Neural plasticity model optimization method based on neural connection group constraint

Abstract

The application discloses a plasticity neural dynamics model optimization method based on neural connection group constraints, and aims to solve the problems of insufficient reality of neural network modeling and limited continuous learning ability. The method comprises the following steps: acquiring neural connection group data to construct network topology; configuring Hodgkin-Huxley and other neural dynamics models; activating the network through external stimulation and calculating local plasticity traces; generating a global modulation signal; optimizing synaptic weights by using a three-factor learning mechanism; and iteratively updating network parameters. The application retains structural features by using real neural connection group data, realizes online parameter update by fusing local plasticity and global modulation, reduces computational overhead by using a three-factor learning rule, and takes into account structural reality, biological rationality and task learning ability, and can be applied to the fields of neural system digital twinning, neuromorphic computing and brain-like intelligence research.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to an optimization method for plastic neurodynamic models based on neural connectome constraints. Background Technology

[0002] Neural connectomics is an important research direction in neuroscience, aiming to systematically analyze the structural connections between neurons within the nervous system. With the rapid development of electron microscopy reconstruction techniques, high-throughput imaging technologies, and neural structure measurement techniques, researchers have been able to obtain relatively complete connectome data, including from Caenorhabditis elegans, Drosophila melanogaster, and some mammalian brain regions. This precise neural structural information provides an important data foundation for connectome-based neural system modeling research, driving the in-depth development of computational modeling in neuroscience.

[0003] Among these, neurodynamic modeling based on connectomes is one of the core research areas in this field. This method typically utilizes the neuronal topology provided by connectomes to construct neural network models, and combines these with biophysical neuron models such as the Hodgkin-Huxley model, the Leaky Integrate-and-Fire model, the Morris-Lecar model, or multi-compartment neuron models to numerically simulate nervous system activity, thereby studying the information encoding mechanisms and behavior generation mechanisms of the nervous system. This type of method can effectively preserve the structural and dynamic characteristics of the nervous system and is widely used in neuroscience research.

[0004] However, existing modeling methods based on neural connectomes have significant shortcomings. First, current connectome models typically set the connection parameters between neurons to fixed values ​​or obtain connection weights and dynamic parameters through offline fitting. After parameter fitting, the model's structure and parameters remain unchanged, only able to reproduce neural activity patterns under specific experimental conditions. They lack the ability to continuously adjust connection parameters according to environmental changes, making it difficult to reflect the continuous learning and environmental adaptation capabilities of real biological neural systems. Second, existing research often uses optimization methods such as gradient descent, error backpropagation, and time backpropagation to train model parameters. While these methods have achieved good results in artificial neural networks, they suffer from high computational costs and storage requirements when applied to large-scale neural dynamics models. Furthermore, the backpropagation algorithm relies on global gradient information, which differs significantly from the local synaptic update mechanism in real biological neural systems, resulting in weak biological plausibility. Furthermore, while researchers have proposed local plasticity mechanisms such as the Hebbian learning rule and the STDP learning rule, as well as learning methods that can introduce reward or error signals for regulation, such as the three-factor learning rule, these methods are mostly applied to artificially constructed spiking neural networks or randomly generated network structures. They do not fully integrate the structural prior information provided by real neural connectomes, making it difficult to simultaneously consider structural realism, biological plausibility, and continuous learning ability. In summary, current technologies lack an optimization framework that can unify and integrate neural dynamics models, local plasticity mechanisms, and global task feedback under the constraints of real neural connectomes, making it difficult to achieve online adaptive optimization of neural network connection parameters and dynamic parameters. Summary of the Invention

[0005] The purpose of this invention is to provide an optimization method for plastic neurodynamic models based on neural connectome constraints, which solves the problems mentioned in the background art.

[0006] This invention is implemented as follows: a method for optimizing a plastic neurodynamic model based on connectome constraints, comprising the following specific steps: Step 1: Acquire connectome data of the target nervous system and construct a neural network topology based on the structural connections between neurons. The connectome data is obtained from electron microscopy reconstruction, connectome databases, or other neural structure measurement techniques. A network connection matrix is ​​established based on chemical synaptic connections, electrical synaptic connections, and connection strength information between neurons, so that the model retains the structural features of the real nervous system. Step 2: Configure neurodynamic models for neurons in the network. The neurodynamic models adopt Hodgkin-Huxley models, Leaky Integrate-and-Fire models, Morris-Lecar models, multi-compartment neuron models, or other models that can describe the dynamic characteristics of neural activity. These models describe the dynamic changes of neuronal membrane potential and ion channel states over time. A biologically constrained neural network simulation system is constructed by combining the neuronal dynamic model with the connectome topology. Step 3: Activate the neural network through external sensory stimuli, environmental input signals, or task-driven signals. This causes neurons to generate dynamic responses, obtaining the neuronal activity state and synaptic connections during network operation. The neural network calculates local plasticity traces based on the activity correlation between neurons during operation, considering the synaptic activity state and network output results. This trace is used to characterize the historical activity information and potential plasticity change trends of synaptic connections within a certain time window. Step 4 generates a global modulation signal based on task objectives, environmental feedback, behavioral rewards, prediction errors, or output error information. This global modulation signal reflects the degree of difference between the current network output result and the target state, and serves as a regulatory factor in the network learning process to participate in synaptic parameter updates. It introduces global task information into the synaptic plasticity process, enabling the network to adjust its learning direction according to the task execution results. Step 5 uses... The three-factor learning mechanism optimizes neural network parameters: it uses local plasticity traces as the basis for synaptic updates, modulates them with global modulation signals, and calculates the weight changes of each synaptic connection. This allows synaptic parameter updates to be influenced by both local neural activity and global task objectives, combining biological rationality and task-oriented capabilities. Step 6 applies the updated synaptic weights back to the neural network model and continues neurodynamic simulation. The above process is executed iteratively, allowing network connection parameters and related dynamic parameters to continuously and adaptively adjust according to environmental input and task requirements, ultimately obtaining an optimized network state that meets the target task requirements.

[0007] Preferably, the neural connectivity data in step 1 includes chemical synaptic connectivity information and electrical synaptic connectivity information. The chemical synaptic connectivity includes attributes such as synapse type, synapse strength, and neurotransmitter type, while the electrical synaptic connectivity includes attributes such as connection resistance and conductivity. The network connectivity matrix established based on the connectivity information has a dimension of N multiplied by N, where N is the number of neurons in the neural network.

[0008] Preferably, when the neurodynamic model in step 2 adopts the Hodgkin-Huxley model, it includes parameters such as sodium ion channel conductance, potassium ion channel conductance, leakage channel conductance, sodium ion reversal potential, potassium ion reversal potential, and leakage reversal potential. The model is numerically solved using the fourth-order Runge-Kutta method or the adaptive step-size integration method, and the time step is set to a preset time accuracy range.

[0009] Preferably, when the neurodynamic model in step 2 adopts a multi-compartment neuron model, the neuron is modeled as a combination of soma compartment, dendritic compartment and axon compartment, and the compartments are connected by intercompartmental resistance. The membrane capacitance parameter is set to a preset membrane capacitance value, and the axial resistance parameter is set to a preset axial resistance value.

[0010] Preferably, the calculation of local plasticity traces in step 3 adopts the time window method. The width of the time window is set within a preset time window range. Specifically, the calculation method is to integrate the correlation between the activity of presynaptic neurons and the activity of postsynaptic neurons within the time window, and the plasticity trace value is obtained after nonlinear transformation of the integration result.

[0011] Preferably, the calculation of local plasticity traces in step 3 can also be performed using a difference method based on the spike timing of presynaptic and postsynaptic neurons. When the spike of the presynaptic neuron is earlier than the spike of the postsynaptic neuron, enhanced plasticity traces are generated; when the spike of the presynaptic neuron is later than the spike of the postsynaptic neuron, weakened plasticity traces are generated. The time difference threshold is set within a preset time difference threshold range.

[0012] Preferably, the generation of the global modulation signal in step 4 is carried out in different ways depending on the task type: for reinforcement learning tasks, the modulation signal is calculated by the reward function, and the reward value is processed by the time-series difference algorithm to generate the modulation signal; for supervised learning tasks, the modulation signal is obtained by the error between the network output and the target label after error scaling and normalization; for unsupervised learning tasks, the modulation signal is obtained by the prediction error or reconstruction error after adaptive gain adjustment.

[0013] Preferably, the mathematical expression of the three-factor learning rule in step 5 is as follows: the change in synaptic weight is equal to the global modulation signal multiplied by the local plasticity trace and then multiplied by the learning rate, wherein the learning rate is set to a preset learning rate range, and the change in weight is applied to the synaptic weight update after being constrained by the upper and lower limits. The upper limit of the weight is set to a preset multiple range of the original weight, and the lower limit of the weight is set to a preset minimum weight value.

[0014] Preferably, in addition to synaptic connection weights, the parameters in the neurodynamic model are also optimized, including neuronal membrane conductance parameters, ion channel parameters, time constant parameters, and neuronal excitability parameters. A collaborative optimization strategy is adopted for connection parameters and dynamic parameters. In each iteration, synaptic weights are updated first and then dynamic parameters are updated, or an alternating update method is adopted.

[0015] Preferably, the termination conditions of the iteration process in step 6 include: the error between the network output and the target output is lower than a preset error threshold, the error improvement amount of multiple consecutive iterations is lower than a preset improvement threshold, the number of iterations reaches a preset iteration limit or a preset training time is reached.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes real neural connectome data to construct network topology structures, enabling the model to retain the structural features of the nervous system, improving the realism and interpretability of nervous system modeling, and providing a more accurate computational model for neuroscience research.

[0017] This invention introduces local plasticity and global modulation mechanisms to achieve online updates of neural network connection parameters, enabling the model to have continuous learning and environmental adaptability, which is closer to the learning process of real biological neural systems and can dynamically adjust the network connection strength according to environmental changes.

[0018] This invention employs a parameter optimization method based on a three-factor learning rule, which does not rely on a complete error backpropagation process. This reduces the computational resource consumption and storage overhead during the training of large-scale neurodynamic models, making it possible to perform long-term simulations of large-scale networks containing neurons of a specific size.

[0019] This invention achieves a unified integration of neural connectome structural constraints, neurodynamic modeling, and plasticity learning mechanisms, improving the model's biological plausibility and task learning ability, while taking into account structural realism, biological plausibility, learning ability, and computational efficiency.

[0020] This invention can be applied to the modeling and optimization of nervous systems at different scales, such as Caenorhabditis elegans, Drosophila melanogaster, and mammalian brain regions. It can be widely used in fields such as digital twins of nervous systems, biological neural network simulation, neuromorphic computing, and intelligent control, providing new tools and methods for brain-like intelligence research and the exploration of brain disease mechanisms. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical architecture of the plastic neurodynamic model optimization method based on neural connectome constraints proposed in this invention. Detailed Implementation

[0022] Example 1 To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0023] In one specific embodiment of the present invention, a method for optimizing a plastic neurodynamic model based on neural connectome constraints is disclosed. The method includes six core steps, from step 1 to step 6. The technical details of each step are described below.

[0024] Step 1: Obtain neural connectome data of the target nervous system and construct the neural network topology. In step 1, the first step is to acquire connectome data of the target nervous system. This connectome data can be derived from various techniques, including electron microscopy reconstruction, high-throughput neuroimaging, connectome databases, or other neural structure measurement techniques. Taking electron microscopy reconstruction as an example, this technique involves ultrathin sections and continuous imaging of neural tissue, followed by image registration and segmentation algorithms to identify neuronal cell bodies, dendrites, axons, and synaptic connections, ultimately reconstructing a complete neuronal morphology and synaptic connection matrix. Connectome databases contain pre-organized lists of neurons, connections between neurons, and connection strength information; researchers can directly export the necessary data from the database for subsequent modeling.

[0025] After acquiring the neural connectome data, it is necessary to extract information on the structural connections between neurons. These structural connections include two types: chemical synaptic connections and electrical synaptic connections. Chemical synaptic connections are the most common mode of information transmission between neurons, and their characteristic parameters include synapse type (e.g., excitatory or inhibitory synapses), synaptic strength (usually characterized by the amount of neurotransmitters released from the presynaptic membrane and the efficiency of receptor response at the postsynaptic membrane), and neurotransmitter type (e.g., glutamate, γ-aminobutyric acid, acetylcholine, etc.). Electrical synaptic connections, also known as gap connections, allow ionic currents to flow directly between adjacent neurons, and their characteristic parameters include connection resistance and conductivity, reflecting the efficiency of current transmission between two neurons.

[0026] Based on the extracted connection information, a network connection matrix is ​​constructed. This matrix is ​​a two-dimensional array of dimension N multiplied by N, where N represents the number of neurons in the neural network. Each element W(i,j) in the matrix represents the synaptic connection weight from neuron i to neuron j. When W(i,j) is greater than zero, it indicates the existence of an excitatory chemical synaptic connection from neuron i to neuron j; when W(i,j) is less than zero, it indicates the existence of an inhibitory chemical synaptic connection from neuron i to neuron j; and when W(i,j) is equal to zero, it indicates the absence of a direct chemical synaptic connection. For electrical synaptic connections, a separate electrical synaptic connection matrix can be used, where each element represents the conductance between corresponding neuron pairs. In this way, the network connection matrix comprehensively describes the topological structure of the neural network, enabling the subsequently constructed model to retain the structural characteristics of the real nervous system.

[0027] When constructing the network topology, each neuron corresponds to a node in the network, and the connections between neurons correspond to edges in the network. For chemical synaptic connections with a clear direction, directed edges are used to represent the direction of information transmission; for bidirectional electrical synaptic connections, undirected edges are used to represent the connection. The weight of the edge is the corresponding synaptic connection strength value. In this way, the structural connection information of the biological nervous system is transformed into a computer-processable graph theory representation, providing a basic data structure for subsequent dynamic simulations and parameter optimization.

[0028] Step 2: Configure neurodynamic models for neurons in the network In step 2, a corresponding neurodynamic model needs to be configured for each neuron in the network topology constructed in step 1. The neurodynamic model describes the dynamic changes in neuronal membrane potential and ion channel states over time and is the basic computational unit for simulating neural activity. This invention supports various types of neurodynamic models, including but not limited to the Hodgkin-Huxley model, the Leaky Integrate-and-Fire model, the Morris-Lecar model, and multi-compartment neuron models. Several commonly used neurodynamic models are described in detail below.

[0029] When the Hodgkin-Huxley model is used, it is a biophysical model containing multiple state variables, with its core equation describing the relationship between changes in cell membrane potential and ion channel activity. Specifically, the Hodgkin-Huxley model includes parameters such as sodium ion channel conductance, potassium ion channel conductance, leakage channel conductance, sodium ion reversal potential, potassium ion reversal potential, and leakage reversal potential. Sodium ion channel conductance is typically set to a high value (e.g., 120 mSiemens per square centimeter) to generate a rapid rise in membrane potential depolarization; potassium ion channel conductance is set to a medium value (e.g., 36 mSiemens per square centimeter) to generate subsequent membrane potential repolarization; and leakage channel conductance is set to a low value (e.g., 0.3 mSiemens per square centimeter) to maintain the resting membrane potential. Sodium ion reversal potential is typically set to a positive value (e.g., 50 mV), potassium ion reversal potential is typically set to a negative value (e.g., -77 mV), and leakage reversal potential is typically set to a value close to the resting membrane potential (e.g., -54.4 mV). In addition, the model includes state variables such as sodium ion channel activation variable *m*, sodium ion channel inactivation variable *h*, and potassium ion channel activation variable *n*. The kinetics of these variables changing over time are described by corresponding differential equations.

[0030] Numerical solutions to the Hodgkin-Huxley model typically employ either the fourth-order Runge-Kutta method or an adaptive step-size integration method. The fourth-order Runge-Kutta method improves accuracy by repeatedly calculating the rate of change of the state variable within each time step, with a typical time step ranging from 0.01 to 0.05 milliseconds. The adaptive step-size integration method automatically adjusts the time step based on the rate of change of the state variable, using a larger step size for smooth changes to improve computational efficiency and a smaller step size for rapid changes to ensure accuracy; the time accuracy is typically set between 0.001 and 0.1 milliseconds.

[0031] When using a multi-compartment neuron model, the neuron is modeled as a combination of multiple electrical compartments. Each compartment can represent a different anatomical location of the neuron, such as the soma compartment, dendritic compartment, and axonal compartment. The compartments are connected by intercompartmental resistance, simulating axial current flow between the cell body and processes. Membrane capacitance parameters are set to preset values, typically around 1 microfarad per square centimeter, to store charge and generate membrane potential changes; axial resistance parameters are set to preset values, typically in the range of 100 to 1000 ohms per square centimeter, to control the current transfer efficiency between compartments. Each compartment can be configured with independent ion channel types and parameters to simulate the electrophysiological characteristics of different brain regions and neuron types. Numerical solutions for multi-compartment models typically employ implicit Euler methods or the Crank-Nicolson method to address numerical stability issues arising from the fast timescale dynamics between compartments.

[0032] By configuring a neurodynamic model for each neuron and combining it with the network topology constructed in step 1, a biologically constrained neural network simulation system is built. This simulation system can simulate the dynamic evolution of neural activity while preserving the structural characteristics of the real nervous system, providing a computing platform for subsequent online learning.

[0033] Step 3: Activate the neural network through external stimuli and calculate local plasticity traces In step 3, the neural network constructed in step 2 needs to be activated by external sensory stimuli, environmental input signals, or task-driven signals to induce dynamic responses in the neurons. During model operation, the system records the neuronal activity state, synaptic activity state, and network output results in real time, and calculates local plasticity traces based on this activity information.

[0034] External stimuli can originate from a variety of sources. In sensory information processing tasks, external stimuli correspond to physical stimuli received by sensory organs, such as visual images, auditory sounds, and tactile pressure. These stimuli are encoded and input into the sensory input neurons of the network. In environmental interaction tasks, external stimuli correspond to state information obtained by the agent during interaction with the environment, such as the position of the robotic arm's end effector, the position of the target object, and the distribution of obstacles. In a pre-defined task-driven mode, external stimuli are pre-designed sequences of input patterns used to drive the network to produce specific output responses. The encoding method of external stimuli can be designed according to specific task requirements; common encoding methods include frequency encoding, time encoding, and group encoding.

[0035] Neural networks evolve dynamically under the influence of external stimuli. The membrane potential of each neuron updates over time according to its configured neurodynamic model, and synaptic connections between neurons transmit information from one neuron to another. When a neuron's membrane potential reaches a threshold, it generates an action potential (spike) and transmits the signal to the next level neuron through synaptic connections. In this way, the network generates neural activity patterns and network outputs corresponding to the input signals.

[0036] During the operation of the neural network, the system records the activity state of each neuron in real time, including membrane potential values, ion channel state variables, and action potential firing times. Simultaneously, the system records the activity state of each synaptic connection, including the firing time of the presynaptic neuron, the reception time of the postsynaptic neuron, and the intensity of synaptic currents. This activity state information forms the basis for calculating local plasticity traces.

[0037] Local plasticity traces are used to characterize the historical activity and potential plasticity trends of synaptic connections within a certain time window. This invention provides two methods for calculating local plasticity traces.

[0038] The first method is the time window approach. The time window width is set within a preset range, typically from 10 milliseconds to 500 milliseconds. Specifically, the correlation between presynaptic and postsynaptic neuronal activity within the time window is integrated. First, the activity sequence of the presynaptic neuron within the time window is recorded and converted into a corresponding binary time series or continuous activity value; similarly, the activity sequence of the postsynaptic neuron within the time window is recorded. Then, the correlation coefficient or covariance between the two binary time series or continuous activity values ​​is calculated as a preliminary measure of plasticity correlation. The integration result is then subjected to a nonlinear transformation to obtain the plasticity trace value. The nonlinear transformation function can be an exponential function, a sigmoid function, or other monotonically increasing functions. When the presynaptic and postsynaptic neuronal activities are strongly correlated, the corresponding synaptic connection will accumulate a high plasticity trace value; when the correlation is low, the plasticity trace value is low or close to zero.

[0039] The second method is based on the difference in spike timing between presynaptic and postsynaptic neurons, also known as the time-dependent plasticity method. This method considers the impact of the relative temporal order of action potential firing between presynaptic and postsynaptic neurons on plasticity. When the presynaptic neuron spikes earlier than the postsynaptic neuron spikes (the time difference is positive and within a preset range), an enhanced plasticity trace is generated, indicating that the synaptic connection may contribute to neuronal activation; when the presynaptic neuron spikes later than the postsynaptic neuron spikes (the time difference is negative and within a preset range), a diminished plasticity trace is generated, indicating that the synaptic connection may be irrelevant or harmful. The time difference threshold is set within a preset temporal difference threshold range, typically ±20 ms to ±100 ms. The smaller the time difference, the stronger the enhancement effect; the larger the time difference, the stronger the diminishing effect. This method allows adjustment of synaptic plasticity based on the precise temporal order of neuronal activity, realizing the core idea of ​​the Hebbian learning rule that "neurons firing together connect together."

[0040] Step 4: Generate a global modulation signal In step 4, a global modulation signal needs to be generated based on the task objective, environmental feedback, behavioral rewards, prediction error, or output error information. This global modulation signal reflects the degree of difference between the current network output and the target state, and participates in synaptic parameter updates as a regulatory factor in the network learning process. By introducing global task information into the synaptic plasticity process, the network can adjust its learning direction based on the task execution results, rather than solely relying on the correlation of local neural activity.

[0041] The generation of the global modulation signal is carried out in different ways depending on the task type.

[0042] For reinforcement learning tasks, the modulated signal is calculated from the reward function. After an agent performs a specific action during interaction with the environment, the environment returns a corresponding reward value. The magnitude of the reward value reflects the contribution of the current action to achieving the task objective. The reward value is processed by a temporal difference algorithm to generate the modulated signal. Specifically, the temporal difference algorithm calculates the difference between the reward estimate at the current time step and the reward estimate at the next time step. This difference reflects the change in the value of the current state-action pair. The temporal difference error can be used as a global modulated signal. A positive error indicates that the current behavior is better than expected, generating positive modulation; a negative error indicates that the current behavior is worse than expected, generating negative modulation. The parameters of the temporal difference algorithm, such as the discount factor and the baseline function, can be adjusted according to the specific task.

[0043] For supervised learning tasks, the modulated signal is obtained by scaling and normalizing the error between the network output and the target label. The network output is the actual response of the neural network to a given input, and the target label is the expected correct output. The difference between the two is adjusted by an error scaling factor to map to a suitable numerical range, and then normalized to obtain a standardized modulated signal. The error scaling factor can be adaptively adjusted according to the difficulty of the task and the learning state of the network. The normalization process can employ batch normalization, layer normalization, or a simple numerical range scaling method.

[0044] For unsupervised learning tasks, the modulated signal is obtained by adaptively adjusting the prediction error or reconstruction error. In autoencoder-type tasks, the network attempts to reconstruct the input signal, and the reconstruction error reflects the network's ability to encode and decode the input pattern. In predictive learning tasks, the network attempts to predict future input signals, and the prediction error reflects the network's ability to predict future states. These error values ​​are processed by the adaptive gain adjustment module to generate a modulated signal proportional to the error magnitude. The adaptive gain adjustment module can dynamically adjust the gain coefficient based on historical error statistics, ensuring that the amplitude of the modulated signal remains within a suitable range at different learning stages.

[0045] The global modulation signal is typically set between -1 and +1 or 0 and 1. Positive values ​​indicate modulation that enhances synapses, negative values ​​indicate modulation that weakens synapses, and zero values ​​indicate no modulation. The modulation signal can act on all synaptic connections in the network, or it can act on only specific types of synapses or neurons in specific brain regions.

[0046] Step 5: Optimize the neural network parameters using a three-factor learning mechanism. In step 5, a three-factor learning mechanism is used to optimize the neural network parameters. This mechanism uses local plasticity traces as the basis for synaptic updates and global modulation signals as the learning regulatory factors; both jointly determine the direction and magnitude of synaptic weight changes. In this way, synaptic parameter updates are simultaneously influenced by both local neural activity and global task objectives, making the learning process both biologically plausible and task-oriented.

[0047] The mathematical expression of the three-factor learning rule can be described as follows: the change in synaptic weights equals the global modulation signal multiplied by the local plasticity trace, and then multiplied by the learning rate. The learning rate is set to a preset range, typically between 0.001 and 0.1, with the specific value adjusted according to task difficulty and network size. The learning rate determines the step size of each parameter update; an excessively large learning rate may lead to instability in the learning process, while an excessively small learning rate may result in slow convergence.

[0048] After calculating the change in synaptic weights, upper and lower bound constraints need to be applied to them before updating the synaptic weights. The upper bound is set to a preset multiple range of the original weights, typically 1.5 to 3 times the original weights, to prevent synaptic weights from growing indefinitely. The lower bound is set to a preset minimum weight value, which can be zero or a small positive value, to prevent synaptic weights from becoming negative (for excitatory synapses) or decreasing indefinitely (for inhibitory synapses). By setting upper and lower bound constraints, the numerical stability of the network can be maintained, and the physiological range of synaptic weights in biological neural systems can be simulated.

[0049] In some implementations, in addition to synaptic connection weights, parameters in the neurodynamic model are also optimized. These dynamic parameters include neuronal membrane conductance parameters, ion channel parameters, time constant parameters, and neuronal excitability parameters. By co-optimizing the connection and dynamic parameters, the model can more accurately adapt to changes in the external environment and learn richer dynamic characteristics. A co-optimization strategy is employed for the connection and dynamic parameters; in each iteration, synaptic weights can be updated first, followed by the dynamic parameters, or an alternating update method can be used. The design of the co-optimization strategy needs to consider the different time scales and sensitivities of the two types of parameters to avoid mutual interference during parameter updates.

[0050] Step 6: Iteratively optimize and obtain the optimized network state In step 6, after a parameter update, the updated synaptic weights are reapplied to the neural network model, and the neurodynamic simulation continues. This process is repeated iteratively, allowing the network connection parameters and related dynamic parameters to continuously and adaptively adjust according to environmental inputs and task requirements, ultimately obtaining an optimized network state that meets the target task requirements.

[0051] The termination conditions for the iterative process include the following: the error between the network output and the target output is lower than a preset error threshold, indicating that the network has learned a mapping relationship that meets the task requirements; the error improvement after multiple consecutive iterations is lower than a preset improvement threshold, indicating that the network has entered a convergent state and further iterations will yield little benefit; the number of iterations reaches a preset iteration limit, indicating that the specified computational resource limit has been reached; and the preset training time is reached, indicating that the specified time limit has been reached. Once any termination condition is met, the iterative process ends, and the current optimized network state is output.

[0052] Through continuous learning, neural networks can develop connection structures and dynamic characteristics adapted to task objectives. The synaptic connection weights within the network are adjusted according to task requirements, strengthening task-related connections and weakening connections that are irrelevant or even interfere with the task. Simultaneously, if dynamic parameter optimization is enabled, the electrophysiological characteristics of neurons also change accordingly, enabling the network to process specific types of information more efficiently. Through these methods, online learning and environmental adaptation of the nervous system are achieved.

[0053] The above method will be further explained below with a specific application example.

[0054] In this specific application example, the objective is to enable the neural network to learn and perform a pattern classification task. The neural connectome data uses the complete neural connectome of *Caenorhabditis elegans*, containing 302 neurons and approximately 7000 synaptic connections. First, the neural connectome data is imported into the system, and the network topology is constructed based on chemical and electrical synaptic connections. Then, a Hodgkin-Huxley dynamics model is configured for each neuron, with parameters using default values ​​reported in the literature. The external input patterns are a randomly generated set of binary vectors representing different sensory input patterns. Each input pattern corresponds to a target category, resulting in a total of 5 categories.

[0055] During model execution, for each input pattern, the system records the membrane potential trajectory and action potential firing time of neurons, and calculates the local plasticity traces of all synaptic connections. The local plasticity traces are calculated using a time window method with a window width of 100 milliseconds. Task performance is evaluated using a simple linear classifier that takes the collective activity of neurons as input and outputs a class prediction. The classification error is calculated based on the difference between the classification result and the true class; this error is then scaled and normalized and used as a global modulation signal.

[0056] In the three-factor learning phase, for each synaptic connection, the weight change is calculated based on its local plasticity trace and global modulation signal. The learning rate is set to 0.01, and the synaptic weights are updated after the weight change is constrained by upper and lower limits. The iterative process continues until the classification error improvement is less than 0.001 after 100 consecutive iterations, at which point the network is considered to have converged.

[0057] Ultimately, the trained network was able to correctly classify the five input patterns with an accuracy rate exceeding 85%. Compared to before training, some synaptic connection weights in the network underwent significant changes. These changes were primarily concentrated in the interneuron population directly connected to sensory input neurons and motor output neurons, indicating that the network learned task-specific connection patterns. This example validates the effectiveness of the method presented in this invention.

[0058] Example 2 To further illustrate the technical solution of the present invention, another specific embodiment is provided below.

[0059] In this embodiment, a neurodynamic model for visual motion detection is constructed using the optic nerve lobe of Drosophila as the research object. The neural connectome data comes from published connectome reconstruction results of the Drosophila optic nerve lobe, containing approximately 15,000 neurons and 80,000 synaptic connections.

[0060] In step 1, spatial location information, chemical synaptic connections, and electrical synaptic connections of neurons are extracted from the connectivity dataset to establish a network connectivity matrix. For specific neuron types in the optic lobe, corresponding neuron models are assigned based on their morphological characteristics.

[0061] In step 2, the Leaky Integrate-and-Fire model is used to model the major neuron types in the optic lobe. The Leaky Integrate-and-Fire model is a simplified neuron model where the membrane potential over time is determined by both leakage current and synaptic input current. The model's parameters include membrane capacitance, membrane resistance, leakage reversal potential, threshold potential, and refractory period parameters. Different neuron types correspond to different parameter values ​​to simulate the electrophysiological characteristics of different neurons.

[0062] In step 3, the input signal is a simulated visual-motor stimulus pattern, including stimuli with different motion directions and speeds. The visual stimuli are encoded and input to the first layer of neurons in the network, namely the photoreceptor neurons of the optic nerve lobe. The system records the membrane potential and firing state of the neurons at each time step and calculates local plasticity traces. This embodiment uses a spike timing-based difference method to calculate plasticity traces, with a time difference threshold set at ±50 milliseconds.

[0063] In step 4, the global modulation signal is generated by the task performance evaluation module. Task performance is defined as the accuracy of the network output neurons in recognizing the direction of motion. The recognition accuracy is processed by error scaling to generate the global modulation signal, with positive values ​​indicating correct motion direction detection and negative values ​​indicating incorrect detection.

[0064] In step 5, the synaptic weights are updated using a three-factor learning rule. The learning rate is set to 0.005, and the absolute value of the weight change does not exceed 50% of the original weight. Simultaneously, the threshold potentials of some neurons are fine-tuned to optimize the network's response characteristics.

[0065] In step 6, the iterative process continues until the motion direction recognition accuracy reaches 90% or higher, or the maximum number of iterations is reached. Ultimately, the network develops specific responses to eight different motion directions, validating the effectiveness of the method in complex visual information processing tasks.

[0066] Example 3 In the third embodiment, a neurodynamic model for studying working memory is constructed using local circuits in the mammalian cortex as the research object.

[0067] In step 1, the neural connectome data was derived from the reconstruction results of the mouse primary visual cortex, containing approximately 1000 excitatory neurons and 200 inhibitory neurons, with a synaptic connection density of approximately 10%. A network topology containing 1200 neurons was constructed based on the connectome data.

[0068] In step 2, a multi-compartment neuron model is used to model each neuron. Each neuron contains one soma compartment and 2 to 4 dendritic compartments. Excitatory and inhibitory neurons have different membrane capacitance and axial resistance parameters. NMDA and AMPA receptors are arranged on the dendritic compartments to simulate the voltage-dependent characteristics required for synaptic plasticity.

[0069] In step 3, the working memory task is presented as a delayed response task. During the stimulus presentation phase, specific patterned stimulus information is input into the network; during the delayed phase, the network remains autonomously active; and during the response phase, the network is required to output a specific pattern related to the stimulus. The system records the activity state of neurons during both the stimulus presentation and delayed phases, calculating local plasticity traces. A time window method is combined with a spike timing method to balance the extraction of relevance and temporal information.

[0070] In step 4, the global modulation signal is generated based on task completion. Task completion is defined as the similarity between the network output and the target output during the response phase. After temporal difference processing, the similarity is used to generate the initial modulation signal during the stimulus presentation phase and the feedback modulation signal during the response phase.

[0071] In step 5, in addition to updating synaptic weights, the excitability parameters of neurons are also optimized. Through a collaborative optimization strategy, synaptic weights and neuronal excitability parameters are updated sequentially in each iteration cycle, enabling the network to better maintain its working memory state.

[0072] In step 6, the iterative process continues until the network can maintain a stable activity pattern during the delay period and can output the correct memory content in the response phase. Ultimately, the network forms a population of neurons with sustained activity characteristics, simulating the neural mechanisms of working memory in the biological brain and validating the application potential of the method of this invention in the study of higher cognitive functions.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing a plastic neurodynamic model based on neural connectome constraints, characterized in that, Includes the following steps: Step 1: Obtain neural connectome data of the target nervous system and construct a neural network topology based on the structural connections between neurons. The neural connectome data includes chemical synaptic connection information and electrical synaptic connection information. Establish a network connection matrix of dimension N multiplied by N based on this connection information, where N is the number of neurons in the neural network. Step 2: Configure neurodynamic models for the neurons in the network to describe the dynamic changes in neuronal membrane potential and ion channel states over time. Construct a biologically constrained neural network simulation system by combining the neuronal dynamic models with the connectome topology. Step 3: Activate the neural network through external sensory stimuli, environmental input signals, or task-driven signals to induce dynamic responses in neurons. Obtain the neuronal and synaptic activity states during network operation. Calculate local plasticity traces based on the activity correlations between neurons to characterize synaptic connections within a certain time window. Step 4: Generate a global modulation signal based on task objectives, environmental feedback, behavioral rewards, prediction errors, or output errors. This global modulation signal reflects the degree of difference between the current network output and the target state, and serves as a regulatory factor in the network learning process to participate in synaptic parameter updates. Step 5: Optimize the neural network parameters using a three-factor learning mechanism. Use local plasticity traces as the basis for synaptic updates and modulate them with the global modulation signal to calculate the weight changes of each synaptic connection. This ensures that synaptic parameter updates are simultaneously influenced by both local neural activity and the global task objective. Step 6: Reapply the updated synaptic weights to the neural network model and continue neurodynamic simulation. The above process is executed iteratively, allowing the network connection parameters to continuously and adaptively adjust according to environmental input and task requirements, ultimately obtaining an optimized network state that meets the target task requirements.

2. The method according to claim 1, characterized in that, In step 3, the calculation of local plasticity traces adopts the time window method. The width of the time window is set within a preset time window range. Specifically, the calculation method is to integrate the correlation between the activity of presynaptic neurons and the activity of postsynaptic neurons within the time window. The plasticity trace value is obtained after nonlinear transformation of the integration result.

3. The method according to claim 1, characterized in that, In step 3, the calculation of local plasticity traces is based on the difference in the firing time of presynaptic and postsynaptic neuron action potentials. When the presynaptic neuron action potential is earlier than the postsynaptic neuron action potential, enhanced plasticity traces are generated. When the presynaptic neuron action potential is later than the postsynaptic neuron action potential, weakened plasticity traces are generated. The time difference threshold is set within a preset temporal difference threshold range.

4. The method according to claim 1, characterized in that, In step 4, the generation of the global modulation signal is carried out in different ways depending on the task type: for reinforcement learning tasks, the modulation signal is calculated by the reward function, and the reward value is processed by the temporal difference algorithm to generate the modulation signal; for supervised learning tasks, the modulation signal is obtained by the error between the network output and the target label after error scaling and normalization. For unsupervised learning tasks, the modulation signal is obtained by adaptively adjusting the prediction error or reconstruction error.

5. The method according to claim 1, characterized in that, The mathematical expression of the three-factor learning rule in step 5 is as follows: the change in synaptic weight is equal to the global modulation signal multiplied by the local plasticity trace and then multiplied by the learning rate, where the learning rate is set to a preset learning rate range. The change in weight is applied to the synaptic weight update after being constrained by the upper and lower limits. The upper limit of the weight is set to a preset multiple range of the original weight, and the lower limit of the weight is set to a preset minimum weight value.

6. The method according to claim 1, characterized in that, In addition to synaptic connection weights, parameters in the neurodynamic model are also optimized, including neuronal membrane conductance parameters, ion channel parameters, time constant parameters, and neuronal excitability parameters. A collaborative optimization strategy is adopted for connection parameters and dynamic parameters. In each iteration, synaptic weights are updated first and then dynamic parameters are updated, or an alternating update method is used.

Images