Memristive neuron neuromorphic dynamics prediction method based on hybrid machine learning framework

Abstract

The invention discloses a memristive neuron neuromorphic dynamics prediction method based on a hybrid machine learning framework. The memristive neuron neuromorphic dynamics prediction method is suitable for third-order and above memristive neuron systems with external stimulation. According to the method, an improved next-generation reserve pool is integrated to calculate MNGRC and XGBoost regression algorithms, and a dual-path prediction architecture specially designed for partial state observable scenes is constructed. The architecture can realize high-precision prediction of complex neuromorphic dynamic behaviors of third-order memristive neurons under the condition that only a single state variable is observed. According to the technical scheme, the double-layer secondary reserve pool design and the XGBoost state estimation mechanism work cooperatively, the dependence of a traditional method on complete state information is effectively overcome, and the accuracy and stability of long-term prediction in the chaos marginal area are remarkably improved. The method shows excellent cross-mode generalization ability, and accurate prediction of multiple types of neuromorphic behaviors including resting, periodic oscillation, chaotic oscillation and complex burst modes can be realized by only needing a small amount of training data of several representative behaviors; and a key technical support is provided for the practical application of the neuromorphic computing system in resource-constrained environments such as edge computing and the like.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of neuromorphic computing and artificial intelligence, specifically relating to a method for predicting the neuromorphic dynamics of third-order and higher-order memristor neurons with applied stimuli based on a hybrid machine learning framework. This method can accurately predict and control the complex neuromorphic dynamic behavior of third-order and higher-order memristor neurons in chaotic edge regions under conditions where some state variables are observable. Background Technology

[0002] As traditional transistor computing architectures approach their physical limits and the von Neumann bottleneck increasingly restricts computational efficiency, neuromorphic computing has emerged as a promising paradigm for next-generation information processing. Memristor neurons, especially third-order circuits operating at the edge of chaos, exhibit rich neuromorphic dynamics and can highly realistically simulate biological neural activity. However, due to their complex nonlinear behavior, accurate prediction faces significant challenges.

[0003] In existing technologies, traditional simulation methods predict dynamic behavior by solving nonlinear differential equations, which is computationally intensive and difficult to handle long-term dynamics or input optimization tasks. Machine learning techniques, especially reservoir computation and gradient boosting methods, offer efficient alternatives for modeling such complex systems. However, most existing methods require complete state information of the system, which is often impractical in practical neuromorphic hardware implementations because only a subset of state variables is typically available.

[0004] In recent years, while research progress on memristor chaotic networks has continuously pushed the boundaries of their hardware implementation, related work has mainly focused on forward dynamics analysis. Currently, there is still a lack of a comprehensive method that can accurately predict the complex neuromorphic behavior of third-order and higher-order memristor neurons with applied stimuli based only on some state variables (such as membrane potential), and it is particularly difficult to achieve reliable prediction and regulation of target responses under partially observable conditions. Summary of the Invention

[0005] 1. Purpose of the invention:

[0006] This invention aims to propose a method for predicting the neuromorphic dynamics of memristor neurons based on a hybrid machine learning framework, addressing the challenge of accurately predicting the complex dynamic behavior of third-order memristor neurons under partially observable conditions. This method innovatively integrates an improved next-generation reservoir computation (MNGRC) and XGBoost regression algorithms, constructing an efficient and robust prediction framework. It achieves high-precision prediction of system polymorphism and transient dynamics even with only partial observation of state variables, thus providing crucial technical support for the design and control of practical neuromorphic systems.

[0007] 2. Technical Solution

[0008] To achieve the above-mentioned objectives, this invention provides a method for predicting the neuromorphic dynamics of third-order and higher-order memristor neurons with applied stimuli based on a hybrid machine learning framework, mainly comprising the following steps:

[0009] Step 1: Construct a two-step prediction architecture for MNGRC and a hybrid machine learning framework with a dual-path data flow mechanism. The framework is characterized by integrating an improved next-generation reservoir computation MNGRC and an XGBoost regression algorithm, specifically designed for scenarios with partial state observability.

[0010] MNGRC, as the core module, contains two levels of reservoir computational units (NGRC A and NGRC B) to capture dynamic features at different time scales. XGBoost, as an auxiliary regression module, is used to handle classification and regression tasks, i.e., predicting another unknown state variable. The overall prediction framework takes as input partial state variables of memristor neurons (such as membrane potential) and input stimuli, and outputs the predicted neuromorphic dynamic behavior.

[0011] MNGRC, as the core prediction engine of the framework, adopts a two-tiered reservoir computing architecture, containing two parallel but structurally complementary NGRC units, NGRC A and NGRC B:

[0012] ●NGRC A unit: uses standard interval parameter s1=1, focusing on capturing the short-term time dependence and high-frequency oscillatory characteristics of neuronal dynamics.

[0013] ●NGRC B unit: Configured with differential interval parameter s2=2, specifically for extracting long-term time dependence and slow-varying dynamic features.

[0014] Both units share the same basic parameter settings, including polynomial order P, input delay k, and ridge regression regularization coefficient λ, but achieve multi-scale feature extraction through different time length strategies. This dual-path design enables the framework to simultaneously and accurately capture the full spectrum of neuromorphic behaviors, from fast unimodal spikes to complex multimodal burst patterns.

[0015] Each NGRC unit processes the input signal through the following mathematical process:

[0016] For a d-dimensional input vector u at time T consisting of partial state variables x and input voltage... t =[V in [x]∈R d u i =[u 1,i ,u 2,i ,…,u d,i ] T It is first mapped to a linear delayed embedding.

[0017]

[0018] Among them, O lin O represents the linear feature components generated through discrete sampling. linA and O linB The linear feature components generated for NGRC A and NGRC B are respectively, with each item corresponding to the current state and the previous steps, and the interval is s.

[0019] Subsequently, nonlinear features are formed by concatenating these linear vectors across all time steps i.

[0020]

[0021] in, This represents the outer product between vectors. The number of linear eigenvectors that accept the outer product operation is determined by the order P of the chosen polynomial, which also determines the maximum monomial degree generated in each operation.

[0022] The final nonlinear polynomial eigenvector O all It is by using the nonlinear feature O nonlin Linear features O lin It is constructed by connecting it in series with a bias term R.

[0023]

[0024] Each reservoir nonlinearly maps its input stream to a high-dimensional feature space, producing different feature vectors O. A and O B Subsequently, these two feature vectors are concatenated into a comprehensive fused feature representation O. fin This indicates that dual time information is integrated for subsequent processing.

[0025]

[0026] During the training phase, the output matrix W is obtained by applying ridge regression to the collected high-dimensional feature vectors (or states). out The target is the corresponding system output.

[0027] W out =U'·O T ·(O·O T +γ·I) -1 (5)

[0028] Where U' and O represent u(t) and O respectively. fin The matrix form; γ>0 is the L2 regularization coefficient; I is the identity matrix.

[0029] Finally, the predicted output is obtained by linearly combining the storage layer states O. fin With readout weight W out Multiplying them together, we get:

[0030] y(t+1)=W out ·O fin (6)

[0031] The XGBoost module acts as an auxiliary intelligent state estimator for the framework, addressing key partial observation issues. When the system cannot obtain complete state variables (a common situation in actual hardware implementations), this module operates through the following mechanism:

[0032] The input data, consisting of observable partial state variables (such as membrane potential) and input stimulus signals, is used to construct a decision tree ensemble using the extreme gradient boosting algorithm. The ensemble is then trained by optimizing the loss function through second-order Taylor expansion to more accurately estimate key state variables (such as internal state variables x) that cannot be directly observed.

[0033] The objective function of XGBoost is composed of the loss function. and regularization term Ω(f k This is composed of components to ensure the model's generalization ability.

[0034]

[0035] Where ζ(θ) is the linear space expression, T is the number of leaves in tree k, and ω j Let j represent the number of leaves, and γ and λ be the regularization parameters.

[0036] The model prediction at iteration t is defined as:

[0037]

[0038] Among them, f t (x i () represents the t-th decision tree, and α is the learning rate that controls the contribution of each tree to prevent overfitting. It is the i-th sample x i The predicted value.

[0039] The integrated prediction framework employs an innovative parameter channel integration mechanism to enable the collaborative work of the two modules:

[0040] Parameter channel: Input voltage V in The internal state variable x (measured or XGBoost estimated) is integrated into a joint input vector u(t) = [V in ,x] T .

[0041] Dual-path data flow:

[0042] Direct path: When all state variables are measurable, MNGRC directly processes the joint input vector for prediction.

[0043] Estimation-prediction path: When some states are unpredictable, XGBoost first estimates the missing variables and then sends the complete state information to MNGRC.

[0044] System inputs: Partially observable state x of the memristor neuron and external stimulus signal V in .

[0045] System output: Predicted complete neuromorphic dynamics, including resting states, chaotic oscillations, and different neuromorphic patterns.

[0046] Step 2: Determine the key parameter configurations for the hybrid machine learning framework.

[0047] This step employs Bayesian optimization to globally optimize the key parameters of the hybrid machine learning framework, ensuring optimal prediction performance under partially observable conditions. Bayesian optimization constructs a probabilistic surrogate model of the objective function, intelligently balancing exploration and utilization to efficiently locate the globally optimal parameter combination, making it particularly suitable for computationally expensive hyperparameter optimization problems. The specific implementation process is as follows:

[0048] Bayesian optimization constructs a surrogate model of the objective function based on a Gaussian process (GP), and its optimization process can be mathematically expressed as follows:

[0049] Let x∈χ be the hyperparameter vector to be optimized, f(x) be the evaluation index (such as RMSE), and the optimization objective be:

[0050]

[0051] Gaussian processes establish a probabilistic model of f(x) using the mean function m(x) and the covariance function k(x,x'):

[0052] f(x)~GP(m(x),k(x,x'))(10)

[0053] Given a set of observed points Where y follows a normal distribution i =f(x) i )+ε,ε~Ν(0,σ 2 ), where ε represents the random error term, f(x) i ) represents the input x i The deterministic part obtained by mapping through the true function f, the new point x i The predicted distribution is:

[0054] p(f t+1 |xt+1 D t )=N(u t (x t+1 ),σ t 2 (x t+1 (11)

[0055] Among them, u t (x)=k t T (K t +σ 2 I) -1 y t , σ t 2 (x)=k(x,x)-k t T (K t +σ 2 I) -1 k t Through the acquisition function α t (x) Guiding the selection of new points, such as expected improvements (EI):

[0056]

[0057] Where f * This is the current best observation value.

[0058] Key parameters of the XGBoost module include maximum depth d. max The number of trees, T, the learning rate, η, and the regularization coefficients, λ and γ, are considered. The optimization process is designed as follows:

[0059] The objective function of XGBoost is shown in Equation 7 of Step 1. Add parameter constraints:

[0060]

[0061] Similarly, the hyperparameter optimization of the MNGRC module focuses on the polynomial order p, input delay k, leakage rate α, spectral radius ρ, and ridge regularization coefficient γ. The optimization strategy is as follows:

[0062] The polynomial order P is a key parameter of MNGRC, determining the dimension of the feature space and the model's expressive power. The growth law of the feature dimension with P is as follows:

[0063]

[0064] Where d is the input dimension, dim is the mapping dimension of a single NGRC unit, and k is the delay step.

[0065] The objective function for optimization is to minimize the weighted combination of the Lyapunov time error and RMSE of the predicted trajectory:

[0066]

[0067] Where, λ true and λ pre Let be the actual and predicted Lyapunov exponents, respectively. Weighting coefficient ω1 controls the importance of the Lyapunov exponent relative error term in the total loss, weighting coefficient ω2 controls the importance of the prediction accuracy term in the total loss, and weighting coefficient ω3 controls the importance of the model complexity regularization term in the total loss. The constraints are as follows:

[0068]

[0069] To ensure parameter robustness, a time-series hierarchical cross-validation strategy was implemented, dividing the time-series data into five consecutive blocks, each retaining dynamic characteristics. The "past predicts future" principle was adopted, with the training set strictly preceding the validation set. Optimization terminated when validation performance failed to improve after 10 consecutive iterations. A detailed analysis was conducted specifically for the polynomial order P, ultimately determining the optimal parameter configuration.

[0070] Step 3: Train a hybrid machine learning process based on partial state variables.

[0071] This section systematically describes the training method of the hybrid machine learning framework under partial state variable constraints. Through a cleverly designed training strategy, the framework can achieve high-precision prediction of the complex dynamic behavior of third-order memristor neurons using only limited state information.

[0072] The training framework employs a dual-path collaborative training mechanism to address some observability challenges:

[0073] Main path training: When some state variables are measurable, the input stimulus and the measurable state variables are fused into a joint input vector, and high-dimensional feature representations are extracted through the MNGRC module.

[0074] Auxiliary path training: When key state variables are missing, the XGBoost module estimates the missing states based on measurable variables to form a complete system representation.

[0075] The training process strictly follows the principle of "predicting the future from the past" to ensure that the time series data segmentation conforms to causal relationships and avoids information leakage. Through a systematic reset protocol, the state of the reserve pool is reset to the initial value before each prediction, and a warm-up phase is implemented to eliminate the sensitivity of initial conditions. This is specifically optimized for the high dependence of chaotic systems on the initial state.

[0076] The training data design covers the complete dynamic spectrum of the third-order memristor neuron system:

[0077] Among these features, the periodic-to-chaotic transition includes a training set encompassing the full spectrum of dynamics from stable resting states to self-sustaining oscillations and then to chaotic oscillations. Hopf bifurcation boundaries specifically include sample points at bifurcation boundaries to ensure the model can capture critical transition behaviors.

[0078] For different duty cycle input conditions, an adaptive training strategy is proposed: a basic dynamic model is established using only training data with a single duty cycle (e.g., 10%).

[0079] The training framework innovatively achieves unified modeling of 18 neuromorphic behaviors: requiring only training data from three different representative neuromorphic behaviors.

[0080] This training strategy fully utilizes the inherent dynamic characteristics of memristor neurons at the edge of chaos, capturing the core dynamic principles of the system through partial state variables.

[0081] Step 4: Apply a hybrid machine learning framework to solve the prediction problem.

[0082] This step details the practical application of the trained hybrid machine learning framework in prediction problems. In the prediction scenario, the framework receives partial state variables (such as membrane potential) and input stimulus signals from memristor neurons as input. Working collaboratively through the MNGRC-XGBoost dual-path architecture, it accurately predicts the neuromorphic dynamics of the system. The framework can identify and predict various neuromorphic patterns exhibited by third-order memristor neurons at the edge of chaos, including resting state, periodic oscillation, self-sustaining oscillation, chaotic oscillation, and complex burst patterns, achieving cross-modal generalization with only training data for three representative behaviors. The framework demonstrates excellent cross-regional prediction capabilities near Hopf bifurcation boundaries, accurately capturing the critical transition from a stable resting state to self-sustaining oscillation, and the dynamic transition from periodic oscillation to chaotic oscillation. However, when the required state variables are missing, the process first uses the XGBoost module to estimate the missing state from the measurable state variables, and then iteratively adjusts the input signal using gradient descent to minimize the difference between the predicted output and the target output. For fluctuations occurring at signal transition edges, the framework employs smoothing techniques such as moving average or Kalman filtering for post-processing, significantly improving the quality of the predicted signal. To verify the reliability of the prediction results, numerical validation was performed on MATLAB and Python platforms, calculating the RMSE and phase space topological similarity between the predicted and actual trajectories. Experimental results show that the framework achieves a small normalized RMSE in forward prediction and successfully reconstructs various input signals with duty cycles ranging from 15% to 60%, fully demonstrating the feasibility and effectiveness of this method in practical neuromorphic hardware systems.

[0083] 3. Beneficial effects:

[0084] The proposed method for predicting the neuromorphic dynamics of third-order and higher-order memristor neurons with applied stimulation, based on a hybrid machine learning framework, offers the following significant advantages: First, this method overcomes the dependence of traditional predictions on complete state information, achieving high-precision predictions with only a single state variable and input voltage, significantly reducing the complexity and cost of hardware implementation. Second, the framework exhibits excellent cross-modal generalization capabilities, successfully predicting all behavioral patterns, from periodic oscillations to chaotic oscillations, using only training data from three representative neuromorphic behaviors. Third, through a two-level reservoir design and synergistic optimization with the XGBoost state estimator, the framework effectively overcomes the sensitivity of chaotic edge dynamics to initial conditions, greatly improving the stability of long-term predictions. Furthermore, this method provides a new approach for practical neuromorphic computing applications, particularly suitable for scenarios with limited state observation, such as edge computing, laying an important foundation for building high-performance, low-power brain-like computing systems. Attached Figure Description

[0085] Figure 1 The diagram shown is a flowchart of the implementation process.

[0086] Figure 2 The third-order neuron model used in the verification of this method is shown.

[0087] Figure 3 The diagram shows the structure of the MNGRC model.

[0088] Figure 4 The diagram shows the overall prediction framework of MNGRC and XGBoost.

[0089] Figure 5 The results are shown in the cross-dynamic region experiment.

[0090] Figure 6 The image shows the prediction results of the Hopf bifurcation boundary.

[0091] Figure 7 The results shown are the pulse duty cycle generalization prediction results.

[0092] Figure 8 The results show the predictions of 15 other neuromorphic behaviors using training with three different neuronal morphologies. Detailed Implementation

[0093] To provide a clearer understanding of the objectives, features, and advantages of this invention, the following detailed description is provided in conjunction with the accompanying drawings and specific embodiments:

[0094] This invention discloses a method for predicting the neuromorphic dynamics of memristor neurons based on a hybrid machine learning framework. This method solves the challenge of accurately predicting the complex dynamic behavior of third-order memristor neurons under partially observable conditions by integrating improved next-generation reservoir computation (MNGRC) with the XGBoost regression algorithm.

[0095] First, to verify the effectiveness of the framework, a third-order memristor neuron system was used as a benchmark validation platform. This system consists of locally active memristors (LAMs) and passive components, and its state equation is described as follows:

[0096]

[0097] Where x is the memristor state variable, v c Let i be the capacitor voltage, i be the inductor current, V be the input excitation voltage, and V be the system output voltage. out =v c .

[0098] For each different task prediction, a fourth-order Runge-Kutta (RK4) method was used for numerical simulation, with a time step dt = 0.01 and a total simulation duration of 200 seconds (20,000 data points). To eliminate transient effects, the first 10,000 data points were discarded, and the last 10,000 points were retained for analysis. Data with a mean of 0 and a standard deviation of 1 × 10⁻⁶ were added. -8 Gaussian noise was added to enhance the model's robustness to real-world hardware noise. All data were scaled to the [0,1] range using min-max normalization in Python.

[0099] The training process employs a systematic reset protocol: before each prediction, the MNGRC reservoir state is reset to its initial value, and approximately 10% of the initial transient data (about 1,000 data points) is used as warm-up data, effectively addressing the sensitivity of chaotic systems to initial conditions. The framework adopts a dual-path training strategy: the main path trains the input voltage V... in The measurable state variable x is combined as a joint input; when it is unmeasurable, the auxiliary path utilizes the XGBoost module based on V. in The current i is used to estimate the state variable x.

[0100] Experimental verification was conducted on four levels:

[0101] 1. Generalization across dynamic regions: For the MNGRC module, the key parameters determined by Bayesian optimization are: polynomial order P = 4, system parameters L = 0.08H, C = 0.04F, and initial values ​​[x(0), i(0), v]. c[0] = [0,0,0], with a delay of k = 2, the model was trained under the condition of V = 7.202V (periodic -2 oscillations) and successfully predicted the behavior under V = 7.3V (periodic -4 oscillations) and V = 7.8V (chaotic oscillations). In the chaotic region, the predicted trajectory maintained a Lyapunov exponent of 0.1, which is basically consistent with the 0.09 of the real system, and the RMSE was 0.276, demonstrating the model's ability to capture complex dynamics.

[0102] 2. Hopf bifurcation boundary prediction: Similarly, with the same parameter settings as in 1, after training at V = 8.59V and V = 8.79V, the model accurately predicts the behavior at V = 8.6V (self-sustaining oscillation) and V = 8.8V (steady resting point). The predicted steady equilibrium point coordinates (0.23, 0.69, 0.57) are in high agreement with the actual system coordinates (0.24, 0.67, 0.56), with RMSE values ​​of 0.276 and 0.171, respectively.

[0103] 3. For the MNGRC module, the key parameters determined through Bayesian optimization are: polynomial order P = 2 and delay k = 2. Using only training data with a duty cycle of E = 10%, the model successfully predicted various neuromorphic behaviors in the range of E = 15% to 35%, which highly matched the target signal and completely preserved the key dynamic features of the signal.

[0104] 4.18 Unified Prediction of Neuromorphic Behaviors: MNGRC configuration parameters are the same as in 3, and only three training states are used.

[0105] The framework (using excitable-I, resonator-I, and integrator) successfully predicted all 18 neuromorphic behaviors, including complex four-peak and five-peak burst patterns. The normalized RMSE between the predicted and actual trajectories was 0.298.

[0106] The phase space topology remains highly consistent.

[0107] To verify the reliability of the prediction results, numerical verification was first performed on the MATLAB platform, followed by theoretical analysis using a discretized model. The results show that the framework maintains high prediction accuracy even under partially observable conditions, overcoming the limitation of traditional methods that require complete state information.

[0108] This invention proposes a hybrid machine learning framework that achieves high-precision prediction of complex dynamics in third-order memristor neuron systems at the chaotic edge through synergistic optimization of MNGRC and XGBoost. This framework effectively captures and predicts the polymorphic and transient behaviors of the system using only a limited number of state variables, under conditions where some states are observable. This capability provides crucial technical support for the design and real-time control of neuromorphic computing systems in edge computing and other scenarios where state observation is limited. For those skilled in the art, adjustments and optimizations can be made to the specific implementation without departing from the core ideas of this invention, such as replacing different machine learning algorithms, adjusting model parameters, or adapting to other types of memristor neuron circuits; all such modifications should fall within the protection scope of this invention.

Claims

1. A method for predicting the neuromorphic dynamics of third-order and higher-order memristor neurons with applied stimuli based on a hybrid machine learning framework, characterized in that... Includes the following steps: S1: A method for predicting the neuromorphic dynamics of memristor neurons based on a hybrid machine learning framework. S11. Construct a hybrid machine learning framework with a dual-path prediction architecture. This framework integrates an improved next-generation reservoir computation module and an XGBoost regression module, specifically designed for scenarios where some states are observable. The improved next-generation reservoir computation module contains two reservoir computation units with different time-scale awareness capabilities, which are used to capture the short-term and long-term dynamic characteristics of memristor neurons, respectively. The XGBoost regression module serves as a state estimator, used to estimate missing state variables when some states are unobservable. S12. The two reservoir computing units of the improved next-generation reservoir computing module according to claim S11 have the same polynomial order and input delay parameter, but use different step size parameters. One unit uses the interval parameter s1 = 1 and the other unit uses the differentiated interval parameter s2 = 2, forming complementary multi-scale feature extraction capabilities. S2: Construct an MNGRC model based on next-generation reservoir calculations and determine its key structural parameters; Its characteristic is that the MNGRC includes two parallel and structurally complementary reservoir computing units, namely NGRC A and NGRC B; Each reservoir unit converts the input signal into a high-dimensional feature vector through nonlinear feature mapping, generating different feature vectors O. A and O B Subsequently, these two feature vectors are concatenated into a comprehensive fused feature representation O. fin This indicates that dual time information is integrated for subsequent processing; During the training phase, the output matrix W is obtained by applying ridge regression to the collected high-dimensional feature vectors. out The target is the corresponding system output; IN out =U'·O T ·(O·O T +γ I) -1 , (2) Where U' and O represent u(t) and O respectively. fin The matrix form; γ > 0 is the L2 regularization coefficient; I is the identity matrix; The nonlinear feature mapping is implemented based on polynomial expansion; the two reservoir computing units are configured to capture the short-term and long-term dynamic characteristics of memristor neurons, respectively; the reservoir output is calculated by linearly combining the reservoir state and readout weights. y(t)=W out ·r(t) (3) Where r(t) represents the reservoir state vector, W out This indicates that the weight matrix has been read out. S3: Configure a dual-path data flow mechanism for a hybrid machine learning framework, characterized by including: S31, Direct Path: When all state variables are measurable, the MNGRC module directly processes the joint input vector [V]. in Make predictions using [x]. S32. Estimation-Prediction Path: When some states are unmeasurable, the XGBoost module first estimates the missing states based on measurable variables, and then sends the complete state information to the MNGRC module for prediction. S33. The Bayesian optimization algorithm is used to globally optimize the key parameters of the hybrid framework. The optimization objective is to minimize the weighted combination of Lyapunov time error and RMSE of the predicted trajectory. S4: Using partially observable state variables of third-order and above memristor neurons with input stimuli as frame inputs, and training neuromorphic dynamics behavior through the dual-path prediction architecture, characterized in that: S41. The training process strictly follows the principle of "predicting the future from the past" to segment time series data, and the training set is strictly placed before the time series of the validation set to avoid information leakage. A systematic reset protocol is implemented before each prediction to reset the state of the reserve pool to the initial value, and the sensitivity of the chaotic system to the initial conditions is eliminated through the warm-up phase. S42. The training data covers multiple dynamic spectrums of the memristor neuron system, including the full spectrum of dynamic behaviors from steady resting state to self-sustaining oscillation and chaotic oscillation; sample points at Hopf bifurcation boundaries are specifically included to ensure that the model can capture critical transition behaviors. S43. For input stimulus conditions with different duty cycles, an adaptive training strategy is adopted. A basic dynamic model is established using training data with only a single duty cycle. The optimal parameter configuration is determined by Bayesian optimization algorithm to achieve generalized prediction of neuromorphic behavior under multiple duty cycle conditions. S44. Using three representative neuromorphic behaviors as training data, the hybrid machine learning framework is used to achieve unified modeling of all 18 neuromorphic behaviors, including complex multi-peak burst patterns. S5: Prediction is performed using a hybrid machine learning framework that has been trained, characterized by: S51. Based on the target neuromorphic response, determine the corresponding input stimulus pattern, and iteratively adjust the input signal parameters through optimization algorithms to minimize the difference between the frame prediction output and the target neuromorphic response. S52. During the determination of the stimulus pattern, when fluctuations occur at the signal transition edge, signal smoothing techniques, including moving average filtering or Kalman filtering, are used to improve the quality of the generated stimulus signal. S53. The signal smoothing process is performed while maintaining key dynamic characteristics, and the effectiveness of the smoothing process is verified by preserving the similarity of the phase space topology. S55. A unified model is performed on the input stimulus patterns with different duty cycles, so that the framework can generalize to neuromorphic behavior prediction outside the training data range. Through the above steps, a method for more accurate prediction of various neuromorphic dynamic behaviors of memristor neurons based on a hybrid machine learning framework is presented.

Images