Data identification method and device based on spiking neural network model, storage medium and wearable equipment

Abstract

The invention relates to the technical field of machine learning, and discloses a data recognition method and device based on a spiking neural network model, a storage medium and wearable equipment. In a preset spiking neural network model, at least any two adjacent layers of neurons meet a preset time dependency relationship, and at least one layer of neurons meet a preset spatial dependency relationship, the method comprises the steps of obtaining to-be-identified original feature data; performing feature extraction on the to-be-identified original feature data based on a preset time dependency relationship and a preset spatial dependency relationship to obtain a to-be-identified pulse sequence, the preset time dependency relationship and the preset spatial dependency relationship being determined based on a data mode of the to-be-identified original feature data; and performing data identification on the to-be-identified original feature data according to the to-be-identified pulse sequence. Compared with the prior art, the method can support the recognition of feature data of various modes.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a data recognition method, apparatus, storage medium and wearable device based on a spiking neural network model. Background Technology

[0002] Currently, spiking neural network (SNN) models are a class of computational models inspired by biological nervous systems, simulating the dynamic behavior of biological neurons. A spiking neural network consists of several layers, each containing a number of neurons, connected in a specific way. In practical use, raw feature data is input to the neurons in the form of pulses. The neurons possess an electromotive force (EMF), which decays exponentially over time to a stable level when no input is received. Upon receiving an input pulse, the EMF increases by a certain value. When the rate of increase exceeds the rate of decay, the EMF within the neuron grows larger and larger until it reaches a firing threshold. At this point, the neuron fires an output pulse to the next neuron, and its EMF quickly returns to a resting state. The next neuron then uses this output pulse as its input pulse, similarly increasing its value until it reaches the firing threshold and fires an output pulse. The neurons in the output layer then perform subsequent recognition tasks based on the sequence of output pulses.

[0003] However, the original feature data of different modalities have different characteristics. For example, the original feature data of text modality has bidirectional relationships, while the original feature data of biosignal modality only needs to focus on short-term dependencies. Existing SNN models generally only support the recognition of feature data of one modality. Therefore, how to enable SNN models to support the recognition of feature data of multiple modalities is an urgent problem to be solved. Summary of the Invention

[0004] The main objective of this application is to provide a data recognition method, apparatus, storage medium, and wearable device based on a spiking neural network model, aiming to solve the existing technical problem of how to enable the SNN model to support the recognition of feature data of multiple modalities.

[0005] To achieve the above objectives, embodiments of this application provide a data recognition method based on a spiking neural network model. The method is applied to a preset spiking neural network model, wherein at least two adjacent layers of neurons in the preset spiking neural network model satisfy a preset temporal dependency relationship, and at least one layer of neurons satisfies a preset spatial dependency relationship. The method includes:

[0006] Obtain the original feature data to be identified;

[0007] Based on the preset time dependency and the preset spatial dependency, feature extraction is performed on the original feature data to be identified to obtain the pulse sequence to be identified. The preset time dependency and the preset spatial dependency are determined based on the data mode of the original feature data to be identified.

[0008] Data identification is performed on the original feature data to be identified based on the pulse sequence to be identified.

[0009] In one embodiment, before the step of obtaining the original feature data to be identified, the method further includes:

[0010] The training method of the preset spiking neural network model is determined, and the target analysis operation is determined according to the training method, wherein the target analysis operation is a clustering operation or a classification operation;

[0011] The step of performing data recognition on the original feature data to be identified based on the pulse sequence to be identified includes:

[0012] The pulse sequence to be identified is analyzed according to the target analysis operation, and the original feature data to be identified is identified based on the analysis results.

[0013] In one embodiment, the step of determining the target analysis operation based on the training method includes:

[0014] When the training method is unlabeled training, the clustering operation is used as the target analysis operation;

[0015] The steps of analyzing the pulse sequence to be identified according to the target analysis operation and identifying the original feature data to be identified based on the analysis results include:

[0016] The pulse sequence to be identified is clustered, and the original feature data to be identified is identified based on the obtained clustering results.

[0017] In one embodiment, the step of determining the target analysis operation based on the training method includes:

[0018] When the training method is labeled training, the classification operation is used as the target analysis operation;

[0019] The steps of analyzing the pulse sequence to be identified according to the target analysis operation and identifying the original feature data to be identified based on the analysis results include:

[0020] The pulse sequence to be identified is encoded into a high-dimensional vector to obtain a high-dimensional vector to be identified;

[0021] The high-dimensional vector to be identified is classified using various preset high-dimensional vector templates, and the original feature data to be identified is identified based on the obtained classification results.

[0022] In one embodiment, the step of classifying the high-dimensional vector to be identified using various preset high-dimensional vector templates, and identifying the original feature data to be identified based on the obtained classification results, includes:

[0023] Determine the vector similarity between each of the preset high-dimensional vector templates and the high-dimensional vector to be identified, and sort the vector similarities.

[0024] The original feature data to be identified is identified based on the sorting results.

[0025] In one embodiment, before the step of obtaining the original feature data to be identified, the method further includes:

[0026] Obtain the original feature data of the sample;

[0027] Based on the preset time dependency and the preset spatial dependency, feature extraction is performed on the original feature data of the sample to obtain the sample pulse sequence. The preset time dependency and the preset spatial dependency are determined based on the data mode of the original feature data of the sample.

[0028] The sample pulse sequence is encoded using high-dimensional vector to obtain a sample high-dimensional vector, and the sample high-dimensional vector is superimposed to obtain a preset high-dimensional vector template.

[0029] In one embodiment, the step of superimposing the high-dimensional vectors of the samples to obtain a preset high-dimensional vector template includes:

[0030] The sample high-dimensional vectors are superimposed to obtain an initial high-dimensional vector template;

[0031] Template verification is performed on the initial high-dimensional vector template using the original feature data of the sample.

[0032] If the test result does not meet the preset test requirements, return to the step of superimposing the high-dimensional vector of the sample until the test result meets the preset test requirements, and then use the initial high-dimensional vector template as the preset high-dimensional vector template.

[0033] Furthermore, to achieve the above objectives, embodiments of this application also propose a data recognition device based on a spiking neural network model, the device comprising:

[0034] The data acquisition module is used to acquire the original feature data to be identified;

[0035] The feature extraction module is used to extract features from the original feature data to be identified based on preset time dependencies and preset spatial dependencies to obtain the pulse sequence to be identified. The preset time dependencies and preset spatial dependencies are determined based on the data modes of the original feature data to be identified.

[0036] The data recognition module is used to perform data recognition on the original feature data to be recognized based on the pulse sequence to be recognized.

[0037] Furthermore, to achieve the above objectives, this application also proposes a storage medium storing a data recognition program, which, when executed by a processor, implements the steps of the data recognition method based on the spiking neural network model described above.

[0038] Furthermore, to achieve the above objectives, this application also proposes a wearable device, which includes: a memory, a processor, and a data recognition program stored in the memory and executable on the processor. When the data recognition program is executed by the processor, it implements the steps of the data recognition method based on the spiking neural network model described above.

[0039] This application provides a data recognition method, apparatus, storage medium, and wearable device based on a spiking neural network model. The method is applied to a preset spiking neural network model, in which at least two adjacent layers of neurons satisfy a preset temporal dependency relationship, and at least one layer of neurons satisfies a preset spatial dependency relationship. The method includes: acquiring raw feature data to be recognized; extracting features from the raw feature data to be recognized based on the preset temporal dependency relationship and the preset spatial dependency relationship to obtain a pulse sequence to be recognized, wherein the preset temporal dependency relationship and the preset spatial dependency relationship are determined based on the data modality of the raw feature data to be recognized; and performing data recognition on the raw feature data to be recognized according to the pulse sequence to be recognized.

[0040] Because the pre-defined spiking neural network model of this application has at least two adjacent layers of neurons that satisfy a pre-defined temporal dependency relationship, and at least one layer of neurons that satisfy a pre-defined spatial dependency relationship, in practical use, the original feature data to be identified can be obtained first. Based on the original feature data, the pre-defined temporal dependency relationship and the pre-defined spatial dependency relationship can be determined. Then, feature extraction can be performed on the original feature data to be identified based on these pre-defined temporal dependency relationships and the pre-defined spatial dependency relationship. Finally, data recognition can be performed on the original feature data to be identified based on the obtained pulse sequence. Compared to existing SNN models that only support the recognition of feature data of one modality, this application can first set corresponding pre-defined temporal dependency relationships and pre-defined spatial dependency relationships for the original feature data to be identified of different modalities, and then perform feature extraction based on these pre-defined temporal dependency relationships and the pre-defined spatial dependency relationship to complete data recognition. Therefore, this application can support the recognition of feature data of multiple modalities. Attached Figure Description

[0041] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of the wearable device structure in the hardware operating environment involved in the embodiments of this application;

[0044] Figure 2 This is a flowchart illustrating the first embodiment of the data recognition method based on a spiking neural network model according to this application;

[0045] Figure 3 This is a flowchart illustrating the second embodiment of the data recognition method based on a spiking neural network model of this application;

[0046] Figure 4 This is a schematic diagram of weight update between two layers of neurons in the second embodiment of the data recognition method based on the spiking neural network model of this application;

[0047] Figure 5 This is a schematic diagram of the first weight update strategy in the second embodiment of the data recognition method based on the spiking neural network model of this application;

[0048] Figure 6This is a schematic diagram of the second weight update strategy in the second embodiment of the data recognition method based on the spiking neural network model of this application;

[0049] Figure 7 This is a schematic diagram of the third weight update strategy in the second embodiment of the data recognition method based on the spiking neural network model of this application;

[0050] Figure 8 This is a schematic diagram of the local pulse suppression constraint conditions in the second embodiment of the data recognition method based on the spiking neural network model of this application;

[0051] Figure 9 This is a flowchart illustrating the third embodiment of the data recognition method based on a spiking neural network model of this application;

[0052] Figure 10 This is a flowchart illustrating the fourth embodiment of the data recognition method based on a spiking neural network model of this application;

[0053] Figure 11 This is a schematic diagram of the multimodal clustering task in the fourth embodiment of the data recognition method based on the spiking neural network model of this application;

[0054] Figure 12 This is a flowchart illustrating the fifth embodiment of the data recognition method based on a spiking neural network model of this application;

[0055] Figure 13 This is a schematic diagram of similarity comparison in the fifth embodiment of the data recognition method based on the spiking neural network model of this application;

[0056] Figure 14 This is a schematic diagram of feature classification in the fifth embodiment of the data recognition method based on the spiking neural network model of this application;

[0057] Figure 15 This is a schematic diagram of the superposition operation in the fifth embodiment of the data recognition method based on the spiking neural network model of this application;

[0058] Figure 16 This is a before-and-after comparison diagram of the data recognition method based on the spiking neural network model in the fifth embodiment of this application;

[0059] Figure 17 This is a structural block diagram of the first embodiment of the data recognition device based on the spiking neural network model of this application.

[0060] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0061] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0062] Reference Figure 1 , Figure 1 This is a schematic diagram of the wearable device structure of the hardware operating environment involved in the embodiments of this application.

[0063] like Figure 1 As shown, the wearable device may include: a processor 1001, such as a central processing unit (CPU) or a neuromorphic chip; a communication bus 1002; a user interface 1003; a network interface 1004; and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, and optionally, it may also include a standard wired interface or a wireless interface. In this application, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or non-volatile memory (NVM), such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0064] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on wearable devices and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0065] like Figure 1 As shown, the memory 1005, which is identified as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a data recognition program based on a spiking neural network model.

[0066] exist Figure 1 In the wearable device shown, the network interface 1004 is mainly used to connect to the backend server and communicate with the backend server; the user interface 1003 is mainly used to connect to the user device; the wearable device calls the data recognition program based on the spiking neural network model stored in the memory 1005 through the processor 1001 and executes the data recognition method based on the spiking neural network model provided in the embodiments of this application.

[0067] It's important to note that current Spiking Neural Network (SNN) models are computational models inspired by biological nervous systems, simulating the dynamic behavior of biological neurons. A SNN consists of several layers, each containing a number of neurons, connected in a specific way. In practice, raw feature data is input to the neurons in the form of pulses. Each neuron possesses an electromotive force (EMF), which decays exponentially over time to a stable level when no input is received. Upon receiving an input pulse, the EMF increases by a certain value. When the rate of increase exceeds the rate of decay, the EMF within the neuron grows larger and larger until it reaches a firing threshold. At this threshold, the neuron fires an output pulse to the next neuron, after which its EMF quickly returns to a resting state. The next neuron then uses this output pulse as its input pulse, similarly increasing its value until it reaches the firing threshold and fires an output pulse. The neurons in the output layer then use the sequence of output pulses to complete subsequent recognition tasks.

[0068] However, the original feature data of different modalities have different characteristics. For example, the original feature data of text modality has bidirectional relationships, while the original feature data of biosignal modality only needs to focus on short-term dependencies. Existing SNN models generally only support the recognition of feature data of one modality. Therefore, how to enable SNN models to support the recognition of feature data of multiple modalities is an urgent problem to be solved.

[0069] Therefore, in order to solve the above-mentioned defects, this embodiment provides a data recognition method based on a spiking neural network model. It can first set corresponding preset time dependencies and preset spatial dependencies for the original feature data to be identified in different modalities, and then perform feature extraction based on the preset time dependencies and preset spatial dependencies to complete data recognition. Thus, this embodiment can support the recognition of feature data in multiple modalities.

[0070] For ease of understanding, the following is combined with Figures 2 to 17 The data recognition method based on the spiking neural network model provided in the embodiments of this application will be described in detail.

[0071] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the data recognition method based on a spiking neural network model according to this application. The first embodiment of the data recognition method based on a spiking neural network model according to this application is presented as follows: Figure 2As shown, in this embodiment, the above method is applied to a preset spiking neural network model, in which at least two adjacent layers of neurons satisfy a preset time dependency relationship, and at least one layer of neurons satisfies a preset spatial dependency relationship.

[0072] It is understood that the method of this embodiment can be applied to wearable devices, wherein the wearable device can be an edge device worn on the user's body, such as a smart wristband or smart bracelet, or other wearable devices, and this embodiment does not limit it. In actual use, the aforementioned preset spiking neural network model can be set on the wearable device. The preset spiking neural network model can be used to identify the user's feature data, thereby completing human-computer interaction. The execution subject of this method embodiment can be the aforementioned wearable device (hereinafter referred to as the device). This embodiment and the following embodiments will be described using the device as an example.

[0073] It should be understood that the above-mentioned preset spiking neural network model can have several layers, and each layer can have several neurons. The specific number of layers and the number of neurons in each layer can be set according to the actual situation. This embodiment does not limit this.

[0074] It should also be understood that the aforementioned preset time dependency can be a relationship that adjusts the degree of association between neurons in two adjacent layers; the aforementioned preset spatial dependency can be a relationship that adjusts the spiking of neurons in at least one layer. Therefore, before use, the corresponding preset time dependency and preset spatial dependency can be determined in advance according to the task to be performed (i.e., according to the modality of the feature data to be identified), and the aforementioned preset spiking neural network model can be constructed based on the preset time dependency and preset spatial dependency.

[0075] The method provided in this embodiment may specifically include:

[0076] Step S10: Obtain the original feature data to be identified.

[0077] It should be noted that the aforementioned raw feature data to be identified can be the feature data that the current user needs to identify. In this embodiment, sensors for collecting raw feature data of different modalities can be set within the device. For example, when the user needs to recognize gestures, electromyography (EMG) signals from the hand can be collected by an EMG sensor within the device as the raw feature data to be identified; alternatively, inertial measurement sensors within the device can collect inertial data generated when the user makes a gesture as the raw feature data to be identified; or, for example, an image of the user's gesture can be captured by a camera within the device as the raw feature data to be identified; of course, the raw feature data to be identified can also be the user's voice data, text data, etc., and sensors for collecting corresponding data can be set within the device. This embodiment will not elaborate on these specific examples.

[0078] Step S20: Based on the preset time dependency and the preset spatial dependency, feature extraction is performed on the original feature data to be identified to obtain the pulse sequence to be identified. The preset time dependency and the preset spatial dependency are determined based on the data mode of the original feature data to be identified.

[0079] Understandably, before use, corresponding preset temporal dependencies and preset spatial dependencies can be pre-constructed according to the requirements of the scenario task. When constructing preset temporal dependencies, it is sufficient that the neurons in at least any two adjacent layers of the preset spiking neural network model satisfy the condition, although there can be more layers. When constructing preset spatial dependencies, it is sufficient that the neurons in at least one layer of the preset spiking neural network model satisfy the condition, although there can be more layers. The specific number of layers can be set by the user, and this embodiment does not limit this. The above-mentioned preset temporal dependencies and preset spatial dependencies can be set in the input layer, intermediate layer, or output layer of the preset spiking neural network model. This embodiment uses the intermediate layer as an example.

[0080] In this embodiment, the connection structure between neurons in each layer can be configured with corresponding connection modes according to different task requirements, including but not limited to residual connections and bidirectional connections. For example, taking biological signals as an example, a three-layer SNN network structure can be configured, including an input layer of 64 neurons, an intermediate layer of 128 neurons, and an output layer of 64 neurons. Of course, other structures are also possible, and this embodiment does not limit them.

[0081] In practical use, when the original feature data to be identified is input into the input layer of the preset spiking neural network model, the electromotive force in each neuron can be accumulated. When the electromotive force reaches the firing threshold corresponding to the neuron, it will fire an output pulse to the next layer of neurons. The next layer of neurons continues to operate in the same way until the output layer, thus obtaining the above-mentioned pulse sequence to be identified.

[0082] In the intermediate layers, for neurons with preset time dependencies, the correlation between neurons in the two layers can be adjusted using these preset time dependencies, thereby adjusting the output pulses emitted by neurons in the next layer. For neurons with preset spatial dependencies, the pulse emission of neurons in that layer can be adjusted using these preset spatial dependencies. Furthermore, by using the corresponding preset time and spatial dependencies, the preset spiking neural network model can adapt to different modalities of the original feature data to be identified, thus enabling the preset spiking neural network model to support the identification of feature data in different data modalities.

[0083] Step S30: Perform data recognition on the original feature data to be identified based on the pulse sequence to be identified.

[0084] After obtaining the pulse sequence to be identified, data identification can be performed based on the pulse sequence to complete the identification or classification task. Data identification can be performed by setting sequence templates and comparing the similarity between the pulse sequence to be identified and each sequence template. Other identification methods can also be used, and this embodiment does not limit this approach.

[0085] In this embodiment, a preset time dependency and a preset spatial dependency can be set for the original feature data to be identified in different modalities. Then, feature extraction is performed based on the preset time dependency and the preset spatial dependency to complete the data identification. Thus, this embodiment can support the identification of feature data in multiple modalities.

[0086] Furthermore, in order to obtain the aforementioned preset spiking neural network model, in this embodiment, before the step of obtaining the original feature data to be identified, the following steps are also included:

[0087] Step S01: Obtain the original feature data of the sample;

[0088] Step S02: Based on the original feature data of the sample, the initial spiking neural network model is trained through the preset time dependency and the preset spatial dependency to obtain the preset spiking neural network model.

[0089] It should be noted that the original feature data of the above samples can be the feature data used during training. During training, the corresponding original feature data of the samples can be selected according to the scenario task to be used later, such as biosignals, speech data, text data or image data, etc. This embodiment does not impose any restrictions on this.

[0090] It is understandable that the aforementioned initial spiking neural network model can be a model prior to training a pre-defined spiking neural network model. This initial spiking neural network model can be a model after incorporating the aforementioned pre-defined temporal and spatial dependencies into the SNN model. Furthermore, when selecting the aforementioned pre-defined temporal and spatial dependencies to incorporate into the SNN model, they can be determined based on the data modality of the original feature data of the samples.

[0091] In practical use, after the aforementioned device inputs the original feature data of the sample into the initial spiking neural network model, it can train the initial spiking neural network model based on the preset time dependencies and preset spatial dependencies in the initial spiking neural network model, thereby obtaining the aforementioned preset spiking neural network model. Since the preset time dependencies and preset spatial dependencies are determined according to different data modalities, the obtained preset spiking neural network model can support the recognition of feature data from multiple modalities.

[0092] Furthermore, considering the different characteristics of the original feature data of samples under different modalities, different processing methods are required due to the different characteristics of the original feature data of samples under different modalities. Therefore, in this embodiment, the step of training the initial spiking neural network model based on the original feature data of the samples through the preset time dependency and the preset spatial dependency to obtain the preset spiking neural network model includes:

[0093] Step S021: Determine the corresponding target preprocessing strategy, target encoding strategy, and target input rules based on the data modality of the original feature data of the sample.

[0094] It should be understood that the aforementioned target preprocessing strategy can be a strategy for preprocessing the original feature data of the samples; the aforementioned target encoding strategy can be a strategy for encoding the original feature data of the samples; and the aforementioned target input rule can be a rule for inputting the original feature data of the samples into the initial preset spiking neural network model according to a certain format. The aforementioned target preprocessing strategy, target encoding strategy, and target input rule can all be determined based on the corresponding data modality.

[0095] Step S022: Preprocess the original feature data of the sample according to the target preprocessing strategy, and perform pulse coding on the preprocessed original feature data of the sample according to the target coding strategy.

[0096] In practical use, after the above-mentioned devices obtain the original feature data of the samples, they can be processed according to the corresponding preprocessing strategies. The above-mentioned preprocessing strategies may include, but are not limited to: regularization, standardization, filtering, rectification, data cleaning, adaptive scaling, etc.

[0097] Next, the preprocessed original sample feature data is subjected to spiking neural coding to obtain encoded original sample feature data containing the original feature information. Spiking neural coding can also be performed according to a corresponding target coding strategy. In this embodiment, the target coding strategy may include, but is not limited to, frequency-based pulse coding, differential pulse coding, or adaptive differential pulse coding. Furthermore, the encoded original sample feature data obtained after coding can be in the form of a pulse sequence.

[0098] For example, when the original feature data to be identified is a biological signal, bandpass filtering, signal rectification, outlier detection, missing value repair, signal smoothing, and standardization can be performed first. Then, differential pulse coding is used for encoding. Since the sensor has multiple channels (e.g., N channels) during acquisition and will continuously acquire data for a period of time (e.g., T duration), N pulse sequences can be obtained after encoding, and the duration of each pulse sequence is T.

[0099] Step S023: Adjust the input of the encoded original feature data of the sample according to the target input rule, and train the initial spiking neural network model based on the adjusted original feature data of the sample through the preset time dependency and the preset spatial dependency to obtain the preset spiking neural network model.

[0100] After pulse encoding, since the input to the initial spiking neural network model needs to be in a certain format, the aforementioned device can adjust the original feature data of the encoded samples according to the target input rules. These target input rules may include, but are not limited to, rules regarding the number of neurons in the input layer, the input stride, and the window length.

[0101] The number of neurons in the input layer can be the number of neurons set in the input layer of the initial spiking neural network model. During input, the number of input pulse sequences needs to be consistent with the number of neurons in the input layer. The input step size can be the step size of the input pulse sequence. This parameter is generally related to frequency encoding and is generally required when the modal data is image data. For other modalities, it can be set to the default value (e.g., 1). For example, if the input step size is 32, then 32 sets of data can be input sequentially each time. The window length can be the length of the pulse sequence when input to the neurons in the input layer. It can generally be set when the modal data is biological signal and can be set to 1. Of course, it can also be other values. This embodiment does not limit this.

[0102] After adjusting the original feature data of the encoded sample according to the above target input rules, it can be input into the initial spiking neural network model for training, thereby obtaining the above-mentioned preset spiking neural network model.

[0103] It should be emphasized that when in use, the above-mentioned device can also follow the above process, that is, adjust the original feature data to be identified according to the above-mentioned target preprocessing strategy, target encoding strategy and target input rules before inputting it into the input layer of the preset spiking neural network model. The specific process is the same as that during training, so this embodiment will not elaborate on it.

[0104] This embodiment can set corresponding target preprocessing strategies, target encoding strategies, target input rules, preset time dependencies, and preset spatial dependencies according to different data modalities, thereby enabling the spiking neural network model to support the recognition of feature data of multiple modalities.

[0105] Reference Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the data recognition method based on a spiking neural network model of this application. Based on the first embodiment described above, a second embodiment of the data recognition method based on a spiking neural network model of this application is proposed.

[0106] like Figure 3 As shown, in order to establish the aforementioned preset temporal dependencies and preset spatial dependencies in the initial spiking neural network model, in this embodiment, the step of training the initial spiking neural network model based on the original feature data of the samples through the preset temporal dependencies and preset spatial dependencies to obtain the preset spiking neural network model includes:

[0107] Step S0231: Select a target weight update strategy from each initial weight update strategy according to the data mode of the original feature data of the sample, and determine the local impulse suppression constraint condition according to the data mode of the original feature data of the sample.

[0108] It should be noted that the initial weight update strategy described above can be a strategy for updating the connection weight matrix between neurons in each layer of the SNN model. When the SNN model is in use, since neurons in the previous layer output pulses to neurons in the next layer, the next layer neurons increase their electromotive force based on the received output pulses and their own spatiotemporal dynamics model (e.g., Leaky Integrate-and-Fire (LIF) model or Integrate-and-Fire (IF) model) until they reach the firing threshold, and then continue to output pulses to the next layer of neurons. However, it is important to emphasize that when calculating neuron dynamics, neuron dynamics characteristics, including but not limited to membrane potential decay, membrane potential summation, threshold firing, post-firing membrane potential decay, and refractory period, can be used depending on the scenario and model characteristics.

[0109] When neurons in one layer transmit output pulses to neurons in the next layer, the connection weight matrix between the two layers can be used to characterize the connection strength between them. This connection weight matrix determines how the output pulse is transmitted to the next layer. During training, the connection weight matrix between each layer of neurons is continuously adjusted, thereby changing the intensity of the output pulse transmitted to the next layer. The initial weight update strategy described above is a strategy for updating the connection weight matrix during training. In this embodiment, several different initial weight update strategies can be set, and a suitable initial weight update strategy can be selected as the target weight update strategy based on the current data modality.

[0110] It is important to emphasize that the above-mentioned method of updating the connection weight matrix by selecting an appropriate target weight update strategy can enable spike-timing-dependent plasticity (STDP) between two layers of neurons, that is, a time-dependent relationship. Traditional STDP presents two scenarios: first, the firing pulse of the presynaptic membrane (i.e., the neuron in the previous layer) precedes the firing pulse of the postsynaptic membrane (i.e., the neuron in the next layer), resulting in increased synaptic strength, known as Hebbian Long-Term Potentiation (LTP); second, the firing pulse of the presynaptic membrane follows the firing pulse of the postsynaptic membrane, resulting in decreased synaptic strength, known as Long-Term Depression (LTD).

[0111] For ease of understanding, please refer to Figure 4 , Figure 4 This is a schematic diagram illustrating the weight update between two layers of neurons in the second embodiment of the data recognition method based on a spiking neural network model in this application, as shown below. Figure 4As shown, 401 represents the output pulse, 402 represents the presynaptic membrane, 403 represents the postsynaptic membrane, and 404 represents the spatiotemporal dependency between two adjacent neurons, i.e., the connection weight. In w+δw, w is the initial weight, δw is the increment of the connection weight, t1 is the time point when the presynaptic membrane emits the output pulse, and t2 is the time point when the postsynaptic membrane emits the output pulse. When t1 is less than t2, it indicates that the presynaptic membrane emits the output pulse earlier than the postsynaptic membrane; when t1 is greater than t2, it indicates that the presynaptic membrane emits the output pulse later than the postsynaptic membrane.

[0112] In this embodiment, several different initial weight update strategies are set. Specifically, the initial weight update strategies include at least a first weight update strategy or a second weight update strategy.

[0113] The first weight update strategy is to increase the connection weight between the presynaptic membrane and the postsynaptic membrane when the presynaptic membrane emits a pulse earlier than the postsynaptic membrane, and to decrease the connection weight between the presynaptic membrane and the postsynaptic membrane when the presynaptic membrane emits a pulse later than the postsynaptic membrane.

[0114] The second weight update strategy is to increase the connection weight between the presynaptic membrane and the postsynaptic membrane when the presynaptic membrane emits a pulse earlier than the postsynaptic membrane, or when the presynaptic membrane emits a pulse later than the postsynaptic membrane within a first duration, and to decrease the connection weight between the presynaptic membrane and the postsynaptic membrane when the presynaptic membrane emits a pulse later than the postsynaptic membrane within a second duration, wherein the second duration is not less than the first duration.

[0115] For ease of understanding, please refer to Figure 5 as well as Figure 6 , Figure 5 This is a schematic diagram of the first weight update strategy in the second embodiment of the data recognition method based on the spiking neural network model of this application. Figure 6 This is a schematic diagram of the second weight update strategy in the second embodiment of the data recognition method based on the spiking neural network model of this application;

[0116] like Figure 5As shown, the horizontal axis represents the pulse time interval, i.e., the difference between t2 and t1. When t2-t1 is greater than 0, it indicates that the presynaptic membrane emits an output pulse first, followed by the postsynaptic membrane. When t2-t1 is less than 0, it indicates that the postsynaptic membrane emits a pulse first, followed by the presynaptic membrane. The vertical axis represents the value of the connection weight increment δw. In this embodiment, the above-mentioned first weight update strategy can be the traditional STDP. Specifically, if the presynaptic membrane emits a pulse earlier than the postsynaptic membrane, the δw between the presynaptic and postsynaptic membranes is increased, where δw greater than 0 indicates an increase. If the presynaptic membrane emits a pulse later than the postsynaptic membrane, the δw between the presynaptic and postsynaptic membranes is decreased, where δw less than 0 indicates a decrease.

[0117] However, it needs to be emphasized that, Figure 5 In this context, if the presynaptic membrane fires a pulse and the postsynaptic membrane fires a pulse very quickly afterward, it indicates a high correlation between the presynaptic and postsynaptic neurons, requiring increased weighting. Conversely, as the presynaptic and postsynaptic membrane pulse firing times increase, it indicates a smaller influence of the presynaptic neuron on the postsynaptic neuron, necessitating less weighting (i.e.,...). Figure 5 When t2-t1 is greater than 0, the larger t2-t1 is, the smaller δw is. Conversely, if the presynaptic membrane fires a pulse shortly after the postsynaptic membrane fires a pulse, it indicates a high correlation between the two neurons, requiring a reduction in weight. As the pulse firing times of the postsynaptic and presynaptic membranes increase, it indicates a smaller influence of the postsynaptic neuron on the presynaptic neuron, and the less weight needs to be reduced (i.e., ...). Figure 5 When t2-t1 is less than 0, the smaller t2-t1 is, the larger δw is.

[0118] The above Figure 5 This weight update strategy is too absolute, only representing a strict sequential relationship between neurons. However, since this embodiment contains feature data from different modalities, the relationship is not necessarily this strict. For example, text has bidirectional correlation, and biological signals need to focus more on short-term dependencies. Therefore, in this embodiment, the aforementioned second weight update strategy is proposed.

[0119] like Figure 6 As shown, the definitions of the x-axis and y-axis are the same as those described above. Figure 5 Consistency, in Figure 6 In the weight update strategy shown, if the presynaptic membrane emits a pulse earlier than the postsynaptic membrane, or the presynaptic membrane emits a pulse later than the postsynaptic membrane for a first duration (i.e., ... Figure 6 When a pulse is emitted within the interval from the 0 point to the intersection of the curve and the horizontal axis, the δw between the presynaptic and postsynaptic membranes is increased; if the presynaptic membrane is later than the postsynaptic membrane by the second duration (i.e., Figure 6 The point where the curve intersects the horizontal axis emits a pulse in the negative direction of the horizontal axis, reducing the δw between the presynaptic membrane and the postsynaptic membrane;

[0120] Specifically, Figure 6 This can be understood as, within a certain time frame, when the postsynaptic membrane emits a pulse before the presynaptic membrane, it can also indicate a strong correlation between the preceding and following neurons. It can express a certain bidirectional relationship and is applicable to data with bidirectional correlation properties, such as text.

[0121] Furthermore, considering that only a short period of time is needed to focus on dependencies, this embodiment also proposes a weight update strategy, specifically, the initial weight update strategy includes at least a third weight update strategy;

[0122] The third weight update strategy is to increase the connection weight between the presynaptic membrane and the postsynaptic membrane when the pulse is emitted within a third time period earlier than the postsynaptic membrane, or within a fourth time period later than the presynaptic membrane; and to decrease the connection weight between the presynaptic membrane and the postsynaptic membrane when the pulse is emitted within a fifth time period earlier than the postsynaptic membrane, or within a sixth time period later than the presynaptic membrane. The fifth time period is not less than the third time period, and the sixth time period is not less than the fourth time period.

[0123] Reference Figure 7 , Figure 7 This is a schematic diagram of the third weight update strategy in the second embodiment of the data recognition method based on the spiking neural network model of this application. Figure 7 In the above, the definitions of the x-axis and y-axis are the same as those mentioned above. Figure 5 Consistency, in Figure 7 In the weight update strategy shown, if the presynaptic membrane is earlier than the postsynaptic membrane by the third time interval (i.e. Figure 7 The pulse is emitted within the range from point 0 in the positive direction of the horizontal axis to the intersection of the curve and the horizontal axis, or the presynaptic membrane emits a pulse four time intervals later than the postsynaptic membrane (i.e., Figure 7 When a pulse is emitted within the range from point 0 in the negative direction of the horizontal axis to the intersection of the curve and the horizontal axis, the δw between the presynaptic and postsynaptic membranes is increased; if the presynaptic membrane is earlier than the postsynaptic membrane by five time intervals (i.e., ... Figure 7 The pulse is emitted from the intersection of the curve and the abscissa in the positive direction of the abscissa (or from the presynaptic membrane to the postsynaptic membrane six time intervals later). Figure 7 When a pulse is emitted from the intersection of the curve and the horizontal axis in the negative direction of the horizontal axis, the δw between the presynaptic membrane and the postsynaptic membrane is reduced.

[0124] Specifically, Figure 7This can be understood as follows: within a certain time range, if the time interval between the firing of pulses by the preceding and following neurons is too long, the connection weight should increase negatively to ensure that there is a correlation between the preceding and following neurons within a certain time step. This can avoid interference from many irrelevant factors and enhance robustness. It can be applied to data such as biological signals that require more attention to short-term dependencies.

[0125] It should be emphasized that the second duration is not less than the first duration, the fifth duration is not less than the third duration, and the sixth duration is not less than the fourth duration; specifically, the first to sixth durations can be set according to the actual situation, and this embodiment does not impose any restrictions on them. Furthermore, the specific weight update strategy selected can also be set according to the use case, and this embodiment also does not impose any restrictions on it.

[0126] In practical use, when establishing time dependencies, the above-mentioned device can select a suitable target weight update strategy from each initial weight update strategy based on the data mode of the acquired original feature data of the sample; and can determine the local impulse suppression constraint conditions based on the data mode of the original feature data of the sample.

[0127] It is understandable that the aforementioned local impulse inhibition constraint can be a strategy for inhibiting the firing of impulses by neurons within a layer. The aforementioned local impulse inhibition constraint can achieve a horizontal inhibition effect, thereby forming the aforementioned preset spatial dependency.

[0128] For ease of understanding, please refer to Figure 8 , Figure 8 This is a schematic diagram of the local pulse suppression constraint conditions in the second embodiment of the data recognition method based on the spiking neural network model of this application, as shown below. Figure 8 As shown, the traditional SNN model (i.e. Figure 8 In the traditional SNN model, there is no local impulse suppression constraint within a single layer. Here, 801 represents the neuron impulse firing rule of the traditional SNN model, specifically, when the membrane potential in the neuron is greater than or equal to the firing threshold (i.e., ... Figure 8 When the parameter τ is specified, the neuron fires a pulse. Example 802 illustrates neuron pulse firing in a traditional SNN model. Specifically, if there are 9 neurons with membrane potentials of 4, 5, 7, 2, 5, 3, 0, 1, and 3 respectively, then the neurons firing pulses are those with membrane potentials of 4, 5, 7, 5, 3, and 3 respectively. Figure 8 If the membrane potential is greater than or equal to τ, then discharge occurs.

[0129] Continue as Figure 8 As shown, 803 represents the neuron firing rule after adding local impulse suppression constraints in this embodiment, 804 represents an example of neuron firing after adding local impulse suppression constraints, and 805 represents an example of a local impulse suppression constraint. Exemplarily, in... Figure 8In this context, local impulse inhibition constraints may include neuron connection conditions, membrane potential threshold conditions, and impulse emission quantity conditions; they may also include constraint time, meaning that the above local impulse inhibition constraints occur within the constraint time and do not occur outside the constraint time. The specific constraints can be set according to the actual situation.

[0130] The neuronal connectivity condition can be the condition for the connection relationship between neurons in this layer, specifically including the number of connected neurons and the connection relationship; the membrane potential threshold condition can be the condition for the firing threshold corresponding to the firing pulse; and the pulse emission quantity condition can be the condition for the number of neurons in this layer that are allowed to fire pulses. Figure 8 As shown, if there are n neuron connections, the number of pulses emitted is set to k.

[0131] Figure 8 There are two local impulse suppression constraints; one is n=5, k=2, meaning there are 5 neuron connections, allowing 2 neurons to fire impulses. Figure 8 In the first method, neurons 1, 3, 7, 8, and 9 are connected sequentially from left to right. Since neurons 1, 3, and 9 would normally fire pulses in this connection, but the number of firing pulses is limited, two neurons are randomly selected to fire pulses, allowing neurons 1 and 3 to fire. Alternatively, if n=4 and k=2, there are four neuron connections, allowing two neurons to fire pulses. Figure 8 The second, fourth, fifth, and sixth neurons are connected sequentially from left to right. Since the second and fifth neurons originally emit pulses, they can be directly allowed to emit pulses. However, after applying the aforementioned local pulse inhibition constraint, the sixth and ninth neurons do not emit pulses (i.e., ...). Figure 8 If the membrane potential is greater than or equal to τ, and the membrane potential is among the k largest membrane potentials in the locally impulse-inhibited neurons, then firing occurs.

[0132] It should be emphasized that a number of local impulse inhibition constraints can be set within the aforementioned layer of neurons, and the neuron connection conditions, membrane potential threshold conditions, and pulse emission quantity conditions can all be set according to the task scenario. This embodiment does not impose any restrictions on these.

[0133] It should also be emphasized that when the number of neurons that originally emitted pulses is higher than the set pulse emission number condition, the membrane potentials of each neuron can be sorted first, and neurons that meet the pulse emission number condition can be selected in descending order for pulse firing. Of course, custom firing rules can also be defined, and this embodiment does not restrict this. When the number of neurons that originally emitted pulses is not higher than the set pulse emission number condition, all neurons that originally emitted pulses will fire pulses.

[0134] By setting the local impulse suppression constraint as described above, we can avoid too many neurons firing impulses in the same layer. This allows the model to learn feature data from more modalities and avoids overload after learning feature data from fewer modalities, as is the case with traditional SNN models. It also avoids too many overlapping impulses firing in different modalities.

[0135] Step S0232: Based on the target weight update strategy, establish the preset temporal dependency between neurons in at least any two adjacent layers of the initial spiking neural network model, and establish the preset spatial dependency on neurons in at least one layer of the initial spiking neural network model.

[0136] In practical use, after determining the target weight update strategy and local impulse suppression constraint based on the data modality of the original feature data of the sample, the target weight update strategy can be set between at least any two adjacent layers of neurons in the initial spiking neural network model to establish a preset time dependency, and the local impulse suppression constraint can be set between neurons in at least one layer of the initial spiking neural network model to establish a preset spatial dependency.

[0137] Step S0233: Based on the original feature data of the sample, train the established initial spiking neural network model through the preset time dependency and the preset spatial dependency to obtain the preset spiking neural network model.

[0138] After establishing the aforementioned preset time dependency and preset space dependency, the initial spiking neural network model can be trained using the original sample feature data, specifically the original sample feature data adjusted according to the aforementioned target input rules, thereby obtaining the aforementioned preset spiking neural network model.

[0139] Reference Figure 9 , Figure 9 This is a flowchart illustrating the third embodiment of the data recognition method based on a spiking neural network model of this application. Based on the above embodiments, the third embodiment of the data recognition method based on a spiking neural network model of this application is proposed.

[0140] When training the initial spiking neural network model using the aforementioned preset time dependencies and preset spatial dependencies, such as Figure 9 As shown, the specific process is as follows: Based on the original feature data of the samples, the initial spiking neural network model is trained using the preset time dependency and the preset spatial dependency to obtain the preset spiking neural network model. This includes:

[0141] Step S02331: Update the initial connection weight matrix in the established initial spiking neural network model based on the original feature data of the sample and according to the target weight update strategy.

[0142] It should be noted that the initial connection weight matrix mentioned above can be the connection weight matrix established in the initial spiking neural network model. This initial connection weight matrix can be consistent with that in the traditional SNN model, and this embodiment will not elaborate on this.

[0143] Step S02332: Accumulate the membrane potential of the neuron using the updated initial connection weight matrix, and locally suppress the pulse emission of the neuron using the local impulse suppression constraint.

[0144] During training, for neurons with a target weight update strategy, the initial connection weight matrix can be continuously updated using this strategy. In each new training iteration, the updated initial connection weight matrix can be used to accumulate membrane potentials for the pulses fired by neurons in the previous layer. For neurons with a local impulse suppression constraint, this constraint can be used to locally suppress neurons capable of firing pulses, reducing the number of neurons firing pulses.

[0145] Specifically, the step of locally suppressing the pulse emission of the neuron through the local pulse suppression constraint condition includes:

[0146] The neurons to be constrained are determined by the neuron connectivity conditions, and target neurons are selected from the neurons to be constrained by the membrane potential threshold conditions; the pulse emission of the target neurons is locally suppressed based on the number of pulse emissions.

[0147] It is understood that the above local impulse inhibition constraints may include neuron connection conditions, membrane potential threshold conditions, and pulse emission quantity conditions.

[0148] The neurons to be constrained can be neurons connected within a single layer. The target neurons can be neurons whose membrane potential satisfies the firing threshold.

[0149] In practical use, the connected neurons can be determined first based on the neuron connection conditions, and these neurons can be used as the neurons to be constrained. It can also be determined whether the membrane potential of each neuron to be constrained reaches the firing threshold. The neurons to be constrained that reach the firing threshold are selected as the target neurons. Then, neurons that meet the number of pulses to be emitted are selected from the target neurons for pulse emission, thereby completing local inhibition. For neurons that are not subject to local pulse inhibition constraints, normal emission is sufficient. This embodiment will not elaborate on this.

[0150] Step S2333: Train the established initial spiking neural network model based on the cumulative results and the suppression results to obtain the preset spiking neural network model.

[0151] After accumulating the membrane potential using the updated initial connection weights and firing pulses under local pulse suppression constraints, the initial spiking neural network model can be trained based on the accumulation and suppression results to obtain the aforementioned preset spiking neural network model. Similarly, during subsequent recognition, the membrane potential can be accumulated using the updated initial connection weight matrix, and the emitted pulses can be locally suppressed using local pulse suppression constraints to obtain the pulse sequence to be identified corresponding to the original feature data. Furthermore, the initial connection weight matrix can be continuously updated using the aforementioned target weight update strategy. Since the specific process is similar to the training process described above, this embodiment will not elaborate further.

[0152] This embodiment selects the corresponding target weight update strategy from each initial weight update strategy through different data modes, and sets the corresponding local impulse suppression constraints, so that the obtained preset spiking neural network model can support the recognition of feature data under different data modes.

[0153] Reference Figure 10 , Figure 10 This is a flowchart illustrating the fourth embodiment of the data recognition method based on a spiking neural network model of this application. Based on the above embodiments, the fourth embodiment of the data recognition method based on a spiking neural network model of this application is proposed.

[0154] Considering that existing spiking neural network models generally require manual pre-classification and labeling of feature data before training, their flexibility is relatively poor. Therefore, if... Figure 10 As shown, in this embodiment, before the step of obtaining the original feature data to be identified, the method further includes:

[0155] Step S03: Determine the training method of the preset spiking neural network model, and determine the target analysis operation according to the training method. The target analysis operation is a clustering operation or a classification operation.

[0156] Accordingly, the above-mentioned step of performing data recognition on the original feature data to be identified based on the pulse sequence to be identified includes:

[0157] Step S31: Analyze the pulse sequence to be identified according to the target analysis operation, and identify the original feature data to be identified based on the analysis results.

[0158] It should be noted that the above training method can be the method used when training the initial spiking neural network model. In this embodiment, the above training method can include labeled training and unlabeled training, wherein the labeled training can be used to train the initial spiking neural network model using the original feature data of labeled samples, and the unlabeled training can be used to train the initial spiking neural network model using the original feature data of unlabeled samples.

[0159] It is understood that the aforementioned target analysis operation can be the analysis operation used when the preset spiking neural network model performs data recognition on the pulse sequence to be identified. In this embodiment, the corresponding analysis operation can be set through different training methods. Specifically, the aforementioned target analysis operation can be a classification operation or a clustering operation. For training using an unlabeled method, since there are no labels, clustering operations can be used for analysis during data recognition to complete the data recognition of the original feature data to be identified. For training using a labeled method, since it has labels, classification operations can be directly used for analysis during data recognition to complete the data recognition of the original feature data to be identified.

[0160] Furthermore, when it is unlabeled training, the specific process is the steps of determining the target analysis operation according to the training method, including:

[0161] Step S031: When the training method is unlabeled training, the clustering operation is used as the target analysis operation;

[0162] Accordingly, the steps described above for analyzing the pulse sequence to be identified according to the target analysis operation, and for identifying the original feature data to be identified based on the analysis results, include:

[0163] Step S311: Cluster the pulse sequence to be identified, and perform data identification on the original feature data to be identified based on the obtained clustering results.

[0164] It should be understood that when the training method is unlabeled training, the aforementioned device can use clustering as the target analysis operation. In subsequent use, the device can perform unsupervised learning and predictive applications. That is, after obtaining the pulse sequence to be identified, it can perform a clustering operation on the pulse sequence, and then use the category corresponding to the obtained clustering result as the data identification result, that is, the category corresponding to the original feature data to be identified. For example, when performing user habit analysis, a user performs an action, the device collects the action and generates the original feature data to be identified, and after passing through the aforementioned preset spiking neural network model, it can output the pulse sequence to be identified. Then, a clustering operation is performed to obtain the category to which the action belongs.

[0165] It is important to emphasize that when unlabeled training is used, the analysis operation during training can also be the clustering operation described above. Since the implementation process is similar to the usage process described above, this embodiment will not elaborate further. Furthermore, when performing clustering, the aforementioned local impulse suppression constraint can be added to the output layer of the preset spiking neural network model to obtain a sparse impulse sequence, facilitating subsequent clustering tasks.

[0166] It should also be emphasized that, since traditional clustering methods are suitable for clustering single modalities, and since this embodiment supports multimodal data recognition, when at least two modalities exist, reference can be made to... Figure 11 , Figure 11 This is a schematic diagram of the multimodal clustering task in the fourth embodiment of the data recognition method based on the spiking neural network model of this application.

[0167] like Figure 11 As shown, when at least two modes exist, Figure 11 Images are used in the text (i.e.) Figure 11 The MNIST dataset image input and the corresponding text description of the image (i.e., Figure 11 When inputting a one-hot encoded language (e.g., using handwritten images of 0-9 and numerical descriptions of 0-9), encoding can be performed first according to the corresponding preprocessing and encoding operations. For example, the image can be encoded using convolutional parsing combined with fully connected mapping (i.e.,...). Figure 11 (256->64) to obtain a 64-dimensional pulse vector. Encoding the text using the one-hot encoding method can also yield a 64-dimensional pulse vector.

[0168] Next, the pulse vectors of the image and the pulse vectors of the text can be concatenated (i.e., Figure 11 The vector encoding of visual feature inputs (V0-V9) and Figure 11 After inputting one-hot encoded language features (L0-L9), a 128-dimensional impulse vector can be obtained. Figure 11 The preset spiking neural network model shown (i.e. Figure 11 The single-layer coding network can be configured with two intermediate layers, one with 128 neurons and the other with 64 neurons, and the output layer has 10 neurons. After obtaining the 128-dimensional pulse vector, feature extraction can be performed using the aforementioned preset spiking neural network model to obtain the pulse sequence (i.e.,...). Figure 11 The output is encoded in PC / VC, and then a clustering operation is performed on the output pulse sequence (i.e., ...). Figure 11 Multimodal clustering can be used, specifically by overlapping pulse comparison. If there are many neurons with overlapping output pulses among 10 neurons (a threshold for the number of overlapping neurons can be set, and a value higher than the threshold indicates a large number), it can be indicated that the image and the text belong to the same class, such as text and image containing the number 1. For unrelated inputs, there may be no overlapping neurons in the output, such as text containing the number 2 and image containing the number 1.

[0169] Therefore, through the above methods, this embodiment can achieve the understanding and clustering of multimodal data. At the same time, since it can support multimodal classification, it can characterize the correlation and differences between features of different categories.

[0170] Reference Figure 12 , Figure 12 This is a flowchart illustrating the fifth embodiment of the data recognition method based on a spiking neural network model of this application. Based on the above embodiments, the fifth embodiment of the data recognition method based on a spiking neural network model of this application is proposed.

[0171] Furthermore, when performing labeled training, the specific process is as follows: Figure 12 As shown, the steps for determining the target analysis operation based on the training method include:

[0172] Step S032: When the training method is labeled training, the classification operation is used as the target analysis operation;

[0173] Accordingly, the steps described above for analyzing the pulse sequence to be identified according to the target analysis operation, and for identifying the original feature data to be identified based on the analysis results, include:

[0174] Step S312: Perform high-dimensional vector encoding on the pulse sequence to be identified to obtain the high-dimensional vector to be identified;

[0175] Step S313: Classify the high-dimensional vector to be identified using each preset high-dimensional vector template, and identify the original feature data to be identified based on the obtained classification results.

[0176] It should be noted that when the training method is labeled training, in order to improve the classification accuracy, the above-mentioned device can perform high-dimensional vector encoding on the obtained pulse sequence to be identified, increase the dimensionality, and use the encoding result as the high-dimensional vector to be identified.

[0177] Understandably, the above-mentioned high-dimensional vector encoding can be an encoding that transforms low-dimensional vectors into high-dimensional vectors (including thresholding and dimension expansion). Specifically, it can include, but is not limited to, direct output high-dimensional vector encoding, one-hot encoding, vector dimension expansion, and other encoding methods. The obtained high-dimensional vector to be identified can be composed of 0 or 1, and the dimension is generally not less than 10,000.

[0178] It should be emphasized that when no local impulse suppression constraint is added to the output layer, a non-sparse pulse sequence to be identified can be generated, and then high-dimensional vector encoding of the non-sparse pulse sequence can be performed to obtain the high-dimensional vector to be identified; when a local impulse suppression constraint is added to the output layer, a sparse pulse sequence can be generated, and then high-dimensional vector encoding can be performed to obtain the high-dimensional vector to be identified.

[0179] In practical use, when the training method is labeled training, the above-mentioned device can treat the classification operation as the target analysis operation. Furthermore, after obtaining the pulse sequence to be identified, the pulse sequence can be encoded into a high-dimensional vector to obtain the high-dimensional vector to be identified, and then classified using various preset high-dimensional vector templates.

[0180] The aforementioned preset high-dimensional vector templates can be high-dimensional vector templates generated during training for each category. That is, in this embodiment, corresponding preset high-dimensional vector templates can be set for different categories, and then the high-dimensional vector to be identified is compared with each preset high-dimensional vector template. The comparison result is used as the classification result to complete the identification of the original feature data to be identified.

[0181] Specifically, during the comparison, the steps described above—classifying the high-dimensional vector to be identified using various preset high-dimensional vector templates and identifying the original feature data to be identified based on the obtained classification results—include:

[0182] Step S3131: Determine the vector similarity between each of the preset high-dimensional vector templates and the high-dimensional vector to be identified, and sort the vector similarities.

[0183] Step S3132: Identify the original feature data to be identified based on the sorting results.

[0184] In this embodiment, a similarity comparison method can be used when making comparisons. This method determines the similarity between each preset high-dimensional vector template and the high-dimensional vector to be identified, and uses this similarity as the vector similarity. The similarities of each vector are then sorted, and the category corresponding to the preset high-dimensional vector template with the highest similarity is selected as the identification result of the original feature data to be identified.

[0185] It should be emphasized that when determining similarity, specific methods may include, but are not limited to, cosine similarity, dot product similarity, Manhattan distance similarity, Jaccord distance, etc., which can characterize the correlation and differences between different original feature data. This embodiment does not limit the specific similarity method to be selected.

[0186] For ease of understanding, please refer to Figure 13 , Figure 13 This is a schematic diagram of similarity comparison in the fifth embodiment of the data recognition method based on the spiking neural network model of this application; as shown... Figure 13 As shown, 131 is the high-dimensional vector to be recognized obtained from the user's gesture using the above method (i.e., Figure 13 The high-dimensional vector to be classified), 132 is the first preset high-dimensional vector template (i.e. Figure 13 High-dimensional vector template 1), 133 is the second type of preset high-dimensional vector template (i.e. Figure 13 High-dimensional vector template 2), 134 is the third preset high-dimensional vector template (i.e. Figure 13 High-dimensional vector template 3), 135 is a gesture diagram of the user, 136 is a gesture diagram corresponding to the first preset high-dimensional vector template, 137 is a gesture diagram corresponding to the second preset high-dimensional vector template, and 138 is a gesture diagram corresponding to the third preset high-dimensional vector template.

[0187] After obtaining the high-dimensional vector to be identified, its similarity can be compared with the first to third preset high-dimensional vector templates respectively. The vector similarity with the first preset high-dimensional vector template is 71% (i.e., Figure 13 The similarity between the vector and feature 1 is 71.4%, and the similarity with the second preset high-dimensional vector template is 40% (i.e., Figure 13 The similarity between the vector and feature 2 is 40%, and the similarity between the vector and the third preset high-dimensional vector template is 40% (i.e., Figure 13 If the similarity with feature 3 is 40%, then sorting is performed to determine that the first preset high-dimensional vector template has the highest similarity. Therefore, the category corresponding to the first preset high-dimensional vector template can be used as the recognition result (i.e.,...). Figure 13 (Classification results).

[0188] When the input is multimodal, refer to Figure 14 , Figure 14This is a schematic diagram of feature classification in the fifth embodiment of the data recognition method based on the spiking neural network model of this application, as shown below. Figure 14 As shown, 141 represents the raw feature data to be identified, with the input being electromyography (EMG) signals (i.e., ...). Figure 14 The original feature information of electromyography signal), 142 is the original feature data to be identified when the input is an image signal (i.e., Figure 14 The original feature information of the image signal), 143 is the preset spiking neural network model (i.e. Figure 14 In the SNN model, 1431 is the input layer of the preset spiking neural network model, 1432 is the intermediate layer of the preset spiking neural network model, 1433 is the output layer of the preset spiking neural network model, 144 is the output, 1441 is the high-dimensional vector to be recognized, 1442 is the preset high-dimensional vector template, and 1443 is the user's gesture (i.e., the gesture made by the user). Figure 14 (High-dimensional vector to be classified), 1444 is the gesture corresponding to each preset high-dimensional vector template;

[0189] In use, when electromyography (EMG) signals and image signals are input, preprocessing and encoding operations can be performed first, and then the signals are concatenated and input into a preset spiking neural network model. The output obtains the corresponding high-dimensional vector to be identified, and the similarity is compared with each preset high-dimensional vector template. Figure 14 The obtained vector similarities were 75%, 0%, and 50% respectively, and the category corresponding to the preset high-dimensional vector template with a vector similarity of 75% was taken as the final recognition result (i.e., Figure 14 (Classification based on high-dimensional vector similarity of medium templates).

[0190] Furthermore, in order to obtain the aforementioned preset high-dimensional vector templates, in this embodiment, before the step of obtaining the original feature data to be identified, the following steps are also included:

[0191] Obtain the original feature data of the sample;

[0192] Based on the preset time dependency and the preset spatial dependency, feature extraction is performed on the original feature data of the sample to obtain the sample pulse sequence. The preset time dependency and the preset spatial dependency are determined based on the data mode of the original feature data of the sample.

[0193] The sample pulse sequence is encoded using high-dimensional vector to obtain a sample high-dimensional vector, and the sample high-dimensional vector is superimposed to obtain a preset high-dimensional vector template.

[0194] It should be noted that during the training process, after obtaining the original feature data of the samples, the initial spiking neural network model described above can be constructed as previously explained. Feature extraction is then performed on the original feature data of the samples based on preset time and spatial dependencies, and the resulting data serves as the sample spiking sequence. Next, the sample spiking sequence is encoded into a high-dimensional vector to obtain the sample high-dimensional vector. This process is consistent with the above process and will not be elaborated upon in this embodiment.

[0195] After obtaining the high-dimensional vector template of the sample, the high-dimensional vector templates of samples belonging to the same class can be superimposed to obtain the preset high-dimensional vector template corresponding to that class.

[0196] Furthermore, the specific process of superposition is as follows: the step of superimposing the sample high-dimensional vectors to obtain a preset high-dimensional vector template includes:

[0197] The sample high-dimensional vectors are superimposed to obtain an initial high-dimensional vector template;

[0198] Template verification is performed on the initial high-dimensional vector template using the original feature data of the sample.

[0199] If the test result does not meet the preset test requirements, return to the step of superimposing the high-dimensional vector of the sample until the test result meets the preset test requirements, and then use the initial high-dimensional vector template as the preset high-dimensional vector template.

[0200] For easier understanding, please refer to Figure 15 To explain, Figure 15 This is a schematic diagram of the superposition operation in the fifth embodiment of the data recognition method based on a spiking neural network model of this application. Figure 15 As shown, 151 represents the original feature data of the sample, 152 represents the corresponding high-dimensional vector of the sample generated, 153 represents the superposition operation, 154 represents the generated preset high-dimensional vector template, and 155 represents the template verification.

[0201] It is understandable that different overlay strategies can be adopted when overlaying, and the specific strategy can be selected according to the application scenario. For example, this embodiment uses two overlay strategies for illustration: one is positional overlay, and the other is overlay based on the majority voting principle.

[0202] For bitwise superposition, i.e., adding element by element, it can be done by adding the elements at the same position in all sample high-dimensional vectors, finally dividing the whole by N (where N is the number of sample high-dimensional vectors), and then standardizing to restore the vector to a vector of 0 or 1. The standardization strategy can specifically be: if the value at a certain position is greater than a preset superposition threshold (denoted as λ), then output 1; otherwise, output 0. λ can be 0.5, but other values ​​are also possible; this embodiment does not impose any restrictions on this.

[0203] The following is an example of bitwise superposition: For example, sample high-dimensional vector A = [1, 0, 1, 1, 0, 1, 1], sample high-dimensional vector B = [1, 0, 0, 1, 0, 1, 0], sample high-dimensional vector C = [1, 0, 0, 0, 1, 1, 1]; first, add the elements at each phase position, i.e., A + B + C = [3, 0, 1, 2, 1, 3, 2]; then divide by the number of sample high-dimensional vectors, i.e. (A + B + C) / 3 = [1, 0, 0.33, 0.67, 0.33, 1, 0.67]; finally, standardize, i.e., Stand((A + B + C) / 3) = [1, 0, 0, 1, 0, 1, 1]; the result obtained by standardization can be the initial high-dimensional vector template corresponding to this category.

[0204] For the majority voting principle, the number of 1s and 0s in each position of all sample high-dimensional vectors can be counted, and the initial high-dimensional vector template selects the element with the most 1s and 0s as the value for that position. If the number of 1s and 0s is the same, a preset voting rule can be followed, such as prioritizing 1s as the value for that position, prioritizing 0s as the value for that position, or randomly setting the value for that position, etc. This embodiment does not impose any restrictions on this.

[0205] The following is an example of the majority voting principle superposition: For example, sample high-dimensional vector A = [1, 0, 1, 1, 0, 1, 1], sample high-dimensional vector B = [1, 0, 0, 1, 0, 1, 0], sample high-dimensional vector C = [1, 0, 0, 0, 1, 1, 1]; when the number of 1s and 0s is the same, 1 is used as the value of that position, and the result obtained is [1, 0, 0, 1, 0, 1, 1]. This result is used as the initial high-dimensional vector template corresponding to this category.

[0206] Considering that as the number of training samples increases, the similarity between some sample high-dimensional vectors and their corresponding preset high-dimensional vector templates may gradually decrease, this embodiment can also verify the obtained initial high-dimensional vector templates to improve template accuracy. Figure 15As shown, after obtaining the initial high-dimensional vector template, the similarity between each sample high-dimensional vector and the initial high-dimensional vector template can be calculated. A similarity test threshold is set as the aforementioned preset test requirement. If all calculated similarities are higher than the similarity test threshold, it indicates that the generated initial high-dimensional vector template meets the requirements, and the initial high-dimensional vector template can be used as the preset high-dimensional vector template. If at least one of the calculated similarities is not higher than the similarity test threshold, it indicates that the generated initial high-dimensional vector template does not meet the requirements. In this case, the sample high-dimensional vector that does not meet the requirements is superimposed with the corresponding initial high-dimensional vector template, and the similarity is calculated again. This process is repeated until all calculated similarities are higher than the similarity test threshold, at which point the obtained initial high-dimensional vector template is used as the preset high-dimensional vector template.

[0207] This embodiment achieves high-precision classification by using high-dimensional vector encoding and a preset high-dimensional vector template. Furthermore, because this embodiment employs high-dimensional vector encoding for dimensionality increase, it enables small-sample classification and incremental learning.

[0208] In addition, to facilitate understanding the comparison before and after establishing the aforementioned preset temporal and spatial dependencies, refer to Figure 16 , Figure 16 This is a before-and-after comparison diagram of the fifth embodiment of the data recognition method based on the spiking neural network model in this application. Figure 16 As shown, where Figure 16 In the diagram, 'a' represents the spurious firing behavior of neurons in the traditional SNN model (i.e.,...). Figure 16 b represents the initial high-dimensional feature vector, and b represents the pulse emission of the improved SNN model in this embodiment (i.e., ...). Figure 16 In the learning of high-dimensional feature vectors), c is the connection weight between neurons in the traditional SNN model (i.e., Figure 16 The initial connection weight matrix is ​​d, where d is the connection weight of the improved SNN model in this embodiment (i.e., ...). Figure 16 (Learning the post-connection weight matrix).

[0209] In a and b, the black portion is 0, representing no pulse, and the white portion is 1, representing a pulse. For a and b, the horizontal axis represents the neuron number, and the vertical axis represents the time step number. The leftmost part of a and b represents the pulse firing of the 64 neurons in the input layer over time steps, the middle part represents the pulse firing of the 128 neurons in the middle layer of the network (this example has only one middle layer), and the rightmost part represents the pulse firing of the 64 neurons in the output layer. In a and b, at input, assuming the signal is frequency-encoded with a step size of 32, the encoded pulse sequence is input with a step size of 32 (i.e., ...). Figure 16In the case of t=0, t=32, t=64, 32 sets of pulse signals are input sequentially each time, and the output vector length is 64. Based on a and b, it is easy to see that the number of neurons emitting pulses in the intermediate layer is reduced through improvement.

[0210] In diagrams c and d, the black portion (0) represents no connection weights or no connection at all, while the white portion (1) represents the presence of connection weights. For c and d, the horizontal axis represents the presynaptic neuron index, and the vertical axis represents the postsynaptic neuron index. The left middle portion of c and d represents the connection weights from the 64 neurons in the input layer to the 128 neurons in the middle layer; the right middle portion represents the connection weights from the 128 neurons in the middle layer to the 64 neurons in the output layer. Based on c and d, it is clear that the improved connection weight matrix reduces the number of neurons, thus improving recognition accuracy.

[0211] Furthermore, this application also proposes a storage medium storing a data recognition program based on a spiking neural network model. When the data recognition program based on the spiking neural network model is executed by a processor, it implements the steps of the data recognition method based on the spiking neural network model described above.

[0212] In addition, refer to Figure 17 , Figure 17 This is a structural block diagram of the first embodiment of the data recognition device based on the spiking neural network model of this application; as shown... Figure 17 As shown in the embodiments of this application, a data recognition device based on a spiking neural network model is also proposed. This data recognition device includes:

[0213] Data acquisition module 171 is used to acquire the original feature data to be identified;

[0214] Feature extraction module 172 is used to extract features from the original feature data to be identified based on preset time dependency and preset spatial dependency to obtain the pulse sequence to be identified. The preset time dependency and the preset spatial dependency are determined based on the data mode of the original feature data to be identified.

[0215] The data recognition module 173 is used to perform data recognition on the original feature data to be recognized based on the pulse sequence to be recognized.

[0216] In this embodiment, a preset time dependency and a preset spatial dependency can be set for the original feature data to be identified in different modalities. Then, feature extraction is performed based on the preset time dependency and the preset spatial dependency to complete the data identification. Thus, this embodiment can support the identification of feature data in multiple modalities.

[0217] Based on the first embodiment of the data recognition device based on the spiking neural network model described above in this application, a second embodiment of the data recognition device based on the spiking neural network model of this application is proposed.

[0218] In this embodiment, the data acquisition module 171 is further configured to determine the training method of the preset spiking neural network model and determine the target analysis operation based on the training method, wherein the target analysis operation is a clustering operation or a classification operation.

[0219] The data recognition module 173 is also used to analyze the pulse sequence to be identified according to the target analysis operation, and to perform data recognition on the original feature data to be identified based on the analysis results.

[0220] In one implementation, the data acquisition module 171 is further configured to use the clustering operation as the target analysis operation when the training method is unlabeled training.

[0221] The data recognition module 173 is also used to cluster the pulse sequence to be identified and to perform data recognition on the original feature data to be identified based on the obtained clustering results.

[0222] Based on the above embodiments of the data recognition device based on the spiking neural network model of this application, a third embodiment of the data recognition device based on the spiking neural network model of this application is proposed.

[0223] In this embodiment, the data acquisition module 171 is further configured to use the classification operation as the target analysis operation when the training method is labeled training;

[0224] The data recognition module 173 is further configured to perform high-dimensional vector encoding on the pulse sequence to be recognized to obtain a high-dimensional vector to be recognized; classify the high-dimensional vector to be recognized through each preset high-dimensional vector template; and recognize the original feature data to be recognized based on the obtained classification results.

[0225] As one implementation, the data recognition module 173 is further configured to determine the vector similarity between each of the preset high-dimensional vector templates and the high-dimensional vector to be recognized, and to sort the vector similarities; and to recognize the original feature data to be recognized based on the sorting results.

[0226] In one implementation, the data acquisition module 171 is further configured to acquire original sample feature data; extract features from the original sample feature data based on the preset time dependency and the preset spatial dependency to obtain a sample pulse sequence, wherein the preset time dependency and the preset spatial dependency are determined based on the data mode of the original sample feature data; encode the sample pulse sequence into a high-dimensional vector to obtain a high-dimensional sample vector, and superimpose the high-dimensional sample vector to obtain a preset high-dimensional vector template.

[0227] In one implementation, the data acquisition module 171 is further configured to superimpose the high-dimensional vectors of the samples to obtain an initial high-dimensional vector template; perform template verification on the initial high-dimensional vector template using the original feature data of the samples; and if the verification result does not meet the preset verification requirements, return to the step of superimposing the high-dimensional vectors of the samples until the verification result meets the preset verification requirements, and then use the initial high-dimensional vector template as the preset high-dimensional vector template.

[0228] Other embodiments or specific implementations of the data recognition device based on the spiking neural network model described in this application can be found in the above-described method embodiments, and will not be repeated here.

[0229] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0230] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0231] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory image (ROM) / random access memory (RAM), magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0232] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A data recognition method based on a spiking neural network model, characterized in that, The method is applied to a preset spiking neural network model, wherein at least two adjacent layers of neurons in the preset spiking neural network model satisfy a preset temporal dependency relationship, and at least one layer of neurons satisfies a preset spatial dependency relationship. The method includes: Obtain the original feature data to be identified; Based on the preset time dependency and the preset spatial dependency, feature extraction is performed on the original feature data to be identified to obtain the pulse sequence to be identified. The preset time dependency and the preset spatial dependency are determined based on the data mode of the original feature data to be identified. Data identification is performed on the original feature data to be identified based on the pulse sequence to be identified.

2. The method as described in claim 1, characterized in that, Before the step of obtaining the original feature data to be identified, the method further includes: The training method of the preset spiking neural network model is determined, and the target analysis operation is determined according to the training method, wherein the target analysis operation is a clustering operation or a classification operation; The step of performing data recognition on the original feature data to be identified based on the pulse sequence to be identified includes: The pulse sequence to be identified is analyzed according to the target analysis operation, and the original feature data to be identified is identified based on the analysis results.

3. The method as described in claim 2, characterized in that, The step of determining the target analysis operation based on the training method includes: When the training method is unlabeled training, the clustering operation is used as the target analysis operation; The steps of analyzing the pulse sequence to be identified according to the target analysis operation and identifying the original feature data to be identified based on the analysis results include: The pulse sequence to be identified is clustered, and the original feature data to be identified is identified based on the obtained clustering results.

4. The method as described in claim 2, characterized in that, The step of determining the target analysis operation based on the training method includes: When the training method is labeled training, the classification operation is used as the target analysis operation; The steps of analyzing the pulse sequence to be identified according to the target analysis operation and identifying the original feature data to be identified based on the analysis results include: The pulse sequence to be identified is encoded into a high-dimensional vector to obtain a high-dimensional vector to be identified; The high-dimensional vector to be identified is classified using various preset high-dimensional vector templates, and the original feature data to be identified is identified based on the obtained classification results.

5. The method as described in claim 4, characterized in that, The step of classifying the high-dimensional vector to be identified using various preset high-dimensional vector templates, and identifying the original feature data to be identified based on the obtained classification results, includes: Determine the vector similarity between each of the preset high-dimensional vector templates and the high-dimensional vector to be identified, and sort the vector similarities. The original feature data to be identified is identified based on the sorting results.

6. The method as described in claim 5, characterized in that, Before the step of obtaining the original feature data to be identified, the method further includes: Obtain the original feature data of the sample; Based on the preset time dependency and the preset spatial dependency, feature extraction is performed on the original feature data of the sample to obtain the sample pulse sequence. The preset time dependency and the preset spatial dependency are determined based on the data mode of the original feature data of the sample. The sample pulse sequence is encoded using high-dimensional vector to obtain a sample high-dimensional vector, and the sample high-dimensional vector is superimposed to obtain a preset high-dimensional vector template.

7. The method as described in claim 6, characterized in that, The step of superimposing the high-dimensional vectors of the samples to obtain a preset high-dimensional vector template includes: The sample high-dimensional vectors are superimposed to obtain an initial high-dimensional vector template; Template verification is performed on the initial high-dimensional vector template using the original feature data of the sample. If the test result does not meet the preset test requirements, return to the step of superimposing the high-dimensional vector of the sample until the test result meets the preset test requirements, and then use the initial high-dimensional vector template as the preset high-dimensional vector template.

8. A data identification device, characterized in that, The device includes: The data acquisition module is used to acquire the original feature data to be identified; The feature extraction module is used to extract features from the original feature data to be identified based on preset time dependencies and preset spatial dependencies to obtain the pulse sequence to be identified. The preset time dependencies and preset spatial dependencies are determined based on the data modes of the original feature data to be identified. The data recognition module is used to perform data recognition on the original feature data to be recognized based on the pulse sequence to be recognized.

9. A storage medium, characterized in that, The storage medium stores a data recognition program, which, when executed by a processor, implements the steps of the data recognition method based on a spiking neural network model as described in any one of claims 1 to 7.

10. A wearable device, characterized in that, The wearable device includes: a memory, a processor, and a data recognition program stored in the memory and executable on the processor, wherein the data recognition program, when executed by the processor, implements the steps of the data recognition method based on a spiking neural network model as described in any one of claims 1 to 7.

Images