Liquid state machine dynamics optimization system and method based on self-feedback mask

Abstract

The invention discloses a liquid state machine dynamics optimization system based on a self-feedback mask, and the system comprises a liquid state machine reservoir which is provided with a plurality of spike neurons; the liquid state machine reservoir updates the membrane potential state of the spiking neuron by receiving the time sequence input signal; the mask generation module is connected with a membrane potential state of a spiking neuron in a reservoir of the liquid state machine and is used for generating a self-feedback mask according to the membrane potential state; and the feedback adjusting unit is respectively connected with the mask generation module and the liquid state machine reservoir and is used for adjusting the circulating connection and the self-feedback calculation of the liquid state machine reservoir according to the self-feedback mask. The dynamic stability of the liquid state machine and the sufficiency of time sequence information utilization can be optimized, and the performance and robustness of the liquid state machine can be improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of neuromorphic computing and artificial intelligence, and particularly relates to a liquid state machine dynamics optimization system and method based on self-feedback mask. BACKGROUND

[0002] The existing liquid state machine (LSM) still has some optimization spaces worth attention in the field of time series signal processing, especially in the application of biological electrical signal (such as electrocardiogram ECG) analysis: Firstly, the reservoir dynamics may face stability challenges under continuous input. The neuron reservoir of the existing LSM is prone to enter a saturated state when processing continuous and information-rich analog signals. Network saturation will reduce the sensitivity of the reservoir to subtle changes in the signal and lose dynamic computing ability, thereby seriously affecting the performance of feature extraction and classification. The traditional method of maintaining network stability by pre-setting a fixed excitation-inhibition connection ratio couples feature mapping and network stability in the reservoir weight, which is a static and suboptimal compromise solution that limits the expression ability of the reservoir and leads to the need to balance feature extraction and stability maintenance, thereby failing to fully exert the computing ability.

[0003] Secondly, there is potential for improvement in the spatiotemporal dimension utilization of time series feature extraction. The traditional LSM readout strategy usually only utilizes the final state of the reservoir after processing a time window for judgment and classification. This readout method ignores the rich temporal evolution information excited in the reservoir during the entire processing process. A large amount of intermediate state that can represent the dynamic characteristics of the signal is discarded, resulting in insufficient capture of long-time series dependence and limiting the final performance and robustness of the model. SUMMARY

[0004] The purpose of the present application is to provide a liquid state machine dynamics optimization system and method based on self-feedback mask. The present application can optimize the stability of liquid state machine dynamics and the sufficiency of time series information utilization, and can improve the performance and robustness of the liquid state machine.

[0005] The technical solution of the present application: a liquid state machine dynamics optimization system based on self-feedback mask, comprising: a liquid state machine reservoir provided with a plurality of spike neurons; the liquid state machine reservoir updates the membrane potential state of the spike neurons by receiving a time series input signal; a mask generation module connected to the membrane potential state of the spike neurons in the liquid state machine reservoir, for generating a self-feedback mask according to the membrane potential state; a feedback adjustment unit, connected with the mask generation module and the liquid state machine reservoir respectively, for adjusting the recurrent connection and self-feedback calculation of the liquid state machine reservoir according to the self-feedback mask.

[0006] The self-feedback mask-based liquid state machine dynamics optimization system as claimed in the preceding claim, wherein the mask generation module comprises: an encoder for compressing the membrane potential state into a low-dimensional feature vector; a decoder for expanding the low-dimensional feature vector into the self-feedback mask.

[0007] The self-feedback mask-based liquid state machine dynamics optimization system as claimed in the preceding claim, further comprising a readout module for recording the spike output of the liquid state machine reservoir at each time step within a preset processing time window, splicing all the spike outputs in chronological order to form a single enhanced feature vector, and finally inputting the enhanced feature vector into a classifier to obtain a final classification result.

[0008] The method of the self-feedback mask-based liquid state machine dynamics optimization system as claimed in the preceding claim, comprising the following steps: Step 1: receiving a time sequence input signal through the liquid state machine reservoir; Step 2: updating the membrane potential state of the spiking neuron in the liquid state machine reservoir according to the time step in the time sequence input signal; Step 3: inputting the membrane potential state into the mask generation module, and generating a corresponding feedback gain value for the spiking neuron by the mask generation module, thereby forming a self-feedback mask; Step 4: applying the self-feedback mask to the recurrent connection and self-feedback calculation of the liquid state machine reservoir at the subsequent time step in the feedback adjustment unit to adjust the dynamic behavior of the spiking neuron; Step 5: generating a spike output sequence according to the adjusted dynamic behavior of the spiking neuron and performing classification.

[0009] The method of the self-feedback mask-based liquid state machine dynamics optimization system as claimed in the preceding claim, wherein the discrete-time dynamic equation for updating the membrane potential state of the spiking neuron in the liquid state machine reservoir is as follows: ; ; wherein, is the membrane potential of the spiking neuron the residual value obtained after the decay of the membrane potential at the previous time step, is the spike firing state of the spiking neuron at the previous time step, is the membrane potential decay factor, is the current from the input layer, It is the circulating current originating from within the liquid state machine reservoir. Indicated in each spike neuron At time step The membrane potential; When membrane potential Exceeding the threshold At that time, the spike neuron fires a spike. At the same time, the membrane potential of the spike neuron is reset to 0, as shown in the following formula: .

[0010] In the aforementioned method for optimizing the dynamics of a liquid state machine based on a self-feedback mask, the process of forming the self-feedback mask involves compressing the membrane potential state into a low-dimensional feature vector, and then expanding the low-dimensional feature vector to obtain the feedback gain value corresponding to the peak neuron generation. The specific formula is as follows: ; ; In the formula, This represents the generated low-dimensional feature vector. , , and These are the learnable parameters of the mask generation module. This represents the membrane potential vector of all spike neurons in the liquid state machine reservoir. Represents the mask vector. Each element in The feedback gain corresponds to a spike neuron.

[0011] In the aforementioned method for optimizing the dynamics of a liquid state machine based on a self-feedback mask, the calculation formula for applying the self-feedback mask to the cyclic connection and self-feedback calculation of the liquid state machine reservoir in subsequent time steps is as follows: ; In the formula, Represents the next time step The circulating current, This represents the pre-defined cyclic connection weight matrix within the reservoir of the liquid state machine. Represents all spike neurons At the current peak distribution stage, This represents element-wise multiplication. This represents the mask vector.

[0012] In the aforementioned method for optimizing a liquid state machine dynamics system based on a self-feedback mask, the process of generating the peak output sequence employs a full-time peaking strategy. After the input window of one time step is processed, the spike firing sequence of all the spiking neurons in the liquid state machine reservoir is recorded, and then spliced into a single, high-dimensional augmented feature vector The formula is as follows: ; The classification is performed through a trainable linear readout layer The augmented feature vector The formula is as follows: ; In the formula, represents the feature vector before classification after linear readout layer processing of the full-time spike sequence.

[0013] Compared with the prior art, the present application has the following beneficial effects: The present application solves the saturation and stability problems of the existing liquid state machine in processing continuous signals by introducing a liquid state machine based on membrane potential and self-feedback mask, and constructs an end-to-end processing framework with high performance, high robustness and suitable for neuromorphic hardware deployment. The self-feedback mask learned by the present application through data-driven learning can actively maintain the dynamic balance of the liquid state machine reservoir network, effectively prevent network saturation, and make the liquid state machine reservoir always in a high-efficiency computing state sensitive to the details of the input signal. The present application decouples the functions of feature mapping (liquid state machine reservoir weight) and network stability (self-feedback mask), so that the liquid state machine reservoir can focus on complex feature extraction, thereby significantly improving the classification performance of the entire system. In addition, the present application fully utilizes the dynamic information generated by the liquid state machine reservoir in the entire time evolution process through the full-time spike readout strategy, avoids the information loss of traditional methods, not only improves the ability to capture long-term dependencies, but also greatly reduces the sensitivity of the model to liquid state machine reservoir superparameters, and has stronger robustness. The entire framework of the present application is set as a lightweight computing module, which is suitable for efficient deployment on low-power neuromorphic hardware with limited resources, and provides a feasible technical path for real-time, online biological signal processing. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is a schematic diagram of the present application; Figure 2 is a schematic diagram of the baseline liquid state machine model without an excitation-inhibition balance self-feedback mask for comparison with the present application; Figure 3 is a schematic diagram of the calculation ability and attenuation coefficient of the present application; Figure 4 is a heart rate preprocessing data graph of the present application; Figure 5 ​Figure 1 shows the loss curve (cross-entropy loss) of the self-feedback mask based liquid state machine model of the present application during the training process as a function of the training epoch; Figure 6 Figure 2 shows the accuracy curve (classification accuracy) of the self-feedback mask based liquid state machine model of the present application during the training process as a function of the training epoch; Figure 7 Figure 3 shows the effect of different numbers of reservoir neurons on the recall rate of the model of the present application at a fixed training epoch; Figure 8 Figure 4 shows the normalized confusion matrix for the five-class classification of heart rate abnormalities of the present application; Figure 9 Figure 5 shows a visualization of the activation level (network saturation level) of neurons for different synaptic weight scaling parameters for the baseline liquid state machine model as a comparison to the present application; Figure 10 Figure 6 shows a visualization of the effect of the self-feedback mask based liquid state machine model on alleviating network saturation. DETAILED DESCRIPTION

[0015] The present application will be further described below in conjunction with the accompanying drawings and examples, but should not be limited as a basis for the present application.

[0016] Example 1: A self-feedback mask based liquid state machine dynamics optimization system, comprising an input layer, a liquid layer, and a readout layer, the input layer is responsible for receiving the original continuous analog ECG signal. The signal is segmented using the sliding window method, and the signal in the window is preprocessed through a fixed linear transformation layer, then the processed continuous value is taken as the input current fed to the liquid layer, this process does not require spike encoding. The liquid layer as the core computing unit, the liquid layer includes a liquid state machine reservoir containing N leaky integrate-and-fire (LIF) spiking neurons, a mask generation module for excitation-inhibition balance, and a feedback adjustment unit. In this example, the liquid state machine reservoir is provided with a plurality of spiking neurons, and the liquid state machine reservoir updates the membrane potential state of the spiking neurons by receiving the time sequence input signal. The mask generation module is connected to the membrane potential state of the spiking neurons in the liquid state machine reservoir, and is used to generate a self-feedback mask according to the membrane potential state. The function of the mask generation module is to generate a self-feedback gain for each spiking neuron according to the current global state of the liquid state machine reservoir, which is used to adjust the circulating current , and specifically implemented as a two-layer fully connected network, including an encoder and a decoder, the encoder is used to compress the membrane potential state into a low-dimensional feature vector; the decoder is used to expand the low-dimensional feature vector into the self-feedback mask. The feedback adjustment unit is connected with the mask generation module and the liquid state machine reservoir respectively, and is used to adjust the cycle connection and self-feedback calculation of the liquid state machine reservoir according to the self-feedback mask. The readout layer is responsible for interpreting the spike sequence generated by the liquid layer and classifying, this layer adopts the full-time spike strategy, and inputs the final result into a trainable linear classifier. The readout module is integrated in the readout layer, and the readout module is used to record the spike output of the liquid state machine reservoir at each time step within a preset processing time window, and splice all spike outputs recorded within the preset processing time window in time sequence to form a single enhanced feature vector, and finally input the enhanced feature vector into the trainable classifier to obtain the final classification result.

[0017] The liquid state machine model with self-feedback mask (LHSF) in the embodiment is implemented using the PyTorch2.2.0 framework. The optimal hyperparameters are determined through an initial optimization process, which involves systematically changing each parameter one by one while keeping other parameters unchanged. The final key settings are summarized in Table 1. All computational experiments were performed on a workstation equipped with an Intel Core i5-13400 CPU, 32GB of RAM, and an NVIDIA GeForce RTX 4070 GPU.

[0018] Table 1. Hyperparameter settings of the model

[0019] Embodiment 2: Based on the system in embodiment 1, this embodiment takes the ECG signal classification of MIT-BIH arrhythmia database as the application scenario, as shown in Figure 1 , specifically including the following steps: Step one: receiving a time sequence input signal through a liquid state machine reservoir; In this embodiment, the time sequence input signal is a biological electrical signal, specifically an electrocardiogram (ECG) signal. The source of the time sequence input signal uses the MIT-BIH arrhythmia database, which contains 48 half-hour electrocardiogram records sampled at 360Hz. According to the standard AAMI recommendation, records with paced heartbeats are excluded, and the original annotations are mapped to 5 heart beat categories, as shown in Table 2. For all electrocardiogram analysis, only the modified limb lead II (MLII) signal is used as input.

[0020] Table 2. Correspondence between MIT-BIH annotations and AAMI beat classes

[0021] A standard pre-processing pipeline was applied for the primary arrhythmia classification task using the MIT-BIH dataset. First, for each annotated heartbeat, a 180-sample segment centered at the R-peak was extracted, as shown in Figure 4 Subsequently, a sliding window approach was used to divide these segments into overlapping sub-segments. The specific length of these windows (window size) and the degree of overlap (step size) are hyperparameters of the self-feedback mask liquid state machine, which were adjusted during the optimization and analysis phase. The above segments were divided into overlapping sub-segments using a sliding window approach, with the window size and step size being hyperparameters of the self-feedback mask liquid state machine, which were adjusted during the subsequent optimization phase (optimal values are shown in Table 1).

[0022] The MIT-BIH arrhythmia database exhibits severe class imbalance, with the normal beat (N) class containing 83,268 samples, while the underrepresented classes (e.g., fused beats (F)) contain only 801 samples. A balanced training dataset was created by applying random oversampling to the minority classes and random undersampling to the majority classes, mitigating potential bias of the classifier towards the majority classes. The final balanced training dataset used for training consisted of 5,000 instances from each of the five AAMI classes, totaling 25,000 instances.

[0023] Step two: update the membrane potential state of the spiking neuron in the liquid state machine reservoir according to the time step in the time-series input signal; In this step, the discrete-time dynamic equation for updating the membrane potential state of the spiking neuron in the liquid state machine reservoir is as follows: ; ; where is the residual value of the spiking neuron after the decay of the membrane potential at the previous time step, is the spiking state of the spiking neuron at the previous time step, is the membrane potential decay factor, is the input current from the input layer, is the recurrent current from within the liquid state machine reservoir, denotes the membrane potential of each spiking neuron at time step ; When the membrane potential exceeds a threshold , the spiking neuron fires a spike , while the membrane potential of the spiking neuron is reset to 0, as follows: ; Since the spiking function in the formula is not derivable, during the backpropagation process of BPTT, a surrogate gradient strategy is adopted, that is, when calculating the gradient, the derivative of the spiking function is approximated as a continuous function (such as a rectangular window function) that has a non-zero value near the threshold . This enables the gradient to be propagated back from the final classification loss function to the excitation-inhibition balance self-feedback mask and readout layer, thereby enabling effective optimization of their parameters.

[0024] Step three: input the membrane potential state into the mask generation module, and the mask generation module generates the corresponding feedback gain value for the spiking neuron, thereby forming a self-feedback mask; In this step, the process of forming the self-feedback mask is to compress the membrane potential state into a low-dimensional feature vector, and then expand the low-dimensional feature vector to obtain the corresponding feedback gain value for the spiking neuron, and the specific formula is as follows: ; ; In the formula, represents the generated low-dimensional feature vector, , , and are learnable parameters of the mask generation module, represents the membrane potential vector of all spiking neurons in the liquid state machine reservoir, represents the mask vector, each element in the mask vector corresponds to the feedback gain of a spiking neuron.

[0025] Step four: in the feedback adjustment unit, apply the self-feedback mask to the recurrent connection and self-feedback calculation of the liquid state machine reservoir in the subsequent time step to adjust the dynamic behavior of the spiking neuron; In this step, the calculation formula for applying the self-feedback mask to the recurrent connection and self-feedback calculation of the liquid state machine reservoir in the subsequent time step is as follows: ; In the formula, represents the recurrent current of the next time step, represents the preset recurrent connection weight matrix inside the liquid state machine reservoir, represents the recurrent current of the next time step, represents all spiking neurons in the current time step, represents element-wise multiplication, denotes the mask vector.

[0026] Step five: generate spike output sequences and perform classification according to the dynamic behavior of the adjusted spiking neurons.

[0027] In this step, the process of generating spike output sequences adopts a full-time spike strategy, and after the input window of time steps is processed, the spike firing sequence of all spiking neurons in the liquid state machine reservoir is recorded, and then spliced into a single, high-dimensional enhanced feature vector , as follows: ; The classification is performed by a trainable linear readout layer on the enhanced feature vector , as follows: ; wherein represents the feature vector before classification of the full-time spike sequence after linear readout layer processing.

[0028] This embodiment adopts an end-to-end joint optimization training strategy, and the parameters of the self-feedback mask are obtained through end-to-end joint optimization learning with the readout layer parameters of the liquid state machine in the training phase. The parameters of the self-feedback mask , , and and the parameters of the readout layer are learned through a unified training process using the time backpropagation algorithm (Backpropagation Through Time, BPTT). In the inference phase of the mask generation module, the parameters of the self-feedback mask are fixed.

[0029] ​The application solves the saturation and stability problems of the existing liquid state machine in processing continuous signals by introducing a liquid state machine based on membrane potential and self-feedback mask, and builds an end-to-end processing framework with high performance, high robustness and suitable for neuromorphic hardware deployment. The self-feedback mask learned by data-driven can actively maintain the dynamic balance of the liquid state machine reservoir network, effectively prevent network saturation, and make the liquid state machine reservoir always in a high-efficiency computing state sensitive to the details of the input signal. The application decouples the functions of feature mapping (liquid state machine reservoir weight) and network stability (self-feedback mask), so that the liquid state machine reservoir can focus on complex feature extraction, thereby significantly improving the classification performance of the entire system. In addition, the application fully utilizes the dynamic information generated by the liquid state machine reservoir in the whole time evolution process through the all-time spike readout strategy, avoids the information loss of the traditional method, not only improves the ability to capture long-term dependencies, but also greatly reduces the sensitivity of the model to the liquid state machine reservoir parameters, and has stronger robustness. The whole framework of the application is set as a lightweight computing module, which is suitable for efficient deployment on low-power neuromorphic hardware with limited resources, and provides a feasible technical path for real-time and online biological signal processing.

[0030] Verification example: compare the baseline liquid state machine without excitation-inhibition balance self-feedback mask ( Figure 2 ) and the liquid state machine dynamics optimization system based on self-feedback mask ( Figure 3 ). The experimental results show that the generalization rank (GR) of the baseline liquid state machine increases sharply with the increase of the decay factor, indicating that the network tends to be saturated and chaotic. While the liquid state machine dynamics optimization system based on self-feedback mask of the application maintains a very low and stable GR value in the whole parameter range, which proves the significant effect of the excitation-inhibition balance self-feedback mask in actively maintaining the dynamic balance of the network and preventing saturation. The baseline liquid state machine only shows high computing ability (CA) in a narrow optimal parameter range. While the liquid state machine dynamics optimization system based on self-feedback mask of the application decouples the functions of feature mapping and network stability, it can maintain high computing ability and high expression ability (KQ) in a very wide parameter range, and the robustness is significantly enhanced.

[0031] Figure 5 and Figure 6 The loss curve (cross-entropy loss) and accuracy curve (classification accuracy) of the liquid state machine model based on self-feedback mask during the training process are shown, and the stable performance of the loss and accuracy curves in the figure proves the dynamic stability of the reservoir actively maintained by the self-feedback mask, which ensures that the liquid state machine model based on self-feedback mask can learn efficiently and stably during the whole training process.

[0032] Figure 7 This study demonstrates the impact of varying numbers of reservoir neurons (from 16 to 4096) on the recall of a self-feedback mask-based liquid state machine model under a fixed number of training epochs. It shows that the self-feedback mask-based liquid state machine model can maintain high performance with fewer neurons, thanks to its low sensitivity to changes in data size and the algorithm's strong stability. The effect of the number of neurons on model performance is similar to the biological characteristics of the brain; the more neurons in the reservoir, the more complex the temporal patterns and dynamic features the network can express, and the better the classification performance.

[0033] Figure 8 The normalized confusion matrix (true label vs. predicted label) of the liquid state machine model based on self-feedback mask on five types of cardiac beat classification tasks is shown. It shows extremely high diagonal values ​​and extremely low off-diagonal values, which directly proves the high performance and high accuracy of the classification effect achieved by the method of the present invention, especially its excellent performance in distinguishing clinically important premature beats and abnormal categories.

[0034] The self-feedback mask-based liquid state machine model achieved an accuracy of 98.79% on the MIT-BIH dataset. The performance improvement is attributed to its ability to alleviate network saturation within the reservoir. To investigate this mechanism, the global synaptic scaling factor (…) was systematically changed. A control experiment was conducted to adjust the input signal strength, and the results are as follows: Figure 9 and Figure 10 As shown.

[0035] Figure 9 This indicates that at high input scaling ( At input scales ≥0.6, the Baseline Liquid State Machine (LSM) exhibits significant saturation, characterized by most of the network reaching uniform maximum activation. This homogenization of neuronal activity indicates a compressed dynamic range. Conversely, even at the maximum input scale (≥0.6), the dynamic range remains relatively constant. At (=1.0), a liquid state machine based on a self-feedback mask ( Figure 10 It can also maintain the active mode. At this maximum scale, the saturation observed in the self-feedback mask-based liquid state machine is comparable to the saturation of the baseline liquid state machine at a significantly lower scale (approximately 0.4). These results intuitively confirm that the self-feedback mask-based liquid state machine maintains the dynamic range of the reservoir by preventing large-scale saturation.

[0036] To verify the robustness of the self-feedback mask-based liquid state machine model, we conducted a comprehensive hyperparameter analysis to compare its performance with that of the baseline liquid state machine model from both intrinsic neural dynamics parameters and extrinsic learning strategy parameters, as shown in Table 3, to evaluate the difference in sensitivity to parameter changes between the two types of models.

[0037] The intrinsic neural dynamics parameters mainly include parameters controlling the E-I balance ( , ), parameters determining the network structure (connection coefficient , reservoir neuron number ), and parameters affecting time integration (decay coefficient ). When these parameters undergo a large range of changes, the self-feedback mask-based liquid state machine model consistently maintains good performance stability, with its classification accuracy remaining above 0.90, indicating that its effectiveness does not depend on specific, finely tuned internal dynamics parameters and that it has strong adaptability to changes in intrinsic neural dynamics.

[0038] In sharp contrast, the performance of the baseline liquid state machine model is highly sensitive to intrinsic neural dynamics parameters: it can only achieve high classification accuracy within a very narrow parameter range (e.g., connection coefficient <0.2), and once the parameters exceed this range, its accuracy drops sharply, often falling below 0.60, showing strong dependence on intrinsic neural dynamics parameters.

[0039] The extrinsic learning strategy parameters mainly include window-related parameters (window size , step size ) and synaptic weight scaling parameters (a ). Experimental results show that the self-feedback mask-based liquid state machine model is not sensitive to the selection of these parameters: within most of the tested parameter ranges, its classification accuracy consistently remains close to the optimal level (about 0.95), without significant performance fluctuations due to parameter adjustments.

[0040] The baseline liquid state machine model again shows significant parameter sensitivity: its accuracy is closely related to the selection of extrinsic learning strategy parameters, for example, when the synaptic weight scaling parameter a >0.5, the classification performance of the baseline liquid state machine model will decrease significantly, and it cannot maintain stable accuracy within a wide parameter range.

[0041] Table 3 Summary of parameter robustness analysis data

[0042] In summary, this invention can optimize the stability of liquid state machine dynamics and the sufficiency of timing information utilization, thereby improving the performance and robustness of liquid state machines.

Claims

1. A self-feedback mask based liquid state machine dynamics optimization system, characterized by: The application relates to a liquid state machine reservoir, comprising: a liquid state machine reservoir provided with a plurality of spiking neurons; the liquid state machine reservoir updates the membrane potential state of the spiking neurons by receiving a time sequence input signal; a mask generation module connected with the membrane potential state of the spiking neurons in the liquid state machine reservoir, used for generating a self-feedback mask according to the membrane potential state; a feedback adjustment unit connected with the mask generation module and the liquid state machine reservoir respectively, used for adjusting the recurrent connection and self-feedback calculation of the liquid state machine reservoir according to the self-feedback mask.

2. The self-feedback mask based liquid state machine dynamics optimization system of claim 1, wherein: The mask generation module comprises: an encoder used for compressing the membrane potential state into a low-dimensional feature vector; a decoder used for expanding the low-dimensional feature vector into the self-feedback mask.

3. The self-feedback mask based liquid state machine dynamics optimization system of claim 1, wherein: The application further comprises a readout module used for recording the spiking output of the liquid state machine reservoir at each time step within a preset processing time window, splicing all the spiking outputs in time sequence to form a single enhanced feature vector, and finally inputting the enhanced feature vector into a classifier to obtain a final classification result.

4. The method of optimization of a self-feedback mask based liquid state machine dynamics optimization system according to any of claims 1-3, characterized in that, The application comprises the following steps: Step one: receiving a time sequence input signal through the liquid state machine reservoir; Step two: updating the membrane potential state of the spiking neurons in the liquid state machine reservoir according to the time step in the time sequence input signal; Step three: inputting the membrane potential state into the mask generation module, and the mask generation module generates corresponding feedback gain values for the spiking neurons to form a self-feedback mask; Step four: in the feedback adjustment unit, the self-feedback mask is applied to the recurrent connection and self-feedback calculation of the liquid state machine reservoir at the subsequent time step to adjust the dynamic behavior of the spiking neurons; Step five: generating a spiking output sequence according to the dynamic behavior of the adjusted spiking neurons and performing classification.

5. The method of optimization of a self-feedback mask based liquid state machine dynamics optimization system of claim 4, wherein: In step two, the discrete-time dynamic equation for updating the membrane potential state of the spiking neurons in the liquid state machine reservoir is as follows: ； ； wherein is a spiking neuron is the residual value obtained after decay of the membrane potential at the previous time step, is a spiking neuron is the spiking state at the previous time step, is a membrane potential decay factor, is an input current, is a circulating current from the reservoir interior of the liquid state machine, denotes the membrane potential at each spiking neuron at time step ; When the membrane potential exceeds a threshold the spiking neuron emits a spike while the membrane potential of the spiking neuron is reset to 0, as follows: 。 6. The method of optimization of a self-feedback mask based liquid state machine dynamics optimization system of claim 4, wherein: In step three, the process of forming the self-feedback mask is to compress the membrane potential state into a low-dimensional feature vector, and then expand the low-dimensional feature vector to obtain corresponding feedback gain values for the spiking neurons, and the specific formula is as follows: ； ； wherein, represents the generated low-dimensional feature vector, , , and are learnable parameters of the mask generation module, represents the membrane potential vector of all spiking neurons in the liquid state machine reservoir, represents the mask vector, each element in corresponds to the feedback gain of one spiking neuron.

7. The method of optimization of a self-feedback mask based liquid state machine dynamics optimization system of claim 4, wherein: In step four, the calculation formula for applying the self-feedback mask to the recurrent connection and self-feedback calculation of the liquid state machine reservoir at the subsequent time step is as follows: ； wherein represents the recurrent current of the next time step represents the recurrent current of the next time step represents the recurrent connection weight matrix preset inside the liquid state machine reservoir, represents all spiking neurons in the spiking firing state at the current time step, represents an element-wise multiplication, denotes a mask vector.

8. The method of optimization of a self-feedback mask based liquid state machine dynamics optimization system of claim 4, wherein: In step five, the process of generating spike output sequence adopts a full-time spike strategy, in which the input window of one time step is processed, and the spike firing sequence of all spike neurons in the liquid state machine reservoir is recorded , and then spliced into a single, high-dimensional enhanced feature vector , as follows: ； The classification is by a trainable linear readout layer The enhanced feature vector is computed as follows: The enhanced feature vector is computed as follows: ； In the formula, represents the feature vector before classification after the full-time spike sequence is processed by the linear readout layer.

Images