Pulse neural network modeling method based on synaptic heterogeneity and application thereof

Abstract

The invention provides a pulse neural network modeling method based on synaptic heterogeneity and application thereof. The method comprises the following steps: S1, constructing a synaptic heterogeneity neuron model with generalization ability; s2, integrating the synaptic heterogeneity neuron model into a pulse neural network, and constructing a network structure with a synaptic heterogeneity mechanism; s3, training and optimizing the constructed pulse neural network based on task data; according to the method, the key characteristic of synaptic heterogeneity in biological neurons is introduced, so that the time sequence information processing capability of the spiking neural network under multiple time scales is effectively improved while the modeling universality is kept.

Description

Technical Field

[0001] The present invention relates to the fields of brain-like computing and spiking neural networks, and in particular to a spiking neural network modeling method based on synaptic heterogeneity and its application. Background Art

[0002] Spiking Neural Networks (SNNs) are a new type of neural network model inspired by biological neural systems. They feature event-driven, sparse activation, and natural temporal dynamic modeling capabilities. They are widely considered a key approach to achieving low-power, efficient neural computing. With the development of neuromorphic chips and the growing demand for sequential task processing, SNNs are showing broad application prospects in areas such as image recognition, speech processing, and robotic control.

[0003] To further enhance the temporal modeling capabilities of SNNs, previous studies have attempted to introduce various neuronal dynamic mechanisms, such as adjusting membrane time constants and thresholds, to improve their responsiveness to complex temporal structures. However, these improvements have mostly focused on the neuronal level and have not fully considered the diversity of synaptic connections in terms of temporal dynamics.

[0004] Numerous neuroscience studies have shown that synaptic time constants vary significantly across brain regions and cell types, demonstrating a high degree of "synaptic heterogeneity." This heterogeneity is considered a key foundation for the neural system to integrate and process information across multiple timescales, and is particularly important in complex cognitive functions. However, most current SNN models still generally adopt a unified synaptic time constant, ignoring this important biological characteristic, which limits their performance in multi-timescale tasks. Therefore, how to construct neural models that can reflect the diversity of synaptic temporal dynamics and improve the ability to model complex temporal structures has become a key issue that needs to be addressed. Summary of the Invention

[0005] To address the challenges of existing technologies, this paper proposes a spiking neural network modeling method based on synaptic heterogeneity. By introducing a heterogeneous mechanism for synaptic time constants, this method can flexibly simulate the differences in synaptic dynamics across different brain regions and neuron types, significantly enhancing the modeling and expressive capabilities of SNNs for diverse temporal tasks.

[0006] The synaptic heterogeneity modeling framework proposed in this invention has the characteristics of simple structure and strong adaptability. It can be seamlessly embedded in the existing SNN structure, effectively improving the biological rationality and engineering feasibility of its temporal modeling.

[0007] In order to solve the problems of the prior art, the present invention adopts the following technical solution: a pulse neural network modeling method based on synaptic heterogeneity, comprising:

[0008] S1: Constructing a synaptic heterogeneity neuron model with generalization capability; the membrane potential of the synaptic heterogeneity neuron model is determined by the difference between the sum of the synaptic current and the reset current;

[0009] S2: Integrate the synaptic heterogeneity neuron model into a spiking neural network to construct a network structure with a synaptic heterogeneity mechanism;

[0010] S3: Train and optimize the constructed spiking neural network based on task data.

[0011] Furthermore, the synaptic heterogeneity neuron model consists of a synaptic module and a cell body module; wherein:

[0012] The synaptic module is used to correspond each synaptic module to a presynaptic neuron, and receive the input pulse from the neuron at the current time t Through the synaptic weight w i weighted and compared with the synaptic current of the previous moment According to the attenuation coefficient r i Perform weighted accumulation to update the current synaptic current All synaptic currents will be funneled into the cell bodies of downstream neurons;

[0013] The cell body module is used for the current output by all synaptic modules The total input current is formed by summing; at the same time, the neuron is also subjected to a reset current J t The influence of the current simulates the membrane potential reset process after the pulse is released; after combining the two, the membrane potential V of the neuron at the current time t is calculated. t If the membrane potential exceeds the set threshold θ, the neuron emits a pulse and updates J at the same time. t ;in:

[0014] The synaptic heterogeneity neuron model is obtained according to the following formula:

[0015]

[0016] Among them, V t is the membrane potential at time t, is the synaptic current from the i-th afferent neuron at time t, with a time constant τ specific to the synapse. s，i Attenuation, J t Represents the reset current at time t, which is a neuron-level value used to simulate the reset effect triggered by a pulse, and is expressed with a time constant τ J Attenuation, w i represents the weight of the i-th synapse, θ represents the threshold, and denote the time of the jth input pulse of the i-th presynaptic neuron and the jth output pulse of the postsynaptic neuron, respectively, and δ(·) denotes the Dirac function;

[0017] Furthermore, the discrete-time form of the synaptic current is as follows:

[0018]

[0019] in, represents the synaptic current decay coefficient from the i-th afferent neuron, Δt represents the time step, represents the input pulse from the i-th afferent neuron at time t;

[0020] The discrete-time form of the reset current is as follows:

[0021] J t =κ·J t-1 +θ·z t-1

[0022] in, represents the attenuation coefficient of the neuronal reset current, z t-1 represents the output pulse of the neuron at time t-1.

[0023] Furthermore, based on the synaptic heterogeneity neuron model, all presynaptic neurons i and postsynaptic neurons j satisfy r ji =κ j =ρ, the generalization model of neurons with first synaptic heterogeneity can be established according to the following formula:

[0024]

[0025] in, and are the membrane potential values ​​of postsynaptic neuron j at time t-1 and time t, ρ is the membrane potential attenuation coefficient, and w ji is the synaptic weight between presynaptic neuron i and postsynaptic neuron j, is the input pulse of presynaptic neuron i at time t, is the output pulse of postsynaptic neuron j at time t-1, and θ is the pulse emission threshold.

[0026] The spiking neural network modeling method based on synaptic heterogeneity according to claim 2 is characterized in that, based on the first synaptic heterogeneity neuron generalization model, when the reset current of the neuron model is further decomposed into a standard component and an additional adaptive current triggered by the pulse, a second synaptic heterogeneity neuron generalization model can be established according to the following formula:

[0027]

[0028] in, is the additional adaptive current at time t.

[0029] Furthermore, based on the synaptic heterogeneity neuron model, all presynaptic neurons i and postsynaptic neurons j satisfy r ji =κ j =ρ j When , the third synaptic heterogeneity neuron generalization model can be established according to the following formula:

[0030]

[0031] Among them, ρ j The membrane potential decay coefficient is independent for each postsynaptic neuron.

[0032] The present invention can also be implemented by the following technical solutions:

[0033] A delay matching task processing system, wherein the system adopts any one of the models of claims 1-5 to construct a spiking neural network, wherein: the input layer includes left / right cue channel and noise channel neurons, and the output layer includes matching decision neurons.

[0034] Beneficial effects

[0035] 1. The synaptic heterogeneity neuron model proposed in this paper effectively enhances the spiking neural network's ability to model multi-scale temporal dynamic information by introducing synaptic-specific time constants, significantly improving the network's overall performance and accuracy in a variety of time series tasks. Furthermore, the model exhibits excellent configurability and compatibility, and can be flexibly generalized to other common neuron models, such as standard LIF neurons, through parameter settings. This facilitates integration into various spiking neural network frameworks, supports rapid adaptation to a variety of network structures and learning algorithms, and possesses high engineering application value.

[0036] 2. The model has good noise robustness and working memory retention capabilities. It can maintain output stability in high-noise or frequently disturbed environments, effectively identify and retain task-related information, and has good task generalization capabilities.

[0037] 3. Under the condition of limited neuronal resources, the model still maintains efficient operation, has strong parameter utilization efficiency and resource adaptability, reduces dependence on large-scale network structures, and is suitable for deployment on edge computing devices, mobile terminals or other resource-constrained platforms.

[0038] 4. The core of the present invention is to construct a neuron model in which synaptic currents have independent time constants. Combined with the separation modeling mechanism of reset current, it effectively improves the expressiveness and stability of neurons in the dynamic response process. The model can be generalized to standard LIF, threshold adaptation and neuron heterogeneous models. In the delay matching task, the constructed network maintains an accuracy rate of >90% in a long delay (3000ms) and high noise (10Hz) environment, showing excellent robustness and memory retention ability, and is suitable for accelerated processing of cognitive tasks on neuromorphic computing chips. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The process of the pulse neural network modeling method based on synaptic heterogeneity of the present invention;

[0040] Figure 2 This is a schematic diagram of the structure of the synaptic heterogeneity neuron model proposed in the present invention;

[0041] Figure 3 Graph showing experimental results in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following combination Figure 1 and Figure 2 The present invention is described in detail as follows:

[0043] like Figure 1 As shown, the present invention provides a pulse neural network modeling method based on synaptic heterogeneity, comprising:

[0044] S1: Construct a synaptic heterogeneity neuron model with generalization ability; the synaptic heterogeneity neuron model is established according to the following formula:

[0045]

[0046] Among them, V t is the membrane potential at time t, is the synaptic current from the i-th afferent neuron at time t, with a time constant τ specific to the synapse. s，i Attenuation, J t Represents the reset current at time t, which is a neuron-level value used to simulate the reset effect triggered by a pulse, and is expressed with a time constant τ J Attenuation, w i represents the weight of the i-th synapse, θ represents the threshold, and denote the time of the jth input pulse of the i-th presynaptic neuron and the jth output pulse of the postsynaptic neuron, respectively, and δ(·) denotes the Dirac function.

[0047] Furthermore, the discrete-time form of the synaptic current is as follows:

[0048]

[0049] in, represents the synaptic current decay coefficient from the i-th afferent neuron, Δt represents the time step, represents the input pulse from the i-th afferent neuron at time t.

[0050] The discrete-time form of the reset current is as follows:

[0051] J t =κ·J t-1 +θ·z t-1

[0052] in, represents the attenuation coefficient of the neuronal reset current, z t-1 represents the output pulse of the neuron at time t-1.

[0053] The structural diagram of the synaptic heterogeneity neuron model is shown in FIG. Figure 2 As shown in the figure, it mainly includes several synaptic modules and a neuron cell body module. The modules work together to realize the dynamic integration of input pulses and pulse output mechanism. The specific process is as follows:

[0054] (1) Synaptic modules (multiple parallel structures above): Each synaptic module corresponds to a presynaptic neuron and receives input pulses from the neuron at the current time t Through the synaptic weight w i weighted and compared with the synaptic current of the previous moment According to the attenuation coefficient r i Perform weighted accumulation to update the current synaptic current All synaptic currents will be transported to the cell bodies of downstream neurons.

[0055] (2) Cell body module (lower half of the figure): the current output by all synaptic modules The total input current is formed by summing; at the same time, the neuron is also subjected to a reset current J t The influence of the current simulates the membrane potential reset process after the pulse is released; after combining the two, the membrane potential V of the neuron at the current time t is calculated. t If the membrane potential exceeds the set threshold θ, the neuron emits a pulse and updates J at the same time. t .

[0056] S2: Integrate the synaptic heterogeneity neuron model into a spiking neural network to construct a network structure with a synaptic heterogeneity mechanism;

[0057] S3: Train and optimize the constructed spiking neural network based on specific tasks.

[0058] Example 1: The following describes a specific application of the present invention using a delay matching task as an example, which mainly includes the following steps:

[0059] 101: Prepare the dataset and divide it into training set and test set;

[0060] 102: Construct a neuronal model with synaptic heterogeneity, including:

[0061] The synaptic heterogeneity neuron model was established according to the following formula:

[0062]

[0063] Among them, V t is the membrane potential at time t, is the synaptic current from the i-th afferent neuron at time t, with a time constant τ specific to the synapse. s，i Attenuation, J t Represents the reset current at time t, which is a neuron-level value used to simulate the reset effect triggered by a pulse, and is expressed with a time constant τ J Attenuation, w i represents the weight of the i-th synapse, θ represents the threshold, and denote the time of the jth input pulse of the i-th presynaptic neuron and the jth output pulse of the postsynaptic neuron, respectively, and δ(·) denotes the Dirac function.

[0064] Furthermore, the discrete-time form of the synaptic current is as follows:

[0065]

[0066] in, represents the synaptic current decay coefficient from the i-th afferent neuron, Δt represents the time step, represents the input pulse from the i-th afferent neuron at time t.

[0067] The discrete-time form of the reset current is as follows:

[0068] J t =κ·J t-1 +θ·z t-1

[0069] in, represents the attenuation coefficient of the neuronal reset current, z t-1 represents the output pulse of the neuron at time t-1.

[0070] In addition, through parameter configuration, the synaptic heterogeneity neuron model can be generalized into the following three common neuron models:

[0071] (1) The first synaptic heterogeneity neuron generalization model is the standard LIF model: when all presynaptic neurons i and postsynaptic neurons j satisfy r ji =κ j =ρ, the synaptic heterogeneity neuron model can be transformed into a standard LIF model, and its dynamic equation is:

[0072]

[0073] (2) The second generalized model of synaptic heterogeneity neurons: a neuron model with threshold adaptability: Based on case (1), the reset current of the synaptic heterogeneity neuron model is further decomposed into a standard component and an additional adaptive current triggered by a pulse, so that the model is transformed into a neuron model with threshold adaptability, and its dynamic equation is:

[0074]

[0075] in is the additional adaptive current at time t.

[0076] (3) The third synaptic heterogeneity neuron generalization model is the neuron hierarchical heterogeneity model: when all presynaptic neurons i and postsynaptic neurons j satisfy r ji =κ j =ρ j When , the model realizes membrane time constant heterogeneity, that is, each neuron has an independent time decay factor ρ j , forming a neuron model with heterogeneity at the neuron level, whose dynamic equation is:

[0077]

[0078] 103: Based on the proposed synaptic heterogeneity neuron model and its three generalizable common neuron models, pulse neural network structures are constructed. Each network consists of an input layer, a hidden layer, and an output layer. The input layer has 30 neurons, corresponding to the left prompt channel, the right prompt channel, and the noise channel, with 10 neurons for each channel; the output layer contains two neurons, representing the classification results of "match" and "mismatch" respectively;

[0079] 104: Using a back propagation through time (BPTT) algorithm to train the spiking neural network and optimize network parameters;

[0080] 105: Use the test set to evaluate the performance of the trained network. The specific results are as follows: Figure 3 As shown: Figure 3 (AB) show the comparison results of the classification accuracy of the synaptic heterogeneity neuron model proposed in the present invention and three types of generalized neuron models in two network architectures under different settings of the number of hidden layer neurons. The compared models include: feedforward and recurrent networks constructed by synaptic heterogeneity neuron models (HetSynLifFSNN, HetSynLifRSNN), feedforward and recurrent networks constructed by standard LIF neurons (HomNeuLifFSNN, HomNeuLifRSNN), feedforward and recurrent networks constructed by threshold adaptive neurons (HomNeuALifFSNN, HomNeuALifRSNN), and feedforward and recurrent networks constructed by neuron hierarchical heterogeneity models (HetNeuLifFSNN, HetNeuLifRSNN). The experimental results show that among all network structures, the synaptic heterogeneity neuron model proposed in the present invention achieved the highest accuracy, further verifying its effectiveness in performance improvement.

[0081] Figure 3 Figure 3 (CD) further demonstrates the performance of the HetSynLifRSNN model under varying delays and noise intensities. The data in the figure shows the model's classification accuracy under each condition, as well as the number of training rounds required to converge to that accuracy. The results show that the model maintains high accuracy even in long-delay and high-noise environments, while maintaining no significant decrease in training efficiency, demonstrating excellent robustness and learning stability.

Claims

1. A pulse neural network modeling method based on synaptic heterogeneity, characterized in that: include: S1: Constructing a neuronal model with synaptic heterogeneity and generalization capability; The membrane potential of the synaptic heterogeneity neuron model is determined by the difference between the sum of the synaptic currents and the reset current; S2: Integrate the synaptic heterogeneity neuron model into a spiking neural network to construct a network structure with a synaptic heterogeneity mechanism; S3: Train and optimize the constructed spiking neural network based on task data.

2. The spiking neural network modeling method based on synaptic heterogeneity according to claim 1, characterized in that: The synaptic heterogeneity neuron model consists of a synapse module and a cell body module; wherein: The synaptic module is used to correspond each synaptic module to a presynaptic neuron, and receive the input pulse from the neuron at the current time t Through the synaptic weight w i weighted and compared with the synaptic current of the previous moment According to the attenuation coefficient r i Perform weighted accumulation to update the current synaptic current All synaptic currents will be funneled into the cell bodies of downstream neurons; The cell body module is used for the current output by all synaptic modules The total input current is formed by summing; at the same time, the neuron is also subjected to a reset current J t The influence of the current simulates the membrane potential reset process after the pulse is released; after combining the two, the membrane potential V of the neuron at the current time t is calculated. t If the membrane potential exceeds the set threshold θ, the neuron emits a pulse and updates J at the same time. t ;in: The synaptic heterogeneity neuron model is obtained according to the following formula: Among them, V t is the membrane potential at time t, is the synaptic current from the i-th afferent neuron at time t, with a time constant τ specific to the synapse. s,i Attenuation, J t Represents the reset current at time t, which is a neuron-level value used to simulate the reset effect triggered by a pulse, and is expressed with a time constant τ J Attenuation, w i represents the weight of the i-th synapse, θ represents the threshold, and denote the time of the jth input pulse of the i-th presynaptic neuron and the jth output pulse of the postsynaptic neuron, respectively, and δ(·) denotes the Dirac function; Furthermore, the discrete-time form of the synaptic current is as follows: in, represents the synaptic current decay coefficient from the i-th afferent neuron, Δt represents the time step, represents the input pulse from the i-th afferent neuron at time t; The discrete-time form of the reset current is as follows: J t =k·J t-1 +θ·z t-1 in, represents the attenuation coefficient of the neuronal reset current, z t-1 represents the output pulse of the neuron at time t-1.

3. The spiking neural network modeling method based on synaptic heterogeneity according to claim 2, characterized in that: Based on the synaptic heterogeneity neuron model, all presynaptic neurons i and postsynaptic neurons j satisfy r ji =κ j =ρ, the generalization model of neurons with first synaptic heterogeneity can be established according to the following formula: in, and are the membrane potential values ​​of postsynaptic neuron j at time t-1 and time t, ρ is the membrane potential attenuation coefficient, and w ji is the synaptic weight between presynaptic neuron i and postsynaptic neuron j, is the input pulse of presynaptic neuron i at time t, is the output pulse of postsynaptic neuron j at time t-1, and θ is the pulse emission threshold.

4. The spiking neural network modeling method based on synaptic heterogeneity according to claim 2, characterized in that: Based on the first synaptic heterogeneity neuron generalization model, when the reset current of the neuron model is further decomposed into a standard component and an additional adaptive current triggered by a pulse, the second synaptic heterogeneity neuron generalization model can be established according to the following formula: in, is the additional adaptive current at time t.

5. The spiking neural network modeling method based on synaptic heterogeneity according to claim 2, characterized in that: Based on the synaptic heterogeneity neuron model, all presynaptic neurons i and postsynaptic neurons j satisfy r ji =κ j =ρ j When , the third synaptic heterogeneity neuron generalization model can be established according to the following formula: Among them, ρ j The membrane potential decay coefficient is independent for each postsynaptic neuron.

6. A delay matching task processing system, characterized by: The system adopts any one of the models of claims 1-5 to construct a spiking neural network, wherein: the input layer includes left / right cue channel and noise channel neurons, and the output layer includes matching decision neurons.

Images