A selective activation-based spiking neural network continuous learning target recognition system

Abstract

The application discloses a kind of based on selective activation's pulse neural network continuous learning target identification system, including computer memory, computer processor and the computer program stored in the computer memory and can be executed on the computer processor, the computer memory has the selective activation SNNs model of well-trained;Selective activation SNNs model includes the feature extraction part of pulse neural network front end, and trajectory-based K-WTA mechanism and variable threshold mechanism are used in rear end to carry out continuous learning;The computer processor executes the computer program and realizes the following steps: the image to be identified is input into well-trained selective activation SNNs model, and the identification classification result is obtained.The present application reduces the catastrophic forgetting by enhancing the neural dynamic characteristics in intelligent learning network (SNN), improves the target recognition effect of pulse neural network in continuous learning scene, and expands the application scene of pulse neural network.

Description

Technical Field

[0001] This invention belongs to the field of image recognition, and in particular relates to a target recognition system based on selective activation spiking neural network continuous learning. Background Technology

[0002] Biological brains can continuously learn from interactions with the environment throughout their lifespan, possessing lifelong learning capabilities for multitasking. Current artificial neural networks (ANNs) have demonstrated remarkable machine intelligence. However, next-generation edge applications, such as autonomous vehicles and wearable devices, require new machine intelligence with continuous learning capabilities—capable of acquiring new knowledge without forgetting old knowledge, while operating efficiently under computational constraints. Inspired by biological systems, spiking neural networks (SNNs) possess richer spatiotemporal dynamics compared to ANNs (Maass, 1997; Subbulakshmi Radhakrishnan et al., 2021; Yin, Corradi, and Bohte, 2023), and, when combined with neuromorphic hardware, hold greater potential for enabling next-generation machine intelligence with low power consumption. The goal of this continuous learning is to acquire knowledge in an ordered manner while ensuring that the agent can only access data for the current task without affecting its ability to recall previously learned tasks.

[0003] Current mainstream continuous learning techniques can be broadly categorized into three types: regularization-based techniques, replay-based techniques, and architecture-based techniques. Regularization-based techniques protect synaptic connections by adding regularization terms to the loss function, such as EWC (Kirkpatrick et al., 2017), MAS (Aljundi et al., 2018), and SI (Zenke, Poole, and Ganguli, 2017). They calculate the importance of each parameter according to certain rules and generate penalty terms to limit variations in important parameters. Replay-based techniques aim to improve knowledge retention by preserving a few key samples (or intermediate representations) from each task and propagating them through the network after mixing them with the data from the current task. These methods are mostly geared towards artificial intelligence networks, for example (Van de Ven, Siegelmann, and Tolias 2020; Arani, Sarfraz, and Zonouz 2022; Rebuffi et al. 2017). By definition, these methods inevitably require additional storage space to store additional information and expand the model. Architecture-based approaches (Kang et al., 2022; Yoon et al., 2017) that enhance a network's ability to perform different tasks by continuously adjusting the architecture also encounter similar problems. Due to the purposeful partitioning / expansion of subnetworks, these methods typically exhibit relatively stable performance across different tasks. However, to differentiate the applicability of different subnetworks, these methods often require prior knowledge of the task information.

[0004] Regarding other SNN-based continuous learning methods, (Antonov, Sviatov, and Sukhov 2022) determined the importance of synaptic weights through stochastic Langevin dynamics of local STDP and achieved continuous learning through unsupervised learning. (Skatchkovsky, Jang, and Simeone 2022) introduced an online rule-based SNN model based on Bayesian methods. (Hammouamri, Masquelier, and Wilson 2022) achieved continuous learning of SNNs by generating the firing threshold of the classifier through training an external network using an evolutionary strategy. (Tadros et al. 2022) achieved local plasticity by switching between pulse frequency encoding and pulse time encoding using a transformation algorithm to help the model correct biases after learning new tasks.

[0005] In short, although the aforementioned technologies have obvious performance advantages, they often require complex algorithms, additional storage space, or task-specific knowledge, thus deviating significantly from the inherent sequential learning capabilities of biopharmaceuticals. Summary of the Invention

[0006] This invention discloses a target recognition system based on selective activation spiking neural networks for continuous learning. By enhancing the neural dynamics of intelligent learning networks (SNNs), it mitigates catastrophic forgetting, improves image recognition performance in continuous learning scenarios, and expands the application scenarios of spiking neural networks.

[0007] A target recognition system based on selective activation spiking neural networks (SNNs) for continuous learning includes a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor. The computer memory stores a trained selective activation SNN model.

[0008] The selective activation SNNs model includes a feature extraction part at the front end of a spiking neural network, and uses a trajectory-based K-WTA mechanism and a variable threshold mechanism at the back end for continuous learning;

[0009] Specifically, a trajectory-based K-WTA mechanism is applied to the dynamic changes of neurons in the hidden layer of a spiking neural network to reduce mutual interference between different tasks. At the same time, a variable spiking threshold is used for neurons in the hidden layer to encourage silent neurons to participate in learning and to inhibit the reactivation of neurons that have been activated during training of old tasks.

[0010] When the computer processor executes the computer program, it performs the following steps:

[0011] The image to be identified is input into a trained selective activation SNN model to obtain the image classification result.

[0012] This invention ensures the selective activation of spiking neurons at different time steps, thereby enabling the selective activation of specific spiking neuron subgroups for different tasks during continuous learning. Based on the unique structure of spiking neural networks and the principle of spatiotemporal dynamics calculation, this system realizes a continuous learning method for spiking neural networks based on structure adaptive selection. It has stronger adaptability to different tasks and overcomes the catastrophic forgetting problem of the original spiking neural network model in the process of multi-task continuous learning, thus expanding the application scenarios of spiking neural networks.

[0013] Furthermore, the trajectory-based K-WTA mechanism is specifically as follows:

[0014] An internal variable called Trace is introduced into the neuron as an indicator of K-WTA in the time dimension. The trajectory of the spiking neuron is calculated as follows:

[0015]

[0016] Where τ is the time constant, which determines the decay rate of the trace; tr[t] is the trace of the neuron at time step t; and S[t+1] represents the pulse output of the neuron at step t.

[0017] Furthermore, when using a variable spike firing threshold for neurons in the hidden layer, the spike firing threshold is set to increase slowly as the activation time increases, and the spike firing threshold does not decay. This irreversible threshold change enhances the ability to maintain memories related to old tasks.

[0018] Furthermore, by incorporating a trajectory-based K-WTA mechanism and a variable threshold mechanism into neurons, the following neuronal dynamics model was obtained:

[0019] H(t) = f(V[t-1], X[t]),

[0020] S[t]=Θ(H(t)-V th ),

[0021] Mask[t = TopK(tr[t]),

[0022] S * [t] = S[t]·Mask[t],

[0023] V[t]=H[t]-V th ·S[t],

[0024] Where X[t] is the neuron's input at time step t; S[t] represents the neuron's original pulse output; V[t] and H(t) are the membrane potentials before and after the neuron fires a pulse, respectively; Θ(-) is the function that controls the firing of a pulse when the membrane voltage exceeds a threshold; TopK(-) is the function used in the trajectory-based K-WTA component to generate a mask to obtain the mask corresponding to the top K largest trajectories; S * V represents the actual activation impulse output after the TopK function is applied; f(·) represents the neuron's state update equation; V th This represents the variable threshold of the neuron model.

[0025] Variable threshold V of neuron model th This can be achieved using linear variable or other nonlinear mapping functions. Taking the linear variable method as an example, the calculation method for the variable threshold is as follows:

[0026]

[0027] Among them, Th max and Th minC is the upper and lower limits of the trigger threshold, C is a counter that records the number of triggers received by the neuron, and p is a hyperparameter that controls the rate of change of the threshold.

[0028] Furthermore, the neuron model can be of different types, such as the LIF neuron model or the IF neuron model.

[0029] Furthermore, the L2 regularization method was used to control sparsity in the input layer and its associated weight matrix of the selectively activated SNNs model.

[0030] Furthermore, when training the selective activation SNNs model, during the training process of each learning stage, the cross-entropy between the output and the true label is calculated as the loss function, and STBP (spatio-temporal backpropagation) is used to train the learnable parameters in the network.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. This invention, based on the unique structure and spatiotemporal dynamics of spiking neural networks (SNs), designs a continuous learning method for SNs from the perspective of structure-adaptive selective activation. The accuracy curve of the model in this invention maintains a relatively balanced performance across various tasks during continuous learning. This balance between different tasks benefits from the model's selective activation of specific subgroups of neurons and their related connections. Therefore, the spiking neural network model of this invention exhibits stronger stability compared to other baseline methods in multi-task continuous learning scenarios.

[0033] 2. In the model proposed in this invention, the selective distribution of spiking neurons in response to different tasks is more uniform during the learning process. This prevents the selective spiking neurons for certain tasks from being too few or too many, which would interfere with the overall continuous learning process. Therefore, the method of this invention has been proven to be more robust.

[0034] 3. Compared to K-WTA based on pulse firing frequency, the Trace-based K-WTA proposed in this invention has a certain tolerance for neurons with similar functions, and therefore has better mixed selectivity between different tasks. Thus, the method of this patent can still maintain a relatively uniform selective distribution. This invention applies a variable threshold component to the spiking neural network, which to a certain extent inhibits the reactivation of neurons that have been activated during the learning process of old tasks, and promotes the activation of specific neuron clusters during the training process of new tasks, thereby avoiding catastrophic forgetting in the continuous learning process. Attached Figure Description

[0035] Figure 1This is a framework diagram of a target recognition system based on selective activation spiking neural network continuous learning according to the present invention;

[0036] Figure 2 This is a schematic diagram illustrating the performance of the system of the present invention on the Cifar10 dataset;

[0037] Figure 3 This invention provides an embodiment for analyzing the selectivity of neurons for different tasks. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0039] A target recognition system based on selective activation spiking neural networks through continuous learning, the structure of which is as follows: Figure 1 As shown, the system includes a feature extraction part at the front end of a spiking neural network, used to mine deep features in complex tasks, and a continuous learning part at the back end based on Trace-based K-WTA and variable threshold components, enabling the network to have continuous learning capabilities.

[0040] The Trace-based K-WTA mechanism applied to the dynamic changes of hidden neurons in SNNs reduces mutual interference between different tasks. Simultaneously, neurons with variable firing thresholds encourage silent neurons to participate in learning to some extent, while inhibiting the reactivation of neurons already activated during training on older tasks. This invention is applicable to scenarios such as class-IL, task-IL, and domain-IL.

[0041] This invention takes the most challenging incremental scenario as an example, and its implementation steps are as follows:

[0042] Assume this type of incremental learning process includes N learning stages (i.e., a total of N sequential tasks). Each stage's task contains all c... i Training sample data D for each category i , where i represents the index of the learning stage. During the training phase, each learning stage (e.g., the i-th stage, i = 1, 2, ..., N) can only access data relevant to the current task (i.e., D). i It belongs to category c i All data (including all data) are collected and the model is trained using the method proposed in this patent. Once training is complete, the model will be evaluated on a test set containing all categories across all learning stages.

[0043] The key technologies of the selective activation-based spiking neural network model involved in this invention are as follows:

[0044] (1) Trace-based K-WTA method

[0045] As some studies (Lin et al. 2014) and (Stevens 2015) have noted, inhibitory neurons exist in animal neural circuits. These neurons collect the excitation of a subset of neurons and emit feedback inhibition, ultimately preventing most neurons from firing. This mechanism is often referred to as the "winner-takes-all" (WTA) mechanism, which may help maintain the robustness of the network (this mechanism is also known as K-WTA, where K represents the number of winners). However, the homogeneous firing of neurons within a time window in SNNs makes it difficult to make appropriate direct comparisons between neurons.

[0046] Directly integrating pulses across the entire time window before deploying the K-WTA mechanism may not be reasonable. Firstly, due to the discrete nature of pulses, many neurons may fire the same number of pulses (especially with relatively small time steps), thus easily leading to confusion. Secondly, using future information to generate gating signals for past events seems unreasonable. During the research, it was noted that synaptic time-related plasticity updates weights based on the synaptic time interval between presynaptic and postsynaptic neurons. To bridge the gap between the time scale and action potential in plasticity theory, an internal variable called "Trace" is introduced for neurons. Since this variable can estimate the average firing rate in the pulse sequence online, it can also be used as an indicator of K-WTA in the time dimension. The calculation method for the trajectory of the aforementioned spiking neurons is as follows:

[0047]

[0048] Where τ is a time constant that determines the decay rate of the trace; tr[t] is the trace of the neuron at time step t; S[t+1] represents the pulse output of the neuron at step t; the trace can be calculated at each time step and is relatively easy to compare and obtain the Top-K value, so it is applied to the stepwise deployment K-WTA calculation. Figure 1 Figure (A) shows a possible trajectory example and its corresponding Top-1 selection. Another potential benefit of this trajectory-based K-WTA method is that it does not strictly adhere to the constraints of K from the perspective of the entire time window, which may improve the expressiveness of the subnetwork under the Top-K function.

[0049] (2) Variable threshold

[0050] Top-K activation functions often lead to the death of a large number of neurons (Ahmad and Scheinkman, 2019; Fedus, Zoph, and Shazeer, 2022). This is because randomly initialized weights can easily activate a group of neurons whose synaptic weights are constantly updated, while other neurons are never activated and therefore never receive feedback signals. This invention proposes using a variable pulse firing threshold for spiking neurons in the hidden layer. Figure 1 As shown in (B), the pulse firing threshold is set to increase slowly with increasing activation time. It is worth noting that the pulse firing threshold in this invention does not decay, thus enhancing the maintenance of memories related to old tasks through irreversible threshold changes, while being simple to implement and computationally efficient. Through this method, the probability of reactivating the most frequently activated neurons decreases, making it more likely that neurons with relatively low activation frequencies will be activated and gradually participate in the network's learning process. Adding these two features to the basic neuron model yields the following neuron dynamics model:

[0051] H(t) = f(V[t-1], X[t]),

[0052] S[t]=Θ(H(t)-V th ),

[0053] Mask[t] = TopK(tr[t]),

[0054] S * [t] = S[t]·Mask[t],

[0055] V[t]=H[t]-V th ·S[t],

[0056] Where X[t] is the neuron's input at time step t; S[t] represents the neuron's original pulse output; V[t] and H(t) are the membrane potentials before and after the neuron fires a pulse, respectively; Θ(-) is the function that controls the firing of a pulse when the membrane voltage exceeds a threshold; TopK(-) is the function used in the trajectory-based K-WTA component to generate a mask to obtain the mask corresponding to the top K largest trajectories; S * This represents the actual activation pulse output after the TopK function is applied; f(·) represents the neuron's state update equation. The method of this invention is applicable to different types of neuron models, such as the Leaky integrate-and-fire (LIF) neuron model or the Integrate-and-fire (IF) neuron model. The variable threshold of the neuron model proposed in this invention is implemented using a linearly variable or other nonlinear mapping function, such as... Figure 1 As shown in the middle (B left). Taking a simple linear variable method as an example, the calculation method of the variable threshold is as follows:

[0057]

[0058] Among them, Th max and Th min C is the upper and lower limits of the trigger threshold, C is a counter that records the number of triggers received by the neuron, and p is a hyperparameter that controls the rate of change of the threshold.

[0059] While the two key methods mentioned above can facilitate the formation of subnetworks and guide neurons to participate in continuous learning, excessively dense inputs can still lead to performance anomalies. This is because the greedy kernel of backpropagation and the limited precision and activation of SNNs can cause these neurons to fire spikes too frequently, thus impairing memory retention. Therefore, this invention employs an L2 regularization method to control sparsity in the input layer and its associated weight matrix. Furthermore, this invention applies several proven continuous learning techniques, such as Dale rules and the SGD optimizer, to avoid the "stale momentum" problem. During training at each learning stage, the SNN uses a training method consistent with the standard image classification algorithm for SNNs. That is, this invention calculates the cross-entropy between the output and the true label as the loss function and uses STBP (spatio-temporal backpropagation) to train the learnable parameters in the network.

[0060] This invention proposes a continuous learning target recognition system based on selective activation spiking neural networks, applicable to image target recognition using traditional cameras and temporal target recognition using dynamic event cameras. To verify the effectiveness of the proposed framework in continuous learning problems, the original image recognition dataset CIFAR10 was randomly divided into five tasks, each containing two classes, hereinafter referred to as the "splitCIFAR10" dataset. As described above, each model was trained sequentially on these tasks. This invention uses a pre-trained feature extractor from (Bricken et al. 2023) to transform each sample in the CIFAR dataset into a 256-dimensional latent embedding. Furthermore, this invention evaluates the model's performance on the splitMNIST, splitN-MNIST, and split-CIFAR100 datasets. The model was trained on each subset of the dataset, with a batch size of 256 samples and 500 training epochs. Performance tests were performed on each class learned by the model. The final accuracy is the average of three experimental results using a random seed. In Trace-based K-WTA, the K value is set to 10, and the p value is set to 2*1e6.

[0061] As shown in Table 1, the proposed SASNN achieves accuracies of 60.06% and 77.73% on the splitMNIST and splitCIFAR10 datasets, respectively, outperforming other baseline methods.

[0062] Table 1

[0063]

[0064] In addition, from Figure 2 As can be seen, the method proposed in this invention maintains a higher accuracy rate compared to other methods throughout the entire multi-task continuous learning process. Furthermore, the SA-SNN model also demonstrates superior performance compared to two advanced ANN models with similar continuous learning principles (i.e., SDMLP (73.27%) and FlyModel (70.09%)). Moreover, even when directly utilizing the rate mask (denoted by SA-SNN(rate)), the model proposed in this invention exhibits relatively high performance, with an accuracy higher than the SDMLP algorithm (76.78%). This proves that the method proposed in this invention, by considering the dynamic computation of neurons at different time steps in the spiking neural network, can effectively improve the ability of the spiking neural network to handle multi-task continuous learning tasks. In particular, the introduction of the Trace-based K-WTA mechanism enables the spiking neural network model to better mitigate the catastrophic forgetting problem.

[0065] At the same time, such as Figure 2 As shown, the accuracy variation curve of the model of the present invention maintains a relatively balanced performance across tasks during continuous learning. This balance between different tasks benefits from the model's selective activation of specific subgroups of neurons and their related connections. Therefore, the spiking neural network model of the present invention exhibits stronger stability and robustness compared to other baseline methods in multi-task continuous learning scenarios.

[0066] Furthermore, the spiking neural network model of this invention does not utilize the importance of weights, such as by introducing a weight regularization term to avoid forgetting. Therefore, it is compatible with weight regularization-based methods (such as EWC, MAS, and SI), which typically achieve this by adding a penalty term to the loss function. As shown in Table 1, when the method of this invention is used in conjunction with EWC, the spiking neural network model of this invention improves the accuracy by approximately 2.66% and 22.12% on the splitCIFAR10 and splitMNIST datasets, respectively.

[0067] Finally, the selectivity of neurons for different tasks in the spiking neural network model of this invention was analyzed.

[0068] This selectivity refers to the class selectivity of a neuron when it primarily responds to inputs of only one class. To coarsely classify neurons with different selectivity, using splitCIFAR10 as an example, this invention regresses hidden layer activity against the different class selectivity (i.e., neurons responding only to specific classes) of 10 expected datasets. "Activity" refers to the neuron's output (after ReLU), while in SNNs it refers to the number of impulses. One result is as follows... Figure 3 As shown. In Figure 3 As can be seen in the following section, the distribution of neurons with different selectivity is relatively biased in the SDMLP method, while in the method of this invention, the selective distribution of neurons facing different tasks is more uniform during the learning process. This phenomenon is also directly reflected in... Figure 3 In the final confusion matrix of the model (above), the proportion of neurons with selectivity for categories 2, 4, and 5 is relatively small during the learning process of the SDMLP model. Therefore, the ability to identify these categories in subsequent learning processes is easily interfered with, ultimately leading to relatively low accuracy for several categories. However, using the method of this invention, due to the stronger robustness of the spiking neural network method designed in this invention, the method of this invention is largely unaffected by the aforementioned interference. Furthermore, compared to using trace-based K-WTA masks at different time steps, using a pulse firing frequency-based mask results in a larger number of neurons with specific selectivity, but this invention exhibits a more stable uniform distribution. This may be because multi-step masks have a certain tolerance for neurons with similar functions, thereby encouraging neurons to exhibit mixed selectivity for different tasks, especially similar tasks. Even so, this invention still maintains a relatively uniform selectivity distribution.

[0069] This invention explores how to mitigate catastrophic forgetting by enhancing the neural dynamics of Smart Learning Networks (SNNs). A Selective Activation SNN (SA-SNN) model is proposed, employing a trajectory-based K-Winner-Take-All (Trace-based K-WTA) and variable thresholding mechanism for continuous learning. This model alleviates catastrophic forgetting by enhancing the neural dynamics of the SNN, without requiring task labels or memory replay. In the SA-SNN model, a biologically feasible time-trajectory-based K-WTA method is first employed to reduce interference between different tasks. The trajectory-based K-WTA method itself converges in connectivity with many brain regions using inhibitory interneurons; this method is further modified to accommodate spiking neurons operating across multiple time steps. Then, a simple and effective variable thresholding method is designed to modify the threshold of spiking neurons, thereby encouraging all neurons to participate and enhancing the gating effect of the population of neurons when facing different tasks. This invention is applicable to scenarios such as class increment (Class-IL), task increment (Task-IL), and domain increment (Domain-IL).

[0070] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A target recognition system based on selective activation spiking neural network continuous learning, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that: The computer memory contains a trained selective activation SNN model. The selective activation SNNs model includes a feature extraction part at the front end of a spiking neural network, and uses a trajectory-based K-WTA mechanism and a variable threshold mechanism at the back end for continuous learning; Specifically, a trajectory-based K-WTA mechanism is applied to the dynamic changes of neurons in the hidden layer of a spiking neural network to reduce mutual interference between different tasks. At the same time, a variable spiking threshold is used for neurons in the hidden layer to encourage silent neurons to participate in learning and to inhibit the reactivation of neurons that have been activated during training of old tasks. The trajectory-based K-WTA mechanism specifically involves introducing an internal variable called "Trace" into the neuron as an indicator of K-WTA in the time dimension. The trajectory calculation method for the spiking neuron is as follows: in, It is the time constant, which determines the decay rate of the trace; The neuron at a time step of The trace of time; This indicates that the neuron is at a step size of Pulse output at time; When using a variable spike firing threshold on neurons in the hidden layer, the spike firing threshold is set to increase slowly as the activation time increases. The spike firing threshold does not decay, and the ability to maintain old task-related memories is enhanced through irreversible threshold changes. When the computer processor executes the computer program, it performs the following steps: The image to be identified is input into a trained selective activation SNN model to obtain the recognition and classification results.

2. The target recognition system based on selective activation spiking neural network continuous learning according to claim 1, characterized in that, By incorporating trajectory-based K-WTA and variable threshold mechanisms into neurons, the following neuronal dynamics model is obtained: in, It is the neuron in time step Input on; This represents the original impulse output of the neuron; and These are the membrane potentials before and after the neuron fires its pulse, respectively. (-) is a function that controls the pulse to be emitted when the membrane voltage exceeds a threshold; (-) is a function used in the trajectory-based K-WTA component to generate a mask that corresponds to the K largest trajectories; express The actual activation pulse output after the function is applied; (·) represents the state update equation of a neuron; This represents the variable threshold of the neuron model.

3. The target recognition system based on selective activation spiking neural network continuous learning according to claim 2, characterized in that, Variable threshold of neuron model This can be achieved using linear variable or other nonlinear mapping functions. Taking the linear variable method as an example, the calculation method for the variable threshold is as follows: in, and These are the upper and lower limits of the trigger threshold. It is a counter that records the number of times a neuron receives a trigger. It is a hyperparameter that controls the rate of change of the threshold.

4. The target recognition system based on selective activation spiking neural network continuous learning according to claim 2, characterized in that, The neuron model is either the LIF neuron model or the IF neuron model.

5. The target recognition system based on selective activation spiking neural network continuous learning according to claim 1, characterized in that, L2 regularization was used to control sparsity in the input layer and its associated weight matrix of the selectively activated SNNs model.

6. The target recognition system based on selective activation spiking neural network continuous learning according to claim 1, characterized in that, When training a selectively activated SNN model, the cross-entropy between the output and the true label is calculated as the loss function during the training process of each learning stage, and STBP is used to train the learnable parameters in the network.

Images