Target recognition method based on three-section pulse neural network

Abstract

The application discloses a target recognition method based on a three-section pulse neural network, constructs a three-section neuron structure divided into a cell body, a proximal dendrite and a distal dendrite, independently supervises injection through the distal dendrite, makes teaching signals independently input into the neural network through the distal dendrite, adopts a local pulse pattern matching learning rule, synchronously updates neural network weights based on a weight update mode of likelihood maximization, and performs Softmax normalization processing on neural network outputs on the whole time window through a time dimension joint decoding layer, and adjusts decoding layer weights through reverse propagation of an obtained cross entropy average loss value. The method runs in a GPU, is used for recognizing image data sets, splits the image data sets into a training set and a test set, and then trains a neural network model, so that the processing capacity of the model for complex information in image recognition is improved, and meanwhile, the wide applicability of the model to different intelligent tasks is maintained.

Description

Technical Field

[0001] This invention relates to a method for identifying a target, and more particularly to a target identification method based on a three-segment spiking neural network, belonging to the field of neural network optimization technology. Background Technology

[0002] In the field of object recognition, deep neural networks have made continuous breakthroughs through a series of innovative techniques. In 2012, AlexNet significantly improved the accuracy of image classification by introducing techniques such as ReLU activation function, Dropout, and data augmentation, driving a revolution in deep learning in computer vision. In 2015, ResNet effectively solved the gradient vanishing problem by utilizing residual connections, enabling classification accuracy to surpass human levels for the first time. Building on this, the YOLO series of models further achieved end-to-end real-time multi-object detection and localization, and has been widely used in scenarios such as surveillance, autonomous driving, and drone navigation.

[0003] Despite the significant achievements of deep neural networks in many fields, their large-scale application still faces core technological challenges:

[0004] 1) Traditional models are massive in scale, and the computational resources and energy consumption required for training grow exponentially. Modern networks often contain hundreds of millions or even trillions of parameters; for example, the training cost of GPT-3 (175 billion parameters) reached millions of dollars, and the resource and energy consumption of GPT-4 soared even further. The resource requirements vary significantly between different tasks and architectures, easily leading to excessively high training costs or difficulty in deployment in resource-constrained environments. How to automatically and robustly balance model performance and resource consumption in various applications remains an open challenge.

[0005] 2) Traditional deep neural networks rely on the von Neumann architecture during training, which is limited by the slowdown of Moore's Law and data transmission bottlenecks. The gap between computation speed and storage retrieval speed is gradually widening, significantly restricting training efficiency. This architectural bottleneck results in a significant lack of deployment capability between high-performance computing and low-power devices (e.g., mobile devices, embedded systems, or IoT devices), especially in scenarios with high real-time requirements or limited resources. The difficulty in optimization severely restricts solution efficiency and system integration capabilities.

[0006] In the field of spiking neural networks, existing research mainly follows two paths: one is to unfold the spiking neural network into a computational graph in the time dimension and use backpropagation for supervised learning. This path performs well in terms of accuracy, but the computational and memory overhead increases linearly with the time step, making it difficult to adapt to resource-constrained environments; the other is to rely on local rules such as pulse temporal dependence plasticity to update weights. This path has high biological rationality, but its processing capacity is limited in complex supervised tasks, and its classification accuracy is far lower than that of mainstream artificial neural networks.

[0007] The two paths described above present a clear trade-off between accuracy and efficiency. Currently, there is no spiking neural network training method that can simultaneously take into account supervised learning capabilities, local update mechanisms, and the fact that it does not require expansion along the time dimension. Summary of the Invention

[0008] The purpose of this invention is to solve the above-mentioned problems and provide a target recognition method based on a three-segment spiking neural network.

[0009] The technical solution of this invention is a target recognition method based on a three-segment spiking neural network. Its key feature is the construction of a three-segment neuron structure. This structure provides a physically isolated channel for distal supervision, eliminating the need for learning rules to unfold along the time axis. Functionally, the three-segment recognition structure is divided into a cell body, proximal dendrites, and distal dendrites, with their functional structures mutually separated. Specifically, the cell body processes input pulse signals, the proximal dendrites regulate local synaptic plasticity, and the distal dendrites independently detect injected teaching signals. By employing independent supervised injection through distal dendrites, the teaching signal is independently input into the neural network via distal dendrites. A local spiking pattern matching learning rule is used, allowing for synchronous updates of neural network weights based on likelihood maximization. A temporal joint decoding layer is incorporated, whose design, in turn, imposes requirements on the front-end local spiking pattern matching, forming a closed-loop constraint. This temporal joint decoding layer performs Softmax normalization on the neural network output over the entire time window, calculates the cross-entropy loss with the target label, and obtains the average cross-entropy loss value over the time window. The decoding layer weights are then adjusted using backpropagation of the corresponding average loss value, thus realizing a spiking neural network training framework that is "supervised, locally updated, and requires no temporal backpropagation."

[0010] This target recognition method based on a three-segment spiking neural network runs on a GPU and includes the following steps:

[0011] Step (1) Obtain the image dataset to be identified and split it into a training set and a test set;

[0012] Step (II) Training the neural network model. The specific steps are as follows:

[0013] 1) Set the structure and training limit N of the local spiking neural network, and initialize the weight matrix W of the neural network, including the number of excitatory and inhibitory neurons, membrane potential constant and threshold in the spiking neural network. The experiment adopts the structure of Input1-LIF500-FC10 (Input represents the input layer; LIF represents the three-segment spiking neural network layer, FC represents the fully connected layer; the number after it represents the number of neurons in this layer), and initialize the weight matrix W to a zero matrix.

[0014] 2) Dynamic modeling of multi-segmented, multi-input neuron models: The LIF model is introduced as a framework for describing the membrane potential dynamics of the model. The equation form of the LIF model is as follows:

[0015]

[0016] Where, v=[v soma ,v prox ,v dist Each segment has independent capacitance, leakage conductance, threshold, and reset potential; when the cell body, proximal dendrites, and distal dendrites are all activated, each variable is transformed into a vector form corresponding to the three-part structure.

[0017] 3) Probabilistic modeling of pulse emission, defining the pulse emission probability within each time step Δt as:

[0018]

[0019] As k approaches 0, the model degenerates into a deterministic step function. .

[0020] 4) Modeling of the strong impulse mechanism: Based on the cellular mechanisms of the cerebral cortex, a strong impulse mechanism is introduced, where neurons generate a high-frequency, continuous sequence of impulses within a certain time window. The exponential filter function is defined in a loop:

[0021]

[0022] Based on this filtering function, postsynaptic potential filtering, cell body target filtering, and weight update filtering (used for smoothing pulse signals) are introduced respectively; when a neuron emits a pulse, a time window of length Δt is opened:

[0023]

[0024] in, To control the threshold parameter of the time window length, when the proximal or distal dendrites generate strong pulses above the threshold, the cell body can obtain additional current input from the dendritic strong pulses; when strong pulses caused by both proximal and distal dendrites occur simultaneously, the strong pulse window is opened.

[0025] 5) Modeling of neuronal input current. The cell membrane potential v i t The current input mainly consists of current input from other neuronal cell bodies, external information input, and adaptive current input generated by adaptive changes; the membrane potential u of the proximal dendrites i tThe electrical input mainly comes from the cyclic connections generated by itself and all neurons' cell bodies; the membrane potential of distal dendrites The current input mainly consists of the target signal (i.e., image label) input, and its mathematical expression can be simply expressed as:

[0026]

[0027] in, It is the target signal (tag). These are the weights of the target signal. This structure takes the target signal current as input and uses the difference between it and the final output of the biological neural network as the optimization objective.

[0028] 6) Construct neuron connection patterns and learning rules, and adjust connection strength by maximizing the log-likelihood function:

[0029]

[0030] in, The pulse signal is the target signal (tag). This represents the pulse signal output by the neural network. The update rule relies only on the local pulse state and does not require backpropagation along the time axis.

[0031] 7) For model input processing, the original image is compressed or cropped according to the input layer size requirements to match the number of input neurons. The processed image is then flattened into a one-dimensional vector, and the grayscale value of each pixel is linearly normalized to the [0,1] interval. This normalization value represents the probability that the corresponding input neuron will fire a pulse in a single time step. Within the simulated time window (T time steps), a Bernoulli trial is performed independently at each time step, and the probability determines whether to fire a pulse at the current time step. Therefore, high-brightness pixels (normalized values ​​close to 1) will fire pulses with a higher probability at each time step, resulting in a dense pulse sequence; conversely, low-brightness pixels (normalized values ​​close to 0) have an extremely low probability of firing pulses, producing only a sparse pulse sequence. This probability-based pulse coding method converts image grayscale information into a pulse distribution in the time dimension, providing a continuous spatiotemporal input signal for subsequent network processing.

[0032] 8) Generate teaching signals by converting the original classification labels of the dataset into a one-hot encoding format, encoding them through a pulse coding process similar to that of the input, and then injecting the encoded signals into the distal dendrites of the output layer neurons.

[0033] 9) Design a decoding layer based on temporal softmax and cross-entropy loss for joint temporal decoding. For each time step t, the output layer is independently fed into the softmax function:

[0034]

[0035] in, This refers to the output term of the output layer neurons. The cross-entropy loss function is used.

[0036]

[0037] Calculate the gradients of the weights and biases:

[0038]

[0039] Then the weights and biases are updated as follows:

[0040]

[0041] By adopting the technical solution of this invention, the three-segment structure of cell body, proximal dendrite and distal dendrite is introduced to improve the intelligence level of machine perception, learning, decision-making and execution from the fundamental level of simulating biological neural information processing mechanism. This three-segment structure can enable neurons to achieve functional differentiation in different links such as sensory input, instruction generation and execution output, thereby reducing functional coupling and information redundancy between links, and enabling the network to obtain better information processing efficiency with a limited number of neurons.

[0042] Compared with existing spiking neuron models applied to image recognition, the three-segment structure of this invention does not require pre-setting specific task scenarios, simplifies the assumptions, and has stronger bio-inspiredness and versatility. Compared with traditional single-segment or double-segment models, the three-segment structure, as a biomimetic design, can achieve greater flexibility and modularity when applied to the neural network architecture without the need to introduce additional complex external control mechanisms.

[0043] Therefore, this invention improves the model's ability to process complex information and its adaptability in image recognition tasks while maintaining broad applicability to different intelligent tasks and hardware environments. Compared with traditional artificial neural network methods, this solution exhibits lower classification error and lower power consumption on the MNIST standard image recognition dataset. This method not only effectively complements existing spiking neural network models and traditional artificial neural network technology paths, but also provides a novel and practical construction idea for the development of energy-efficient machine learning. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the model training process of the present invention;

[0045] Figure 2 This is a schematic diagram illustrating the image classification process of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will be further described below with reference to the accompanying drawings to make it easier to understand and master. The accompanying drawings are intended to comprehensively and systematically elucidate the purpose, core features, and practical application value of the present invention, thereby facilitating a deeper understanding and mastery of its essence.

[0047] like Figures 1-2 As shown, the present invention provides a target recognition method based on a three-segment spiking neural network, including a model training stage and an image classification stage.

[0048] The model training phase includes the following steps: 1) Initialization: Set the initial weights of the neural network and specify the total number of training iterations; 2) Pulse coding: Convert the input image into a pixel pulse sequence; 3) Input network: Input the pulse sequence and its corresponding image label into the neural network; 4) Membrane potential update: Update the distal dendritic membrane potential, cell body membrane potential, and proximal dendritic membrane potential sequentially; 5) Pulse firing and filtering: If the membrane potential exceeds the threshold, fire a pulse and perform filtering; when multiple neurons fire pulses simultaneously, a strong pulse signal is generated; 6) Learning window and parameter update: Once a strong pulse is generated, the learning window is opened, and the synaptic weights and fully connected layer parameters are updated within this window; 7) Time step loop: Determine whether the current time step is less than the total time step (t < T). If so, repeat steps 4 to 6; otherwise, end the simulation of the current sample; 8) Training iteration judgment: Check whether the preset total number of training iterations has been reached. If not, start the next round of training. If the total number of training iterations has been reached, save the trained neural network model, and the training is complete.

[0049] The image classification stage includes the following steps: 1) Preprocessing: preprocessing the input image, including grayscale normalization; 2) Pulse coding: converting the processed image into a pixel pulse sequence; 3) Forward propagation: inputting the pulse sequence into the trained neural network; 4) Potential update: updating only the cell body membrane potential and the proximal dendrite membrane potential; 5) Pulse firing and filtering: if the membrane potential exceeds the threshold, a pulse is fired and filtered; if multiple neurons fire pulses simultaneously, a strong pulse is generated; 6) Time step loop: determining whether the current time step is less than the total time step (t < T). If so, repeating steps 4-5; otherwise, ending the simulation; 7) Classification output: applying the softmax function to the collected strong pulse time series to obtain the final classification result.

[0050] To verify the applicability and performance of this embodiment, the experimental parameters and values ​​are shown in Table 1 below:

[0051] Table 1: Experimental Parameters and Values

[0052]

[0053] This embodiment analyzes the three-segment spiking neural network method on the MNIST dataset. The experimental results are shown in Table 2.

[0054] Table 2: Classification accuracy on the training and test sets after five epochs of model operation.

[0055]

[0056] As shown in Table 2, the classification accuracy on the training set exhibits a continuous and steady upward trend with increasing training cycles. This phenomenon fully demonstrates that the cell body-proximal dendrite-distal dendrite three-segment structure constructed in this invention, combined with the learning rule based on the consistency of pulse spatiotemporal patterns, enables the model to effectively extract discriminative features from image pulse sequences during training and continuously optimize the connection weights between neurons, achieving continuous learning and good fitting of the training data. Meanwhile, although the accuracy on the test set also gradually increases with the advancement of the training cycle, its fluctuation range is significantly smaller than that on the training set, and the overall trend is more stable. This indicates that the method of this invention maintains relatively stable and reliable classification performance even when facing test samples that have not participated in training, without significant overfitting. In summary, the three-segment spiking neural network method of this invention not only achieves efficient learning and fitting on the MNIST dataset but also demonstrates excellent generalization stability and anti-overfitting ability.

[0057] As described above, compared to traditional artificial neural network methods, the technical solution of this invention exhibits lower classification error and lower power consumption on the MNIST standard image recognition dataset. Specifically, the three-segment neuron structure divides the spiking neuron into three functional modules: the cell body, proximal dendrites, and distal dendrites. This design decouples the functions within the neuron. The proximal dendrites primarily receive current input from other neuron cell bodies, while the distal dendrites independently receive supervisory signals. The cell body integrates current input from other neuron cell bodies, external information input, and adaptive current input generated by adaptive changes to produce a pulse output. This structural separation physically separates information from different sources, reducing information redundancy caused by functional coupling and improving information representation efficiency. Functional decoupling also makes the gradient update path clearer, avoiding interference between different signals in traditional models, thus effectively maintaining training stability and laying a structural foundation for efficient training of deep spiking neural networks.

[0058] Thus, by employing the technical solution of this invention, a strong impulse mechanism is introduced, in which neurons generate high-frequency continuous impulse sequences within a certain time window, and the impulse pattern is dynamically controlled through precise modeling of the time window. This mechanism significantly enhances the network's ability to model temporal information: traditional spiking neural networks rely only on the temporal relationships of single impulses, while the strong impulse mechanism can capture longer-range and more complex impulse pattern dependencies. Therefore, the model has stronger adaptability to dynamic inputs and can maintain stable representation and decision-making capabilities in time-varying environments.

[0059] In addition, this invention employs a design where distal dendrites independently receive teaching signals, enabling supervisory information to be directly injected into neurons without traversing the forward propagation path. This mechanism fundamentally avoids the gradient vanishing and gradient exploding problems that accumulate with time steps in traditional backpropagation methods. Simultaneously, the independent input of distal dendrites allows weight updates to be completed based on local information, realizing local learning rules. Each neuron only needs to adjust its connection strength based on its own spiking activity and the distal teaching signal, without accessing the global computation graph. This design significantly reduces computational and memory overhead during training, substantially improving training efficiency, and is particularly suitable for resource-constrained edge computing scenarios.

[0060] Furthermore, compared to the decoding methods in traditional spiking neural networks that rely solely on pulse counting or single-step membrane potential, this invention proposes a decoding layer design based on time-dimensional softmax. This time-dimensional softmax-based decoding layer collects the cumulative responses of output neurons throughout the entire simulation time window, performs global softmax normalization in the time dimension to obtain the class probability distribution, and uses cross-entropy as the loss function. The core advantage of this design is that the integration of global temporal information significantly improves classification accuracy, because the network can learn the decision boundary from the complete spatiotemporal dynamics, rather than relying on transient information from local time steps; moreover, the combination of softmax and cross-entropy provides a smooth gradient surface, which helps stabilize gradient updates during backpropagation and avoids drastic gradient fluctuations.

[0061] The technical solution, working process, and implementation effects of the present invention have been described in detail above. It should be noted that the described examples are only typical examples of the present invention. In addition, the present invention may have many other specific implementation methods. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A target recognition method based on a three-segment spiking neural network, characterized in that: A three-segment neuron structure is constructed, which is functionally divided into a cell body, proximal dendrites, and distal dendrites, with each segment functionally separate from the others. Independent supervised injection through distal dendrites allows the teaching signal to be independently input into the neural network. A local impulse pattern matching learning rule is adopted, and the neural network weights are updated synchronously based on likelihood maximization. A temporal joint decoding layer is set up, which performs Softmax normalization on the neural network output over the entire time window, calculates the cross-entropy loss with the target label, obtains the average cross-entropy loss value over the time window, and adjusts the decoding layer weights using the backpropagation of the corresponding average loss value.

2. The target recognition method based on a three-segment spiking neural network according to claim 1, characterized in that: This method runs on a GPU and includes the following steps: Step 1: Obtain the image dataset to be identified and split it into a training set and a test set; Step two: Train the neural network model.

3. The target recognition method based on a three-segment spiking neural network according to claim 2, characterized in that, The process of training a neural network model includes the following steps: (a) Set the structure and training limit N of the local spiking neural network, and initialize the neural network weight matrix W, including the number of excitatory and inhibitory neurons, membrane potential constant and threshold in the spiking neural network; (b) Dynamic modeling of multi-segmented multi-input neuron model, introducing the LIF model as the membrane potential dynamic description framework of the model; (c) Probabilistic modeling of pulse emission; (d) Modeling of the strong impulse mechanism, introducing the strong impulse mechanism based on the cellular mechanism of the cerebral cortex; (e) Modeling of neuronal input current and cell body membrane potential v i t The current input is composed of current input from other neuronal cell bodies, external information input, and adaptive current input generated by adaptive changes. The membrane potential u of the proximal dendrites i t The current input mainly comes from the cyclic connections generated by the cell bodies of the neurons themselves and all neurons, and the membrane potential of the distal dendrites. The current input mainly consists of the target signal (i.e., image tag) input; (f) Constructing the connection patterns and learning rules of neurons; (h) For model input processing, the original image is compressed or cropped according to the input layer size requirements to match its size with the number of input neurons. The processed image is then flattened into a one-dimensional vector. The gray value of each pixel is linearly normalized to the [0,1] interval. This normalization value is the probability that the corresponding input neuron fires a pulse in a single time step. Within the simulation time window, a Bernoulli trial is performed independently at each time step. The probability determines whether to fire a pulse at the current time step. (i) Generate teaching signals by converting the original classification labels of the dataset into a one-hot encoding format, encoding them through a pulse coding process similar to that of the input, and then injecting the encoded signals into the distal dendrites of the output layer neurons; (j) Design a decoding layer, and perform joint decoding in the temporal dimension based on the decoding layer of temporal dimension Softmax and cross-entropy loss.

4. The target recognition method based on a three-segment spiking neural network according to claim 3, characterized in that: In step (b), the equation form of the LIF model is: Where, v=[v soma ,v prox ,v dist Each segment has independent capacitance, leakage conductance, threshold, and reset potential; when the cell body, proximal dendrites, and distal dendrites are all activated, each variable is transformed into a vector form corresponding to the three-segment structure.

5. The target recognition method based on a three-segment spiking neural network according to claim 3, characterized in that: In step (c), the pulse emission probability within each time step Δt is defined as: As k approaches 0, the model degenerates into a deterministic step function. .

6. The target recognition method based on a three-segment spiking neural network according to claim 3, characterized in that: In step (d), the cyclically defined exponential filter function is defined as follows: Based on this filtering function, postsynaptic potential filtering, cell body target filtering, and weight update filtering for smoothing pulse signals are introduced respectively; when a neuron emits a pulse, a time window of length ΔT is opened: in, To control the threshold parameter of the time window length, when the proximal or distal dendrites generate strong pulses above the threshold, the cell body can obtain additional current input from the dendritic strong pulses; when strong pulses caused by both proximal and distal dendrites occur simultaneously, the strong pulse window is opened.

7. The target recognition method based on a three-segment spiking neural network according to claim 3, characterized in that: In step (e), the cell membrane potential v i t The current input mainly consists of current input from other neuronal cell bodies, external information input, and adaptive current input generated by adaptive changes; the membrane potential u of the proximal dendrites i t The electrical input mainly comes from the cyclic connections generated by itself and all neurons' cell bodies; the membrane potential of distal dendrites The current input mainly consists of the target signal input, and its mathematical expression can be simply expressed as: in, It is the target signal (tag). It is the weight of the target signal; the structure takes the target signal current as input and uses the difference between it and the final output of the biological neural network as the optimization target.

8. The target recognition method based on a three-segment spiking neural network according to claim 3, characterized in that: In step (f), link strength is adjusted by maximizing the log-likelihood function: in, The pulse signal is the target signal (tag). This refers to the pulse signal output by the neural network.

9. The target recognition method based on a three-segment spiking neural network according to claim 3, characterized in that: In step (j), the output layer is independently fed into the Softmax function for each time step t: In the formula, It is the output term of the output layer neuron; the cross-entropy loss function is used: Calculate the gradients of the weights and biases: Then the weights and biases are updated as follows: 。

Images