An image classification method based on mesoscale adjustment pulse neural network

Abstract

The present application relates to a kind of image classification method based on mesoscale adjustment pulse neural network, belong to image classification technical field.The method includes: according to LIF neuron model, build target pulse neural network model, and obtain the pulse firing rate of each neuron in target pulse neural network model by pre-training, based on pulse firing rate assessment neuron excitation degree, neuron is divided core interval, steady interval and silent interval and different routing mask is generated respectively in gradient in the phase of back propagation, form the mesoscale regulation mechanism to pulse neuron connection and update gradient and neural network gradient, the neuron connection of silent interval is implemented pruning, obtain the updated target pulse neural network model, for image classification.It is aimed at solving the lack of mesoscale regulation mechanism in prior art, lead to the extraction ability of image feature is limited, and the problem of computational redundancy due to a large number of invalid pulses of irrelevant neuron.

Description

Technical Field

[0001] This invention relates to an image classification method based on a mesoscale-tuned spiking neural network, belonging to the field of image classification technology. Background Technology

[0002] Spiking neural networks (SNNs), due to their unique event-driven characteristics, offer advantages such as high processing speed and extremely low energy consumption, demonstrating enormous application potential in areas such as low-power image classification and edge vision computing. However, when processing image data with complex spatiotemporal features, mainstream SNN models primarily rely on global loss functions such as alternative gradients for macroscopic training. This "black box" optimization approach severely neglects the crucial local and mesoscopic topological plasticity in biological visual systems, making it difficult to achieve true brain-like dynamic structural evolution.

[0003] Existing image classification models based on spiking neural networks lack coordination between microscopic neuronal firing and macroscopic visual cognition, leading to limitations in visual feature extraction and high computational redundancy in image classification tasks. Therefore, a biologically based mechanism is urgently needed to control and organize neuronal groups at the mesoscopic level, forming a stable topological structure to handle complex image classification tasks.

[0004] Therefore, this invention proposes an image classification method based on mesoscale-tuned spiking neural networks. Drawing inspiration from the mesoscale memory imprinting mechanism of the biological brain, mesoscale tuning is introduced into the process of image feature extraction by the spiking neural network. By dynamically evaluating and controlling the spiking of neurons, the network is encouraged to form a stable topology highly correlated with image categories, thereby enhancing the neural network's ability to recognize and represent image features. Simultaneously, by suppressing the ineffective firing of non-critical neurons, computational redundancy in the image classification process is significantly reduced. Summary of the Invention

[0005] The purpose of this invention is to provide an image classification method based on mesoscale-scale modulated spiking neural networks, which aims to solve the problems of existing spiking neural network-based image classification methods lacking mesoscale-level modulation mechanisms, resulting in limited image feature extraction capabilities, and computational redundancy caused by a large number of irrelevant neurons' invalid pulses.

[0006] To achieve the above objectives, the technical solution of the present invention is: an image classification method based on a mesoscale-tuned spiking neural network, comprising the following steps:

[0007] Step 1: Introduce a rectangular surrogate gradient and build the target spiking neural network model based on the LIF neuron model;

[0008] Step 2: Pre-train the image classification task based on the target spiking neural network model. Obtain the pulse firing rate of each neuron in the target spiking neural network model through a preset number of epochs of pre-training. Evaluate the excitation level of the neurons based on the pulse firing rate. Divide the neurons into core intervals, stable intervals, and silent intervals according to the neuron excitation level evaluation results and a preset ratio.

[0009] Step 3: Generate different routing masks for the core interval, the stable interval, and the silent interval respectively. Apply the routing masks to the gradient during the backpropagation stage to form a mesoscopic regulation mechanism for the connection of spiking neurons. Update the gradient and the neural network gradient according to the mesoscopic regulation mechanism, and prune the neuron connections in the silent interval to obtain the updated target spiking neural network model.

[0010] Step 4: Input the image to be classified into the updated target spiking neural network model to obtain the image classification result.

[0011] Optionally, Step 1 specifically includes:

[0012] Step 1.1: Initialize the neurons of the spiking neural network according to the LIF neuron model. The neuron receives upper-layer pulse input, converts the upper-layer pulse input into synaptic current, and updates its own membrane potential. The process of accumulating membrane potential is as follows:

[0013]

[0014] In the formula, This refers to the membrane potential. This is a membrane potential attenuation factor, used to achieve potential attenuation in neural networks when there is no input. Presynaptic neuron and postsynaptic neurons Synaptic weights, For the current time step The received pulse sequence input;

[0015] When the membrane potential in a neuron exceeds a set trigger threshold The neuron is then activated and fires a discrete binary impulse to the next layer. The expression is:

[0016]

[0017] in, Let be the Heaviside step function, when At that time, the Heaviside step function Assigning output 1 to At this point, the neuron fires a pulse; otherwise... Neurons remain in a resting state;

[0018] After a neuron fires a pulse, the accumulated membrane potential is reset, and the neuron enters a refractory period. The membrane potential momentarily degrades, as expressed by:

[0019]

[0020] Step 1.2: In the backpropagation process of the LIF neuron model, a rectangular surrogate gradient is introduced to approximate the calculation of the derivative of the step function. The expression is:

[0021]

[0022] in, To effectively proxy the half-width of the gradient window;

[0023] Step 1.3: Using Step 1.1 and Step 1.2 as the LIF neuron computation mechanism, and defining the LIF neuron computation mechanism as... Operator, according to Operators form the target spiking neural network model.

[0024] Optionally, the pulse firing rate is specifically:

[0025]

[0026] In the formula, Represents the pulse firing rate. The total time step, For training batch size, For the first Layer The neuron in the processing batch For each sample, the time step The output pulse belongs to {0, 1}.

[0027] Optionally, the step of dividing neurons into core regions, stable regions, and silent regions based on the neuronal excitation level assessment results and a preset ratio specifically involves:

[0028] The firing rate of each neuron As an indicator for assessing neuronal excitability, neurons are sorted in reverse chronological order according to their excitability, and the division ratios of the core interval, the stable interval, and the silent interval are set as follows: ,and Based on the excitability of neurons and a predetermined ratio, subgroups are formed, and the total number of neurons is defined as follows: Then, they are sorted according to the degree of neuronal excitation, and those belonging to the sorting interval are... The neurons within the range are the core region, and those belonging to the sorting interval are... The range of neurons is the stable interval, and the sorting interval belongs to... The range of neurons is the silent interval.

[0029] Optionally, generating different routing masks for the core interval, the stability interval, and the silent interval specifically involves:

[0030]

[0031] In the formula, For the routing mask, , , These are the core region, the homeostasis region, and the silent region of the neuron, respectively. and These are the fundamental amplitude parameter matrices for the core interval and the fundamental amplitude parameter matrices for the stable interval, respectively. and These are the strengthening factor for the core interval and the stabilization factor for the stable interval, respectively. For the silent interval, the gradient change is cut off and the mask is set to 0.

[0032] Optionally, the mesoscopic regulatory mechanism specifically includes:

[0033] During backpropagation, the fundamental error gradients between the connections of neurons in the neural network are obtained, through the first... Presynaptic neurons in the layer and postsynaptic neurons The synaptic weight is , obtained by The original spatiotemporal gradient of time backpropagation unfolds within a time window consisting of several time steps. The expression is:

[0034]

[0035] In the formula, For the global loss function, For membrane potential, according to Obtain the global loss function at time step Partial derivative with respect to membrane potential, For the current time step The received pulse sequence input;

[0036] The original gradient is modulated using the routing mask to update the effective gradient propagated backward in the spiking neural network. The expression is as follows:

[0037]

[0038] In the formula, The updated gradient;

[0039] The weights of the silent intervals are pruned using the following expression:

[0040]

[0041] In the formula, This represents the weighted connection pruning rate.

[0042] The beneficial effects of this invention are:

[0043] (1) This invention obtains the excitation level of spiking neurons through pre-training and divides them into core intervals, stable intervals, and silent intervals. In image classification tasks, this enables the network to accurately locate and strengthen core neurons that are highly sensitive to key image features; by controlling the gradient of backpropagation to perform hierarchical weight updates on neurons in different intervals, the interference of complex image backgrounds on network learning is effectively avoided, thereby improving the model's feature extraction ability for different categories of images and the final classification accuracy.

[0044] (2) The present invention performs a pruning strategy on neurons evaluated as silent intervals, cuts off invalid impulse responses to irrelevant image features, filters image noise while reducing the consumption of computing resources and improving the computational efficiency of image classification algorithms.

[0045] (3) This invention breaks away from the traditional “black box” global optimization of spiking neural networks and introduces the mesoscale adjustment mechanism into the image classification scenario, so that the network topology can adapt and dynamically evolve according to the input image category features, thereby enhancing the biological interpretability in visual classification tasks. Attached Figure Description

[0046] Figure 1 This is a flowchart of the steps of the present invention. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0048] Example 1: As Figure 1 As shown, an image classification method based on a mesoscale-tuned spiking neural network includes the following steps:

[0049] Step 1: Introduce rectangular surrogate gradients and build the target spiking neural network model based on the Leaky Integrate and Fired (LIF) neuron model;

[0050] Step 1.1: Initialize the neurons of the spiking neural network according to the LIF neuron model. The neuron receives upper-layer pulse input, converts the upper-layer pulse input into synaptic current, and updates its own membrane potential. The process of accumulating membrane potential is as follows:

[0051]

[0052] In the formula, For membrane potential, This is a membrane potential attenuation factor, used to achieve potential attenuation in neural networks when there is no input. Presynaptic neuron and postsynaptic neurons Synaptic weights, For the current time step The received pulse sequence input;

[0053] The membrane potential in a neuron will continuously accumulate, and when the membrane potential in the neuron exceeds a set trigger threshold... The neuron is then activated and fires a discrete binary impulse to the next layer. The expression is:

[0054]

[0055] in, Let be the Heaviside step function, when At that time, the Heaviside step function Assigning output 1 to At this point, the neuron fires a pulse; otherwise... Neurons remain in a resting state;

[0056] After a neuron fires a pulse, the accumulated membrane potential is reset, and the neuron enters a refractory period. The membrane potential momentarily degrades, as expressed by:

[0057]

[0058] Step 1.2: Introduce a rectangular surrogate gradient into the LIF neuron model to solve the step function. The non-differentiability of the gradient leads to vanishing gradients when directly applied to backpropagation. Therefore, a rectangular surrogate gradient is introduced during the backpropagation process of the LIF neuron model to approximate the calculation of the derivative of the step function, guiding the stable backpropagation of the error. The expression is as follows:

[0059]

[0060] in, To effectively proxy the half-width of the gradient window; when the membrane potential When the gradient approaches a set threshold, the surrogate gradient will approximate a constant of 1, allowing the error signal to be smoothly passed on to the weights of the previous layer. Otherwise, the gradient stage is 0.

[0061] Step 1.3: Using Step 1.1 and Step 1.2 as the LIF neuron computation mechanism, and defining the LIF neuron computation mechanism as... Operator, according to Operators form the target spiking neural network model.

[0062] Optionally, in this embodiment, one of the following four structures can be selected as the target spiking neural network framework: the classic Leaky Integrate and Fired-based Spiking Neural Network (LIF-SNN), the Spiking Visual Geometry Group Network (SpikingVGG), the Spike-Element-Wise Residual Network (SEW-ResNet), and the Spiking Self-Attention Mechanism Network (Spikingformer).

[0063] Specifically, in the LIF-SNN structure, the convolutional feature hidden layer and the fully connected hidden layer are connected through... The operator extracts the global semantic features of the target, expressed as:

[0064]

[0065]

[0066]

[0067] In the formula, For the first Layer in time step The pulse characteristics after activation output For convolution operations, For spatial flattening operations, the shift pulse characteristics of the first layer are obtained. , For fully connected hidden layers At time step The output high-dimensional hidden layer impulse vector, This is the weight matrix for fully connected synapses;

[0068] The calculated high-dimensional hidden layer pulse vector is summed with the membrane potential, and the predicted output is obtained by time averaging. The expression is as follows:

[0069]

[0070] In the formula, To predict the output, This represents the membrane potential of the output layer neurons. The synaptic weight matrix of the output layer.

[0071] Specifically, in the SpikingVGG structure, the deep feature extraction layer and the spatial downsampling layer are executed sequentially, through... After extracting local spatiotemporal semantic features, the operator is passed to the downsampling layer for spatial dimension compression and receptive field expansion. The expression is:

[0072]

[0073]

[0074] In the formula, For normalization operations, For time step Next The pulse tensor fired by neurons in a layered spiking neural network. For the first The convolution kernel weight parameters of the layer, This is a max pooling operation;

[0075] Furthermore, all that has passed The features computed by the operator and downsampled are spatially flattened and processed by a fully connected layer, mapping the extracted features into a one-dimensional impulse feature vector. The expression is:

[0076]

[0077] The one-dimensional pulse feature vector is mapped to current after passing through the output layer. The final classification prediction output is calculated based on the accumulation of membrane potential and time averaging, and the expression is:

[0078]

[0079] Specifically, in the SEW-ResNet structure, two-dimensional spatial pulses are processed by serially stacking multiple sets of pulse residual blocks. Residual feature extraction and identity mapping are performed within each block, and the original features within the block are aligned with the extracted residual features and then added together. The expression is as follows:

[0080]

[0081]

[0082]

[0083] In the formula, This is the intermediate pulse feature tensor extracted from the hidden layer of the first-level network. The residual impulse feature tensor is further extracted by the second-level network. and These are the batch normalization operations applied to the output currents of the first and second stage convolutional networks, respectively. and These are the independent convolution kernel weight parameters corresponding to the spatial convolution operations of the first and second level networks, respectively. This is a pulse aggregation operator that performs element-wise addition on the target object.

[0084] Furthermore, the multidimensional features extracted and passed from multiple residual blocks are processed through spatial flattening and fully connected layers to integrate the global semantic features into a one-dimensional impulse feature vector, expressed as:

[0085]

[0086] The flattened one-dimensional features are then used by the output layer to calculate the final prediction result, expressed as:

[0087]

[0088] Specifically, in the Spikingformer structure, the image is segmented by pulses, converted into serialized tokens, and interacts with features through multiple sets of serially stacked Spikingformer blocks. Each block consists of a pulse self-attention submodule and a pulse feedforward network submodule.

[0089] Furthermore, in the pulse self-attention submodule, the input features are first processed by three parallel linear mappings and... Operators that generate binary queries. Key Sum (Value) The pulse is expressed as:

[0090]

[0091]

[0092]

[0093] In the formula, For linear fully connected mapping operations, , and These are separate batch normalization operations for generating queries, keys, and values, respectively. , and These generate independent weight matrices for the query, key, and value, respectively.

[0094] Furthermore, the generated binary query, key, and value pulses are multiplied by matrix to extract contextual information, and then residual features are obtained through linear operations, expressed as:

[0095]

[0096]

[0097]

[0098] In the formula, The global context features generated by aggregation within the self-attention submodule, and the indices. Represents the matrix transpose operation. The characteristics of the self-attention impulses output by the self-attention module. This represents the pulse characteristics of the intermediate hidden layer;

[0099] Furthermore, in the pulse feedforward network submodule, the pulse features of the hidden layer are nonlinearly upscaled and feature extracted, as expressed by:

[0100]

[0101]

[0102]

[0103] In the formula, This represents the time step after the first-level linear mapping and spiking neuron activation. The characteristics of the intermediate pulses in the output feedforward network. For the second-order linear mapping and spiking neuron activation at time step Output feedforward network residual pulse characteristics and These are independent weight parameters for the first-order linear mapping and the second-order linear mapping, respectively.

[0104] Furthermore, after feature extraction and secondary residual fusion, the first... The pulse output sequence of a Spikingformer block ,sequence After flattening and fully connected processing, a one-dimensional pulse feature vector is obtained, expressed as:

[0105]

[0106] The one-dimensional pulse output is passed through the output layer to obtain the predicted classification result. .

[0107] Step 2: Pre-train the image classification task based on the target spiking neural network model. Obtain the pulse firing rate of each neuron in the target spiking neural network model through a preset number of epochs of pre-training. Evaluate the excitation level of the neurons based on the pulse firing rate. Divide the neurons into core intervals, stable intervals, and silent intervals according to the neuron excitation level evaluation results and a preset ratio.

[0108] Optionally, network pre-training is enabled for the image classification task, and the pulse firing rate of neurons in each layer is statistically analyzed during the forward transmission. The specific pulse firing rate is as follows:

[0109]

[0110] In the formula, Represents the pulse firing rate. The total time step, For training batch size, For the first Layer The first neuron in the processing batch For each sample, the time step The output pulse belongs to {0, 1}.

[0111] Optionally, the step of dividing neurons into core regions, stable regions, and silent regions based on the neuronal excitation level assessment results and a preset ratio specifically involves:

[0112] The firing rate of each neuron As an indicator for assessing neuronal excitability, neurons are sorted in reverse chronological order according to their excitability, and the division ratio of the core interval, the stable interval, and the silent interval is set as follows: ,and Based on the excitability of neurons and a predetermined ratio, subgroups are formed, and the total number of neurons is defined as follows: Then, they are sorted according to the degree of neuronal excitation, and those belonging to the sorting interval are... The neurons within the range are the core region, and those belonging to the sorting interval are... The range of neurons is the stable interval, and the sorting interval belongs to... The range of neurons is the silent interval.

[0113] Understandably, neurons in the core region have high excitability and a high pulse firing rate, and are mainly task-related encodings; neurons in the stable region maintain a moderate level of excitability, providing representational diversity and structural robustness for the neural network; neurons in the silent region consist of neurons that have failed to establish an effective stimulus-response mapping for the current task, and most of them are in a low-metabolic inhibitory state during task execution.

[0114] Step 3: Generate different routing masks for the core interval, the stable interval, and the silent interval respectively. Apply the routing masks to the gradient during the backpropagation stage to form a mesoscopic regulation mechanism for the connection of spiking neurons. Update the gradient and the neural network gradient according to the mesoscopic regulation mechanism, and prune the neuron connections in the silent interval to obtain the updated target spiking neural network model.

[0115] Optionally, generating different routing masks for the core interval, the stability interval, and the silent interval specifically involves:

[0116]

[0117] In the formula, For the routing mask, , , These are the core region, the homeostasis region, and the silent region of the neuron, respectively. and These are the fundamental amplitude parameter matrices for the core interval and the fundamental amplitude parameter matrices for the stable interval, respectively. and These are the strengthening factor for the core interval and the stabilization factor for the stable interval, respectively. For the silent interval, the gradient change is cut off and the mask is set to 0.

[0118] Optionally, the mesoscopic regulatory mechanism specifically includes:

[0119] During backpropagation, the fundamental error gradients between the connections of neurons in the neural network are obtained, through the first... Presynaptic neurons in the layer and postsynaptic neurons The synaptic weight is , obtained by The original spatiotemporal gradient of time backpropagation unfolds within a time window consisting of several time steps. The expression is:

[0120]

[0121] In the formula, For the global loss function, For membrane potential, according to Obtain the global loss function at time step Partial derivative with respect to membrane potential, For the current time step The received pulse sequence input;

[0122] The original gradient is modulated using the routing mask to update the effective gradient propagated backward in the spiking neural network. The expression is as follows:

[0123]

[0124] In the formula, The updated gradient ensures that the core region gains an advantage in weight updates.

[0125] The weights of the silent intervals are pruned using the following expression:

[0126]

[0127] In the formula, This represents the weighted connection pruning rate.

[0128] It is understood that the mesoscopic regulation mechanism proposed in this embodiment uses task-driven adaptive regulation of neuronal activity and global task objectives to act on neuronal activity, forming mesoscopic neural activity between the microscopic and macroscopic levels, optimizing the cognitive abilities of different spiking neural network architectures, and reducing redundant computation.

[0129] Step 4: Input the image to be classified into the updated target spiking neural network model to obtain the image classification result.

[0130] Based on the specific implementation details, the effectiveness of the technical solution of the present invention will be demonstrated through experiments.

[0131] Specifically, this experiment selected three datasets—MNIST, CIFAR10, and DVS Gesture—as the experimental verification objects, respectively addressing multi-class classification verification of images, complex tasks, and dynamic event stream images. The training set, test set, and corresponding class labels for each dataset were obtained. This experiment used the Python programming language, and the results of performing the classification task on the three datasets according to this invention are shown in Table 1.

[0132] Table 1. Classification accuracy and comparative experiments of the present invention on MNIST, CIFAR10, and DVS Gesture datasets.

[0133]

[0134] As shown in Table 1, to demonstrate the optimization of spiking neural networks by Topology-Centric Soft Plasticity (TCSP) inspired by memory imprints in this invention, this experiment evaluated the MNIST, CIFAR10, and DVS Gesture datasets using four metrics: accuracy, spike firing rate, floating-point operations per second (FLOPs), and theoretical energy consumption. The results show improvements over four different algorithmic frameworks: classic shallow feedforward spiking neural networks, spiking visual geometry networks, spiking element-level residual networks, and spiking self-attention mechanism networks. Table 1 also lists spike-timing-dependent plasticity (STDP) and reward-modulated spike-timing-dependent plasticity. Comparing the performance of three existing plasticity rules, including Plasticity (RSTDP) and Pulse Time-dependent Plasticity (QSTDP) under Q-learning regulation, on various metrics of the LIF-SNN network architecture, it can be seen that the present invention maintains high accuracy and low energy consumption in performing classification tasks.

[0135] Furthermore, the results in Table 1 demonstrate that this invention can improve the accuracy of image classification tasks while reducing the computational redundancy of spiking neural network models. This invention enhances the ability of different neural networks to extract core image features, primarily reflected in the improved classification accuracy of four network architectures on the MNIST, CIFAR10, and DVS Gesture datasets. On the other hand, this invention suppresses ineffective firing caused by image background noise by implementing a pruning strategy on neurons in silent regions, as evidenced by the decrease in three indicators in Table 1: pulse firing rate, floating-point operations per second, and theoretical energy consumption.

[0136] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An image classification method based on a mesoscale-tuned spiking neural network, characterized in that, The method includes the following steps: Step 1: Introduce a rectangular surrogate gradient and build the target spiking neural network model based on the LIF neuron model; Step 2: Pre-train the image classification task based on the target spiking neural network model. Obtain the pulse firing rate of each neuron in the target spiking neural network model through a preset number of epochs of pre-training. Evaluate the excitation level of the neurons based on the pulse firing rate. Divide the neurons into core intervals, stable intervals, and silent intervals according to the neuron excitation level evaluation results and a preset ratio. Step 3: Generate different routing masks for the core interval, the stable interval, and the silent interval respectively. Apply the routing masks to the gradient during the backpropagation stage to form a mesoscopic regulation mechanism for the connection of spiking neurons. Update the gradient and the neural network gradient according to the mesoscopic regulation mechanism, and prune the neuron connections in the silent interval to obtain the updated target spiking neural network model. Step 4: Input the image to be classified into the updated target spiking neural network model to obtain the image classification result.

2. The image classification method based on a mesoscale-tuned spiking neural network according to claim 1, characterized in that, Step 1 specifically refers to: Step 1.1: Initialize the neurons of the spiking neural network according to the LIF neuron model. The neuron receives upper-layer pulse input, converts the upper-layer pulse input into synaptic current, and updates its own membrane potential. The process of accumulating membrane potential is as follows: ； In the formula, For membrane potential, This is a membrane potential attenuation factor, used to achieve potential attenuation in neural networks when there is no input. Presynaptic neuron and postsynaptic neurons Synaptic weights, For the current time step The received pulse sequence input; When the membrane potential in a neuron exceeds a set trigger threshold The neuron is then activated and fires a discrete binary impulse to the next layer. The expression is: ； in, Let be the Heaviside step function, when At that time, the Heaviside step function Assigning output 1 to At this point, the neuron fires a pulse; otherwise... Neurons remain in a resting state; After a neuron fires a pulse, the accumulated membrane potential is reset, and the neuron enters a refractory period. The membrane potential momentarily degrades, as expressed by: ； Step 1.2: In the backpropagation process of the LIF neuron model, a rectangular surrogate gradient is introduced to approximate the calculation of the derivative of the step function. The expression is: ； in, To effectively proxy the half-width of the gradient window; Step 1.3: Using Step 1.1 and Step 1.2 as the LIF neuron computation mechanism, and defining the LIF neuron computation mechanism as... Operator, according to Operators form the target spiking neural network model.

3. The image classification method based on a mesoscale-tuned spiking neural network according to claim 1, characterized in that, The pulse firing rate is specifically: ； In the formula, Represents the pulse firing rate. The total time step, For training batch size, For the first Layer The first neuron in the processing batch For each sample, the time step The output pulse belongs to {0, 1}.

4. The image classification method based on a mesoscale-tuned spiking neural network according to claim 3, characterized in that, The specific steps of dividing neurons into core, stable, and silent regions based on neuronal excitation level assessment results and preset ratios are as follows: The firing rate of each neuron As an indicator for assessing neuronal excitability, neurons are sorted in reverse chronological order according to their excitability, and the division ratio of the core interval, the stable interval, and the silent interval is set as follows: ,and Based on the excitability of neurons and a predetermined ratio, subgroups are formed, and the total number of neurons is defined as follows: Then, they are sorted according to the degree of neuronal excitation, and those belonging to the sorting interval are... The neurons within the range are the core region, and those belonging to the sorting interval are... The range of neurons is the stable interval, and the sorting interval belongs to... The range of neurons is the silent interval.

5. The image classification method based on a mesoscale-tuned spiking neural network according to claim 1, characterized in that, The specific steps for generating different routing masks for the core interval, the stability interval, and the silent interval are as follows: ； In the formula, For the routing mask, , , These are the core region, the homeostasis region, and the silent region of the neuron, respectively. and These are the fundamental amplitude parameter matrices for the core interval and the fundamental amplitude parameter matrices for the stable interval, respectively. and These are the strengthening factor for the core interval and the stabilization factor for the stable interval, respectively. For the silent interval, the gradient change is cut off and the mask is set to 0.

6. The image classification method based on a mesoscale-tuned spiking neural network according to claim 5, characterized in that, The mesoscopic regulatory mechanism is specifically as follows: During backpropagation, the fundamental error gradients between the connections of neurons in the neural network are obtained, through the first... Presynaptic neurons in the layer and postsynaptic neurons The synaptic weight is , obtained by The original spatiotemporal gradient of time backpropagation unfolds within a time window consisting of several time steps. The expression is: ； In the formula, For the global loss function, For membrane potential, according to Obtain the global loss function at time step Partial derivative with respect to membrane potential, For the current time step The received pulse sequence input; The original gradient is modulated using the routing mask to update the effective gradient propagated backward in the spiking neural network. The expression is as follows: ； In the formula, The updated gradient; The weights of the silent intervals are pruned using the following expression: ； In the formula, This represents the weighted connection pruning rate.

Images