Liquid state machine design method combined with sliding window
By combining sliding window and Bayesian optimization technology, the liquid state machine design is solved, and the problem of inaccurate feature transfer and poor classification performance in image classification is achieved, which achieves higher accuracy and stability, surpassing the entire The performance of the training model.
Patent Information
- Application Number
- CN202411077370.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-07
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-08-07
AI Technical Summary
In the image classification task, existing liquid state machines have problems such as inaccurate feature transfer, low training overhead, poor classification performance, and model performance depends on parameter settings, which affects their performance in practical applications.
The liquid state machine design method combined with sliding window is adopted to construct the connection between input layer neurons and liquid layer neurons through sliding windows, and the model parameters are automatically adjusted using Bayesian optimization technology to enhance the model's information extraction ability and classification performance.
It improves the accuracy and stability of image classification, reduces training overhead, and even exceeds the performance of fully trained MLP models, and has strong versatility and scalability.
Smart Images

Figure CN119048879A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer science and brain-like intelligence technology, and in particular, relates to a liquid state machine design method combined with a sliding window. Background Art
[0002] Artificial neural networks (ANN) have made significant progress in the fields of computer vision and natural language processing. In particular, in image classification tasks, ANN can accurately identify and classify various types of images. However, ANN has the obvious defect of high energy consumption in the training and inference stages, which is particularly prominent when processing large-scale data and complex models. Spiking neural networks (SNNs) are known as the third generation of artificial neural networks. They transmit signals between neural nodes in the form of pulses, have event-driven characteristics, and are superior to traditional artificial neural networks in terms of computational overhead. In addition, SNNs have the advantages of biological interpretability and efficient spatiotemporal feature extraction capabilities, and also have potential application value in tasks such as image classification.
[0003] Liquid state machine (LSM), as one of the traditional models of pulse neural networks, is a special recurrent neural network with good dynamic information processing capabilities. 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 by a simple readout mechanism. In the application of liquid state machines, 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 exploding in recurrent neural network training in image classification, and significantly reduces the training overhead.
[0004] However, the current application of liquid state machines in image classification tasks mainly has the following problems:
[0005] (1) Images contain rich and diverse features, such as shape and texture. The random connection used in the input layer cannot accurately transmit key features to the liquid layer neurons in a targeted manner, which affects the model's effective capture and analysis of image features and reduces the accuracy of classification.
[0006] (2) Since the model is only trained on the output layer, although it has the advantage of low training overhead, its classification performance is lower than that of the fully trained MLP model, and its image classification accuracy and generalization ability are poor.
[0007] (3) Model performance is highly dependent on parameter settings. Since different image classification datasets have their own unique feature distribution and complexity, the model parameters 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 expanding its application in practical scenarios such as image classification and better realizing its potential. Therefore, it is urgent to study a new liquid state machine design method. Summary of the invention
[0009] In order to solve the problems existing in the existing liquid state machine in the image classification task, the present invention provides a liquid state machine design method combined with a sliding window, which can improve the accuracy of image classification.
[0010] A method for designing a liquid state machine combined with a sliding window comprises the following steps:
[0011] (1) Obtaining an image data set, preprocessing the image data as training data, and performing pulse coding conversion to obtain a pulse signal;
[0012] (2) Construct a liquid state machine model combined with a sliding window, as follows:
[0013] Select neurons from the liquid layer that need to be connected to the input layer neurons; construct connections from the input layer neurons to the selected liquid layer neurons through a sliding window; establish internal connections between the liquid layer neurons; and implement data classification through a trainable output layer;
[0014] (3) Based on the training data in step (1), the key parameters of the liquid state machine are automatically adjusted by Bayesian optimization technology to find a set of better parameter settings to optimize the classification performance of the model;
[0015] (4) After data processing, the image to be classified is input into the trained model to obtain the corresponding classification result.
[0016] The method of the present invention improves the liquid state machine and enhances its performance in image classification tasks: by introducing a sliding window mechanism, the model's ability to extract local information is enhanced; and by using Bayesian optimization technology, automatic adjustment of key parameters of the liquid state machine is achieved.
[0017] In step (2), neurons that need to be connected to the input layer neurons are selected from the liquid layer. The specific process is as follows:
[0018] Assume the number of neurons in the liquid layer is N liquid , select one of neurons are connected to the input layer, and when Equal to N liquid When , all liquid layer neurons are selected; when Less than N liquid When selecting, the selection is done by random selection or by sorting and screening according to the liquid layer connection topology index.
[0019] In step (2), the connection from the input layer neurons to the selected liquid layer neurons is constructed by means of a sliding window. The specific process is as follows:
[0020] 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;
[0021] Assume that the number of input mapping blocks extracted by a sliding window traversal process is q, then at least Only one sliding window traversal can ensure that the number of extracted input mapping blocks meets the requirements;
[0022] For each selected liquid layer neuron, a structured input connection is constructed according to its corresponding input mapping block and the set input layer weights.
[0023] Furthermore, 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×W; let the sliding window size be k×k, the step size be s, mark its position with the upper left corner of the window, let the initial position be (0,0), the end position be (Hk,Wk), and the position of the window with sequence number i be (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, with a step length of s each time; during each movement, an input mapping block F is divided slide , and then build the connection from the input layer neurons to the corresponding liquid layer neurons according to the mapping block; when the right axis of the window exceeds the right edge of the input layer, it moves 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 completed.
[0025] The specific process of building a structured input connection is:
[0026] Let liquid neuron h be the selected liquid layer neuron, and its corresponding input mapping block extracted by the sliding window is It includes k×k input layer neurons, and the input layer weights corresponding to the nodes are The connection weights are set by randomly selecting weights from candidate weights, generating weights according to Gaussian distribution, or using fixed weights; the connection matrix from the input layer neurons to the liquid layer neurons h is constructed based on the weights and mapping blocks. in
[0027] In step (2), in the process of establishing the internal connection of the liquid layer neurons, the neurons are connected according to the set topological connection rules, which include random connection rules, small-world network connection rules, scale-free network connection rules and connection rules based on distance probability.
[0028] Preferably, the liquid layer maintains a sparsely connected characteristic.
[0029] Preferably, during the connection establishment process, it is ensured that there is no self-connection, thereby ensuring efficient transmission of information.
[0030] Preferably, heterogeneity can be introduced into neurons inside the liquid layer, including different dynamic behaviors and connectivity modes, and the heterogeneity of neurons can be achieved by selecting different membrane potential time constants, connection weights, release delays and other setting parameters to simulate the natural differences in the characteristics of biological neurons.
[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 passed to the output layer and mapped to the expected output through training.
[0033] Preferably, before the liquid layer state vector is transmitted to the output layer, the data may be standardized to reduce the impact 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 adopt trainable structures such as convolution and Transformer.
[0035] In step (3), the key parameters of the liquid state machine are automatically adjusted by Bayesian optimization technology, as follows:
[0036] Define the search space, such as the sliding window size k, topological connection scheme parameters, and the connection weight W from the input neuron to the liquid layer neuron IL , the connection weight W from the liquid layer excitatory neurons to the excitatory neurons EE , inhibitory neuron to inhibitory neuron connection weight W II , the connection weight W from inhibitory neurons to excitatory neurons IE , the connection weight W from excitatory neurons to inhibitory neurons EI ; Determine the optimal parameter combination through Bayesian optimization to achieve the best performance on the corresponding task.
[0037] Preferably, the process can use part of the data set for training and testing, quickly explore different parameters in the early stages, reduce the consumption of computing resources in the evaluation phase, and help avoid overfitting.
[0038] Compared with the prior art, the present invention has the following beneficial effects:
[0039] 1. The present invention uses a sliding window method to construct the connection between the input layer neurons and the liquid layer neurons, allowing the liquid layer neurons to focus on the local area of the input data, simulating the processing of biological vision. Compared with the search-based method, the overhead of the input layer connection design is simplified; compared with random connections, the information extraction ability of the model is enhanced.
[0040] 2. The present invention adopts Bayesian optimization technology to optimize the key parameters of the model, which is beneficial to improve the accuracy of the model in scenarios such as classification tasks, so that the model can fully learn the patterns and laws in the data.
[0041] 3. The present invention performs image classification tasks through an improved liquid state machine, which reduces the training overhead compared to the fully trained model, and can more effectively extract features in the image through the design of the sliding window, thereby improving the accuracy and stability of classification, reaching or even exceeding the classification performance of the fully trained MLP model with the same structure.
[0042] 4. The method of the present invention has strong versatility and scalability. It is not only suitable for common image classification tasks, but also can be applied to other tasks related to spatiotemporal features, such as language recognition, robot control, abnormal diagnosis, biological signals and other scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 This is a flow chart of a method for designing a liquid state machine combined with a sliding window in an embodiment of the present invention;
[0044] Figure 2 Schematic diagram of the sliding window technology used in the input layer in an embodiment of the present invention;
[0045] Figure 3 Schematic diagram of the Bayesian optimization process in an embodiment of the present invention. DETAILED DESCRIPTION
[0046] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.
[0047] like Figure 1As shown in the figure, a design method of a liquid state machine combined with a sliding window is shown, wherein the liquid state machine structure includes input data, a coding processing part, an input layer, a liquid layer, and an output layer. The data is converted into the corresponding input layer neuron signal after coding processing; the connection relationship between the input layer neurons and the liquid layer neurons is constructed by a sliding window method; the neurons inside 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 in a reasonable proportion; and standardize the data at the same time.
[0051] Specifically, the following three datasets are mainly used. The MNIST dataset is a classic handwriting dataset, which contains 70,000 grayscale images showing handwriting from 0 to 9; the Fashion-MNIST dataset is more challenging than the MNIST dataset, containing 70,000 grayscale images showing 10 different types of clothing including skirts and shoes; N-MNIST is a neuromorphic MNIST dataset that records light events caused by changes in image pixel brightness and is widely used to study pulse neural networks.
[0052] Preferably, data enhancement processing such as rotation, flipping, translation, and noise addition can be performed on the data set 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 broad application prospects and potential.
[0054] 1-2, select a suitable encoding method to convert it into a pulse signal to obtain a preprocessed pulse signal.
[0055] Specifically, the direct encoding method maintains the original input value and directly inputs it multiple times through a fixed time step; Poisson encoding generates a Poisson distributed pulse sequence for each input value, and the frequency of the pulse is proportional to the input value.
[0056] Step 2: Build 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] Assume the number of neurons in the liquid layer is N liquid , select one of neurons are connected to the input layer, and when Equal to N liquid When , the corresponding input layer connections need to be set for all liquid layer neurons; when Less than N liquid When selecting, it can be selected by random selection, sorting and screening according to the liquid layer connection topology index, etc.
[0059] 2-2, combined with sliding window technology: a variation of the traditional model, by moving the sliding window in the input layer to build the connection between the input layer neurons and the selected liquid layer neurons one by one, the principle can be seen Figure 2 ;
[0060] Specifically, traditional models usually use random connections to determine the connection of each liquid layer neuron, and each liquid layer neuron randomly selects 15% of the neurons from the input layer as pre-neurons. The setting method of the traditional model is that the information mapped to the liquid layer neurons is too scattered, which cannot achieve efficient input information extraction and is not conducive to generating efficient liquid layer state vectors. The sliding window technology in step 2 improves it: the sliding window technology is used to divide the input layer into multiple input mapping blocks to capture local characteristics of 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 corresponding structured inputs are constructed for the liquid layer neurons to capture the local features of space and time. The implementation 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) Assuming that the number of input mapping blocks extracted by one traversal of the sliding window is q, at least A sliding window traversal is required to ensure that the number of extracted input mapping blocks meets the requirements; 3) For each selected liquid layer neuron, a structured input connection is constructed based on its corresponding input mapping block and the set input layer weights.
[0062] The traversal process of the sliding window is as follows: For the case where the input layer neurons are two-dimensional, the sliding window size is k×k, the step size is s, and the number of input layer neurons is N input =H×W. The position of the sliding window is marked with the upper left corner of the window where the sliding window is located. 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 default initial position is (0,0). The sliding window moves horizontally to the right from the initial position, with each movement step of s; 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, the sliding window is moved to the window position (r+s,0); during each movement of the sliding window, an input mapping block F is divided slide , and then build the connection from the input layer neurons to the corresponding liquid layer neurons 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), it indicates that the traversal process of the sliding window is completed.
[0063] The process of constructing a structured input for the selected liquid layer neuron is as follows: Let the liquid neuron h be the selected liquid layer neuron, and its corresponding input mapping block extracted by the sliding window be It includes k×k input layer neurons, and the input layer weights corresponding to the nodes are The connection weights can be set by randomly selecting weights from candidate weights, generating weights according to Gaussian distribution, or using fixed weights. The connection matrix from the input layer neurons to the liquid layer neurons h is constructed based on the weights and mapping blocks. in
[0064] like Figure 2 As shown, the sliding window technology allows liquid layer neurons to focus on local areas of input data, simulating the processing process of biological vision, simplifying the cumbersome input connection design process, and has the advantages of high flexibility and strong information extraction capabilities.
[0065] It is worth noting that the application of the sliding window proposed in the present invention is significantly different from convolution. Convolution is a widely used technology in deep learning. It slides the convolution kernel on the image, performs element-wise multiplication and accumulation operations, and extracts features from each local area of the image. The sliding window in the present invention divides the input layer neurons into multiple input layer mapping blocks by sliding, and constructs the connection from the corresponding input layer neurons to the liquid layer neurons according to each mapping block and weight setting, providing the liquid layer neurons with "local" input features.
[0066] 2-3. Construct the internal connections of liquid layer neurons:
[0067] The neurons inside the liquid layer are divided into excitatory neurons and inhibitory neurons according to their attributes. The ratio of excitatory neurons to inhibitory neurons in the liquid layer is 8:2. The connection weight W corresponding to the outgoing edge of the excitatory neuron is EE , W EIare all positive values, which amplify neural activity and enable the network to distinguish different input patterns; the inhibitory neuron corresponds to the connection weight W of the outgoing edge IE , W II All are negative values, weakening neural activity and regulating the stable network state. Each neuron in the liquid layer is randomly set as an excitatory neuron and an inhibitory neuron in a ratio of 8:2. Connections are established between neurons according to certain topological connection rules.
[0068] Preferably, the interior of the liquid layer can maintain a sparsely connected characteristic.
[0069] Preferably, the small world model is used as the liquid layer topology rule. Initialize the topology, all liquid layer neuron connections obey the regular ring topology, and each liquid layer neuron is connected to its first N adjacent neurons. neighbor / 2 neurons are connected to their adjacent N neurons neighbor / 2 neurons are connected. Each current connection is traversed, and reconnection probability p is used to detect whether reconnection is needed. If reconnection is needed, 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 liquid layer connection matrix is updated.
[0070] Preferably, during the connection establishment process, it is ensured that there is no self-connection, thereby ensuring efficient transmission of information.
[0071] Preferably, the liquid layer neurons use different membrane potential time constants, connection weights, and delay settings 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; the transmission delay setting affects the speed at which the signal is transmitted to the next neuron. Using different membrane potential time constants, connection weights, and delays can enhance the nonlinear capabilities of the liquid layer and improve its ability to handle complex inputs. Specifically, in response to the heterogeneity of the membrane potential time constant, the time constant is set to be randomly selected from a preset range; in response to the heterogeneity of the connection weights within the liquid layer, the connection weights are set to add Gaussian noise on the basis of the original weights; in response to the heterogeneity of the release delay, the release delay is set to be 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 frequency vector, the pulse number vector, the pulse interval vector, the pulse time vector, etc., which can be selected as the liquid layer state. The liquid layer state vector is passed to the input layer and mapped to the expected output through training.
[0074] Preferably, for ten types of tasks, the output layer can be taken from the liquid layer (in_features = N liquid ) to the classification output (out_features = 10), and softmax is used as the activation function. The mean square error is used as the loss function for model training, and the Adam optimizer is used for weight optimization. The number of training epochs is 200, and the initial learning rate lr = 0.001.
[0075] Preferably, before the liquid layer state vector is transmitted to the output layer, the data may be standardized 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 adopt trainable structures such as convolution and Transformer.
[0077] Preferably, some liquid layer neurons can be selected according to the index sorting and their states can be collected as the liquid layer state.
[0078] Step 3: Use Bayesian optimization technology to automatically adjust the key parameters of the liquid state machine and find a set of better parameter settings to optimize the performance of the liquid state machine. The principle can be seen in Figure 3 :
[0079] (a) Define the search space and determine the key parameters to be optimized and their value ranges, including the sliding window size k and the number of neighbors in the small world model N. neighbor , the connection weight W from the input neuron to the liquid layer neuron IL , the connection weight W from the excitatory neuron to the excitatory neuron in the liquid layer EE , the inhibitory neuron to inhibitory neuron connection weight W EI , the connection weight W from inhibitory neurons to excitatory neurons IE , the connection weight W from excitatory neurons to inhibitory neurons EI ; (b) Initialize the proxy model and train the proxy model using a small number of initial observations; (c) Select the next set of parameters to be evaluated based on the proxy model; (d) Train and verify the liquid state machine based on the currently selected parameters and record the model performance; (e) Update the proxy model and improve the proxy model using the newly evaluated performance data; (f) Repeat steps (c)-(e) and iteratively optimize until the predetermined maximum number of evaluations is reached or the convergence condition is met; (g) Finally determine the optimal parameter combination to achieve the best performance of the liquid state machine in the classification task.
[0080] Preferably, according to experience, the parameter search space can be set to k∈[6,18], N neighbor ∈[2,16],W IL ∈[1.0,10],W EE∈[0.6,10],W EI ∈[0.6,10],W IE ∈[-5.0,-0.6],W II ∈[-5.0,-0.6]. The number of evaluations is set to 200.
[0081] Preferably, note that in order to reduce the computational overhead in the evaluation phase, the model uses part of the data set for training and verification, 6000 images in the training data set are randomly selected for training in the evaluation phase, and 1000 images in the training data set are randomly selected for testing in the evaluation phase. The number of training epochs in the evaluation phase is limited to 10.
[0082] Step 4, reasoning test phase:
[0083] The model uses parameters obtained by Bayesian optimization, inputs the processed image data into the trained model to obtain the corresponding classification results, and verifies the model's image classification performance and robustness on the test set.
[0084] Preferably, in order to compare and illustrate the effectiveness of the liquid state machine design method combined with a sliding window proposed in the present invention, an ablation experiment was carried out on the MNIST dataset. Four model settings were adopted, setting one is a heterogeneous liquid state machine combined with a sliding window, setting two is a heterogeneous liquid state machine using random connections, setting three is a non-heterogeneous liquid state machine combined with a sliding window, and setting four is a non-heterogeneous liquid state machine using random connections. The experimental results are as follows: liquid = 1000, the model with setting 1 achieved an accuracy of 97.38%, which was 1.76%, 0.15%, and 1.52% higher than that with setting 2, setting 3, and setting 4 respectively. liquid =8000, the model with setting 1 achieved an accuracy of 98.41%, which exceeded the MLP model with the same structure that required full training. From the above results, it can be seen that the method proposed in the present invention is more stable in image classification, and the setting of the sliding window can effectively improve the accuracy of image classification.
[0085] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for designing a liquid state machine combined with a sliding window, characterized in that: The following steps are involved: (1) Obtaining an image data set, preprocessing the image data as training data, and performing pulse coding conversion to obtain a pulse signal; (2) Construct a liquid state machine model combined with a sliding window, as follows: Select neurons from the liquid layer that need to be connected to the input layer neurons; construct connections from the input layer neurons to the selected liquid layer neurons through a sliding window; establish internal connections between the liquid layer neurons; and implement data classification through a trainable output layer; (3) Based on the training data in step (1), the key parameters of the liquid state machine are automatically adjusted by Bayesian optimization technology to find a set of better parameter settings to optimize the classification performance of the model; (4) After data processing, the image to be classified is input into the trained model to obtain the corresponding classification result.
2. The method for designing a liquid state machine in combination with a sliding window according to claim 1, characterized in that: In step (2), neurons that need to be connected to the input layer neurons are selected from the liquid layer. The specific process is as follows: Assume the number of neurons in the liquid layer is N liquid , select one of neurons are connected to the input layer, and when Equal to N liquid When , all liquid layer neurons are selected; when Less than N liquid When selecting, the selection is done by random selection or by sorting and screening according to the liquid layer connection topology index.
3. The method for designing a liquid state machine in combination with a sliding window according to claim 2, characterized in that: In step (2), the connection from the input layer neurons to the selected liquid layer neurons is constructed by means of a sliding window. The specific process is as follows: 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; Assume that the number of input mapping blocks extracted by a sliding window traversal process is q, then at least Only one sliding window traversal can ensure that the number of extracted input mapping blocks meets the requirements; For each selected liquid layer neuron, a structured input connection is constructed according to its corresponding input mapping block and the set input layer weights.
4. The method for designing a liquid state machine in combination with a sliding window according to claim 3, characterized in that: 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×W; let the sliding window size be k×k, the step size be s, mark its position with the upper left corner of the window, let the initial position be (0,0), the end position be (Hk,Wk), and the position of the window with sequence number i be (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, with a step length of s each time; during each movement, an input mapping block F is divided slide , and then build the connection from the input layer neurons to the corresponding liquid layer neurons according to the mapping block; when the right axis of the window exceeds the right edge of the input layer, it moves 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 completed.
5. The method for designing a liquid state machine in combination with a sliding window according to claim 3, characterized in that: Construct a structured input connection. The specific process is: Let liquid neuron h be the selected liquid layer neuron, and its corresponding input mapping block extracted by the sliding window is It includes k×k input layer neurons, and the input layer weights corresponding to the nodes are The connection weights are set by randomly selecting weights from candidate weights, generating weights according to Gaussian distribution, or using fixed weights; the connection matrix from the input layer neurons to the liquid layer neurons h is constructed based on the weights and mapping blocks. in 6. The method for designing a liquid state machine in combination with a sliding window according to claim 1, characterized in that: In step (2), in the process of establishing the internal connection of the liquid layer neurons, the neurons are connected according to the set topological connection rules, which include random connection rules, small-world network connection rules, scale-free network connection rules and connection rules based on distance probability.
7. The method for designing a liquid state machine in combination with a sliding window according to claim 1, characterized in that: In step (2), data classification is achieved through a trainable output layer, as follows: 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 passed to the output layer and mapped to the expected output through training.
8. The method for designing a liquid state machine in combination with a sliding window according to claim 1, characterized in that: In step (3), the key parameters of the liquid state machine are automatically adjusted by Bayesian optimization technology, as follows: Define the parameter search space, including the sliding window size k, topological connection scheme parameters, and the connection weight W from the input neuron to the liquid layer neuron IL , the connection weight W from the excitatory neurons in the liquid layer to the excitatory neurons EE , inhibitory neuron to inhibitory neuron connection weight W II , the connection weight W from inhibitory neurons to excitatory neurons IE , the connection weight W from excitatory neurons to inhibitory neurons EI ; Determine the optimal parameter combination through Bayesian optimization to achieve the best performance on the corresponding task.
Citation Information
Patent Citations
Pulse neural network image segmentation method based on FPGA
CN114612493A
LSM data merging and sorting method and device based on sliding window
CN118069891A
System and method of gesture recognition using a reservoir based convolutional spiking neural network
US20210397878A1
