A method for modeling action potentials in giant squid axons

By tuning the hyperparameters and convolutional layer structure of the neural network through simulated annealing method, the problems of high calculation cost and slow convergence speed of the FitzHugh-Nagumo model parameter estimation are solved, and more efficient parameter estimation and noise impact analysis are achieved, which significantly improves the estimation. Accuracy.

CN113361688BActive Publication Date: 2024-12-06ZHEJIANG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110623080.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-06-04
Publication Date
2024-12-06
Estimated Expiration
2041-06-04

AI Technical Summary

Technical Problem

The prior art has high calculation cost, slow convergence speed when estimating FitzHugh-Nagumo model parameters, lacks analysis on the impact of noise, and the training set size is fixed, which cannot effectively reduce the calculation cost and improve the estimation accuracy.

Method used

The hyperparameters of the neural network are tuned by simulated annealing method, and the parameters of the FitzHugh-Nagumo model are fitted through the neural network to reduce the training set size, and the impact of noise on the data is analyzed. The neural network is used to use the convolutional layer, pooling layer and fully connected layer. The network performs parameter estimation.

Benefits of technology

The cumbersome process of selecting hyperparameters in neural networks is avoided, and the nonlinearity and non-convexity of the objective function is overcome, which significantly reduces the training set size and improves the accuracy and calculation efficiency of parameter estimation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113361688B_ABST
    Figure CN113361688B_ABST
Patent Text Reader

Abstract

The present invention provides a modeling method for action potentials in giant axons of squid. First, the FitzHugh-Nagumo model is used to describe the ionic mechanism of initiation and transmission of action potentials in giant axons of squid. Then, based on the weak prior knowledge of the parameters of the FitzHugh-Nagumo model, a data set is generated in combination with the established FitzHugh-Nagumo model. A neural network composed of a convolutional layer, a pooling layer, and a fully connected layer is established for parameter estimation of the FitzHugh-Nagumo model. The established neural network hyperparameters are adjusted using a simulated annealing algorithm. The neural network with adjusted hyperparameters is used for parameter estimation of the FitzHugh-Nagumo model. The action potential in the giant axon of the squid is modeled. The present invention avoids the problem of selecting hyperparameters of the neural network, thereby saving tedious manual trial and error processes, greatly reducing the size of the training set, analyzing the influence of noise in training and / or test data, and also considering and analyzing the influence of the size of the training set on the parameter estimation results, and the obtained parameter estimation results are good.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the fields of biological neuroscience and ordinary differential equation parameter estimation, and in particular to a modeling method for action potentials in giant axons of squid. Background Art

[0002] The Hodgkin-Huxley model is a model established based on the measurement of the temporal characteristics of the cell membrane and ion conductance of the giant squid axon. It is a mathematical model that describes how to initiate and transmit neuronal action potentials, thereby explaining the ionic mechanism of the initiation and transmission of action potentials in the giant squid axon. The Hodgkin-Huxley model is a set of nonlinear differential equations and a continuous-time model. The FitzHugh-Nagumo model consists of two coupled nonlinear ordinary differential equations, one of which describes the rapid evolution of the neuronal membrane potential, and the other represents the slow "recovery" effect of sodium channel inactivation and potassium channel inactivation. The FitzHugh-Nagumo model is a simplified model of the Hodgkin-Huxley model, but it retains many of the characteristics of the Hodgkin-Huxley model.

[0003] Common parameter estimation methods in FitzHugh-Nagumo models use heuristics and trial and error and may involve complex regression steps. Techniques used include differential evolution, genetic algorithms, and brute force grid search. The disadvantages of these methods are that they are computationally expensive or have slow convergence. On the other hand, using gradient descent to find the minimum of a FitzHugh-Nagumo-based regression problem has strong nonlinearities in the objective function and requires a good initial guess.

[0004] Although machine learning techniques are often widely used in prediction and classification tasks, they are rarely used to estimate the parameters of mathematical models. Some scholars have proposed neural network-based neural dynamics parameter estimation methods, but the cost of their methods is a large number of training sets (10 5 order), which is about two orders of magnitude larger than our work, and each training sample requires the solution of an ODE model, which incurs a large computational cost if performed frequently. Moreover, previous studies did not analyze the impact of noise in the training and / or test data, and the size of the training set was not varied. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention proposes a modeling method for action potentials in giant axons of squid. The method uses simulated annealing to adjust the established neural network hyperparameters, avoiding the problem of selecting the neural network hyperparameters, thereby saving the tedious manual trial and error process. Through the fitting of the neural network, the high nonlinearity and non-convexity of the objective function (or loss function) of the inversion problem, the strong nonlinearity and dependence between parameters, and the potential challenges of multiple local minima are overcome. The present invention greatly reduces the size of the training set, analyzes the influence of noise in training and / or test data, and also considers and analyzes the influence of the size of the training set on the parameter estimation results, and obtains better parameter estimation results.

[0006] To achieve the above object, the present invention adopts the following technical solution: a modeling method for action potentials in giant axons of squid, comprising the following steps:

[0007] S1: Generate FitzHugh-Nagumo model parameters based on weak prior knowledge of FitzHugh-Nagumo model parameters, then use the generated model parameters and combine with the FitzHugh-Nagumo model to generate noise-free membrane potential data, add first-order autoregressive noise simulation to the noise-free membrane potential data to generate noisy membrane potential measurement data, and construct the FitzHugh-Nagumo model parameters and the corresponding noisy membrane potential measurement data into a data set; wherein, the FitzHugh-Nagumo model is expressed as:

[0008]

[0009]

[0010] The membrane potential u = u(t) and the recovery variable v = v(t) are the unknowns of the FitzHugh-Nagumo ODE. γ is a constant value, ζ represents the total membrane current, and θ0 and θ1 are model parameters. The weak prior knowledge of the model parameters is specifically that the value interval of the model parameters is greater than [0,1].

[0011] S2: Establish a neural network composed of a convolutional layer, a pooling layer and a fully connected layer for FitzHugh-Nagumo model parameter estimation; wherein the input of the neural network is a noisy membrane potential measurement value, and the output is a FitzHugh-Nagumo model parameter estimation value.

[0012] S3: training the neural network using the data set constructed in step S1, and adjusting the hyperparameters of the established neural network using simulated annealing method;

[0013] S4: The actual measured membrane potential data is input into the neural network with adjusted hyperparameters to estimate the parameters of the FitzHugh-Nagumo model.

[0014] S5: Modeling action potentials in squid giant axons using estimated FitzHugh-Nagumo model parameters.

[0015] 3. Further, the step S1 specifically includes the following sub-steps:

[0016] S1.1: Assume that the parameters θ0, θ1 of the FitzHugh-Nagumo model are independent and identically distributed normally:

[0017] θ0~N(0.4,0.3), θ1~N(0.4,0.4)

[0018] And the interval range of θ0,θ1 is:

[0019] θ0∈[-0.2,1.0],θ1∈[-0.4,1.2]

[0020] Solve the FitzHugh-Nagumo ODE by sampling the parameters θ from the prior distribution and storing the membrane potential u in equal steps θ (t i ), i=1,...,N t .

[0021] S1.2: Adding first-order autoregressive noise to the noise-free membrane potential data to generate noisy membrane potential measurement data d(t i ), expressed as:

[0022] d(t i ):=u θ (t i )+η(t i )

[0023] Among them, d(t i ) is the simulated noise-containing t i The membrane potential measurement data at the time, η(t i ) is the first-order autoregressive colored noise. η(t i ) size is the original u θ (t i ) is about 5 to 10%.

[0024] S1.3: Finally, the sampled parameters θ and the corresponding noisy membrane potential measurement data form a data set.

[0025] Furthermore, in step S3, the neural network hyperparameters adjusted using the simulated annealing method include the number of convolutional layer filters (filters), convolution kernel size (kernel_size), convolution step (convolution_strides), pooling layer type (pool_type), pooling layer filter size (pool_strides), pooling step (pool_strides), the number of fully connected layer nodes (units), activation function (activation), learning rate (learning_rate), epoch (epoch), batch size (batch_size), loss function (loss_function), the number of convolutional layers and the number of fully connected layers.

[0026] Furthermore, in step S3, the simulated annealing utilizes the solid annealing principle to simulate the solid being gradually cooled from a high-temperature disordered state to a low-temperature ordered state. The simulated annealing algorithm accepts a solution that is worse than the current solution with a certain probability, so it may jump out of the current local optimal solution and reach the global optimal solution. According to the Metropolis criterion, the probability that a particle tends to equilibrium at temperature T is exp(-ΔE / kT), where E is the internal energy at temperature T, ΔE is its change, and k is the Boltzmann constant. The Metropolis criterion can be expressed as:

[0027]

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows: the modeling method for action potentials in giant squid axons of the present invention uses simulated annealing to adjust the established neural network hyperparameters, avoiding the problem of selecting the neural network hyperparameters, thereby saving the tedious manual trial and error process; through the fitting of the neural network, the high nonlinearity and non-convexity of the objective function (or loss function) of the inversion problem, the strong nonlinearity and dependence between parameters, and the potential challenges of multiple local minima are overcome; the size of the training set is greatly reduced, the influence of noise in training and / or test data is analyzed, and the influence of the size of the training set on the parameter estimation results is considered and analyzed, and the parameter estimation results are better. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the solution of the two ODEs system of the FitzHugh-Nagumo model when θ0=0.7, θ1=0.8, γ=3.0, ζ=-0.4, u0=v0=0 in the present invention. DETAILED DESCRIPTION

[0030] The present invention adopts the FitzHugh-Nagumo model to describe the ion mechanism of initiation and transmission of action potentials in the giant axons of squids, adopts the FitzHugh-Nagumo model to describe the ion mechanism of initiation and transmission of action potentials in the giant axons of squids, and uses the weak prior knowledge of FitzHugh-Nagumo ODE parameters to generate training sets and test sets without / with first-order autoregressive colored noise. A neural network composed of a convolutional layer, a pooling layer and a fully connected layer is established for parameter estimation of the FitzHugh-Nagumo model. The hyperparameters of the neural network are regarded as parameters to be estimated, and the optimal hyperparameters of the neural network are found by a simulated annealing algorithm. The neural network with adjusted hyperparameters is used for parameter estimation of the FitzHugh-Nagumo model. The action potential in the giant axons of squids is modeled using the estimated FitzHugh-Nagumo model parameters.

[0031] The modeling method of the present invention for action potential in giant axons of squid specifically comprises the following steps:

[0032] (1): The FitzHugh-Nagumo model is used to describe the ionic mechanisms of action potential initiation and transmission in the giant axon of the squid;

[0033] The FitzHugh-Nagumo model consists of two coupled nonlinear ordinary differential equations, one of which describes the rapid evolution of the neuronal membrane potential and the other of which represents the slow “recovery” effect of sodium channel inactivation and potassium channel inactivation:

[0034]

[0035]

[0036] The membrane potential u = u(t) and the recovery variable v = v(t) are the unknowns of the FitzHugh-Nagumo ODE (ordinary differential equation). γ is a constant value, and ζ represents the total membrane current, which is a stimulus applied to the neuron and is assumed to be constant in time. θ0 and θ1 are model parameters. The initial conditions of the FitzHugh-Nagumo ODE are u(0) = u0, v(0) = v0.

[0037] Since usually only the membrane potential is observed in experiments, the second variable v of the FitzHugh-Nagumo ODE is excluded from the data, and we get the measured data:

[0038]

[0039] Among them, d(t) is the measurement data at time t, u θ* (t) is the true value of the unknown parameter θ * =(θ0 * ,θ1 * ) The solution of the FitzHugh-Nagumo ODE obtained by η(t) is a first-order autoregressive colored noise and satisfies the following form:

[0040] η(t i ): = ρη(t i-1 )+ε(t i ), i=2,...,N t ,

[0041] Among them, |ρ|<1, ε(t i ) is a normally distributed white noise and is related to η(t i-1 ) are independent of each other. represents the variance of the normal distribution, N t is the number of time sampling points, Δ t represents the sampling time interval. Since the absolute value of the autocorrelation coefficient |ρ|<1, the noise process is stable, so the colored noise η(t i ) is constant in time, so we can get η(t i ) and ε(t i )distributed:

[0042]

[0043] η(t i ) size is the original u θ (t i ) is about 5 to 10%.

[0044] (2): Generate training and test sets based on the weak prior knowledge of the FitzHugh-Nagumo model parameters and the established FitzHugh-Nagumo ODE model;

[0045] Since we assume that little is known about the distribution of the parameter θ, that is, we only have weak prior knowledge, we choose a relatively large variance for the distribution of θ. Since in practice, the FitzHugh-Nagumo model chooses θ0,θ1∈[0,1], we choose even weaker prior knowledge and assume that θ0,θ1 are independent and identically distributed normal distributions:

[0046] θ0~N(0.4,0.3), θ1~N(0.4,0.4)

[0047] Further restrict the interval range of θ0,θ1:

[0048] θ0∈[-0.2,1.0],θ1∈[-0.4,1.2]

[0049] By sampling the parameters θ from its prior distribution and discarding samples outside the prior bounds. After obtaining samples of θ, the FitzHugh-Nagumo ODE is solved and the membrane potential u is stored at equal steps of Δt = 0.2 ms. θ (t i ), i=1,...,N t Thus, the sampled parameter θ and the corresponding noise-free membrane potential data or the corresponding noisy membrane potential measurement data constitute a data set.

[0050] In the statistical / Bayesian framework, the inversion problem is transformed into finding the posterior probability density π(θ|d) of the parameter θ given the data d, where d is the discretization of d(t). By Bayesian principles, the posterior term π(θ|d) is composed of a likelihood term π(d|θ) and a prior term π(θ):

[0051] π(θ|d)∝π(d|θ)π(θ)

[0052] By minimizing the negative logarithm of the above formula, we can obtain a point estimate of the posterior probability density, called the maximum a posteriori point (MAP), to derive a point estimate of θ from a given time series d:

[0053]

[0054] (3) Establish a neural network consisting of convolutional layers, pooling layers and fully connected layers for parameter estimation of the FitzHugh-Nagumo model;

[0055] The reconstruction mapping of the learning inversion problem is calculated by constructing a neural network. The input of the neural network is the time series of the solution from the FitzHugh-Nagumo ODE, and the output is the corresponding parameter θ that generates the time series. The neural network-based mapping is defined as:

[0056] y l =F l (y l-1 ), 1≤l≤L, y0:=d

[0057] Among them, the input d of the neural network is the discrete form of the time series d(t), F l represents the lth layer of the neural network, is the output of the neural network and serves as the parameter of the prediction model.

[0058] Each fully connected layer consists of nu nodes, and each node in each layer is connected to all nodes in the adjacent layer. A fully connected layer is written as a combination of an affine mapping and a nonlinear function:

[0059] F l (y l-1 )=φ(W l y l-1 +b l )

[0060] Among them, W l ∈R nu×nu is the weight matrix of the lth layer, b l is the bias of the lth layer, and φ is the nonlinear activation function. Through the Adam algorithm and the loss function, stochastic gradient descent is used to optimize W l and b l .

[0061] A convolutional layer consists of n f The n filters of a convolution layer are each associated with a convolution kernel. f Each convolution kernel is connected to all adjacent layers, similar to the nodes in the fully connected layer above. The weights of these connections constitute the neural network parameters to be optimized. Since these weights are shared, the weights that need to be optimized are greatly reduced.

[0062] The convolutional layer and the pooling layer are intertwined. Both the convolutional step and the pooling step convert a small area of ​​the previous layer into a value, which is the same as the input y. l-1 In contrast, the output vector y l The size of is reduced. After the combination of several convolutional layers and pooling layers, a fully connected layer is connected.

[0063] (4): Use the training set constructed in step 2 to perform back propagation training on the neural network, and use the simulated annealing method to adjust the established neural network hyperparameters. The objective function of the neural network training is to minimize the mean square error, where the mean square error is expressed as:

[0064]

[0065] Among them, θ (j) is the true value, θ ^(j) is the predicted value, and j is the sample index.

[0066] Simulated annealing uses the principle of solid annealing to simulate the slow cooling of solids from a high-temperature disordered state to a low-temperature ordered state. The simulated annealing algorithm accepts a solution that is worse than the current solution with a certain probability, so it may jump out of the current local optimal solution and reach the global optimal solution. According to the Metropolis criterion, the probability that a particle tends to equilibrium at temperature T is exp(-ΔE / kT), where E is the internal energy at temperature T, ΔE is its change, and k is the Boltzmann constant. The Metropolis criterion can be expressed as:

[0067]

[0068] In the specific implementation, it is necessary to set the initial temperature and the termination temperature, as well as the learning rate and the objective function. The learning rate is generally 0.99 or 0.98 to gradually reduce the temperature. The objective function is generally the minimum mean squared error (MSE) between the output value of the neural network and the true value, corresponding to the internal energy E at temperature T, k is the Boltzmann constant, which is usually taken as 1 in practice. The training is stopped until the preset number of iterations, the termination temperature or ΔE is less than the preset threshold, and a neural network with tuned hyperparameters is obtained.

[0069] (5): The neural network with tuned hyperparameters is used for parameter estimation of the FitzHugh-Nagumo model.

[0070] According to the actual measured value of the membrane potential, the parameters of the FitzHugh-Nagumo model can be obtained by inputting the neural network in the present invention, and then a FitzHugh-Nagumo model close to the real one can be constructed, which can be applied to the research of random resonance, discharge rhythm, bifurcation phenomenon, chaotic behavior, dynamic behavior, etc. of giant squid axons.

[0071] (6): Modeling the action potential in the squid giant axon.

[0072] The FitzHugh-Nagumo model parameters obtained from neural network training were substituted into the FitzHugh-Nagumo model to obtain the dynamic model of action potentials in the giant axon of the squid.

[0073] The example performs parameter estimation for the FitzHugh-Nagumo model used to describe action potentials in squid giant axons.

[0074] Step 1: Experimental setup;

[0075] The algorithm is implemented in Python, using the explicit Runge-Kutta (RK23) method of order 3(2) to integrate the FitzHugh-Nagumo ODE in time, and using Tensorflow / Keras (2.4.1) to implement the neural network. In all experiments, M=2000 randomly selected test samples are used to evaluate the parameter estimation performance of different neural networks. The training samples N are 500, 1000, 2000, 4000, and 8000 respectively to evaluate the impact of different numbers of training samples on parameter estimation. The neural network uses the Adam optimizer, and the padding of the convolutional layer and the pooling layer is fixed to SAME.

[0076] Step 2: Generate training and test sets without or with first-order autoregressive colored noise;

[0077] By sampling the parameters θ from the prior distribution of the FitzHugh-Nagumo model parameters θ and discarding samples outside the prior bounds, 10,000 sets of parameters θ are obtained. After obtaining the samples of θ, the FitzHugh-Nagumo ODE is solved and the membrane potential u is stored at equal steps of Δt = 0.2 ms. θ (t i ), i=1,...,N t , thus generating a N-sized t = 1000 time series. Thus, 8000 and 2000 training and test sets without first-order autoregressive colored noise can be obtained respectively. θ (t i ) to obtain the actual noisy measurement data, and similarly, 8000 and 2000 sets of training and test sets containing first-order autoregressive colored noise can be obtained respectively.

[0078] In the first-order autoregressive colored noise model, ρ and σ are not constant values, but random values. 2 ),σ~N(0.07,0.01 2 ). Randomly sample 100 independent ρ and σ to generate 10,000 pairs of parameters (ρ, σ) used to generate first-order autoregressive colored noise and add to the membrane potential u of 10,000 training and test sets. θ (t i )superior.

[0079] Step 3: Use simulated annealing algorithm to adjust the hyperparameters of the neural network;

[0080] Establish a neural network consisting of convolutional layers, pooling layers, and fully connected layers for parameter estimation of the FitzHugh-Nagumo model. Select the neural network hyperparameters to be adjusted: the number of convolutional layer filters (filters), convolution kernel size (kernel_size), convolution stride (convolution_strides), pooling layer type (pool_type), pooling layer filter size (pool_strides), pooling stride (pool_strides), number of fully connected layer nodes (units), activation function (activation), learning rate (learning_rate), epoch (epoch), batch size (batch_size), loss function (loss_function), number of convolutional layers, and number of fully connected layers.

[0081] Use simulated annealing to adjust the hyperparameters of the neural network selected above. First, set the initial temperature to 10 5 ℃, the end temperature is 10 -5 ℃, the learning rate is 0.98, and the objective function is MSE. Using the Metropolis criterion, a solution that is better than the current solution is accepted, and a solution that is worse than the current solution is accepted with probability exp(-ΔE / kT).

[0082] The advantage of noise-free data is that it is possible to perform a sufficient number of iterations for the optimization algorithm using stochastic gradient descent without overfitting due to the noise in the data. Therefore, the training set and test set selected for tuning the neural network hyperparameters are 1000 / 2000 sets of data without first-order autoregressive colored noise.

[0083] The neural network hyperparameters obtained by simulated annealing are: filters=16, kernel_size=3, convolution_strides=1, pool_size=3, pool_strides=1, units=64, learning_rate=0.0013, epochs=500, batch_size=64, activation is swish, pool_type is averagepooling, loss_function is mean squared error, there are 3 convolutional layers, and 2 fully connected layers.

[0084] Step 4: Use the neural network with tuned hyperparameters for FitzHugh-Nagumo model parameter estimation.

[0085] After the neural network hyperparameters are tuned, the parameters of the FitzHugh-Nagumo model are estimated. The experimental results are as follows Figure 1 As shown in Figure 2, the evaluation criterion is the median of the absolute percentage error (Median-APE), where the absolute value of the relative error is represented by |θ (j) -θ ^(j) | / |θ (j) |,θ (j) is the true value, θ ^(j) is the predicted value, j is the sample index, and the smaller the value of Median-APE, the higher the prediction accuracy. There are three columns in the table representing different noise configurations: both the test set and the training set do not contain noise (the second column), only the test set contains noise (the third column), and both the test set and the training set contain noise (the fourth column). The first column of the table shows the impact of different training set numbers N on the model parameter estimation. In all experiments, the test set size of M = 2000 remains unchanged for comparison.

[0086] Table 1 Parameter estimation results of FitzHugh-Nagumo model

[0087]

[0088] The results in Table 1 show that when both the test set and the training set do not contain / contain noise, the prediction accuracy gradually improves with the increase of the number of training set samples N. When only the test set contains noise, the prediction accuracy decreases with the increase of the number of training set samples N. The prediction accuracy is best when both the test set and the training set do not contain noise. When the training set N = 8000, the Median-APE is only 1.30E-2; when only the test set has noise, the prediction accuracy is poor, and when N = 8000, the Median-APE reaches 2.25E-1; when both the training set and the test set contain noise, the prediction accuracy is good. When N = 8000, the Median-APE is 6.45E-2. In addition, the number of training set samples N is 10 higher than that of previous works. 5 It has been reduced by at least an order of magnitude, greatly reducing the neural network training time.

[0089] Step 5: Modeling the action potential in the squid giant axon.

[0090] Substitute the FitzHugh-Nagumo model parameters obtained by neural network training into the FitzHugh-Nagumo model to obtain the dynamic model of action potential in the giant axon of the squid. Since the number of test set samples is M = 2000, the average value is taken to obtain θ = [0.404391454119409, 0.4001243252087388] and the true value is θ = [0.4, 0.4], which shows that the effect is very good. Therefore, the following action potential model in the giant axon of the squid can be established:

[0091]

[0092]

[0093] The above model can simulate the slow "recovery" phenomenon of sodium channel inactivation and potassium channel inactivation of squid giant axons, and study the discharge rhythm of squid giant axons.

[0094] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also involves equivalent technical means that can be thought of by those skilled in the art based on the inventive concept.

Claims

1. A modeling method for action potentials in giant axons of squid, characterized in that: The following steps are involved: S1: Generate FitzHugh-Nagumo model parameters based on weak prior knowledge of FitzHugh-Nagumo model parameters, then use the generated model parameters and combine with the FitzHugh-Nagumo model to generate noise-free membrane potential data, add first-order autoregressive noise simulation to the noise-free membrane potential data to generate noisy membrane potential measurement data, and construct the FitzHugh-Nagumo model parameters and the corresponding noisy membrane potential measurement data into a data set; wherein, the FitzHugh-Nagumo model is expressed as: The membrane potential u=u(t) and the recovery variable v=v(t) are unknowns of the FitzHugh-Nagumo ODE; γ is a constant value, ζ represents the total membrane current, and θ0 and θ1 are model parameters; the weak prior knowledge of the model parameters is specifically: the value interval of the model parameters is greater than [0,1]; S2: Establishing a neural network composed of a convolutional layer, a pooling layer and a fully connected layer for FitzHugh-Nagumo model parameter estimation; wherein the input of the neural network is a noisy membrane potential measurement value, and the output is a FitzHugh-Nagumo model parameter estimation value; S3: Use the data set constructed in step S1 to train the neural network, and use the simulated annealing method to adjust the established neural network hyperparameters; when using the simulated annealing method to adjust the established neural network hyperparameters, the minimized mean square error between the output value of the neural network and the true value is used as the internal energy, and a solution worse than the current solution is accepted with a certain probability according to the Metropolis criterion, which is expressed as: Where p is the probability, E is the internal energy at temperature T, ΔE is the change in E corresponding to the two sets of hyperparameters, and k is the Boltzmann constant; S4: Input the actual measured membrane potential data into the neural network with tuned hyperparameters to estimate the parameters of the FitzHugh-Nagumo model; S5: Modeling action potentials in squid giant axons using estimated FitzHugh-Nagumo model parameters.

2. A modeling method for action potentials in giant axons of squid as claimed in claim 1, characterized in that: The step S1 specifically includes the following sub-steps: S1.1: Assume that the parameters θ0, θ1 of the FitzHugh-Nagumo model are independent and identically distributed normally: θ0~N(0.4,0.3), θ1~N(0.4,0.4) And the interval range of θ0,θ1 is: θ0∈[-0.2,1.0],θ1∈[-0.4,1.2] Solve the FitzHugh-Nagumo ODE by sampling the parameters θ from the prior distribution and storing the membrane potential u in equal steps θ (t i ), i=1,...,N t ; N t is the number of time sampling points; S1.2: Adding first-order autoregressive noise to the noise-free membrane potential data to generate noisy membrane potential measurement data d(t i ), expressed as: d(t i ):=u θ (t i )+η(t i ) Among them, d(t i ) is the simulated noise-containing t i The membrane potential measurement data at the time, η(t i ) is the first-order autoregressive colored noise; η(t i ) size is the original u θ (t i ) 5-10%; S1.3: Finally, the sampled parameters θ and the corresponding noisy membrane potential measurement data form a data set.

3. The method for modeling action potentials in giant axons of squid as claimed in claim 1, characterized in that: In step S3, the neural network hyperparameters adjusted using the simulated annealing method include the number of convolutional layer filters, the size of the convolution kernel, the convolution step, the pooling layer type, the pooling layer filter size, the pooling step, the number of fully connected layer nodes, the activation function, the learning rate, the period, the batch size, the loss function, the number of convolutional layers, and the number of fully connected layers.