotion space. It is the user-personalized bias of cognitive agent i, which encodes the cumulative interaction tendencies of a specific user.

[0009] Optionally, in the above method, the drive variable is the current search impulse intensity of the cognitive agent, and the drive variable update rule includes: , in, Let i be the driving force value of the cognitive agent i at time step t. For passive decay rate, constant temperature regulation is provided, with a value range of (0,1). The stagnation accumulation rate refers to the rate at which the drive increases when the cognitive agent's state remains unchanged for a long period of time. To meet the decay rate, the rate at which the driving force decreases upon the arrival of a new external input, Let be the absolute value of the rate of change of the state of cognitive agent i, so that the state change is normalized to the range [0,1]. The absolute value of the change in the external input signal; The driving force variable represents the current search impulse intensity of cognitive agent i, with a lower limit of 0 and an upper limit of the driving force of cognitive agent i.

[0010] Optionally, in the above method, the cognitive separation neurodynamic system continuously evolves in the form of a system of ordinary differential equations, including: , Where B is the somatic state vector, including but not limited to energy, arousal, comfort, information hunger, social needs, and circadian rhythm phase. The effective decay rate of cognitive agent i is modulated by the energy dimension of the body state vector B. The effective gain of cognitive agent i is modulated by the arousal dimension of the somatic state vector B, η i The driving force perturbation direction variable, with a value range of [0,1], aims to make the perturbation direction tend to pull the cognitive agent's state back to a neutral value, preventing state saturation. For energy pumping terms, spontaneous activity driven by the energy dimension when the cognitive agent's state approaches a neutral value; The physical state evolves on a timescale that is 100 times slower than the neural state.

[0011] Optionally, in the above method, the large model call gate value is the maximum value of the trigger signal, and the trigger signal includes, but is not limited to, trajectory chaos signal, dynamic dilemma signal, inter-layer conflict signal, and input novelty signal; The preset threshold is an adjustable hyperparameter; The pulse types include, but are not limited to, chaotic, dilemma-based, conflict-based, and novel pulses. The corresponding vector is an N-dimensional vector or an N×N matrix; The injection into the cognitive separation neurodynamic system includes, but is not limited to: multiplying an N-dimensional vector by a scaling factor and merging it into the current state vector, and multiplying an N×N matrix by a scaling factor matrix and merging it into the influence matrix.

[0012] Optionally, in the above method, the impulse learning rule includes: , in, To influence the update amount of the element in the i-th row and j-th column of the matrix, The learning rate controls the magnitude of each update. The difference in energy function before and after pulse injection. The state value of cognitive agent i after pulse injection. The state value of cognitive agent i before pulse injection. The state value of cognitive agent i before pulse injection.

[0013] According to another aspect of this application, a large-model cognitive separation system is provided, characterized in that it includes: The module is used to construct a cognitive separation neurodynamic system, which models N neurons as N cognitive agents in a non-cooperative game. The N neurons are distributed in the cognitive function layer according to their cognitive functions. An influence matrix is ​​constructed, and the initial parameters of the influence matrix are encoded according to the relationship between the cognitive functions. In response to continuous data stream input, the parameters of the influence matrix are updated in real time, the equilibrium parameters of the influence matrix are solved, and the cognitive strategy is output. The state update module is used to update the current activation state of the cognitive agent based on the state update rules, correct the equilibrium parameters of the influence matrix, and update the cognitive strategy. The intrinsic drive module is used to monitor the drive variables of the cognitive agent in real time and maintain the activity of the cognitive dissociation neural dynamic system based on the drive variable update rules. An evolution module is used for the continuous evolution of the cognitive separation neurodynamic system in the form of a system of ordinary differential equations; The gating judgment module is used to determine in real time whether the gating value of the large model call of the cognitive separation neurodynamic system exceeds a preset threshold based on the local features of the trajectory of the cognitive separation neurodynamic system. If so, the pulse type is determined, the large model is called to obtain the response vector corresponding to the pulse type, and the response vector is injected into the cognitive separation neurodynamic system. The pulse learning module is used to calculate the energy function difference of the cognitive separation neurodynamic system before and after pulse injection, determine whether the energy decreases after pulse injection, and if so, update the influence matrix parameters based on the pulse learning rules and update the cognitive strategy.

[0014] Optionally, in the above system, the construction module is also used to construct an exact potential game by forming an N×N real symmetric matrix of influence matrix; The cognitive functional layers include, but are not limited to: the perception layer, the cognition layer, the evaluation layer, the action layer, the constraint layer, the attention layer, the basic level, the social layer, and the metacognitive layer; The utility function of the cognitive agent i is: , in, Let represent the current activation strength of cognitive agent i, with a value ranging from [0,1]. For the strategy combination of all cognitive agents other than the current cognitive agent i, The element in the i-th row and j-th column of the influence matrix represents the influence coefficient of cognitive agent j on cognitive agent i. The external signal input received by the cognitive agent is provided by a data stream from outside the system. The activation function of the cognitive agent i is the sigmoid function. Let the adjustment cost function of cognitive agent i be defined as the activation function. The integral of the inverse function; The process of solving the equilibrium parameters of the influence matrix includes: solving the spectral radius of the influence matrix in real time, determining whether the spectral radius of the influence matrix is ​​less than the convergence threshold, and if so, determining that the cognitive separation neurodynamic system is in a Nash equilibrium state of the game, solving the local minimum value of the energy function, and obtaining the equilibrium parameters of the influence matrix. The cognitive strategy is a combination of strategies in which all cognitive agents have no unilateral deviation motivation, including the current state vector.

[0015] Optionally, in the above system, the state update module is further configured to update the current activation state of the cognitive agent based on the state update rules, including: The state update rule is as follows: , in, It is the inertial parameter of cognitive agent i, with a value range of (0,1). It is the basic gain parameter of the cognitive agent i, which controls the sensitivity to the input signal. It is the temperature parameter of the l-th layer at time step t, which decreases according to the exponential annealing rule, and is the initial temperature multiplied by the decay coefficient raised to the power of t. The lower limit is not lower than the minimum temperature value. The pre-activation input aggregates four signal sources, including: , in, Let be the effective fan-in degree of agent i, and let be the number of non-zero elements in the i-th row of the influence matrix. The fan-in normalization coefficient ensures that cognitive agents with different connection densities have comparable signal amplitudes. This is the inherent bias of cognitive agent i, reflecting the basic activity level of this cognitive process. The value of the k-th emotional component at time step t is calculated by multiplying the emotional readout matrix R by the current state vector s(t). It is the element in the i-th row and k-th column of the emotion feedback matrix, defined as the feedback intensity coefficient multiplied by the transpose of the emotion readout matrix R, where E is the dimension of the emotion space. It is the user-personalized bias of cognitive agent i, which encodes the cumulative interaction tendencies of a specific user.

[0016] Optionally, in the above system, the intrinsic drive module is further configured to define the drive variable as the current search impulse intensity of the cognitive agent, and the drive variable update rule includes: , in, Let i be the driving force value of the cognitive agent i at time step t. For passive decay rate, constant temperature regulation is provided, with a value range of (0,1). The stagnation accumulation rate refers to the rate at which the drive increases when the cognitive agent's state remains unchanged for a long period of time. To meet the decay rate, the rate at which the driving force decreases upon the arrival of a new external input, Let be the absolute value of the rate of change of the state of cognitive agent i, so that the state change is normalized to the range [0,1]. The absolute value of the change in the external input signal; The driving force variable represents the current search impulse intensity of cognitive agent i, with a lower limit of 0 and an upper limit of the driving force of cognitive agent i.

[0017] Optionally, in the above system, the evolution module is further used for the continuous evolution of the cognitive separation neurodynamic system in the form of a system of ordinary differential equations, including: , Where B is the somatic state vector, including but not limited to energy, arousal, comfort, information hunger, social needs, and circadian rhythm phase. The effective decay rate of cognitive agent i is modulated by the energy dimension of the body state vector B. The effective gain of cognitive agent i is modulated by the arousal dimension of the somatic state vector B, η i The driving force perturbation direction variable has a value range of [0,1], which makes the perturbation direction tend to pull the cognitive agent's state back to a neutral value, preventing state saturation. For energy pumping terms, spontaneous activity driven by the energy dimension when the cognitive agent's state approaches a neutral value; The physical state evolves on a timescale that is 100 times slower than the neural state.

[0018] Optionally, in the above system, the gating judgment module is also used to call the gating value of the large model as the maximum value of the trigger signal, wherein the trigger signal includes, but is not limited to, trajectory chaos signal, dynamic dilemma signal, inter-layer conflict signal, and input novelty signal; The preset threshold is an adjustable hyperparameter; The pulse types include, but are not limited to, chaotic, dilemma-based, conflict-based, and novel pulses. The corresponding vector is an N-dimensional vector or an N×N matrix; The injection into the cognitive separation neurodynamic system includes, but is not limited to: multiplying an N-dimensional vector by a scaling factor and merging it into the current state vector, and multiplying an N×N matrix by a scaling factor matrix and merging it into the influence matrix.

[0019] Optionally, in the above system, the impulse learning module is further configured to include the impulse learning rules as follows: , in, To influence the update amount of the element in the i-th row and j-th column of the matrix, The learning rate controls the magnitude of each update. The difference in energy function before and after pulse injection. The state value of cognitive agent i after pulse injection. The state value of cognitive agent i before pulse injection. The state value of cognitive agent i before pulse injection.

[0020] According to another aspect of this application, an electronic device is provided, characterized in that it includes a processor and a memory, the processor being coupled to the memory, the processor being configured to execute a computer program stored in the memory to cause the electronic device to perform any of the methods described above.

[0021] According to another aspect of this application, a computer-readable storage medium is provided, characterized in that it includes a computer program or instructions that, when executed on a computer, cause the computer to perform any of the methods described above.

[0022] The beneficial effects of this application are as follows: most cognition is completed by the local cognitive separation neurodynamic system, which does not consume the reasoning resources of the large language model, and the operating cost is decoupled from the cognitive complexity; the decision trajectory has Lyapunov convergence guarantee, and each equilibrium state can be traced back to a specific cognitive dimension; the local parameter matrix is ​​learned from each external call, the dependence on the large language model decreases with the accumulation of experience, and the system's functional expressive power continues to grow.

[0023] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application described in detail with reference to the accompanying drawings. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0025] Figure 1 A schematic diagram of a large-model cognitive separation method according to an embodiment of this application is shown.

[0026] Figure 2This paper presents a panoramic view of the continuous ODE evolution under four input conditions of a large-model cognitive separation method according to an embodiment of this application.

[0027] Figure 3 The diagram illustrates a comparison of the α-gated triggering mode of a large model cognitive separation method according to an embodiment of this application across three data streams.

[0028] Figure 4 A schematic diagram of the pulse learning judgment process of a large model cognitive separation method according to an embodiment of this application is shown.

[0029] Figure 5 A schematic diagram of the structure of a large-model cognitive separation system according to an embodiment of this application is shown.

[0030] Figure 6 A schematic diagram of the structure of an electronic device according to an embodiment of this application is shown.

[0031] Figure 7 A schematic diagram of the structure of a computer-readable storage medium according to an embodiment of this application is shown. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] The concept of this application is to integrate neurodynamics with non-cooperative game theory, model neurons as independent cognitive agents, construct a multi-level cognitive separation dynamic system, in which each cognitive agent autonomously solves Nash equilibrium and outputs cognitive strategies through influence matrix game iteration, thereby decoupling cognition from large model inference, maintaining system activity through intrinsic drive, realizing dual-timescale biomimetic evolution with ordinary differential equations, and gating the large model to obtain its return value only when the system triggers the large model call threshold, solidifying the cognitive knowledge output by the large model into the system's own parameters through impulse learning, continuously optimizing the influence matrix, and ultimately realizing a cognitive separation dynamic system that is cognitively independent, has low computational cost, autonomous evolution, and decreasing dependency.

[0034] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0035] Figure 1The diagram illustrates a large-model cognitive separation method according to an embodiment of this application, the method comprising: Step 101: Construct a cognitive separation neurodynamic system, modeling N neurons as N cognitive agents in a non-cooperative game. The N neurons are distributed in the cognitive function layer according to their cognitive functions. Construct an influence matrix, encode the initial parameters of the influence matrix according to the relationship between the cognitive functions, update the influence matrix parameters in real time in response to continuous data stream input, solve the equilibrium parameters of the influence matrix, and output the cognitive strategy.

[0036] The cognitive separation neurodynamic system refers to establishing a neurodynamic system that isolates the cognitive functions of a large model from the large model and relies on the cognitive separation neurodynamic system to realize cognitive functions. When an event occurs that is beyond the scope of cognition, the large model is invoked for cognitive processing, and the result of the invocation is fed back to the cognitive separation neurodynamic system to evaluate whether the system's energy is stable. If the invocation improves the system, the invocation process is learned and the learned content is internalized into the cognitive separation neurodynamic system to gradually optimize the system and reduce the dependence of the cognitive process on the large model.

[0037] The cognitive function layer is distributed according to neurodynamics and can be set up in multiple layers, including but not limited to the perception layer, cognition layer, evaluation layer, action layer, constraint layer, attention layer, basic layer, social layer, and metacognitive layer. Each layer is set with different functions. The influence matrix is ​​a weight matrix, which is a tool for quantifying the connection strength between neurons. Each non-zero connection weight in the influence matrix is ​​set with an initial value by the designer based on the known causal relationship between cognitive functions, and maintains bidirectional symmetry.

[0038] The non-zero connections in the influence matrix are derived from known cross-process causal relationships in cognitive science, including but not limited to: cortical hierarchical feedforward pathways, bidirectional loops in cognitive appraisal theory, decision-making pathways in the somatic marking hypothesis, winner-take-all inhibition in competitive neural networks, prefrontal executive control, neural modulation of attention networks, working memory models, and social adaptation in theory of mind. The initial values ​​of the influence matrix are determined through causal relationships.

[0039] As a continuous data stream is input, the parameter values ​​of the influence matrix iteratively change, and the matrix spectral radius is calculated in real time. When the spectral radius is less than a preset convergence threshold, the system is determined to be in Nash equilibrium. The local minimum of the energy function is calculated, the equilibrium parameters of the influence matrix are obtained, and a stable cognitive strategy without unilateral deviation motivation is output. The cognitive strategy is a vector composed of the activation intensities of all current cognitive agents.

[0040] Step 102: Update the current activation state of the cognitive agent based on the state update rules, correct the equilibrium parameters of the influence matrix, and update the cognitive strategy.

[0041] The activation strength of the cognitive agent is iterated according to the state update rule. The parameters fused by the state update rule include, but are not limited to, inertia parameters, base gain, and exponential annealing temperature parameters. Multiple signal sources are also aggregated through pre-activation input to ensure smooth and stable state updates. The pre-activation input fusion includes, but is not limited to, inherent paranoia, external input, emotional feedback, and user-personalized bias signal sources. The equilibrium parameters of the influence matrix are corrected based on the updated activation state, and the output of the cognitive strategy is optimized synchronously.

[0042] State updates include, but are not limited to, the reaction speed of the cognitive dissociative neurodynamic system, the decisiveness of decision output, feedback speed, and individual bias. The equilibrium state is corrected by the decision performance of the cognitive dissociative neurodynamic system, so that the system remains in a stable state within a relatively range during the continuous correction process.

[0043] Step 103: Monitor the driving variables of the cognitive agent in real time, and maintain the activity of the cognitive dissociation neurodynamic system based on the driving variable update rules.

[0044] The drive variable is defined as the intensity of the cognitive agent's search impulse. The drive value is updated in real time by combining passive decay, stagnation accumulation, and satisfaction decay. The drive increases when the state is stagnant and decreases when new external inputs arrive, maintaining a stable small level of activity in the system.

[0045] Drive parameters are set differently according to the level. For example, the perception layer has high drive to simulate the continuous spontaneous activity of the sensory cortex, the action layer has low drive to maintain the stability of the action plan, and the metacognitive layer has zero drive, where metacognitive judgments are completely determined by the determinism of the network state.

[0046] Step 104: The cognitive separation neurodynamic system continues to evolve in the form of a system of ordinary differential equations.

[0047] The cognitive separation neurodynamic system operates around a set of ordinary differential equations, incorporating somatic state vectors into the evolutionary process. These somatic state vectors include, but are not limited to, energy, arousal, comfort, information hunger, social needs, and circadian rhythm phase. The somatic state vectors are used to modulate the effective decay rate, effective gain, and energy pumping term during the evolutionary process. The somatic state iterates at a timescale 100 times slower than the neural state, thus restoring the physiological rhythm of biological cognition and achieving the co-evolution of cognitive and somatic states.

[0048] Step 105: Based on the local features of the trajectory of the cognitive separation neurodynamic system, determine in real time whether the large model call gate value of the cognitive separation neurodynamic system exceeds a preset threshold. If so, determine the pulse type, call the large model to obtain the response vector corresponding to the pulse type, and inject it into the cognitive separation neurodynamic system.

[0049] The local feature triggering signals of the cognitive separation neurodynamic system trajectory include, but are not limited to, trajectory chaos triggering signals, dynamic dilemma triggering signals, inter-layer conflict triggering signals, and input novelty triggering signals. The local feature triggering signals of the separation neurodynamic system trajectory are collected in real time, and the maximum value is taken as the large model call gate value. The gate value is compared with the corresponding preset threshold. If it exceeds the preset threshold, it is determined that the large model assistance is required. The corresponding pulse type is matched according to the trigger signal type, the large model is called, the response vector or response matrix corresponding to the pulse type is obtained, and merged into the current state vector or influence matrix.

[0050] The process of calling a large model includes: encoding, job search, and decoding.

[0051] Step 106: Calculate the energy function difference of the cognitive separation neurodynamic system before and after pulse injection, determine whether the energy decreases after pulse injection, and if so, update the influence matrix parameters based on the pulse learning rule and update the cognitive strategy.

[0052] The effect of pulse injection is evaluated by calculating the difference in energy function before and after pulse injection. When the energy function decreases, pulse learning is triggered, and the parameters of the influence matrix are updated according to the learning rules. The cognitive experience of the large model is solidified into the system's own parameters. The parameter update amplitude is precisely controlled by the learning rate to avoid over-adjustment. After the knowledge is internalized, the cognitive strategy is updated synchronously to reduce the dependence on the large model.

[0053] Depend on Figure 1 As can be seen from the method shown, the beneficial effects of this application are: cognitive cost decoupling, with most cognition being completed by the local cognitive separation neurodynamic system, without consuming large model inference resources; the system has a global energy function, the decision trajectory has convergence guarantee, and each equilibrium state can be traced back to a specific cognitive dimension; local parameters are learned from external calls each time, and the cognitive dependence on the large model decreases with experience.

[0054] In some embodiments of this application, the above method includes: the influence matrix is ​​an N×N real symmetric matrix, constituting an exact potential game; The cognitive functional layers include, but are not limited to: the perception layer, the cognition layer, the evaluation layer, the action layer, the constraint layer, the attention layer, the basic level, the social layer, and the metacognitive layer; The utility function of the cognitive agent i is: , in, Let represent the current activation strength of cognitive agent i, with a value ranging from [0,1]. For the strategy combination of all cognitive agents other than the current cognitive agent i, The element in the i-th row and j-th column of the influence matrix represents the influence coefficient of cognitive agent j on cognitive agent i. The external signal input received by the cognitive agent is provided by a data stream from outside the system. The activation function of the cognitive agent i is a sigmoid function, which is monotonically increasing and bounded. Let the adjustment cost function of cognitive agent i be defined as the activation function. The integral of the inverse function; The process of solving the equilibrium parameters of the influence matrix includes: solving the spectral radius of the influence matrix in real time, determining whether the spectral radius of the influence matrix is ​​less than the convergence threshold, and if so, determining that the cognitive separation neurodynamic system is in a Nash equilibrium state of the game, solving the local minimum value of the energy function, and obtaining the equilibrium parameters of the influence matrix. The cognitive strategy is a combination of strategies in which all cognitive agents have no unilateral deviation motivation, including the current state vector.

[0055] As a specific carrier of the cognitive strategy described in this application, the cognitive strategy is expressed as an N-dimensional activation vector composed of the activation intensities of all N cognitive agents, where each component of the activation vector takes values ​​in the range [0, 1]. In a specific embodiment, the cognitive strategy simultaneously satisfies the following three equivalent mathematical conditions: (i) The cognitive strategy is the Nash equilibrium of the non-cooperative game, that is, the strategy combination in which all cognitive agents have no unilateral deviation motivation; (ii) The cognitive strategy described is a local maximum of the potential function, which is equivalent to a local minimum of the Hopfield energy function; (iii) The cognitive strategy is the fixed point of the state update rule, that is, the activation value of each cognitive agent remains unchanged after the state update rule is iterated.

[0056] The above three conditions are guaranteed by an exact potential game constructed from symmetric influence matrices. The equivalence relation allows for three interchangeable solution paths to the cognitive strategy: iterative solution through optimal game response, solution through energy function descent search, and solution through fixed-point iterative solution. In practical engineering implementation, the solution path with the lowest computational cost can be selected based on the real-time nature of the data flow scenario, convergence accuracy, and computational resource constraints. The three paths are mathematically equivalent, and the resulting cognitive strategy is the same.

[0057] In a specific embodiment, N is 52 neurons, the potential function is denoted as E(s), and the Nash equilibrium is the strategy combination in which all agents do not deviate from their motivations unilaterally. The Nash equilibrium is also a local minimum of the energy function E(s).

[0058] The 52 neurons are distributed across 9 functional layers: the perception layer (8 neurons) is responsible for extracting features such as the intensity, novelty, and pattern of input signals; the cognition layer (6 neurons) is responsible for cognitive judgments such as attention, certainty, and complexity; the evaluation layer (6 neurons) is responsible for positive / negative evaluation and risk / opportunity judgment; the action layer (8 neurons) is responsible for competitive action strategies such as analysis, synthesis, comparison, prediction, recommendation, explanation, questioning, and creation; the constraint layer (6 neurons) is responsible for output constraints such as accuracy, breadth, depth, and style; the attention layer (4 neurons) is responsible for focusing, inhibition, distraction, and alertness; the memory layer (4 neurons) is responsible for working memory load, long-term memory retrieval, recency, and coherence; the social layer (4 neurons) is responsible for intention recognition, emotion perception, and professionalism judgment; and the metacognitive layer (6 neurons) is responsible for coherence monitoring, confidence, ethical checks, risk assessment, honesty, and boundary perception.

[0059] The spectral radius of the influence matrix (i.e., the absolute value of the largest eigenvalue) multiplied by the maximum derivative of the activation function is less than 1, thus ensuring that the system converges globally to Nash equilibrium. In one specific embodiment, the spectral radius is 0.91, which is much smaller than the upper convergence bound of 4.0.

[0060] By modeling neurons as cognitive agents in a non-cooperative game, each agent pursues the maximization of its own utility during the cognitive decision-making process. When the influence coefficients of all agents form a symmetric matrix, the game constitutes an exact potential game with a Nash equilibrium, and the system converges globally, thus ensuring the stability of the system.

[0061] In some embodiments of this application, the above method includes: updating the current activation state of the cognitive agent based on state update rules includes: The state update rule is as follows: , in, It is the inertial parameter of cognitive agent i, with a value range of (0,1). It is the basic gain parameter of the cognitive agent i, which controls the sensitivity to the input signal. It is the temperature parameter of the l-th layer at time step t, which decreases according to the exponential annealing rule, and is the initial temperature multiplied by the decay coefficient raised to the power of t. The lower limit is not lower than the minimum temperature value. The pre-activation input aggregates four signal sources, including: , in, Let be the effective fan-in degree of agent i, and let be the number of non-zero elements in the i-th row of the influence matrix. The fan-in normalization coefficient ensures that cognitive agents with different connection densities have comparable signal amplitudes. This is the inherent bias of cognitive agent i, reflecting the basic activity level of this cognitive process. The value of the k-th emotional component at time step t is calculated by multiplying the emotional readout matrix R by the current state vector s(t). It is the element in the i-th row and k-th column of the emotion feedback matrix, defined as the feedback intensity coefficient multiplied by the transpose of the emotion readout matrix R, where E is the dimension of the emotion space. It is the user-personalized bias of cognitive agent i, which encodes the cumulative interaction tendencies of a specific user.

[0062] In one specific embodiment, the inertial parameters of different cognitive functional layers have different values: the perception layer uses a smaller parameter, approximately 0.3, to ensure a fast system response, while the metacognitive layer uses a larger value, approximately 0.5, to ensure slow and prudent decision-making. The agent's base gain parameter controls sensitivity to input signals; the perception layer uses a higher value, approximately 0.7, while the metacognitive layer uses a lower value, approximately 0.35. At high temperatures, the system explores multiple strategy combinations; at low temperatures, the system locks into the optimal equilibrium state.

[0063] The temperature annealing rule is T l (t) represents the initial temperature multiplied by the decay coefficient raised to the power of t. The initial temperature is the initial temperature of the first layer. The decay coefficient ranges from (0,1), with a typical value of 0.88 to 0.93. It is used to control the annealing rate. The closer it is to 1, the slower the annealing and the longer the exploration phase. The lower limit is not lower than the minimum temperature value. The minimum temperature value is a preset constant to prevent the activation function from degenerating into a step function when the temperature approaches 0. A typical value is 0.05.

[0064] The emotional space dimension E=12, corresponding to joy, trust, fear, surprise, sadness, disgust, anger, expectation, curiosity, confusion, empathy, and awe.

[0065] The emotion readout vector e(t) is calculated at each time step by multiplying R by the current state vector s(t) and truncated to the range [-1, 1].

[0066] The feedback matrix is ​​defined as the transpose of the feedback intensity coefficient (0.2 in the specific embodiment) multiplied by R. Emotional signals are injected into the pre-activation input of each agent through the feedback matrix, causing the system to generate different equilibrium states for scenarios with varying emotional nuances.

[0067] By using state update rules, the system can be updated through self-updating of inertial parameters, basic gain parameters, temperature parameters, and pre-activation inputs.

[0068] In some embodiments of this application, the above method includes: the drive variable being the current search impulse intensity of the cognitive agent, and the drive variable update rule including: , in, Let i be the driving force value of the cognitive agent i at time step t. For passive decay rate, constant temperature regulation is provided, with a value range of (0,1). The stagnation accumulation rate refers to the rate at which the drive increases when the cognitive agent's state remains unchanged for a long period of time. To meet the decay rate, the rate at which the driving force decreases upon the arrival of a new external input, Let be the absolute value of the rate of change of the state of cognitive agent i, so that the state change is normalized to the range [0,1]. The absolute value of the change in the external input signal; The driving force variable represents the current search impulse intensity of cognitive agent i, with a lower limit of 0 and an upper limit of the driving force of cognitive agent i.

[0069] The accumulated drive is injected into the state update through a perturbation term, which is defined as the drive value multiplied by the exploration direction. The exploration direction includes a central regression bias term, which is defined as the regression strength coefficient multiplied by the difference between the neutral state value and the current state, to prevent the drive from pushing the state into an unrecoverable saturation region.

[0070] Drive parameters are set differently according to the level: high drive in the perception layer is used to simulate the continuous spontaneous activity of the sensory cortex; low drive in the action layer is used to maintain the stability of the user's action plan; and zero drive in the metacognitive layer is used to ensure that metacognitive judgments are completely determined by the deterministic nature of the network state.

[0071] In one specific embodiment, the stagnation accumulation rate parameter is set to 0.8 for the perception layer, 0.02 for the action layer, and 0 for the metacognitive layer. When the cognitive agent's state remains unchanged for a long time, the driving force rises at a higher rate than the perception layer, ensuring that the cognitive agent continues to drive.

[0072] By continuously updating the driving variables, the cognitive agent of the system is ensured not to fall into a static state.

[0073] In some embodiments of this application, the above method includes: the cognitive dissociation neurodynamic system continuously evolves in the form of a system of ordinary differential equations, including: , Where B is the somatic state vector, including but not limited to energy, arousal, comfort, information hunger, social needs, and circadian rhythm phase. The effective decay rate of cognitive agent i is modulated by the energy dimension of the body state vector B. The effective gain of cognitive agent i is modulated by the arousal dimension of the somatic state vector B, η i The driving force perturbation direction variable, with a value range of [0,1], aims to make the perturbation direction tend to pull the cognitive agent's state back to a neutral value, preventing state saturation. For energy pumping terms, spontaneous activity driven by the energy dimension when the cognitive agent's state approaches a neutral value; The physical state evolves on a timescale that is 100 times slower than the neural state.

[0074] The effective decay rate of the cognitive agent accelerates at low energy levels, simulating cognitive sluggishness under fatigue. The effective gain of the cognitive agent increases with high arousal levels, simulating perceptual sharpness under alertness.

[0075] By evolving the cognitive separation neurodynamic system in the form of ordinary differential equations, the neural state and the somatic state are bidirectionally coupled.

[0076] In some embodiments of this application, the above method includes: the large model call gating value is the maximum value of the trigger signal, the trigger signal including but not limited to trajectory chaos signal, dynamic dilemma signal, inter-layer conflict signal, and input novelty signal; The preset threshold is an adjustable hyperparameter; The pulse types include, but are not limited to, chaotic, dilemma-based, conflict-based, and novel pulses. The response vector is an N-dimensional vector or an N×N matrix; The injection into the cognitive separation neurodynamic system includes, but is not limited to: multiplying an N-dimensional vector by a scaling factor and merging it into the current state vector, and multiplying an N×N matrix by a scaling factor matrix and merging it into the influence matrix.

[0077] The large model invocation process includes: First, encoding: Encoding the current state of the cognitive separation neurodynamic system (including a 52-dimensional neural state vector, a 12-dimensional emotion vector, a 6-dimensional somatic state vector, and the most recent memory context) into a structured text sequence, which serves as the input to the large language model. The encoded content includes a state summary, historical information most relevant to the current state, and instructions specifying the output format.

[0078] The second step is evaluation: the encoded text sequence is fed into the large language model for forward inference to obtain the text output.

[0079] The third step, decoding, involves parsing the text output of the large language model into numerical vectors, including direction vectors, escape vectors, connection adjustment matrices, or knowledge supplementation vectors. These vectors are then pruned within a reasonable range and injected into the dynamical system. The text sequence serves as an intermediate representation connecting the numerical state space of the dynamical system and the text parameter space of the large language model. The system does not utilize the language capabilities of the large language model itself, but rather the high-dimensional nonlinear transformations achieved by its internal parameter matrices.

[0080] When the cognitive separation neurodynamic system encounters difficulties, i.e. when the large model is invoked, feedback from the large model is obtained and injected into the cognitive separation neurodynamic system, thereby reducing dependence on the large model and gradually optimizing the cognitive separation neurodynamic system.

[0081] Optionally, in the above method, the impulse learning rule includes: , in, To influence the update amount of the element in the i-th row and j-th column of the matrix, The learning rate controls the magnitude of each update. The difference in energy function before and after pulse injection. The state value of cognitive agent i after pulse injection. The state value of cognitive agent i before pulse injection. The state value of cognitive agent i before pulse injection.

[0082] In one specific embodiment, for a chaotic trajectory signal, the window average of the cosine similarity between the directions of state changes in the most recent several steps (10 steps in the specific embodiment) is calculated. When this value is less than 0.7, it indicates that the trajectory direction is constantly changing and the system is trapped in a state of direction uncertainty.

[0083] The dynamic distress signal is calculated by measuring the difference between the cumulative proportion of the driving force variable and the rate of state change. When the driving force is high and the rate of change is low, it indicates that the system is trapped in a local distress.

[0084] Interlayer conflict signal detection assesses whether opposing neuron pairs (such as positive evaluation neurons and negative evaluation neurons) in the evaluation layer are simultaneously activated above their respective resting levels. The minimum conflict value of the two opposing neuron pairs is summed to indicate that the system has received contradictory input signals.

[0085] Input a novel signal and calculate the mean of the cosine similarity between the current external input vector and several recent historical input vectors. When the mean is low, it indicates that the system has encountered an input pattern that has not appeared before.

[0086] The pulse type is determined based on the largest of the four signals, corresponding to chaotic, dilemma, conflict, and novel types, respectively. The corresponding response vector is obtained by calling the large language model and injected into the system. For chaotic large language models, an N-dimensional direction vector, denoted as d, is returned, representing the suggested state adjustment direction. The system adds the scaled direction vector d to the current state vector s, i.e.: post-injection state = pre-injection state + scaling factor multiplied by d. In a specific implementation, the scaling factor is 0.1.

[0087] For the troubled type, the large language model returns an N-dimensional escape perturbation vector, denoted as p, which is used to push the system away from the currently trapped saddle point. The injection method is the same as that for the chaotic type.

[0088] For conflict-prone, large language models, an N x N connection adjustment matrix is ​​returned. This matrix is ​​used to temporarily modify the influence matrix, allowing conflicting inter-layer signals to resolve naturally. The adjusted influence matrix = original influence matrix + scaling factor multiplied by the adjustment matrix.

[0089] The novel large language model returns an N-dimensional knowledge supplement vector, denoted as k, which is used to enhance the external input signal. The enhanced input = original input + scaling factor multiplied by k.

[0090] By calculating the energy function difference between the cognitive dissociative neurodynamic system before and after pulse injection, if the energy decreases after pulse injection, the pulse injection is effective. The cognitive dissociative neurodynamic system records this state change process, performs pulse learning based on the pulse learning rule, updates the influence matrix parameters, and updates the cognitive strategy. When encountering similar problems again, the large model can be avoided, and the use of the large model can be reduced based on experience.

[0091] In one specific embodiment, the method described in this application achieves cross-modal cross-validation across three continuous data streams. See [link to relevant documentation]. Figure 2 A panoramic view of continuous ODE evolution under four input conditions, including three data streams: financial market monitoring data stream, sensor array anomaly detection data stream, and scientific literature tracking data stream.

[0092] The common implementation parameters for the three data streams are set as follows: the number of cognitive agents N=52, distributed across 9 cognitive functional layers; the emotional space dimension is 12; the somatic state vector dimension is 6; the ordinary differential equation system of the cognitive separation neurodynamic system uses Euler integrals with a time step of 0.05, and each data stream runs for 2000 steps; the evolution timescale of the somatic state vector is about 100 times slower than that of the neural state; the large model call gating threshold is 0.25, and the gating value is compared with the threshold after being modulated by the energy dimension and arousal dimension of the somatic state vector; the effective spectral radius of the influence matrix is ​​about 0.249, the activation function is the standard sigmoid function (maximum derivative is 0.25), the overall contraction factor of the system is about 0.062, which is much smaller than the upper convergence bound of 1, satisfying the convergence condition; the influence matrix satisfies exact symmetry with zero symmetry error; the proportion of functional non-zero connections is about 11.7%.

[0093] Figure 2 The first to fourth rows represent zero input, financial market monitoring data stream, sensor array anomaly detection data stream, and scientific literature tracking data stream, respectively. The first to fourth columns represent neural state, activity level, physical state, and emotional readout, respectively.

[0094] See Figure 3 The diagram shows the comparison of the α-gated triggering mode across three data streams. From top to bottom, the first to third rows represent the financial market monitoring data stream, the sensor array anomaly detection data stream, and the scientific literature tracking data stream, respectively. From left to right, the first to second columns represent the α trajectory, which consists of four underlying signals of α. The red line on the vertical axis of the α trajectory represents the trigger threshold. All three data streams start from the same initial state, and the cognitive separation neurodynamic system has no prior knowledge about the data stream type. Experimental results show that the financial market monitoring data stream triggered the large model call 239 times, including 125 times of trajectory chaos, 99 times of inter-layer conflict, and 15 times of novel input, with triggers concentrated in the financial market crash event range and the market state switching range. The sensor array anomaly detection data stream triggered 37 times, mainly of trajectory chaos and novel input. The scientific literature tracking data stream triggered 219 times, with inter-layer conflict overwhelmingly dominating during the scientific paradigm conflict period.

[0095] The differences in trigger distributions described above indicate that the large model invocation gating mechanism described in this application does not require pre-adaptation to the data stream type. The differences in trigger types are entirely determined by the natural projection of the statistical characteristics of the data stream itself onto the trajectory of the cognitive separation neurodynamic system. The specific embodiments described above verify the cross-modal generalization capability of the method described in this application in different modal continuous data stream scenarios such as financial monitoring, sensor anomaly detection, and scientific literature tracking.

[0096] Figure 4A schematic diagram of the pulse learning judgment process of a large model cognitive separation method according to another embodiment of this application is shown.

[0097] Local features of the cognitive neurodynamic system trajectory are acquired, and trajectory chaos signals, dynamic dilemma signals, inter-layer conflict signals, and input novelty signals are calculated. The maximum value among these signals is selected as the gate value for calling the large model. It is determined whether the gate value exceeds the corresponding preset threshold. If it does not exceed the threshold, the process ends and the large model is not called. If it exceeds the threshold, it means that the cognitive neurodynamic system has encountered difficulties and the large model needs to help. The pulse type is determined according to the gate value of the large model call. The corresponding vector returned by the large model call is injected into the cognitive separation neurodynamic system. The energy function difference of the cognitive separation neurodynamic system before and after pulse injection is calculated. If the energy does not decrease, it means that the pulse injection has no effect and pulse learning is not performed. If the energy decreases, it means that the pulse injection needs to be learned based on the pulse learning rules, and the influence matrix parameters and cognitive strategy are updated at the same time.

[0098] Figure 5 A schematic diagram of a large model cognitive separation system according to an embodiment of this application is shown. The device 500 includes a construction module 501, a state update module 502, an intrinsic drive module 503, an evolution module 504, a gating judgment module 505, and a pulse learning module 506.

[0099] According to another aspect of this application, a large-model cognitive separation system is provided, characterized in that it includes: Module 501 is used to construct a cognitive separation neurodynamic system, modeling N neurons as N cognitive agents in a non-cooperative game. The N neurons are distributed in the cognitive function layer according to their cognitive functions. An influence matrix is ​​constructed, and the initial parameters of the influence matrix are encoded according to the relationship between the cognitive functions. In response to continuous data stream input, the parameters of the influence matrix are updated in real time, the equilibrium parameters of the influence matrix are solved, and a cognitive strategy is output. The state update module 502 is used to update the current activation state of the cognitive agent based on the state update rules, correct the equilibrium parameters of the influence matrix, and update the cognitive strategy. The intrinsic drive module 503 is used to monitor the drive variables of the cognitive agent in real time and maintain the activity of the cognitive dissociation neural dynamic system based on the drive variable update rules. Evolution module 504 is used for the continuous evolution of the cognitive separation neurodynamic system in the form of a system of ordinary differential equations; The gating judgment module 505 is used to determine in real time whether the large model call gating value of the cognitive separation neural dynamic system exceeds a preset threshold based on the local features of the trajectory of the cognitive separation neural dynamic system. If so, the pulse type is determined, the large model is called to obtain the response vector corresponding to the pulse type, and the response vector is injected into the cognitive separation neural dynamic system. The pulse learning module 506 is used to calculate the energy function difference between the cognitive separation neurodynamic system before and after pulse injection, determine whether the energy decreases after pulse injection, and if so, update the influence matrix parameters based on the pulse learning rules and update the cognitive strategy.

[0100] It should be noted that the aforementioned large-model cognitive separation system can implement the aforementioned large-model cognitive separation method one by one, which will not be elaborated here.

[0101] Figure 6 A schematic diagram of the structure of an electronic device according to an embodiment of the present application is shown. The electronic device 600 includes a processor 610 and a memory 620 arranged to store computer-executable instructions (computer-readable program code).

[0102] The memory 620 can be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. The memory 620 has a storage space 630 for storing computer-readable program code 631. For example, the storage space 630 for storing computer-readable program code may include various computer-readable program codes 631 respectively used to implement the various steps in the above method. The computer-readable program code 631 can be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. Such computer program products are typically, for example... Figure 7 The aforementioned computer-readable storage medium.

[0103] Figure 7 A schematic diagram of a computer-readable storage medium according to an embodiment of this application is shown. The computer-readable storage medium 700 stores computer-readable program code 631 for performing the method steps according to this application, which can be read by the processor 610 of an electronic device 600. When the computer-readable program code 631 is run by the electronic device 600, it causes the electronic device 600 to perform the various steps of the method described above. Specifically, the computer-readable program code 631 stored in the computer-readable storage medium can perform the methods shown in any of the above embodiments. The computer-readable program code 631 can be compressed in a suitable form.

[0104] A computer-readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. A computer-readable storage medium can be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof. Specifically, a computer-readable storage medium can be a portable computer disk, a hard disk, a USB flash drive, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), staging random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory stick, floppy disk, optical disk, magnetic disk, mechanical encoding device, or any combination thereof.

[0105] It should be noted that the above embodiments are illustrative of this application and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims.

[0106] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to which steps and units are explicitly listed, but may include other steps or units inherent to these processes, methods, products, or apparatuses that are not explicitly listed. The terms "first," "second," etc., are used to distinguish similar phenomena and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate for the embodiments of the invention described herein. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims that enumerate several means, several of these means may be embodied by the same item of hardware.

Claims

1. A method for separating cognition in a large model, characterized in that, include: A cognitive separation neurodynamic system is constructed, which models N neurons as N cognitive agents in a non-cooperative game. The N neurons are distributed in the cognitive function layer according to their cognitive functions. An influence matrix is ​​constructed, and the initial parameters of the influence matrix are encoded according to the relationship between the cognitive functions. In response to continuous data stream input, the parameters of the influence matrix are updated in real time. The equilibrium parameters of the influence matrix are solved, and the cognitive strategy is output. The cognitive agent's current activation state is updated based on the state update rules, the equilibrium parameters of the influence matrix are corrected, and the cognitive strategy is updated. The driving variables of the cognitive agent are monitored in real time, and the activity of the cognitive dissociation neurodynamic system is maintained based on the driving variable update rules. The cognitive separation neurodynamic system continues to evolve in the form of a system of ordinary differential equations; Based on the local features of the trajectory of the cognitive separation neurodynamic system, it is determined in real time whether the large model call gate value of the cognitive separation neurodynamic system exceeds a preset threshold. If so, the pulse type is determined, the large model is called to obtain the response vector corresponding to the pulse type, and injected into the cognitive separation neurodynamic system. Calculate the energy function difference of the cognitive separation neurodynamic system before and after pulse injection, determine whether the energy decreases after pulse injection, and if so, update the influence matrix parameters based on the pulse learning rule and update the cognitive strategy.

2. The method according to claim 1, characterized in that, The influence matrix is ​​an N×N real symmetric matrix, constituting an exact potential game; The cognitive functional layers include, but are not limited to: the perception layer, the cognition layer, the evaluation layer, the action layer, the constraint layer, the attention layer, the basic level, the social layer, and the metacognitive layer; The utility function of the cognitive agent i is: ， in, Let represent the current activation strength of cognitive agent i, with a value ranging from [0,1]. For the strategy combination of all cognitive agents other than the current cognitive agent i, The element in the i-th row and j-th column of the influence matrix represents the influence coefficient of cognitive agent j on cognitive agent i. The external signal input received by the cognitive agent is provided by a data stream from outside the system. The activation function of the cognitive agent i is the sigmoid function. Let the adjustment cost function of cognitive agent i be defined as the activation function. The integral of the inverse function; The process of solving the equilibrium parameters of the influence matrix includes: solving the spectral radius of the influence matrix in real time, determining whether the spectral radius of the influence matrix is ​​less than the convergence threshold, and if so, determining that the cognitive separation neurodynamic system is in a Nash equilibrium state of the game, solving the local minimum value of the energy function, and obtaining the equilibrium parameters of the influence matrix. The cognitive strategy is a combination of strategies in which all cognitive agents have no unilateral deviation motivation, including the current state vector.

3. The method according to claim 1, characterized in that, The update of the cognitive agent's current activation state based on the state update rule includes: The state update rule is as follows: ， in, It is the inertial parameter of cognitive agent i, with a value range of (0,1). This is the basic gain parameter of agent i, which controls the sensitivity to the input signal. It is the temperature parameter of the l-th layer at time step t, which decreases according to the exponential annealing rule, and is the initial temperature multiplied by the decay coefficient raised to the power of t. The lower limit is not lower than the minimum temperature value. The pre-activation input aggregates four signal sources, including: ， in, Let be the effective fan-in degree of cognitive agent i, and let be the number of non-zero elements in the i-th row of the influence matrix. The fan-in normalization coefficient ensures that cognitive agents with different connection densities have comparable signal amplitudes. This is the inherent bias of cognitive agent i, reflecting the basic activity level of this cognitive process. The value of the k-th emotional component at time step t is calculated by multiplying the emotional readout matrix R by the current state vector s(t). It is the element in the i-th row and k-th column of the emotion feedback matrix, defined as the feedback intensity coefficient multiplied by the transpose of the emotion readout matrix R, where E is the dimension of the emotion space. It is the user-personalized bias of cognitive agent i, which encodes the cumulative interaction tendencies of a specific user.

4. The method according to claim 2, characterized in that, The driving variable is the current search impulse intensity of the cognitive agent, and the update rules for the driving variable include: ， in, Let i be the driving force value of the cognitive agent i at time step t. For passive decay rate, constant temperature regulation is provided, with a value range of (0,1). The stagnation accumulation rate refers to the rate at which the drive increases when the cognitive agent's state remains unchanged for a long period of time. To meet the decay rate, the rate at which the driving force decreases upon the arrival of a new external input, Let be the absolute value of the rate of change of the state of cognitive agent i, so that the state change is normalized to the range [0,1]. The absolute value of the change in the external input signal; The driving force variable represents the current search impulse intensity of cognitive agent i, with a lower limit of 0 and an upper limit of the driving force of cognitive agent i.

5. The method according to claim 4, characterized in that, The cognitive dissociation neurodynamic system evolves continuously in the form of a system of ordinary differential equations, including: ， Where B is the somatic state vector, including but not limited to energy, arousal, comfort, information hunger, social needs, and circadian rhythm phase. The effective decay rate of cognitive agent i is modulated by the energy dimension of the body state vector B. The effective gain of cognitive agent i is modulated by the arousal dimension of the somatic state vector B, η i The driving force perturbation direction variable has a value range of [0,1], which makes the perturbation direction tend to pull the cognitive agent's state back to a neutral value, preventing state saturation. For energy pumping terms, spontaneous activity driven by the energy dimension when the cognitive agent's state approaches a neutral value; The physical state evolves on a timescale that is 100 times slower than the neural state.

6. The method according to claim 5, characterized in that, The large model calls the gate value as the maximum value of the trigger signal, which includes, but is not limited to, trajectory chaos signal, dynamic dilemma signal, inter-layer conflict signal, and input novelty signal; The preset threshold is an adjustable hyperparameter; The pulse types include, but are not limited to, chaotic, dilemma-based, conflict-based, and novel pulses. The corresponding vector is an N-dimensional vector or an N×N matrix; The injection into the cognitive separation neurodynamic system includes, but is not limited to: multiplying an N-dimensional vector by a scaling factor and merging it into the current state vector, and multiplying an N×N matrix by a scaling factor matrix and merging it into the influence matrix.

7. The method according to any one of claims 1 to 6, characterized in that, The impulse learning rules include: ， in, To influence the update amount of the element in the i-th row and j-th column of the matrix, The learning rate controls the magnitude of each update. The difference in energy function before and after pulse injection. The state value of cognitive agent i after pulse injection. The state value of cognitive agent i before pulse injection. The state value of cognitive agent i before pulse injection.

8. A large-scale cognitive separation system, characterized in that, include: The module is used to construct a cognitive separation neurodynamic system, which models N neurons as N cognitive agents in a non-cooperative game. The N neurons are distributed in the cognitive function layer according to their cognitive functions. An influence matrix is ​​constructed, and the initial parameters of the influence matrix are encoded according to the relationship between the cognitive functions. In response to continuous data stream input, the parameters of the influence matrix are updated in real time, the equilibrium parameters of the influence matrix are solved, and the cognitive strategy is output. The state update module is used to update the current activation state of the cognitive agent based on the state update rules, correct the equilibrium parameters of the influence matrix, and update the cognitive strategy. The intrinsic drive module is used to monitor the drive variables of the cognitive agent in real time and maintain the activity of the cognitive dissociation neural dynamic system based on the drive variable update rules. An evolution module is used for the continuous evolution of the cognitive separation neurodynamic system in the form of a system of ordinary differential equations; The gating judgment module is used to determine in real time whether the gating value of the large model call of the cognitive separation neurodynamic system exceeds a preset threshold based on the local features of the trajectory of the cognitive separation neurodynamic system. If so, the pulse type is determined, the large model is called to obtain the response vector corresponding to the pulse type, and the response vector is injected into the cognitive separation neurodynamic system. The pulse learning module is used to calculate the energy function difference of the cognitive separation neurodynamic system before and after pulse injection, determine whether the energy decreases after pulse injection, and if so, update the influence matrix parameters based on the pulse learning rules and update the cognitive strategy.

9. An electronic device, characterized in that, The device includes a processor and a memory, the processor being coupled to the memory, the processor being configured to execute a computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Includes a computer program or instructions that, when run on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

Images