A self-feedback mask based liquid state machine dynamics optimization system and method

Abstract

This invention discloses a liquid state machine dynamics optimization system based on a self-feedback mask, comprising a liquid state machine reservoir with multiple spike neurons; the liquid state machine reservoir updates the membrane potential state of the spike neurons by receiving time-series input signals; a mask generation module, connected to the membrane potential state of the spike neurons in the liquid state machine reservoir, is used to generate a self-feedback mask based on the membrane potential state; a feedback adjustment unit, connected to both the mask generation module and the liquid state machine reservoir, is used to adjust the cyclic connections and self-feedback calculations of the liquid state machine reservoir based on the self-feedback mask. This invention can optimize the stability of liquid state machine dynamics and the sufficiency of time-series information utilization, thereby improving the performance and robustness of the liquid state machine.

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:

[0003] 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 driven by continuous input. 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.

[0004] 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 time 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 number of intermediate states that can represent the dynamic characteristics of the signal are discarded, resulting in insufficient capture of long-time series dependencies and limiting the final performance and robustness of the model. SUMMARY

[0005] 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.

[0006] The technical solution of the present application: a liquid state machine dynamics optimization system based on self-feedback mask, comprising:

[0007] A liquid state machine reservoir is 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;

[0008] A mask generation module is connected to the membrane potential state of the spike neurons in the liquid state machine reservoir and is used to generate a self-feedback mask according to the membrane potential state;

[0009] The feedback adjustment unit is connected to the mask generation module and the liquid state machine reservoir respectively, and is used to adjust the cyclic connection and self-feedback calculation of the liquid state machine reservoir according to the self-feedback mask.

[0010] In the aforementioned liquid state machine dynamics optimization system based on self-feedback masks, the mask generation module includes:

[0011] An encoder is used to compress the membrane potential state into a low-dimensional feature vector;

[0012] A decoder is used to expand the low-dimensional feature vector into the self-feedback mask.

[0013] The aforementioned liquid state machine dynamics optimization system based on self-feedback mask also includes a readout module. The readout module is used to record the peak output of the liquid state machine reservoir at each time step within a preset processing time window, and to concatenate all the peak outputs in chronological order to form a single enhanced feature vector. Finally, the enhanced feature vector is input into the classifier to obtain the final classification result.

[0014] The aforementioned method for optimizing a liquid state machine dynamics system based on a self-feedback mask includes the following steps:

[0015] Step 1: Receive timing input signals through the liquid state machine reservoir;

[0016] Step 2: Update the membrane potential state of the spike neurons in the liquid state machine reservoir according to the time step in the time sequence input signal;

[0017] Step 3: Input the membrane potential state into the mask generation module. The mask generation module generates the corresponding feedback gain value for the spike neuron, thereby forming a self-feedback mask.

[0018] Step 4: In the feedback adjustment unit, the self-feedback mask is applied to the cyclic connections and self-feedback calculations of the liquid state machine reservoir in subsequent time steps to adjust the dynamic behavior of the spike neurons;

[0019] Step 5: Generate and classify the spike output sequence based on the dynamic behavior of the spike neurons after regulation.

[0020] In the aforementioned method for optimizing the dynamics of a liquid state machine based on a self-feedback mask, the discrete-time dynamic equation for updating the membrane potential state of the peak neurons in the liquid state machine reservoir is as follows:

[0021] ;

[0022] ;

[0023] In the formula, It is a spike neuron The residual value obtained after the membrane potential decays in the previous time step. It is a spike neuron In the previous peak distribution phase It is the membrane potential decay factor. It 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;

[0024] 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:

[0025] .

[0026] 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:

[0027] ;

[0028] ;

[0029] 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 corresponding to a spike neuron.

[0030] 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:

[0031] ;

[0032] 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.

[0033] 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 at each time step has been processed, the spike firing sequence of all spike neurons in the liquid state machine reservoir is obtained. It is recorded and then concatenated into a single, high-dimensional enhanced feature vector. The formula is as follows:

[0034] ;

[0035] The classification is performed through a trainable linear readout layer. Enhance feature vectors The procedure is as follows:

[0036] ;

[0037] In the formula, This represents the feature vector of the full-time spike sequence before classification, after processing by the linear readout layer.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] This invention addresses the saturation and stability challenges of existing liquid state machines (LSMs) when processing continuous signals by introducing a liquid state machine based on membrane potential and a self-feedback mask. It constructs a high-performance, robust, and neuromorphic hardware-suitable end-to-end processing framework. Through data-driven learning of the self-feedback mask, this invention actively maintains the dynamic balance of the LSM reservoir network, effectively preventing network saturation and ensuring the LSM reservoir remains in a highly efficient computational state, highly sensitive to input signal details. This invention decouples feature mapping (LSM reservoir weights) from network stabilization (self-feedback mask), allowing the LSM reservoir to focus on complex feature extraction, thus significantly improving the overall system's classification performance. Furthermore, this invention fully utilizes the dynamic information generated by the LSM reservoir throughout its temporal evolution through a full-time spike readout strategy, avoiding information loss common in traditional methods. This not only improves the ability to capture long-term dependencies but also significantly reduces the model's sensitivity to LSM reservoir hyperparameters, resulting in greater robustness. The entire framework of this invention is designed as a lightweight computing module, which is suitable for efficient deployment on resource-constrained low-power neuromorphic hardware, providing a feasible technical path for real-time, online biosignal processing. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the present invention;

[0041] Figure 2 A schematic diagram showing the computational power and attenuation coefficient of a baseline liquid state machine model without an excitation-inhibition balance self-feedback mask, used in contrast to the present invention;

[0042] Figure 3 This is a schematic diagram illustrating the computing power and attenuation coefficient of the present invention;

[0043] Figure 4 This is a graph of heart rate preprocessing data from the present invention;

[0044] Figure 5 This is a graph showing the change of the loss curve (cross-entropy loss) of the liquid state machine model based on self-feedback mask in the training process of the present invention with the training rounds.

[0045] Figure 6 This is a graph showing the change in accuracy (classification accuracy) of the self-feedback mask-based liquid state machine model during the training process as a function of training rounds.

[0046] Figure 7 This is a graph showing the effect of different numbers of reservoir neurons on the model recall rate under a fixed number of training rounds in this invention;

[0047] Figure 8 This is the standardized confusion matrix diagram of the five-category classification of heart rate abnormalities according to the present invention;

[0048] Figure 9 A visualization of the activation level (network saturation) of neurons with different synaptic weight scaling parameters is shown for the baseline liquid state machine model compared with the present invention.

[0049] Figure 10 This is a visualization of how a liquid state machine model based on a self-feedback mask alleviates network saturation. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0051] Example 1: A liquid state machine dynamics optimization system based on a self-feedback mask, comprising an input layer, a liquid layer, and a readout layer. The input layer is responsible for receiving the original continuous analog ECG signal. A sliding window method is used to segment the signal, and a fixed linear transformation layer is used to preprocess the signal within the window. The processed continuous value is then used as the input current. The feed is sent to the liquid layer, a process that does not require spike encoding. The liquid layer serves as the core computational unit, comprising a liquid state machine reservoir containing N leak-integral-fire (LIF) spike neurons, an excitation-inhibition balance mask generation module, and a feedback regulation unit. In this example, the liquid state machine reservoir has multiple spike neurons, which update the membrane potential state of these neurons by receiving time-series input signals. The mask generation module is connected to the membrane potential state of the spike neurons in the liquid state machine reservoir and generates a self-feedback mask based on this state. The function of the mask generation module is to generate a self-feedback gain for each spike neuron based on the current global state of the liquid state machine reservoir. Used to regulate circulating current Specifically, it is implemented as a two-layer fully connected network, including an encoder and a decoder. The encoder compresses the membrane potential state into a low-dimensional feature vector; the decoder expands the low-dimensional feature vector into the self-feedback mask. A feedback adjustment unit is connected to the mask generation module and the liquid state machine reservoir, respectively, and is used to adjust the cyclic connections and self-feedback calculations of the liquid state machine reservoir according to the self-feedback mask. The readout layer is responsible for interpreting and classifying the spike sequences generated by the liquid layer. This layer adopts a full-time spike strategy and inputs the final result into a trainable linear classifier. The readout layer integrates a readout module, which records the spike output of the liquid state machine reservoir at each time step within a preset processing time window. All spike outputs recorded within the preset processing time window are concatenated in chronological order to form a single enhanced feature vector. Finally, the enhanced feature vector is input into the trainable classifier to obtain the final classification result.

[0052] The Liquid State Machine with Self-feedback Mask (LHSF) model in this embodiment is implemented using the PyTorch 2.2.0 framework. The optimal hyperparameters were determined through an initial optimization process, which involved systematically changing each parameter one by one while keeping other parameters constant. The final key settings are summarized in Table 1. All computational experiments were conducted on a workstation equipped with an Intel Core i5-13400 CPU, 32GB RAM, and an NVIDIA GeForce RTX 4070 GPU.

[0053] Table 1. Hyperparameter settings of the model

[0054]

[0055] Example 2: Based on the system in Example 1, this example uses ECG signal classification from the MIT-BIH arrhythmia database as an application scenario, such as... Figure 1 As shown, the specific steps include:

[0056] Step 1: Receive timing input signals through the liquid state machine reservoir;

[0057] In this embodiment, the timing input signal is a bioelectrical signal, specifically an electrocardiogram (ECG) signal. The timing input signal is sourced from the MIT-BIH arrhythmia database, which contains 48 half-hour ECG records sampled at 360Hz. Records with paced heartbeats were excluded according to the standard AAMI recommendations, and the original annotations were mapped to 5 heartbeat categories, as shown in Table 2. For all ECG analyses, only the modified limb lead II (MLII) signal was used as input.

[0058] Table 2. Correspondence between MIT-BIH annotations and AAMI beat categories

[0059]

[0060] For the primary arrhythmia classification task using the MIT-BIH dataset, a standard preprocessing workflow was applied. First, for each annotated heartbeat, 180 sample segments centered on the R peak were extracted, such as... Figure 4 As shown. Subsequently, a sliding window method is used to divide these segments into overlapping sub-segments. The specific length (window size) and overlap (step size) of these windows are hyperparameters of the self-feedback mask's liquid state machine, which will be adjusted during the optimization and analysis phases. The sliding window method is used to divide the above segments into overlapping sub-segments, with the window size ( ) and step size ( ) represents the hyperparameters of the liquid state machine of the self-feedback mask, which are adjusted in subsequent optimization stages (the optimal values ​​are shown in Table 1).

[0061] The MIT-BIH arrhythmia database exhibits severe class imbalance, with 83,268 samples in the normal heartbeat (N) class and only 801 samples in underrepresented classes (such as fused heartbeat (F)). A balanced training dataset was created by applying random oversampling to the minority classes and random undersampling to the majority classes to mitigate the 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, for a total of 25,000 instances.

[0062] Step 2: Update the membrane potential state of the spike neurons in the liquid state machine reservoir according to the time step in the time sequence input signal;

[0063] In this step, the discrete-time dynamic equation for the membrane potential state of the spike neurons in the updated liquid state machine reservoir is as follows:

[0064] ;

[0065] ;

[0066] In the formula, It is a spike neuron The residual value obtained after the membrane potential decays in the previous time step. It is a spike neuron In the previous peak distribution phase It is the membrane potential decay factor. It is the input 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;

[0067] 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:

[0068] ;

[0069] Since the spiked firing function in this formula is not differentiable, a surrogate gradient strategy is adopted during the backpropagation of BPTT. That is, when calculating the gradient, the derivative of the spiked function is approximated as a value at a threshold. The presence of continuous functions with non-zero values ​​in the vicinity (such as rectangular window functions) allows the gradient to propagate from the final classification loss function all the way back to the excitation-inhibition balanced self-feedback mask and readout layer, thus enabling efficient optimization of their parameters.

[0070] Step 3: Input the membrane potential state into the mask generation module. The mask generation module generates the corresponding feedback gain value for the spike neuron, thereby forming a self-feedback mask.

[0071] In this step, the process of forming a 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:

[0072] ;

[0073] ;

[0074] 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 corresponding to a spike neuron.

[0075] Step 4: In the feedback adjustment unit, the self-feedback mask is applied to the cyclic connections and self-feedback calculations of the liquid state machine reservoir in subsequent time steps to adjust the dynamic behavior of the spike neurons;

[0076] In this step, the calculation formula for applying the self-feedback mask to the liquid state machine reservoir in subsequent time steps for cyclic connection and self-feedback calculation is as follows:

[0077] ;

[0078] 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.

[0079] Step 5: Generate and classify the spike output sequence based on the dynamic behavior of the spike neurons after regulation.

[0080] In this step, the process of generating the spike output sequence adopts a full-time spike strategy. After the input window at each time step has been processed, the spike firing sequence of all spike neurons in the liquid state machine reservoir is obtained. It is recorded and then concatenated into a single, high-dimensional enhanced feature vector. The formula is as follows:

[0081] ;

[0082] The classification is performed through a trainable linear readout layer. Enhance feature vectors The procedure is as follows:

[0083] ;

[0084] In the formula, This represents the feature vector of the full-time spike sequence before classification, after processing by the linear readout layer.

[0085] This embodiment employs an end-to-end joint optimization training strategy. 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 during the training phase. The parameters of the self-feedback mask ( , , and ) and readout layer parameters ( The learning process utilizes a unified training procedure and the Backpropagation Through Time (BPTT) algorithm. During the inference phase of the mask generation module, the parameters of the self-feedback mask are fixed.

[0086] This invention addresses the saturation and stability challenges of existing liquid state machines (LSMs) when processing continuous signals by introducing a liquid state machine based on membrane potential and a self-feedback mask. It constructs a high-performance, robust, and neuromorphic hardware-suitable end-to-end processing framework. Through data-driven learning of the self-feedback mask, this invention actively maintains the dynamic balance of the LSM reservoir network, effectively preventing network saturation and ensuring the LSM reservoir remains in a highly efficient computational state, highly sensitive to input signal details. This invention decouples feature mapping (LSM reservoir weights) from network stabilization (self-feedback mask), allowing the LSM reservoir to focus on complex feature extraction, thus significantly improving the overall system's classification performance. Furthermore, this invention fully utilizes the dynamic information generated by the LSM reservoir throughout its temporal evolution through a full-time spike readout strategy, avoiding information loss common in traditional methods. This not only improves the ability to capture long-term dependencies but also significantly reduces the model's sensitivity to LSM reservoir hyperparameters, resulting in greater robustness. The entire framework of this invention is designed as a lightweight computing module, which is suitable for efficient deployment on resource-constrained low-power neuromorphic hardware, providing a feasible technical path for real-time, online biosignal processing.

[0087] Validation example: Baseline fluid 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 baseline liquid state machine was compared with the proposed system. 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 saturate and become chaotic. In contrast, the liquid state machine dynamic optimization system based on a self-feedback mask in this invention maintains an extremely low and stable GR value across the entire parameter range, demonstrating the significant effect of the excitation-inhibition balance self-feedback mask in actively maintaining network dynamic equilibrium and preventing saturation. The baseline liquid state machine only exhibits high computational power (CA) within a narrow optimal parameter range. However, the liquid state machine dynamic optimization system based on a self-feedback mask in this invention, due to the decoupling of feature mapping and network stabilization functions, maintains high computational power and high expressive power (KQ) across a very wide parameter range, significantly enhancing robustness.

[0088] Figure 5 and Figure 6 The diagram illustrates the changes in the loss curve (cross-entropy loss) and accuracy curve (classification accuracy) of the self-feedback mask-based liquid state machine model during the training process, with each training epoch. The stable performance of the loss and accuracy curves in the figure confirms that the reservoir dynamics stability actively maintained by the self-feedback mask in this invention ensures that the self-feedback mask-based liquid state machine model can learn efficiently and robustly throughout the entire training process.

[0089] 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.

[0090] 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.

[0091] 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.

[0092] 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.

[0093] To systematically verify the robustness of the self-feedback mask-based liquid state machine model, this study conducted a comprehensive hyperparameter analysis and compared its performance with that of the baseline liquid state machine model from two dimensions: intrinsic neurodynamic parameters and extrinsic learning strategy parameters, as shown in Table 3, to assess the differences in sensitivity of the two types of models to parameter changes.

[0094] Intrinsic neurodynamic parameters mainly include parameters that control EI equilibrium ( , ), parameters that determine network structure (connection coefficients) Number of reservoir neurons And parameters that affect time integration (attenuation coefficient) When these parameters change over a wide range, the liquid state machine model based on the self-feedback mask maintains good performance stability, with its classification accuracy generally remaining above 0.90. This indicates that its effectiveness does not depend on specific, finely tuned internal dynamic parameters and that it has strong adaptability to changes in intrinsic neurodynamics.

[0095] In stark contrast, the performance of the baseline liquid state machine model is extremely sensitive to intrinsic neurodynamic parameters: only within a very narrow range of parameters (such as connectivity coefficients). A value <0.2) can achieve high classification accuracy, but once the parameter exceeds this range, its accuracy drops sharply, usually to below 0.60, showing a strong dependence on intrinsic neurodynamic parameters.

[0096] External learning strategy parameters mainly cover window-related parameters (window size) Step length ) and synaptic weight scaling parameters ( Experimental results show that the liquid state machine model based on the self-feedback mask is basically insensitive to the selection of such parameters: within most of the tested parameter range, its classification accuracy remains close to the optimal level (about 0.95), and there is no significant performance fluctuation due to parameter adjustment.

[0097] The baseline liquid state machine model again exhibits significant parameter sensitivity: its accuracy is closely related to the choice of external learning policy parameters, such as the scaling parameters of synaptic weights. When the value is greater than 0.5, the classification performance of the baseline liquid state machine model will decrease significantly, and it will be unable to maintain stable accuracy over a wide parameter range.

[0098] Table 3 Summary of Parameter Robustness Analysis Data

[0099]

[0100] 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 liquid state machine dynamics optimization system based on a self-feedback mask, characterized in that, include: The liquid state machine reservoir is equipped with multiple spike neurons; The liquid state machine reservoir updates the membrane potential state of the spike neurons by receiving a timing input signal; the timing input signal is an electrocardiogram (ECG) signal. The mask generation module is connected to the membrane potential state of the spike neurons in the liquid state machine reservoir and is used to generate a self-feedback mask based on the membrane potential state. A feedback adjustment unit, connected to the mask generation module and the liquid state machine reservoir respectively, is used to adjust the cyclic connection and self-feedback calculation of the liquid state machine reservoir according to the self-feedback mask; The mask generation module includes: An encoder is used to compress the membrane potential state into a low-dimensional feature vector; A decoder is used to expand the low-dimensional feature vector into the self-feedback mask; The process of generating a 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 spike neuron. 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 value corresponding to a spike neuron.

2. The liquid state machine dynamics optimization system based on self-feedback mask according to claim 1, characterized in that: It also includes a readout module, which records the peak output of the liquid state machine reservoir at each time step within a preset processing time window, and splices all the peak outputs in chronological order to form a single enhanced feature vector. Finally, the enhanced feature vector is input into the classifier to obtain the final classification result.

3. The optimization method for a liquid state machine dynamics optimization system based on a self-feedback mask according to claim 1 or 2, characterized in that, Includes the following steps: Step 1: Receive ECG signals through the liquid state machine reservoir; Step 2: Update the membrane potential state of the spike neurons in the liquid state machine reservoir according to the time step in the time sequence input signal; Step 3: Input the membrane potential state into the mask generation module. The mask generation module generates the corresponding feedback gain value for the spike neuron, thereby forming a self-feedback mask. Step 4: In the feedback adjustment unit, the self-feedback mask is applied to the cyclic connections and self-feedback calculations of the liquid state machine reservoir in subsequent time steps to adjust the dynamic behavior of the spike neurons; Step 5: Generate and classify the spike output sequence based on the dynamic behavior of the spike neurons after regulation.

4. The optimization method for a liquid state machine dynamics optimization system based on a self-feedback mask according to claim 3, characterized in that: In step two, the discrete-time dynamic equation for the membrane potential state of the spike neurons in the updated liquid state machine reservoir is as follows: ； ； In the formula, It is a spike neuron The residual value obtained after the membrane potential decays in the previous time step. It is a spike neuron In the previous peak distribution phase It is the membrane potential decay factor. It is the input current. 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: 。 5. The optimization method for a liquid state machine dynamics optimization system based on a self-feedback mask according to claim 3, characterized in that: In step four, the calculation formula for applying the self-feedback mask to the liquid state machine reservoir in subsequent time steps for cyclic connection and self-feedback calculation 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. This represents the peak firing state of all spike neurons at the current time step. This represents element-wise multiplication. This represents the mask vector.

6. The optimization method for a liquid state machine dynamics optimization system based on a self-feedback mask according to claim 3, characterized in that: In step five, the process of generating the spike output sequence adopts a full-time spike strategy. After the input window at each time step has been processed, the spike firing sequence of all spike neurons in the liquid state machine reservoir is obtained. It is recorded and then concatenated into a single, high-dimensional enhanced feature vector. The formula is as follows: ； The classification is performed through a trainable linear readout layer. Enhance feature vectors The processing is performed using the following formula: ； In the formula, This represents the feature vector of the full-time spike sequence before classification, after processing by the linear readout layer.

Images