Method for constructing adaptive spiking neuron model and network thereof

Abstract

The invention provides a method for constructing an adaptive spiking neuron model and a network thereof. The method comprises the following steps: S1, constructing an adaptive dynamic spiking neuron model; s2, constructing an initial adaptive dynamic spiking neural network by using the adaptive dynamic spiking neuron model; and S3, training the initial adaptive dynamic spiking neural network through a back propagation method to obtain an optimized adaptive dynamic spiking neural network. According to the method, the energy efficiency of the spiking neural network is remarkably improved, and meanwhile, the learning ability and robustness of the spiking neural network are enhanced.

Description

Technical Field

[0001] This invention relates to the fields of neuromorphic computing and spiking neural networks, and in particular to improving the learning ability, robustness, and energy efficiency of neurons, specifically a method for constructing an adaptive spiking neuron model and its network. Background Technology

[0002] Spiking Neural Networks (SNNs) are third-generation neural network models inspired by the information processing mechanisms of the biological brain. Unlike traditional artificial neural networks, the core feature of SNNs lies in their use of discrete spike events for communication between neurons, relying on the precise timing of these spikes to encode and process information. This event-driven characteristic makes SNNs widely regarded as a key technology for achieving energy-efficient intelligent computing in the field of neuromorphic computing, particularly suitable for applications with stringent power consumption and real-time requirements, such as edge computing, mobile devices, and neuromorphic hardware.

[0003] The computational power of SNNs is rooted in neuronal dynamics, which describes the dynamic process of the evolution of the neuron's membrane potential over time. Among numerous biophysically inspired neuron models, the Leaky Integrate-and-Fire (LIF) model is the most commonly used due to its good balance between complexity and biological plausibility. In the LIF model, the membrane time constant is a crucial parameter that controls the rate of membrane potential leakage, thereby defining the time window for the neuron to integrate the input signal. A smaller membrane time constant makes the neuron sensitive to rapidly changing inputs, resulting in a rapid response but short-lived memory; while a larger membrane time constant allows the neuron to smooth and integrate the input over a longer period, resulting in more persistent behavior but potentially slower response. Therefore, the value of the membrane time constant directly affects the network's temporal information processing capability, including its ability to distinguish signal timing, frequency, and patterns.

[0004] In existing technologies, most research and practice on SNNs employ a simplified static strategy: setting the membrane time constant as a fixed global hyperparameter, or as a globally learnable scalar shared by all neurons in the network and optimized through gradient descent. This static parameter configuration severely limits the ability of neurons to dynamically adjust their integral behavior based on the characteristics of their local input signals (such as intensity, frequency, and pattern) and their own impulse firing history. This inherent inflexibility makes it difficult for networks to achieve optimal dynamic performance when processing complex, variable, and multi-timescale real-world temporal signals, thus limiting the robustness and adaptability of SNNs in practical applications.

[0005] Neuroscience research shows that biological neurons can dynamically adjust their membrane time constant through various biophysical mechanisms. For example, after strong synaptic input or high-frequency pulse firing, the effective time constant of the membrane potential shortens due to increased conductance, thus enabling faster reset and response to subsequent high-frequency inputs. Conversely, at rest or when input is sparse, the membrane time constant is relatively prolonged, which is beneficial for long-term integration of weak signals. This multi-timescale adaptive mechanism is one of the keys to the efficient processing of complex temporal information by biological nervous systems.

[0006] To simulate this dynamic characteristic, some preliminary studies have been conducted. Early biophysical models naturally introduced input-dependent membrane time constant variations through complex differential equations. More recently, models such as liquid time constant networks and brain-inspired adaptive leaky integral firing neurons have also attempted to achieve dynamic modulation of the time constant by adding additional network layers. Although these methods each have advantages in terms of biological rationality or functional adaptability, the introduced dynamic mechanisms bring significant computational overhead and energy consumption, hindering their deployment in large-scale neuromorphic systems. Furthermore, existing methods mainly focus on short-term adaptation based on input, while neglecting the long-term plasticity of the time constant, which can be optimized globally through the task. Summary of the Invention

[0007] To address the technical challenges of existing spiking neural networks, such as the simplistic neuron dynamics, the high computational cost of dynamic adjustment mechanisms for membrane time constants, and the inability to simultaneously achieve long-term plasticity and short-term dynamic adaptation, this invention provides a method for constructing an adaptive spiking neuron model and its network. This method aims to achieve highly adaptive neuron dynamics, improve the network's ability to process multi-timescale temporal signals, enhance its learning capacity and robustness, and maintain low computational cost and high energy efficiency.

[0008] To address the problems in the existing technology, the present invention adopts the following technical solution: A method for constructing an adaptive spiking neuron network, wherein each neuron of the adaptive spiking neuron network learns dynamic characteristics to simulate the working mode of a biological brain, comprising the following steps: An adaptive spiking neuron model is constructed by dynamically changing the static spiking neural network neurons. An initial adaptive spiking neuron network is constructed by fusing the adaptive spiking neuron model with the spiking neuron network. An adaptive spiking neuron network is constructed by training an initial adaptive spiking neuron network using the backpropagation method.

[0009] Furthermore, the process of constructing an adaptive spiking neuron model by dynamically changing static spiking neural network neurons includes: By introducing dynamic parameters Obtain dynamic input information from spiking neurons:

[0010] in: Parameters that continue to affect input; The current input continues to have an impact on the parameters;

[0011] ; By introducing dynamic parameters Obtain dynamic output information of spiking neurons:

[0012] in: Parameters that continue to affect output; The parameters that are continuously affected by the current output; For the neuron Output information for each time step; Based on the dynamic input information of spiking neurons Dynamic output information of spiking neurons and the introduction of dynamic parameters Combined with learnable benchmark parameters Constructing the dynamic time constant of an adaptive spiking neuron:

[0013] Membrane attenuation coefficient of dynamic spiking neurons Constructing adaptive spiking neuron membrane potentials:

[0014] .

[0015] in: Represents the simulation time step. The pulse release threshold represents the neuron.

[0016] Furthermore, the process of training the initial adaptive spiking neuron network using the backpropagation method to construct the adaptive spiking neuron network includes: The initial output of the adaptive spiking neural network is determined to be... ; The classification result is calculated based on the initial output of the adaptive spiking neural network:

[0017] Calculate the loss function based on the classification results:

[0018] in: For the true classification results, : is the cross-entropy loss function; According to the loss function Weights for each layer W and the time dynamic constant of the reference film An adaptive spiking neural network is obtained. .

[0019] The present invention also employs the following technical methods: An adaptive spiking neuron model, wherein the adaptive spiking neuron model is as follows:

[0020]

[0021]

[0022]

[0023] .

[0024] Furthermore, the neuron release pulse in the adaptive spiking neuron model is determined by judging the membrane potential. Does it exceed the threshold? Value release, that is: .

[0025] Beneficial effects 1. This invention significantly reduces the computational complexity and energy consumption of the model by introducing a trainable basic membrane time constant and a lightweight dynamic modulation mechanism to spiking neurons. Specifically, this adaptive mechanism achieves dynamic modulation through only a linear transformation during the forward propagation process, avoiding complex additional computational modules and maintaining the inherent event-driven and sparse computational characteristics of spiking neural networks. Furthermore, this lightweight design allows neurons to achieve highly dynamic adaptation while effectively reducing the pulse firing rate, making it particularly advantageous for efficient deployment on large-scale neuromorphic chips or resource-constrained edge devices, fundamentally improving energy efficiency at the model level.

[0026] 2. This invention significantly enhances the ability of spiking neural networks to handle long-term dependencies, adapt to signals at multiple time scales, and resist complex noise and interference by combining long-term temporal plasticity with short-term dynamic modulation. The trainable basic time constant allows neurons to stably learn and adapt to the global temporal characteristics of the task, while the dynamic modulation driven by the input and pulses enables real-time fine-tuning of the integral window, thereby intelligently controlling the balance between information retention and forgetting. This mechanism allows the network to flexibly switch between various dynamic modes such as "fast response" and "persistent memory" based on the statistical characteristics of the input signal, effectively filtering noise interference, retaining key temporal information, and comprehensively improving the network's robustness and generalization performance when facing complex dynamic signals in the real world. Attached Figure Description

[0027] Figure 1 This is a flowchart of an embodiment of an adaptive dynamic spiking neuron model and its network training in this invention; Figure 2 This is a computational graph of the membrane potential update process of an adaptive dynamic neuron; Figure 3 This demonstrates that the model proposed in this invention can adapt to changes in the membrane time constant and maintain the highest accuracy. Among them, AND-LIF is the model proposed in this invention, and the others are neuron models in spiking neural networks after the adaptive dynamics are removed. Figure 4 This indicates that the model proposed in this invention can still maintain the highest accuracy under different noise types and intensities. Among them, AND-LIF is the model proposed in this invention, and the others are standard LIF neurons and their neuron models with increased dynamic membrane time constant mechanism and learnable membrane time constant. Detailed Implementation

[0028] The following is in conjunction with the appendix Figure 1 -Appendix Figure 4 The present invention will be described in detail below.

[0029] This invention aims to provide an adaptive dynamic spiking neuron model and a method for constructing and training its network. The system is based on a spiking neural network and includes: S1: Construct an adaptive dynamic spiking neuron model; S2: Constructing an adaptive dynamic spiking neural network by fusing the adaptive dynamic spiking neuron model with a spiking neural network; including: Configure a trainable base membrane time constant for each neuron in a spiking neural network. Dynamic parameters driven by the neuron's own input and output activities and and adaptive membrane time constant :

[0030]

[0031]

[0032] in: Controlling the persistence of the effects of previous inputs and outputs, Controlling the effects of current inputs and outputs For the neuron Information transmitted through the synapse at each time step, and For the neuron Output information for each time step.

[0033] Based on this, the membrane attenuation coefficient of the adaptive spiking neuron and its membrane potential Adaptive update dynamics can be represented as:

[0034]

[0035] in: Represents the simulation time step. The pulse release threshold represents the neuron.

[0036] When membrane potential Neurons release impulses after the threshold is exceeded. :

[0037] S3: The initial adaptive spiking neural network is trained using the backpropagation method to obtain an optimized adaptive spiking neural network, including: The initial output of the adaptive spiking neural network is determined to be... ; The classification result is calculated based on the initial output of the adaptive spiking neural network:

[0038] Calculate the loss based on the classification results:

[0039] in: For the true classification results, : is the cross-entropy loss function; According to the loss Weights for each layer W and reference membrane time constant An optimized adaptive spiking neural network was obtained, and detailed computational graphs can be found here. Figure 2 .

[0040] The technical solution of this invention is a complex neuron update computation with time dynamics and error feedback. This invention is a dynamic system with inherent state and temporal processing capabilities. Its core feature is that the output of a neuron depends not only on the current input but also heavily on its own historical state (such as membrane potential and refractory period). This model is closer to simulating the workings of a biological brain and is a core model in the field of brain-inspired computing.

[0041] Figure 2 The left side of the diagram details the complete lifecycle of a neuron within a single time step, depicting the dynamic process of a single neuron. This is an enhanced version of the classic Leaky Integrate-and-Fire (LIF) model. U (Initial Membrane Potential): The membrane potential from the neuron after leakage in the previous time step. H (Post-Charging Membrane Potential): The membrane potential of the neuron after the charging process. S (Output Spike): A binary signal (0 or 1) generated when the neuron "fires". τ (Refractory period time constant): A parameter that controls the recovery rate of neurons after firing.

[0042] Workflow (closed loop): 1. Charge: The synaptic input pulse S charges the initial membrane potential U, increasing its potential to generate H.

[0043] 2. Fire: When the membrane potential H exceeds a certain set firing threshold, the neuron generates an output pulse S=1.

[0044] 3. Reset: After the pulse is delivered, the membrane potential H is immediately reset.

[0045] 4. Leakage: During charging and resetting, the membrane potential H will change in a manner determined by the time constant τ. The controlled rate of "leakage" or decay gradually returns to the resting potential. This simulates the ion channel leakage effect of biological neurons.

[0046] Advanced Feature: A particularly crucial point in the error-driven plasticity graph is that there is an Error signal fed back, pointing to τ. This indicates that the model involves not only forward computation but also a learning mechanism. The refractory period constant is not a fixed value, but a learnable parameter that can be adjusted based on the error in the network output. This means that the model can learn to adjust the dynamic characteristics of each neuron (e.g., make it react faster or slower) to optimize overall performance.

[0047] Figure 2 The right side deepens and abstracts the concept of "time constant" in the left-side model, demonstrating a more general mechanism for generating dynamic time constants. The behavior of neurons (especially the leakage rate) should not be solely regulated by a fixed or error-adjusted τ. It should not be controlled, but rather dynamically determined by its inputs and outputs.

[0048] Workflow: 1. Input & Output Influence: The input and output signals of a neuron are respectively converted into corresponding time constants. and This indicates that the pattern and intensity of the input, as well as the neuron's own activity history, affect its subsequent "reaction speed".

[0049] 2. Integration: and An integrated addition / subtraction module (+ / -) is used to generate a comprehensive, time-varying effective time constant τ(t).

[0050] 3. Dynamic Leak: This integrated τ(t) directly controls the leakage process of the membrane potential H. Therefore, the integral behavior of neurons is highly dynamic and adaptive.

[0051] 4. Backpropagation: The diagram clearly shows a "backpropagation" path from the output S to τ(t). The entire time constant generation mechanism, i.e., how to calculate τ(t) from the input and output, is differentiable and can be trained end-to-end using gradient descent methods similar to the backpropagation algorithm.

[0052] like Figures 3-4 As shown, the embodiment demonstrates the use of this invention to implement a dynamic image classification task, mainly including the following steps: 101: Prepare the dataset by dividing it into a training set and a test set; 102: Constructing an adaptive dynamic neuron mathematical model:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] The definitions of all symbols used have been explained above.

[0059] 103: Construct a spiking neural network based on adaptive dynamic neurons. Using the mathematical model of the above neurons, construct a specific network structure for a specific task.

[0060] 104: Initialize the spiking neural network.

[0061] 105: Training a spiking neural network; the trainable parameters in the mathematical model mainly include the weights of each layer. W and reference membrane time constant .

[0062] 106: Test the performance of the spiking neural network on the test set. After training, test it directly using the original partitioned test set.

[0063] 107: Test the performance of the spiking neural network with added noise interference on the test set. After training, add various noises to the original test set and test the model's performance under different interference intensities.

[0064] This invention adds two types of noise: Gaussian noise: A noise value with zero mean and adjustable variance superimposed on the input analog current or membrane potential.

[0065] Salt and pepper noise: In the input pulse sequence, the pulses are randomly set to either a continuous firing state or a continuous silent state with a specific probability.

[0066] This invention introduces a dynamic time constant dependent on input / output, more precisely simulating the complex adaptive behavior of biological neurons. The powerful temporal information processing capability of this invention stems from the fact that, because the internal state (membrane potential) changes continuously over time, this type of model is inherently adept at processing temporal signals such as speech, video, and dynamic sensor data, capturing the temporal dependencies within the data. Simultaneously, the core idea of ​​this invention's spiking neural network is sparse event-driven computation (consuming energy only during firing), which, compared to traditional always-activated neural networks, exhibits extremely high energy efficiency when implemented on dedicated hardware (neuromorphic chips, such as Intel's Loihi and Tsinghua's Tianji chip). The advanced learning paradigm of this invention, based on a model framework, integrates error signals with the intrinsic dynamic parameters of neurons (such as τ). This invention connects deep learning with the energy-efficient and dynamic advantages of SNNs, enabling end-to-end training via backpropagation. This represents a cutting-edge direction in SNN research, combining the efficient training of deep learning with the energy-saving and dynamic advantages of SNNs. This invention is a computational abstraction of neuroscience theory and a core blueprint for next-generation high-performance, intelligent temporal information processing systems, with enormous application potential in edge computing, robot control, and real-time speech recognition.

Claims

1. A method of constructing a self-adapting spiking neuron network, characterized in that, Each neuron of the adaptive spiking neuron network simulates the working mode of the biological brain by learning, including the following steps: The static spiking neuron network neuron is changed by dynamic parameters to construct an adaptive spiking neuron model; The adaptive spiking neuron model is fused with the spiking neuron network to construct an initial adaptive spiking neuron network; The initial adaptive spiking neuron network is trained by a back propagation method to construct an adaptive spiking neuron network.

2. The method of claim 1, wherein the adaptive spiking neuron model is constructed by changing the static spiking neuron network neuron by dynamic parameters, and the process includes:

3. The method of claim 1, wherein the adaptive spiking neuron network is constructed by training the initial adaptive spiking neuron network by the back propagation method, and the process includes: The classification result is calculated according to the output of the initial adaptive spiking neuron network: By introducing dynamic parameters Obtaining dynamic input information of spiking neurons: ； wherein: is the input previous persistent influence parameter; is the current input persistent influence parameter; is the neuron's tthtime non-synaptic afferent information; By introducing dynamic parameters Obtaining pulse neuron dynamic output information: ； wherein: is an output preceding influence parameter; is a current output influence parameter; is output information of the neuron at the time step; According to the dynamic input information of the pulse neuron With the dynamic output information of the pulse neuron And introducing dynamic parameters . Combined with the learnable reference parameters Constructing an adaptive pulse neuron dynamic time constant: ； Membrane decay coefficient of a dynamic spiking neuron . Constructing adaptive spiking neuron membrane potential: ； ； wherein: represents an analog time step, represents a pulse release threshold of a neuron. The loss function is calculated according to the classification result: The adaptive spiking neuron model is implemented by the method of claim 2, and the adaptive spiking neuron model is: determining an output of the initial adaptive spiking neural network as ; ​ ； ​ ； wherein: is the true classification result, is a cross-entropy loss function; According to the loss function The weights of each layer W And the reference film time dynamic constant Get an adaptive pulse neural network, .

4. An adaptive spiking neuron model, characterized in that, ​ ； ； ； ； 。 5. The self-adapting spiking neuron model according to claim 4, wherein, The neuron of the adaptive spiking neuron model releases a spike by judging whether the membrane potential exceeds a threshold value, i.e.: 。

Images