A method for designing a liquid state machine incorporating a sliding window

Abstract

The application discloses a liquid state machine design method combined with a sliding window, comprising the following steps: (1) acquiring an image data set, pre-processing the image data as training data, and obtaining pulse signals after pulse coding conversion; (2) constructing a liquid state machine model combined with a sliding window, specifically as follows: selecting neurons in a liquid layer which need to be connected with input layer neurons; connecting the input layer neurons to the selected liquid layer neurons in a sliding window mode; establishing internal connections of the liquid layer neurons; realizing data classification through a trainable output layer; (3) based on the training data, automatically adjusting key parameters of the liquid state machine through a Bayesian optimization technique, finding a group of optimal parameter settings, and optimizing the classification performance of the model; (4) inputting an image to be classified into the trained model after data processing, and obtaining a corresponding classification result. According to the application, the accuracy of image classification can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer science and brain-like intelligence, and particularly relates to a liquid state machine design method combined with a sliding window. BACKGROUND

[0002] Artificial neural networks (ANN) have made significant progress in computer vision, natural language processing and other fields, especially in image classification tasks, ANN can accurately identify and classify various images. However, ANN has the obvious defect of high energy consumption in the training and inference stage, which is particularly prominent when dealing with large-scale data and complex models. Spiking neural networks (SNN) are known as the third generation of artificial neural networks, which transmit signals between neurons through pulses, have the characteristics of event-driven, and are superior to traditional artificial neural networks in terms of computational overhead. In addition, SNN has the advantages of biological interpretability, efficient spatiotemporal feature extraction capability, and potential application value in image classification and other tasks.

[0003] Liquid state machine (LSM) as one of the traditional models of spiking neural networks is a special recurrent neural network with good dynamic information processing capability. For image classification tasks, when processing image data, the liquid layer usually uses a randomly connected nonlinear layer to map the input image features to a complex high-dimensional space. In this high-dimensional space, the image classification task can be processed through a simple readout mechanism. In the application of liquid state machine, the liquid layer and the input layer are not trained, and only the output layer needs to be trained. This feature solves the problem of gradient vanishing or gradient explosion in recurrent neural network training in image classification, and significantly reduces the training overhead.

[0004] However, the current application of liquid state machine in image classification tasks mainly has the following problems:

[0005] (1) Images contain a variety of features such as shape, texture, etc. The input layer uses random connection, which cannot accurately deliver key features to liquid layer neurons, affecting the effective capture and analysis of image features by the model, and reducing the classification accuracy.

[0006] (2) Since the model only trains the output layer, although it has the advantage of low training overhead, it also leads to lower classification performance than the fully trained MLP model, and performs poorly in image classification accuracy and generalization ability.

[0007] (3) The performance of the model is highly dependent on parameter settings, as different image classification datasets have their own unique feature distribution and complexity, so the parameters of the model need to be adjusted according to the specific situation.

[0008] In summary, how to design an efficient liquid state machine model is the key to expand its application in image classification and other practical scenarios and better exploit its potential. Therefore, a new liquid state machine design method is urgently needed. SUMMARY

[0009] To solve the problems of the existing liquid state machine in image classification tasks, the application provides a liquid state machine design method combined with a sliding window, which can improve the accuracy of image classification.

[0010] A liquid state machine design method combined with a sliding window, comprising the following steps:

[0011] (1) Obtain an image data set, pre-process the image data as training data, and obtain pulse signals after pulse coding conversion;

[0012] (2) Construct a liquid state machine model combined with a sliding window, specifically as follows:

[0013] Select neurons from the liquid layer that need to be connected to the input layer neurons; build the connection from the input layer neurons to the selected liquid layer neurons through the sliding window; establish the internal connection of the liquid layer neurons; realize data classification through the trainable output layer;

[0014] (3) Based on the training data of step (1), automatically adjust the key parameters of the liquid state machine through Bayesian optimization technology, find a set of optimal parameter settings, and optimize the classification performance of the model;

[0015] (4) Input the data processed image into the trained model to obtain the corresponding classification result.

[0016] The method of the application improves the liquid state machine and improves its performance in image classification tasks: by introducing the sliding window mechanism, the ability of the model to extract local information is enhanced; the Bayesian optimization technology is used to realize the automatic adjustment of the key parameters of the liquid state machine.

[0017] In step (2), the neurons in the liquid layer that need to be connected to the input layer neurons are selected, and the specific process is as follows:

[0018] Let the number of liquid layer neurons be N liquid , select neurons to connect with the input layer, and When is equal to N liquid , all liquid layer neurons are selected; when is less than N liquid , select by random selection or according to the liquid layer connection topology index sorting screening method.

[0019] In step (2), the connection between the input layer neurons and the selected liquid layer neurons is constructed by means of a sliding window, and the specific process is as follows:

[0020] The input layer is divided into a plurality of input mapping blocks by means of the sliding window, and each input mapping block corresponds to a liquid layer neuron.

[0021] Suppose the number of input mapping blocks extracted in one sliding window traversal process is q, then at least sliding window traversals are required to ensure that the number of input mapping blocks extracted meets the requirements.

[0022] For each selected liquid layer neuron, a structured input connection is constructed according to the corresponding input mapping block and the set input layer weight.

[0023] Further, the sliding window traversal process is as follows:

[0024] For the case of two-dimensional input layer neurons, the number of input layer neurons is N input = H x W; suppose the sliding window size is k x k, the step size is s, the top-left corner of the window is marked as its position, and the initial position is (0, 0) and the end position is (H-k, W-k). The window position of order i is (r i , c i ), where r i ∈ [0, H-k+1), c i ∈ [0, W-k+1); the sliding window moves horizontally to the right from the initial position, and each time it moves s steps; during each movement, an input mapping block F slide is divided, and the subsequent connection between the input layer neurons and the corresponding liquid layer neurons is constructed according to the mapping block; when the right axis of the window exceeds the right edge of the input layer, it is moved to (r+s, 0); when the lower axis of the window exceeds the lower edge of the input layer, it means that this traversal is over.

[0025] The specific process of constructing the structured input connection is as follows:

[0026] Suppose the liquid neuron h is the selected liquid layer neuron, and the input mapping block extracted by the corresponding sliding window is including k x k input layer neurons, and the input layer weight of the node is The connection weight setting adopts the way of randomly selecting weights from candidate weights, generating weights according to Gaussian distribution or using fixed weights; according to the weights and the mapping block, the connection matrix between the input layer neurons and the liquid layer neuron h is constructed wherein

[0027] In step (2), in the process of establishing internal connections of liquid layer neurons, connections are established between neurons according to a set of topological connection rules, including random connection rules, small-world network connection rules, scale-free network connection rules, and distance probability-based connection rules.

[0028] Preferably, the liquid layer maintains the characteristics of sparse connections within the liquid layer.

[0029] Preferably, in the process of establishing connections, it is ensured that there is no self-connection, and efficient information transmission is guaranteed.

[0030] Preferably, heterogeneity can be introduced into neurons within the liquid layer, including different dynamic behaviors and connection methods. Heterogeneity of neurons can be achieved by selecting different membrane potential time constants, connection weights, and firing delays to simulate the natural differences in biological neurons in terms of characteristics.

[0031] In step (2), data classification is achieved through a trainable output layer, as follows:

[0032] Within T time steps, the state variables of the liquid layer neurons are collected, and the average pulse firing frequency vector, pulse firing number vector, pulse firing interval vector, or pulse firing time vector is selected as the liquid layer state vector; the liquid layer state vector is transmitted to the output layer, and is mapped to the expected output through training.

[0033] Preferably, before transmitting the liquid layer state vector to the output layer, the data can be standardized to reduce the influence of different dimensions on the results.

[0034] It is worth noting that in addition to the traditional fully connected layer setting, the output layer can also use trainable structures such as convolution and Transformer.

[0035] In step (3), the key parameters of the liquid state machine are automatically adjusted through Bayesian optimization techniques, as follows:

[0036] Define the search space, such as the sliding window size k, the topological connection scheme parameters, the input neuron to liquid layer neuron connection weight W IL , the liquid layer excitatory neuron to excitatory neuron connection weight W EE , the inhibitory neuron to inhibitory neuron connection weight W II , the inhibitory neuron to excitatory neuron connection weight W IE , the excitatory neuron to inhibitory neuron connection weight W EI ; determine the optimal parameter combination through Bayesian optimization to achieve the best performance on the corresponding task.

[0037] Preferably, the process can use partial datasets for training and testing, quickly exploring different parameters at an early stage, reducing the consumption of computational resources in the evaluation phase, while helping to avoid overfitting.

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

[0039] 1、The present application adopts a method of combining sliding windows to construct the connection between input layer neurons and liquid layer neurons, allowing the liquid layer neurons to focus on the local area of the input data, simulating the processing process of biological vision. Compared with the search-based method, the overhead of input layer connection design is simplified; compared with random connection, the information extraction ability of the model is enhanced.

[0040] 2、The present application adopts Bayesian optimization technology to optimize key parameters of the model, which is beneficial to improve the accuracy of the model in classification tasks and other scenarios, so that the model can fully learn the patterns and rules in the data.

[0041] 3、The present application performs image classification tasks through an improved liquid state machine, which reduces the training overhead compared with the fully trained model, and can more effectively extract features in the image through the design of the sliding window, improve the accuracy and stability of the classification, and achieve or even surpass the classification performance of the fully trained MLP model with the same structure.

[0042] 4、The method of the present application has strong universality and scalability, not only suitable for common image classification tasks, but also applicable to other tasks related to spatio-temporal features, such as language recognition, robot control, anomaly diagnosis, biological signal, etc. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flow chart of a liquid state machine design method combined with a sliding window in an embodiment of the present application is shown in the figure.

[0044] Figure 2 A sliding window technology diagram used in the input layer in an embodiment of the present application is shown in the figure.

[0045] Figure 3 A Bayesian optimization process diagram in an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0046] The present application will be further described in detail below in conjunction with the drawings and embodiments, it should be pointed out that the following described embodiments are intended to facilitate the understanding of the present application, and do not have any limiting effect on it.

[0047] As Figure 1As shown, a liquid state machine design method combined with a sliding window, wherein the liquid state machine structure includes input data, encoding processing part, input layer, liquid layer, and output layer. The data is converted into corresponding input layer neuron signals after encoding processing; the connection relationship between the input layer neurons and the liquid layer neurons is constructed by the sliding window method; the internal neurons of the liquid layer are connected according to certain topological rules; finally, the liquid layer state vector within T time steps is transmitted to the output layer to obtain the corresponding output. The model is only trained for the output layer.

[0048] A liquid state machine design method combined with a sliding window, the specific implementation steps are as follows:

[0049] Step 1, data preprocessing process:

[0050] 1-1, obtain data from public data sets, divide the data into training set, validation set and test set according to reasonable proportion; at the same time, standardize the data.

[0051] Specifically, the following three data sets are mainly used. The MNIST data set is a classic handwritten character data set, which contains 70000 gray scale images, showing 0-9 handwritten characters; the Fashion-MNIST data set is more challenging than the MNIST data set, which contains 70000 gray scale images, showing 10 different types of clothes including skirts and shoes; N-MNIST is a neuromorphic MNIST data set, which records the light sensing events caused by the brightness change of image pixels, and is widely used in the study of pulse neural networks.

[0052] Preferably, the data set can be subjected to rotation, inversion, translation, noise addition and other data enhancement processing, so as to expand the number of training data samples and improve the robustness and accuracy of the model.

[0053] It is worth noting that the model proposed in this embodiment is also applicable to processing language recognition data, robot control data, abnormal diagnosis data, biological signal data, financial data, traffic prediction data, etc., and has wide application prospects and potential.

[0054] 1-2, select a suitable encoding method to convert it into a pulse signal, and obtain a preprocessed pulse signal.

[0055] Specifically, the direct encoding method keeps the original input value directly input multiple times through a fixed time step; the Poisson encoding generates a Poisson distribution pulse sequence for each input value, and the frequency of the pulse is proportional to the input value.

[0056] Step 2, construct a liquid state machine model combined with a sliding window:

[0057] 2-1, select neurons from the liquid layer that need to be connected to the input layer neurons:

[0058] Let the number of liquid layer neurons be N liquid , select neurons to connect with the input layer, and when is equal to N liquid , all liquid layer neurons need to be set with corresponding input layer connections; when is less than N liquid , it can be selected by random selection, sorting and filtering according to the liquid layer connection topology index, etc.

[0059] 2-2, combined with sliding window technology: variant of traditional model, by moving sliding window in input layer to build connection from input layer neurons to selected liquid layer neurons, principle can be seen in Figure 2 ;

[0060] Specifically, the traditional model usually uses random connection to determine the connection of each liquid layer neuron, and each liquid layer neuron randomly selects 15% of the neurons in the input layer as the front neuron. The setting method of the traditional model maps the information of the liquid layer neurons too dispersedly, which cannot realize efficient input information extraction and is not conducive to generating efficient liquid layer state vectors. The sliding window technology in step 2 improves it: using sliding window technology to divide the input layer into multiple input mapping blocks to capture local features in space and time. The improved input layer to liquid layer connection setting process is expressed as follows:

[0061] Based on the sliding window technology, the input layer is divided into multiple input mapping blocks, and the corresponding structured input is constructed for the liquid layer neurons to capture local features in space and time, and the process is as follows: 1) move the sliding window to divide the input layer into several input mapping blocks, and each input mapping block corresponds to a liquid layer neuron; 2) set the number of input mapping blocks extracted in one sliding window traversal process as q, then at least sliding window traversal is required to ensure that the number of input mapping blocks extracted meets the requirements; 3) for each selected liquid layer neuron, construct a structured input connection according to its corresponding input mapping block and the set input layer weight.

[0062] The sliding window traversal process is as follows: for the case of two-dimensional input layer neurons, let the sliding window size be k x k, the step size be s, and the number of input layer neurons be N input = H x W. Mark the position of the sliding window with the top left corner of the window, and the position of window No. i is (r i , c i ), where r i∈ [0, H-k+1), c i ∈ [0, W-k+1), the default initial position is (0, 0). The sliding window moves horizontally to the right from the initial position by s steps each time; when the right axis of the sliding window exceeds the right edge of the input layer, that is, the current corresponding window position (r, c), c >= W-k+1, then the sliding window is moved to the window position (r+s, 0); during each movement of the sliding window, an input mapping block F slide is divided, and subsequent connections from the input layer neurons to the corresponding liquid layer neurons are constructed according to the mapping block; when the lower axis of the sliding window exceeds the lower edge of the input layer, that is, the current corresponding window position (r, c), R >= (h-k+1), then it indicates that the sliding window one-pass traversal process is completed.

[0063] The process of constructing structured input for the selected liquid layer neuron is as follows: let liquid neuron h be the selected liquid layer neuron, and the input mapping block extracted by the corresponding sliding window be including k x k input layer neurons, the input layer weight corresponding to the node is The connection weight setting can adopt the way of randomly selecting weight from the candidate weight, generating weight according to Gaussian distribution or using fixed weight. According to the weight and the mapping block, the connection matrix from the input layer neuron to the liquid layer neuron h is constructed wherein

[0064] As shown in Figure 2 , the sliding window technology allows the liquid layer neuron to focus on the local area of the input data, simulates the processing process of biological vision, simplifies the cumbersome input connection design process, and has the advantages of high flexibility and strong information extraction capability.

[0065] It is worth noting that the application of the sliding window proposed in the present application is obviously different from convolution. Convolution is a widely used technology in deep learning, which performs element multiplication and accumulation operation by sliding the convolution kernel on the image to extract features from each local area of the image. However, in the present application, the sliding window divides the input layer neurons into multiple input layer mapping blocks by sliding, and according to each mapping block and the weight setting, the connection from the corresponding input layer neuron to the liquid layer neuron is constructed to provide the liquid layer neuron with input features with "locality".

[0066] 2-3, constructing internal connections of liquid layer neurons:

[0067] The liquid layer internal neurons are divided into excitatory neurons and inhibitory neurons according to the attributes, and the excitatory neurons and the inhibitory neurons in the liquid layer adopt a ratio of 8:2. The excitatory neurons correspond to the connection weight W EE of the out-edge EIare positive, amplifying neural activities, enabling the network to distinguish different input patterns; the inhibitory neurons correspond to the connection weights W IE , W II are negative, weakening neural activities, regulating stable network states. Each neuron in the liquid layer is randomly set as an excitatory neuron or an inhibitory neuron according to a ratio of 8:2. The neurons are connected according to certain topological connection rules.

[0068] Preferably, the liquid layer can maintain the characteristics of sparse connections.

[0069] Preferably, a small-world model is used as the topological rule of the liquid layer. In the initialization of the topology, all the connections of the liquid layer neurons are subject to a ring-type topology, and each liquid layer neuron is connected to its adjacent N neighbor / 2 neurons in front and N neighbor / 2 neurons behind. Each connection is traversed to detect whether it needs to be reconnected with a reconnection probability p. If it needs to be reconnected, another neuron in the liquid layer is randomly selected to establish a connection, the original connection (i, j) is disconnected, a new connection is established, and the connection matrix of the liquid layer is updated.

[0070] Preferably, in the process of establishing connections, it is ensured that there is no self-connection to ensure efficient transmission of information.

[0071] Preferably, the liquid layer neurons use different membrane potential time constants, connection weights, and time delays to simulate the natural differences in the characteristics of biological neurons. For spiking neurons, the membrane potential time constant determines the response speed of the neuron to input changes; the connection weight setting reflects the difference in connection strength; and the transmission time delay setting affects the speed of signal transmission to the next neuron. Using different membrane potential time constants, connection weights, and time delays enhances the nonlinear ability of the liquid layer and improves its ability to process complex inputs. Specifically, for the heterogeneity of the membrane potential time constant, the time constant is randomly selected from a preset range; for the heterogeneity of the connection weight within the liquid layer, the connection weight is added with Gaussian noise based on the original weight; and for the heterogeneity of the firing delay, the firing delay is selected from a preset range.

[0072] 2-4, the model output layer is set as follows:

[0073] In T time steps, the state variables of the liquid layer neurons are collected, such as the average pulse firing frequency vector, the pulse firing number vector, the pulse firing interval vector, and the pulse firing time vector, which are selected as the liquid layer state. The liquid layer state vector is transmitted to the input layer, and is mapped to the expected output through training.

[0074] Preferably, for a classification task, the output layer can employ a fully connected layer from the liquid layer (in_features = N liquid ) to the classification output (out_features = 10) with softmax as the activation function. The model training employs mean square error as the loss function, Adam optimizer for weight optimization, 200 epochs for training, and an initial learning rate of lr = 0.001.

[0075] Preferably, the liquid layer state vector can be standardized before being passed into the output layer to reduce the impact of different dimensions on the results.

[0076] It is worth noting that in addition to the traditional fully connected layer setting, the output layer can also employ trainable structures such as convolution and Transformer.

[0077] Preferably, some liquid layer neurons can be selected based on index ranking to collect their states as the liquid layer state.

[0078] Step 3: Automatically adjust the key parameters of the liquid state machine through Bayesian optimization techniques to find a set of optimal parameter settings to optimize the performance of the liquid state machine. The principle can be found in Figure 3 :

[0079] (a) Define the search space to determine the key parameters to be optimized and their value ranges, including the sliding window size k, the number of neighbors N neighbor of the small-world model, the input neuron to liquid layer neuron connection weight W IL , the liquid layer excitatory neuron to excitatory neuron connection weight W EE , the inhibitory neuron to inhibitory neuron connection weight W EI , the inhibitory neuron to excitatory neuron connection weight W IE , and the excitatory neuron to inhibitory neuron connection weight W EI ; (b) Initialize the surrogate model and train it using a small number of initial observations; (c) select the next set of parameters to be evaluated based on the surrogate model; (d) train and validate the liquid state machine according to the current selected parameters and record the model performance; (e) update the surrogate model using the newly evaluated performance data to improve the surrogate model; (f) repeat steps (c)-(e) iteratively until the maximum number of evaluations is reached or the convergence condition is met; (g) finally determine the optimal parameter combination to make the liquid state machine perform best in the classification task.

[0080] Preferably, based on experience, the parameter search space can be set as k ∈ [6, 18], N neighbor ∈ [2, 16], W IL ∈ [1.0, 10], and W EE∈[0.6,10], W EI ∈[0.6,10], W IE ∈[-5.0,-0.6], W II ∈[-5.0,-0.6]. Set the evaluation number to 200 evaluations.

[0081] Preferably, in order to reduce the computational overhead of the evaluation stage, the model is trained and validated using a partial dataset, 6000 images in the training dataset are randomly selected for training in the evaluation stage, 1000 images in the training dataset are randomly selected for testing in the evaluation stage, and the number of evaluation stage training epochs is limited to 10.

[0082] Step 4, inference test stage:

[0083] The model uses the parameters obtained by Bayesian optimization, inputs the processed image data into the trained model to obtain the corresponding classification results, and verifies the image classification performance and robustness of the model on the test set.

[0084] Preferably, in order to compare the effectiveness of the liquid state machine design method combined with a sliding window proposed in the present application, an ablation experiment was carried out on the MNIST dataset. Four model settings were used, setting one was a heterogeneous liquid state machine combined with a sliding window, setting two was a heterogeneous liquid state machine with random connections, setting three was a non-heterogeneous liquid state machine combined with a sliding window, and setting four was a non-heterogeneous liquid state machine with random connections. The experimental results are as follows: when N liquid = 1000, the model of setting one achieved an accuracy of 97.38%, which was 1.76%, 0.15%, and 1.52% higher than that of settings two, three, and four, respectively, in terms of test set classification accuracy; when N liquid = 8000, the model of setting one achieved an accuracy of 98.41%, which was higher than that of the MLP model with full training of the same structure. From the above results, it can be seen that the method proposed in the present application performs more stably in image classification, and the sliding window setting can effectively improve the accuracy of image classification.

[0085] The above embodiments have described the technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the present application. Any modification, supplement and equivalent replacement made within the principle range of the present application shall be included in the protection scope of the present application.

Claims

1. A method of designing a liquid state machine incorporating a sliding window, characterized by, The method comprises the following steps: (1) obtaining an image data set, pre-processing the image data as training data, and obtaining a pulse signal after pulse coding conversion; (2) constructing a liquid state machine model combined with a sliding window, specifically as follows: selecting neurons in the liquid layer that need to be connected with input layer neurons; constructing the connection of input layer neurons to the selected liquid layer neurons in the form of a sliding window; establishing internal connections of liquid layer neurons; realizing data classification through a trainable output layer; selecting neurons in the liquid layer that need to be connected with input layer neurons, specifically as follows: Let the number of liquid layer neurons be N liquid , select neurons in the input layer to establish a connection, and when is equal to N liquid , all liquid layer neurons are selected; when is less than N liquid , selection is made by random selection or according to the liquid layer connection topology index sorting screening method; constructing the connection of input layer neurons to the selected liquid layer neurons in the form of a sliding window, specifically as follows: moving the sliding window to divide the input layer into a plurality of input mapping blocks, and each input mapping block corresponds to a liquid layer neuron; If the number of input mapping blocks extracted by one pass of the sliding window traversal process is q, then at least q passes of the sliding window traversal are required to ensure that the number of extracted input mapping blocks meets the requirement. for each selected liquid layer neuron, constructing a structured input connection according to the corresponding input mapping block and the set input layer weight; (3) based on the training data of step (1), automatically adjusting the key parameters of the liquid state machine through Bayesian optimization technology to find a set of optimal parameter settings to optimize the classification performance of the model; (4) inputting the image to be classified after data processing into the trained model to obtain the corresponding classification result.

2. The method of claim 1, wherein, The sliding window traversal process is as follows: For the case of two-dimensional input layer neurons, the number of input layer neurons is N input = H x W; assuming the size of the sliding window is k x k, the step length is s, mark the position of the window with the upper left corner, assume the initial position is (0, 0) and the end position is (H-k, W-k), the position of the window with serial number i is (r i , c i ), where r i ∈ [0, H-k+1), c i ∈ [0, W-k+1); the sliding window starts from the initial position and moves horizontally to the right by s step length each time; during each movement, an input mapping block F slide is divided, and subsequent connection of the input layer neurons to the corresponding liquid layer neurons is constructed according to the mapping block; when the right axis of the window exceeds the right edge of the input layer, it is moved to (r+s, 0); when the lower axis of the window exceeds the lower edge of the input layer, it means that this round of traversal is completed.

3. The method of claim 1, wherein, constructing a structured input connection, specifically as follows: Let liquid neuron h be a selected liquid layer neuron, and the input mapping block extracted by the sliding window corresponding to the liquid neuron h be The input layer neuron includes k x k input layer neurons, and the input layer weight corresponding to the node is The connection weight setting adopts a mode of randomly selecting a weight from a candidate weight, generating a weight according to a Gaussian distribution, or adopting a fixed weight; and a connection matrix of the input layer neuron to the liquid layer neuron h is constructed according to the weight and the mapping block wherein 4. The method of claim 1, wherein, In step (2), in the process of establishing internal connections of liquid layer neurons, connections are established between neurons according to the set topological connection rules, which include random connection rules, small-world network connection rules, scale-free network connection rules and distance probability-based connection rules.

5. The method of claim 1, wherein, In step (2), data classification is realized through a trainable output layer, specifically as follows: In T time steps, collect the state variables of the liquid layer neurons, and select the average pulse firing frequency vector, pulse firing number vector, pulse firing interval vector or pulse firing time vector as the liquid layer state vector; the liquid layer state vector is transmitted to the output layer, and is mapped to the expected output through training.

6. The method of claim 1, wherein, In step (3), the key parameters of the liquid state machine are automatically adjusted through Bayesian optimization technology, specifically as follows: Define parameter search space, including sliding window size k, topological connection scheme parameters, input neuron to liquid layer neuron connection weight W IL , liquid layer excitatory neuron to excitatory neuron connection weight W EE , inhibitory neuron to inhibitory neuron connection weight W II , inhibitory neuron to excitatory neuron connection weight W IE , excitatory neuron to inhibitory neuron connection weight W EI ; Determine the optimal parameter combination by Bayesian optimization to achieve the best performance on the corresponding task.

Images