Method, device and equipment for training a brain activity state classification model

Abstract

To provide a training method, device and apparatus for improving efficiency and accuracy for classification of a brain activity state.SOLUTION: A method comprises the steps of: acquiring spike trains of electroencephalogram signal samples corresponding to a plurality of brain activity state classification training tasks; and inputting the spike trains of the electroencephalogram signal samples corresponding to each training task into an initial brain activity state classification model, training the brain activity state classification model based on a target rule, in a forward propagation stage in the target rule, updating Hebb information which corresponds to each synapse in the brain activity state classification model according to the spike trains corresponding to each training task, and is used to represent a degree of association between the training task and the synapse, and in a back propagation stage in the target rule, determining weight values of the synapses in the brain activity state classification model according to the Hebb information corresponding to each synapse and a back propagation result.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to the technical field of medical signal processing, and in particular to a method, device and apparatus for training a brain activity state classification model. [Background technology]

[0002] In the past few decades, research on artificial intelligence has progressed rapidly, and in particular, the connectionist artificial neural network model has achieved great success in tasks such as image recognition, target detection, voice recognition, natural language processing, etc. Preferably, the artificial neural network model can be applied to clinical medical application scenarios, and can classify the signals of different brain regions by monitoring EEG signals, and help doctors determine the signal source and confirm the cerebral state and physical condition of patients so as to provide more accurate treatment.

[0003] In related art, brain activity states are classified using artificial neural networks. However, when the distribution of data is constantly changing, the artificial neural network model suffers from the problem of catastrophic forgetting, just like the traditional method, i.e., the learning of new knowledge interferes with the memory of old knowledge, reducing the accuracy and efficiency of the classification results of brain activity states. Summary of the Invention [Problem to be solved by the invention]

[0004] In view of the problems in the prior art, embodiments of the present invention provide a method, apparatus and device for training a brain activity state classification model. [Means for solving the problem]

[0005] Specifically, the embodiments of the present invention provide the following technical solutions:

[0006] In a first aspect, an embodiment of the invention comprises: obtaining spike trains of electroencephalographic signal samples corresponding to a plurality of brain activity state classification training tasks; inputting spike trains of electroencephalogram signal samples corresponding to each of the training tasks into an initial brain activity state classification model constructed based on a spiking neural network, training the brain activity state classification model based on a target rule, in which, in a forward propagation step in the target rule, Hebbian information corresponding to each synapse in the brain activity state classification model according to the spike trains corresponding to each of the training tasks, which is determined based on a synaptic co-discharge frequency and is used to represent a degree of association between the training task and the synapse, is updated; and in a back propagation step in the target rule, determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and a back propagation result. A method for training a brain activity state classification model is provided.

[0007] Furthermore, updating Hebbian information corresponding to each synapse in the brain activity state classification model in accordance with the spike train corresponding to each training task as described above, Updating the Hebbian information corresponding to each synapse in the brain activity state classification model using the following formula: JPEG0007688860000001.jpg1447 Of these, JPEG0007688860000002.jpg55 represents the Hebbian information of the i-th synapse before the j-th task in the spike train, JPEG0007688860000003.jpg56 represents the Hebbian information of the ith synapse after the jth task in the spike train, and ω represents a preset update rate. JPEG0007688860000004.jpg55 represents the co-discharge frequency of the i-th synapse in the brain activity state classification model corresponding to the j-th task in the spike train, JPEG0007688860000005.jpg55 represents a target list that stores the Hebbian information of synapses corresponding to each training task. JPEG0007688860000006.jpg57 represents the Hebbian information of the i-th synapse corresponding to the j-th task stored in the target list.

[0008] Furthermore, updating the Hebbian information corresponding to each synapse in the brain activity state classification model described above is Updating synaptic Hebbian information based on synaptic co-discharge states in a single time window; and / or updating the Hebbian information of the synapse based on the average discharge rate within a plurality of time windows.

[0009] Furthermore, in the back propagation step of the target rule, determining the weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and the back propagation result; In the backpropagation step, for any one synapse in the brain activity state classification model, if the Hebbian information of the synapse is greater than a first threshold, the synapse is determined to be relevant to the task and a weight value of the synapse in the brain activity state classification model is locked, and if the Hebbian information of the synapse is not greater than the first threshold, the weight value of the synapse is modified according to the backpropagation result.

[0010] In a second aspect, an embodiment of the invention comprises: obtaining a spike train corresponding to a target electroencephalogram signal; inputting a spike train corresponding to the target electroencephalogram signal into a brain activity state classification model obtained by training based on the method for training a brain activity state classification model according to the first aspect, and obtaining a classification result of the brain activity state. A method for brain activity state classification is further provided.

[0011] In a third aspect, an embodiment of the invention comprises: an acquisition module for acquiring spike trains of electroencephalogram signal samples corresponding to a training task of a plurality of brain activity state classifications; a training module for inputting spike trains of electroencephalogram signal samples corresponding to each of the training tasks into an initial brain activity state classification model constructed based on a spiking neural network, training the brain activity state classification model based on a target rule, in which, in a forward propagation step in the target rule, Hebbian information corresponding to each synapse in the brain activity state classification model is updated according to the spike trains corresponding to each of the training tasks, the Hebbian information being determined based on a synaptic co-discharge frequency and being used to represent the degree of association between the training task and the synapse, and in a back propagation step in the target rule, determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and a back propagation result. An apparatus for training a brain activity state classification model is further provided.

[0012] In a fourth aspect, an embodiment of the invention comprises: The present invention further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and operable by the processor, the program being executed by the processor to realize the method for training a brain activity state classification model described in the first aspect or the method for classifying a brain activity state described in the second aspect.

[0013] In a fifth aspect, an embodiment of the invention comprises: The present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being configured to, when executed by a processor, realize the method for training a brain activity state classification model according to the first aspect or the method for classifying brain activity states according to the second aspect.

[0014] In a sixth aspect, an embodiment of the invention comprises: There is further provided a computer program product comprising a computer program which, when executed by a processor, realises the method for training a brain activity state classification model according to the first aspect or the method for brain activity state classification according to the second aspect. Effect of the Invention

[0015] In the method, device and apparatus for training a brain activity state classification model according to an embodiment of the present invention, in the process of continuously learning spike trains of EEG signal samples corresponding to a plurality of training tasks, in the forward propagation step of the target rule, the degree of association between the training task and synapses is recorded using Hebbian information, and in the back propagation step of the target rule, the weights of the synapses are determined using the recorded Hebbian information. Thus, in the process of continuously learning a plurality of training tasks, the Hebbian information is recorded to protect the information of tasks that have already been trained, thereby enabling the tasks that have already been trained to be accurately identified, solving the problem of catastrophic forgetting, and enabling the brain activity state classification model after training to accurately classify brain activity states, thereby improving the efficiency and accuracy of brain activity state classification. [Brief description of the drawings]

[0016] In order to more clearly explain the technical solutions in the present invention or the prior art, the following briefly introduces drawings that need to be used in the embodiments or the description of the prior art. The drawings in the following description are some embodiments of the present invention, and it is obvious to those skilled in the art that, without performing creative labor, other drawings can be further obtained according to these drawings.

[0017] [Figure 1] FIG. 1 is a schematic diagram showing a flow of a method for training a brain activity state classification model according to an embodiment of the present invention. [Diagram 2] 2 is a schematic diagram showing the flow of a method for training a brain activity state classification model according to an embodiment of the present invention; FIG. [Diagram 3] 3 is a schematic diagram showing the flow of a method for training a brain activity state classification model according to an embodiment of the present invention; [Figure 4] FIG. 4 is a schematic diagram showing the flow of a method for training a brain activity state classification model according to an embodiment of the present invention. [Diagram 5] 1 is a structural schematic diagram of a training device for a brain activity state classification model according to an embodiment of the present invention; [Figure 6] 1 is a structural schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be described below clearly and completely with reference to the drawings in the present invention, and it is clear that the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained based on the embodiments in the present invention without the creative labor of those skilled in the art belong to the protection scope of the present invention.

[0019] The method of the embodiment of the present invention can be applied to the scene of medical signal processing to realize accurate classification of brain activity states.

[0020] In related art, brain activity states are classified using artificial neural networks. However, when the distribution of data is constantly changing, the artificial neural network model suffers from the problem of catastrophic forgetting, just like the traditional method, i.e., the learning of new knowledge interferes with the memory of old knowledge, reducing the accuracy and efficiency of the classification results of brain activity states.

[0021] In the training method for a brain activity state classification model according to an embodiment of the present invention, in the process of continuously learning spike trains of EEG signal samples corresponding to a plurality of training tasks, in the forward propagation step of the target rule, the degree of association between the training task and synapses is recorded using Hebbian information, and in the backward propagation step of the target rule, the weights of the synapses are determined using the recorded Hebbian information. Thus, in the process of continuously learning a plurality of training tasks, the Hebbian information is recorded to protect the information of the tasks that have already been trained, thereby enabling the tasks that have already been trained to be accurately identified, solving the problem of catastrophic forgetting, enabling the brain activity state classification model after training to accurately classify brain activity states, and improving the efficiency and accuracy of brain activity state classification.

[0022] The technical solution of the present invention will be described in detail below through specific embodiments with reference to Figures 1 to 6. Some of the following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeatedly described in some embodiments.

[0023] 1 is a flow diagram of an embodiment of a method for training a brain activity state classification model according to an embodiment of the present invention. As shown in FIG. 1, the method according to this embodiment includes:

[0024] In step 101, spike trains of electroencephalogram signal samples corresponding to a plurality of brain activity state classification training tasks are obtained; Specifically, in the related art, brain activity states are classified using artificial neural networks. However, when the distribution of data is constantly changing, the artificial neural network model suffers from the problem of catastrophic forgetting, just like the traditional methods. That is, the learning of new knowledge interferes with the memory of old knowledge, which reduces the accuracy and efficiency of the classification results of brain activity states.

[0025] To solve the above problem, in an embodiment of the present invention, first, the spike trains of EEG signal samples corresponding to multiple brain activity state classification training tasks are obtained, and preferably, in each brain activity state classification training task, for input signals such as heart rate signals, brain signals, and audio, a spike encoder (e.g., Poisson encoder) is used to encode non-spike input signals into new spike trains, and a brain activity state classification model is trained. For example, for a certain heart rate signal input, it is divided into N frames, and each frame is encoded into a spike train of one normal distribution or other distribution.

[0026] In step 102, the spike trains of the EEG signal samples corresponding to each training task are input into an initial brain activity state classification model constructed based on a spiking neural network, and the brain activity state classification model is trained based on a target rule, in which, in the forward propagation step of the target rule, the Hebbian information corresponding to each synapse in the brain activity state classification model is updated according to the spike trains corresponding to each training task, and the Hebbian information is determined based on the synaptic co-discharge frequency and is used to represent the degree of association between the training task and the synapse, and in the back propagation step of the target rule, the weight value of the synapse in the brain activity state classification model is determined according to the Hebbian information corresponding to each synapse and the back propagation result.

[0027] Specifically, after obtaining spike trains of EEG signal samples corresponding to multiple brain activity state classification training tasks, in an embodiment of the present invention, the spike trains of the EEG signal samples corresponding to each brain activity state classification training task are input into an initial brain activity state classification model for continuous learning, and the brain activity state classification model is trained based on a target rule, preferably the target rule includes a forward propagation stage and a back propagation stage, and back propagation is performed according to the error between the actual output value and the desired output value in the forward propagation stage, and cyclic iteration is performed to perform learning training for the parameters of the brain activity state classification model, and after the training is completed, the brain activity state classification model can be used to classify brain activity states. Preferably, in the embodiment of the present invention, in the forward propagation step of the target rule, the Hebbian information corresponding to each synapse in the brain activity state classification model is updated according to the spike train corresponding to each training task, and in the back propagation step of the target rule, the weight value of the synapse in the brain activity state classification model is determined according to the Hebbian information corresponding to each synapse and the back propagation result. That is, in the process of continuously learning the spike trains of the EEG signal samples corresponding to a plurality of training tasks, in the forward propagation step of the target rule, the degree of association between the training task and the synapse is recorded by the Hebbian information, and in the back propagation step of the target rule, the weight value of the synapse is determined according to the recorded Hebbian information and the back propagation result, so that in the process of continuously learning a plurality of training tasks, the Hebbian information is updated according to the Hebbian information and the back propagation result. By recording the Hebbian information of synapses to protect the information of tasks that have already been trained, it is possible to correctly identify tasks that have already been trained; that is, when training multiple tasks, the Hebbian information of synapses is recorded, highly active neurons corresponding to different tasks are found, and these neurons are assigned as subsystems for the tasks, and their weight values ​​are locked so that they cannot be changed in the subsequent learning of other tasks; thus, a new training task does not affect the previously trained tasks, so that the previous tasks are not forgotten and training can be performed efficiently; and a model and method are innovatively realized that can self-adaptively assign neurons to form subsystems when the information of multiple tasks is unknown, thereby solving the problem of catastrophic forgetting.

[0028] For example, the spike train of the EEG signal sample corresponding to the first training task is a spike train corresponding to when the user watches a picture, which corresponds to a first type of brain activity type; in the forward propagation step of the target rule, the Hebbian information of synapse A in the brain activity state classification model is recorded as a according to the first training task; in the back propagation step of the target rule, the amount of change in the weight value of synapse A is jointly determined according to the result of the back propagation and the Hebbian information a corresponding to synapse A; the spike train of the EEG signal sample corresponding to the second training task is a spike train corresponding to when the user listens to audio, which corresponds to a second type of brain activity type; in the forward propagation step of the target rule, the Hebbian information of synapse B in the brain activity state classification model is recorded as b according to the second training task; in the back propagation step of the target rule, the Hebbian information of synapse B in the brain activity state classification model is recorded as b according to the second training task; According to the Hebb information b, the weight value of synapse B in the brain activity state classification model is determined; that is, in the forward propagation step of the target rule, the degree of association between the training task and the synapse is recorded by the Hebb information; in the backward propagation step of the target rule, the change amount of the synapse weight value is jointly determined according to the result of the backward propagation and the Hebb information; after completing the second training task, the brain activity state classification model can still accurately classify the first brain activity type; that is, in the process of continuous learning of multiple training tasks, by recording the Hebb information to protect the information of the already trained tasks, the already trained tasks can also be accurately identified; that is, in the case of multiple tasks, the new training tasks will not affect the already trained tasks of the previous period, so the previous tasks will not be forgotten, and the problem of catastrophic forgetting will be solved.

[0029] In addition, in the related art, the neural network is modularized, and subsystems including a certain number of neurons are randomly assigned to different tasks. From a biological angle, this is more in line with the characteristics of the cerebrum for continuous learning of multiple tasks (for example, memory and motor control belong to the control of different brain regions). However, such a paradigm also has some problems, the first of which is the problem of the training efficiency of the subsystems. Since the neural network randomly assigns a certain number of neurons to each task to form a subsystem for training, if the number of tasks is too large or the task training amount is too large, for example, when inputting a high-throughput multi-mode data stream such as an EEG signal, the training data will be unbalanced with respect to the number of neurons, which means that the training efficiency of the subsystems is too low, and thus the training efficiency of the entire network is too low. The second problem of training efficiency is that such a modularized architecture requires the number of tasks and the training order to be known in advance so as to divide the subsystems for each task. This means that if the number of tasks and the learning order are not clear in advance, for example, in the case of a classification task in which the order and number cannot be determined such as an EEG signal, it is difficult to train multiple tasks using such a modularized architecture paradigm. Meanwhile, in an embodiment of the present invention, the spike trains of EEG signal samples corresponding to multiple training tasks of brain activity state classification are continuously learned, and in the forward propagation step of the target rule, the degree of association between the training task and the synapse is recorded according to the Hebbian information, and in the back propagation step of the target rule, the weight value of the synapse is determined according to the recorded Hebbian information and the back propagation result. Thus, in the process of continuous learning of multiple training tasks, the Hebbian information is recorded to protect the information of the tasks that have already been trained, so that the previous tasks are not forgotten and the tasks that have already been trained can be accurately identified, thereby solving the problem of catastrophic forgetting.The embodiments of the present invention have stronger continuous learning capabilities than the training process of modularizing neural networks, and when allocating subsystems to continuous learning tasks, they are self-adaptively calculated and allocated. Compared with both deep neural networks and the continuous learning paradigm of conventional modular architecture, the embodiments of the present invention have stronger continuous learning capabilities and can complete training for multiple tasks more efficiently. In the embodiments of the present invention, the entire network is used to learn different tasks, and the training efficiency of multiple tasks and large tasks is higher, which is an ability that the conventional modular architecture continuous learning paradigm does not have.

[0030] In addition, spiking neural networks have more complex neuron and synapse structures than deep neural networks, and many biological rules ignored in existing artificial networks are the key to realizing universal intelligence similar to the human brain. When these biological rules are further added to the spiking neural network similar to the brain, the existing network will have stronger computational and adaptive capabilities. The brain activity state classification model in the embodiment of the present application is built based on the spiking neural network, so the design of the brain activity state classification model and the continuous learning method are both more biologically rational. In the embodiment of the present invention, the Hebbian information of synapses is recorded during the training of multiple tasks, high-activity neurons corresponding to different tasks are found, and this part of neurons is assigned as the subsystem of the task, and its weight value is locked so that it cannot be changed in the subsequent learning of other tasks, and a model and method are innovatively realized that can be efficiently trained, and when the information of multiple tasks is unknown, neurons can be self-adaptively assigned to form a subsystem, solving two problems existing in the paradigm of modular architecture and greatly enhancing the continuous learning ability of spiking neural networks.

[0031] In the method of the above embodiment, in the process of continuously learning the spike trains of EEG signal samples corresponding to a plurality of training tasks, in the forward propagation step of the target rule, the degree of association between the training tasks and synapses is recorded by Hebbian information, and in the back propagation step of the target rule, the weight values ​​of the synapses are determined by the recorded Hebbian information and the back propagation result. Thus, in the process of continuously learning a plurality of training tasks, the Hebbian information is recorded to protect the information of the tasks that have already been trained, so that the tasks that have already been trained can also be accurately identified, the problem of catastrophic forgetting is solved, and the brain activity state classification model after training can accurately classify the brain activity state, thereby improving the efficiency and accuracy of the classification of the brain activity state.

[0032] In one embodiment, updating Hebbian information corresponding to each synapse in the brain activity state classification model in response to a spike train corresponding to each training task includes: Updating the Hebbian information corresponding to each synapse in the brain activity state classification model using the following formula: JPEG0007688860000007.jpg1037Of these, JPEG0007688860000008.jpg55 represents the Hebbian information of the i-th synapse before the j-th task in the spike train, JPEG0007688860000009.jpg56 represents the Hebbian information of the ith synapse after the jth task in the spike train, and ω represents a preset update rate. JPEG0007688860000010.jpg55 represents the co-discharge frequency of the i-th synapse in the brain activity state classification model corresponding to the j-th task in the spike train, JPEG0007688860000011.jpg55 represents the target list, which stores the Hebbian information of synapses corresponding to each training task. JPEG0007688860000012.jpg57 represents the Hebbian information of the i-th synapse corresponding to the j-th task stored in the target list.

[0033] Specifically, in an embodiment of the present invention, the spike trains of electroencephalogram signal samples corresponding to multiple training tasks of brain activity state classification are continuously learned, and in the forward propagation step of the target rule, the degree of association between the training task and the synapse is recorded according to the Hebbian information, and in the backward propagation step of the target rule, the weight value of the synapse is determined according to the recorded Hebbian information and the backward propagation result, so that in the process of continuous learning of multiple tasks, the Hebbian information is recorded to protect the information of the task that has already been trained, so that the previous task is not forgotten and the task that has already been trained can be correctly identified, thereby solving the problem of catastrophic forgetting.Preferably, the Hebbian information corresponding to each synapse in the brain activity state classification model is updated and recorded according to the following formula: JPEG0007688860000013.jpg1034Of these, JPEG0007688860000014.jpg55 represents the Hebbian information of the i-th synapse before the j-th task in the spike train, JPEG0007688860000015.jpg56 represents the Hebbian information of the ith synapse after the jth task in the spike train, and ω represents a preset update rate. JPEG0007688860000016.jpg55 represents the co-discharge frequency of the i-th synapse in the brain activity state classification model corresponding to the j-th task in the spike train, JPEG0007688860000017.jpg55 represents the target list, which stores the Hebbian information of synapses corresponding to each training task. JPEG0007688860000018.jpg57 represents the Hebbian information of the i-th synapse corresponding to the j-th task stored in the target list. That is, one variable is defined for describing the frequency of joint discharge phenomenon for synapses, and the variable is called Hebbian information. In the forward propagation stage of each task training, each synapse calculates and updates the Hebbian information corresponding to the task and records it. The specific processing method is as follows. All tasks are input into the network sequentially and learning is performed in a continuous learning paradigm, and in the learning process of each task, no data of the past task appears, and only the data of the task appears. In the forward propagation stage of each task, each synapse calculates and updates the Hebbian information of the corresponding task as shown in the following formula: JPEG0007688860000019.jpg1036, where ω represents the update rate, JPEG0007688860000020.jpg55 and JPEG0007688860000021.jpg55 represents the Hebbian information before and after the update of the i-th synapse in the forward propagation stage of the j-th task, respectively. JPEG0007688860000022.jpg54 represents the joint discharge frequency of each synapse in the forward propagation stage of the current task. ω is an artificially set parameter, The initialization value of JPEG0007688860000023.jpg55 is 0. JPEG0007688860000024.jpg55 is a list that stores Hebbian information corresponding to each history task of the i-th synapse, JPEG0007688860000025.jpg56 is the Hebbian information corresponding to the jth task stored in the list; that is, during training of multiple tasks, the Hebbian information of synapses is recorded for each task, highly active neurons corresponding to different tasks are found, and this part of neurons is assigned as the subsystem of the task, and its weight value is locked so that it cannot be changed in subsequent learning of other tasks. In the process of continuous learning of multiple training tasks, the Hebbian information is recorded to protect the information of the tasks that have already been trained, so that the tasks that have already been trained can also be correctly identified, thereby solving the problem of catastrophic forgetting, enabling the brain activity state classification model after training to accurately classify the brain activity state, and improving the efficiency and accuracy of brain activity state classification.

[0034] In the method of the above embodiment, all tasks are input into the brain activity state classification model in sequence, and learning is performed under a continuous learning paradigm. In the forward propagation stage of each task training, each synapse calculates, updates, and records the Hebbian information corresponding to the task. That is, when training multiple tasks, the Hebbian information of the synapses is recorded for each task, and highly active neurons corresponding to different tasks are found, and these neurons are assigned as subsystems for the tasks, and their weight values ​​are locked so that they cannot be changed in the subsequent learning of other tasks. In the process of continuous learning of multiple training tasks, the Hebbian information is recorded to protect the information of the tasks that have already been trained, thereby enabling the tasks that have already been trained to be correctly identified, solving the problem of catastrophic forgetting, and enabling the brain activity state classification model after training to accurately classify the brain activity state, thereby improving the efficiency and accuracy of brain activity state classification.

[0035] In one embodiment, updating the Hebbian information corresponding to each synapse in the brain activity state classification model includes: Updating synaptic Hebbian information based on synaptic co-discharge states in a single time window; and / or updating the Hebbian information of the synapse based on the average discharge rate within a plurality of time windows.

[0036] Specifically, the Hebbian information of synapses can be updated according to two methods. The first method updates the Hebbian information according to neuronal activity information within several time windows, that is, in the forward propagation stage, the Hebbian information is updated according to the average discharge rate within several time windows; JPEG0007688860000026.jpg54 is represented as follows: JPEG0007688860000027.jpg1043JPEG0007688860000028.jpg59 and JPEG0007688860000029.jpg511 represents the firing states of pre-synaptic and postsynaptic neurons, respectively, in the t-th time window, where the Hebbian information is updated once every T time windows.

[0037] The second method updates and calculates the Hebbian information according to the synaptic joint discharge state in a single time window, that is, in the forward propagation stage, the Hebbian information is updated according to the neuron activity information in a single time window; JPEG0007688860000030.jpg54 is represented as follows: JPEG0007688860000031.jpg625 In this case the Hebb information is updated once per time window.

[0038] Among them, the more active the neurons before and after the i-th synapse are, the more frequent the joint discharge phenomenon becomes. JPEG0007688860000032.jpg54 becomes larger, and the updated Hebbian information also becomes larger, indicating that the i-th synapse is more important for the j-th task.

[0039] The method of the above embodiment realizes timely and accurate updating of Hebb information by updating the Hebb information with neuronal activity information in several time windows, or updating the Hebb information according to the synaptic joint discharge state in a single time window, so that the more active the activity of the synapse corresponding to the training task, the larger the updated Hebb information will be, and the more important the synapse is for the training task; furthermore, it can also find highly active synapses corresponding to different tasks, assigning these synapses as the subsystem of the task, and locking their weight values ​​so that they cannot be changed in the subsequent learning of other tasks; in the process of continuous learning of multiple training tasks, the Hebb information is recorded to protect the information of the tasks that have already been trained, so that the tasks that have already been trained can also be accurately identified, solving the problem of catastrophic forgetting, and enabling the brain activity state classification model after training to accurately classify the brain activity state, thereby improving the efficiency and accuracy of the classification of the brain activity state.

[0040] In one embodiment, in the backpropagation step of the target rule, determining a weight value of a synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and the backpropagation result includes: In the backpropagation step, for any one synapse in the brain activity state classification model, if the Hebbian information of the synapse is greater than a first threshold, the synapse is determined to be relevant to the task and a weight value of the synapse in the brain activity state classification model is locked, and if the Hebbian information of the synapse is not greater than the first threshold, the weight value of the synapse is modified according to the backpropagation result.

[0041] Specifically, in an embodiment of the present invention, the spike trains of electroencephalogram signal samples corresponding to multiple training tasks of brain activity state classification are continuously learned, and in the forward propagation step of the target rule, the degree of association between the training task and the synapse is recorded according to the Hebbian information, and in the back propagation step of the target rule, the weight value of the synapse is determined according to the recorded Hebbian information and the back propagation result, so that in the process of continuous learning of multiple training tasks, the Hebbian information is recorded to protect the information of the already trained task, so that the previous task is not forgotten and the already trained task can be correctly identified, and the problem of catastrophic forgetting is solved. Preferably, in the back propagation step of the neural network, the locking operation of the Hebbian synapse is performed according to the Hebbian information, and in the back propagation step, a mask is generated for the synapse according to the Hebbian information accumulated in the recorded history task to perform masking, thereby protecting the knowledge related to the history task in the network and improving the continuous learning ability of the network. Specifically, in the back propagation step of each task, it is determined whether the synapse is related to a certain history task according to the Hebbian information corresponding to the history task recorded by each synapse in the forward propagation step. The basis for determining the association of the i-th synapse is obtained by calculation as shown below. JPEG0007688860000033.jpg1028JPEG0007688860000034.jpg55 represents the largest Hebbian information value corresponding to the jth task in the list in which the Hebbian information corresponding to each history task of the ith synapse is stored, JPEG0007688860000035.jpg54 is the associated flag and has the largest Hebbiano information value JPEG0007688860000036.jpg55 is the threshold If it is greater than JPEG0007688860000037.jpg57, the i-th synapse is considered to be related to the j-th task. During backpropagation, the change amount of the i-th synapse is masked with a mask to ensure that the weight value of the related synapse i is not changed by the current task, that is, the weight value of the synapse is locked, and the largest Hebbian information value is obtained. JPEG0007688860000038.jpg55 is the threshold If it is not larger than JPEG0007688860000039.jpg57, then perform back propagation according to the error between the actual output value and the expected output value in the forward propagation stage, and perform cyclic iteration to learn and train the parameters of the brain activity state classification model. Among them, the judgment of the association between synapses and tasks and the masking method of synapses are the main contents of the locking of Hebbian synapses, realizing continuous learning for multiple tasks, and in the forward propagation stage of each task training, each synapse calculates and updates the Hebbian information corresponding to the task, finds highly active neurons corresponding to different tasks, assigns this part of neurons as the subsystem of the task, and locks its weight value so that it cannot be changed in the subsequent learning of other tasks, and in the process of continuous learning of multiple training tasks, the Hebbian information is recorded to protect the information of the tasks that have already been trained, thereby enabling the tasks that have already been trained to be correctly identified, solving the problem of catastrophic forgetting, enabling the brain activity state classification model after training to accurately classify the brain activity state, and improving the efficiency and accuracy of the classification of the brain activity state.

[0042] In the method of the above embodiment, in the forward propagation stage of each task training, each synapse calculates, updates and records Hebbian information corresponding to the task; that is, during the training of multiple tasks, the Hebbian information of the synapses is recorded for each task to find highly active neurons corresponding to different tasks; in the backward propagation stage, this part of highly active neurons is assigned as the subsystem of the task, and its weight value is locked so that it cannot be changed in the subsequent learning of other tasks; in the process of continuous learning of multiple training tasks, the Hebbian information is recorded to protect the information of the tasks that have already been trained, so that the tasks that have already been trained can also be correctly identified, thereby solving the problem of catastrophic forgetting, and enabling the brain activity state classification model after training to accurately classify the brain activity state, thereby improving the efficiency and accuracy of brain activity state classification.

[0043] In one embodiment, the method for classifying brain activity states comprises: obtaining a spike train corresponding to a target electroencephalogram signal; The method includes inputting a spike train corresponding to the target electroencephalogram signal into a brain activity state classification model obtained by training based on a training method for the brain activity state classification model, and obtaining a classification result of the brain activity state.

[0044] Specifically, in the process of continuous learning of multiple training tasks, the embodiment of the present invention records Hebbian information to protect the information of the tasks that have already been trained, so that the tasks that have already been trained can also be correctly identified, solving the problem of catastrophic forgetting, and also enabling the brain activity state classification model after training to accurately classify the brain activity state, improving the efficiency and accuracy of the classification of the brain activity state.Preferably, after training the brain activity state classification model, the spike train corresponding to the EEG signal to be identified is input into the brain activity state classification model to obtain the classification result of the brain activity state, so as to realize the accurate identification and classification of the brain activity state, and further, based on the accurately identified and classified brain activity state, it can assist the doctor to determine the signal source and check the cerebral state and physical condition of the patient, so as to more accurately treat the patient.

[0045] The method of the above embodiment inputs spike trains corresponding to EEG signals awaiting classification into a trained brain activity state classification model, thereby obtaining accurate classification results for brain activity states and achieving accurate classification of brain activity states.

[0046] For example, as shown in the flow chart of the training method for a brain activity state classification model in FIG. 2, the continuous learning model and method of a spiking neural network based on Hebbian synapse locking can achieve stronger continuous learning ability and higher training efficiency, and also provide a more biologically reliable neural network learning model and method, specifically as follows:

[0047] (1) Encoding input data into spike trains: for input signals such as heart rate signals, brain signals, and audio, a spike encoder (e.g., a Poisson encoder) encodes non-spiking input signals into new spike trains of a certain distribution format for subsequent spiking neurons to use and process. For example, for a heart rate signal input, it is divided into N frames, and each frame is encoded into a spike train of one normal distribution or other distribution.

[0048] (2) Regarding the dynamical neuron with a predefined threshold processing spike information, the dynamical neuron encodes the input information and determines the dynamical properties according to the predefined neuronal firing threshold. The process of the basic LIF neuron processing the current time information is as follows. JPEG0007688860000040.jpg1965Of these, JPEG0007688860000041.jpg59 is the membrane potential with history consistent state, S is the neuron firing state, and S=1 is the membrane potential of neuron i JPEG0007688860000042.jpg59 is the ignition threshold JPEG0007688860000043.jpg56 represents the spike that reaches the refractory period of the neuron. At the same time, S represents the method of resetting the membrane potential without directly blocking the membrane potential. Simulate JPEG0007688860000044.jpg57.

[0049] Based on the above LIF neuron basis, the neuronal firing threshold is an artificially set static value, specifically determined by the dynamical properties of the required neuron.

[0050] (3) Regarding the construction of a spiking neural network with self-adaptive Hebbian information calculation using dynamic neurons, a variable is defined for synapses to describe the frequency of co-discharge phenomenon, which is called Hebbian information, and in the forward propagation stage of each task training, each synapse calculates, updates and records the Hebbian information corresponding to the task. The specific processing method is as follows:

[0051] As shown in the flow chart of the training method for the brain activity state classification model in Figure 3, all tasks are input into the network in sequence and learning is performed under a continuous learning paradigm, and in the learning process of each task, no data of past tasks appears, and only data of the task itself appears. In the forward propagation stage of each task, each synapse calculates and updates the Hebbian information of the corresponding task, as shown in the following formula: JPEG0007688860000045.jpg1036, where ω represents the update rate, JPEG0007688860000046.jpg55 and JPEG0007688860000047.jpg55 respectively represent the Hebbian information before and after the update of the i-th synapse in the forward propagation stage of the j-th task, JPEG0007688860000048.jpg54 represents the joint discharge frequency of each synapse in the forward propagation stage of the current task, and is calculated through two technical routes of Hebbian information updating. ω is an artificially set parameter, The initialization value of JPEG0007688860000049.jpg55 is 0. JPEG0007688860000050.jpg55 is a list that stores Hebbian information corresponding to each history task of the i-th synapse, JPEG0007688860000051.jpg56 is the head information corresponding to the j-th task stored in the list.

[0052] Specifically, there are two ways to update the Hebbian information of synapses. The first way is to update the Hebbian information according to the neuronal activity information within some time windows, i.e. In this case, JPEG0007688860000052.jpg54 is represented as follows: JPEG0007688860000053.jpg938JPEG0007688860000054.jpg59 and JPEG0007688860000055.jpg511 represents the firing states of pre-synaptic and postsynaptic neurons, respectively, in the t-th time window, where the Hebbian information is updated once every T time windows.

[0053] The second type of method updates and calculates the Hebbian information according to the synaptic co-discharge state in a single time window, i.e. In this case, JPEG0007688860000056.jpg54 is represented as follows: JPEG0007688860000057.jpg731 In this case, the Hebb information is updated once per time window.

[0054] Preferably, the more active the neurons before and after the i-th synapse are, the more frequent the joint discharge phenomenon becomes. JPEG0007688860000058.jpg54 becomes larger, and the updated Hebbian information also becomes larger, indicating that the i-th synapse is more important for the j-th task.

[0055] (4) In the backpropagation stage, locking of Hebbian synapses according to Hebbian information is performed by generating a mask for the synapse according to the accumulated Hebbian information in the recorded historical task in the backpropagation stage, thereby further protecting the knowledge related to the historical task in the network and improving the continuous learning ability. Specifically, in the backpropagation stage of each task, whether or not a synapse is related to a certain historical task is determined according to the Hebbian information corresponding to the historical task recorded by each synapse. The basis for determining the relevance of the i-th synapse is as shown in the formula: JPEG0007688860000059.jpg44 is calculated and obtained. JPEG0007688860000060.jpg1028JPEG0007688860000061.jpg55 represents the largest Hebbian information value corresponding to the jth task in the list of Hebbian information corresponding to each history task of the ith synapse, JPEG0007688860000062.jpg54 is the associated flag and has the largest heptogram value JPEG0007688860000063.jpg55 is the threshold If it is greater than JPEG0007688860000064.jpg57, the i-th synapse is considered to be related to the j-th task, and the change amount of the i-th synapse is masked with a mask during backpropagation, ensuring that the weight value of the related synapse i is not changed by the current task, that is, the locking of the synaptic weight value. Here, we regard such judgment of the association between synapses and tasks and the method of masking synapses as the main content of Hebbian synapse locking.

[0056] (5) A continuous learning model of a spiking neural network based on Hebbian synapse locking is used to identify sequences of heart rate, brain signals, etc. That is, the sequence information of heart rate, brain signals, etc. is identified by the brain activity state classification model after training, and a population decision method is used in the output layer, so that the category with the most responses to one input is the final output category of the model classification.

[0057] As an example, as shown in the flow chart of the training method for the brain activity state classification model in FIG. 4, specifically, the process is as follows.

[0058] In step S1, a spiking neural network with self-adaptive Hebbian information calculation is constructed using dynamic neurons with a predefined threshold, and an initial brain activity state classification model is further established based on the spiking neural network.

[0059] In step S2, for a signal input, i.e., EEG signal samples corresponding to multiple brain activity state classification training tasks, it is divided into N frames, and each frame is encoded into a spike train of one normal distribution or other distribution.

[0060] In step S3, the spike signal of the current task is input to the constructed initial brain activity state classification model, and in the forward propagation stage of task training, each synapse calculates, updates, and records the Hebbian information corresponding to the task.

[0061] In step S4, in the backpropagation stage, a mask is generated for the synapses according to the accumulated Hebbian information in the recorded historical tasks, thereby performing masking to protect the knowledge related to the historical tasks in the network. By protecting the information of the already trained tasks with the Hebbian information, the already trained tasks can also be accurately identified, and the problem of catastrophic forgetting is solved.

[0062] In step S5, it is determined whether or not an unlearned task is encountered. If an unlearned task is present, steps S3 and S4 are repeated until the initial brain activity state classification model has completed all learned tasks, thereby completing the training of the brain activity state classification model.

[0063] For example, the MNIST dataset is selected to verify the task of continuous learning of Task-IL, where Task_IL is task incremental learning. In such a scenario, the model is notified of the current task ID in both the training and testing stages, and different tasks have independent output layers. Using the above classification learning method, the relationship between the average accuracy rate and the network scale, discharge sparsity, and synapse locking rate is verified. The accuracy rate is defined as the number of correctly identified samples divided by the total number of samples. The threshold is defined as the proportion of locked synapses. The verification results show that the use of the method of the present invention has a high accuracy rate predominance in the continuous learning of Task-IL, and the relationship between the average accuracy rate and the change of the three parameters all meet the properties of the network designed by us.

[0064] For example, the MNIST dataset is selected to verify the Domain-IL continuous learning task, where Domain_IL is a domain incremental learning, and compared with Task-IL, new restrictions are added in the test phase, i.e., the ID of the task is not notified in the prediction phase, and different tasks share the same output layer. The model needs to accurately classify data when the task ID is unknown. Using the above classification learning method, the relationship between the average accuracy rate and the network size, discharge sparsity, and synaptic lock rate is verified. The accuracy rate is defined as the number of correctly identified samples divided by the total number of samples. The verification results show that the change relationship between the average accuracy rate and the three parameters is sufficiently clear and is consistent with the nature of the network constructed in the present invention.

[0065] The specific parameter settings in the above two examples are as shown in Table 1.

[0066] Table 1 JPEG0007688860000065.jpg42152, where g is the conductance coefficient, JPEG0007688860000066.jpg56 is the discharge threshold of the neuron, JPEG0007688860000067.jpg57 is the refractory period and T is the time window to simulate a dynamic neuron. Furthermore, in the present invention, the capacitance of the membrane potential C=1 μF / cm 2 and the reset membrane potential V rest =0mV.

[0067] It can be seen that the present invention has the advantages of stronger continuous learning ability, efficient multi-task training, and biological rationality.

[0068] With regard to stronger continuous learning ability, the present invention self-adaptively calculates and assigns subsystems to continuous learning tasks, and has stronger continuous learning ability than both deep neural networks and conventional modular architecture continuous learning paradigms.

[0069] Regarding efficient multi-task training, in the present invention, the whole network is used for learning different tasks, and the training efficiency of multiple tasks and large tasks is higher, which is a capability that the traditional modular architecture continuous learning paradigm cannot possess.

[0070] In terms of biological rationality, in the present invention, the synapse selection according to Hebb's law and the incorporation of Hebbian synapse locking, as well as the self-adaptive allocation of task subsystems, further enhance the biological rationality of both the model design and the continuous learning method.

[0071] The following describes the brain activity state classification model training device of the present invention, and the brain activity state classification model training device described below can be referred to correspondingly with the brain activity state classification model training method described above.

[0072] 5 is a schematic diagram showing the structure of a training device for a brain activity state classification model according to the present invention. The training device for a brain activity state classification model according to the present embodiment is an acquisition module 710 for acquiring spike trains of electroencephalogram signal samples corresponding to a plurality of brain activity state classification training tasks; The present invention further comprises a training module 720 for inputting spike trains of EEG signal samples corresponding to each training task into an initial brain activity state classification model constructed based on a spiking neural network, training the brain activity state classification model based on a target rule, in which, in a forward propagation step in the target rule, Hebbian information corresponding to each synapse in the brain activity state classification model is updated according to the spike trains corresponding to each training task, the Hebbian information is determined based on the synaptic co-discharge frequency and is used to represent the degree of association between the training task and the synapse, and in a back propagation step in the target rule, determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and the back propagation result.

[0073] Preferably, the training module 720 is specifically used to update the Hebbian information corresponding to each synapse in the brain activity state classification model by the following formula: JPEG0007688860000068.jpg1034Of these, JPEG0007688860000069.jpg55 represents the Hebbian information of the i-th synapse before the j-th task in the spike train, JPEG0007688860000070.jpg56 represents the Hebbian information of the ith synapse after the jth task in the spike train, and ω represents a preset update rate. JPEG0007688860000071.jpg55 represents the co-discharge frequency of the i-th synapse in the brain activity state classification model corresponding to the j-th task in the spike train, JPEG0007688860000072.jpg55 represents the target list, which stores the Hebbian information of synapses corresponding to each training task. JPEG0007688860000073.jpg57 represents the Hebbian information of the i-th synapse corresponding to the j-th task stored in the target list.

[0074] Preferably, the training module 720 specifically updates the Hebbian information of synapses based on the synaptic co-discharge states in a single time window; and / or It is used to update the Hebbian information of synapses based on the average discharge rate within multiple time windows.

[0075] Preferably, the training module 720 is specifically used in the backpropagation step for determining, for any one synapse in the brain activity state classification model, that the synapse is relevant to the task if the Hebb information of the synapse is greater than a first threshold, and locking the weight value of the synapse in the brain activity state classification model; and modifying the weight value of the synapse according to the backpropagation result if the Hebb information of the synapse is not greater than the first threshold.

[0076] The apparatus of the embodiment of the present invention can be used to implement the method in any of the method embodiments described above, and the implementation principles and technical effects thereof are similar, so they will not be described again here.

[0077] FIG. 6 exemplarily shows a schematic diagram of the physical structure of an electronic device, which may include a processor 810, a communications interface 820, a memory 830 and a communications bus 840, among which the processor 810, the communications interface 820 and the memory 830 complete communication between each other via the communications bus 840. The processor 810 can call the logical instructions in the memory 830 to execute a method for training a brain activity state classification model, the method including: obtaining spike trains of EEG signal samples corresponding to a plurality of brain activity state classification training tasks; inputting the spike trains of the EEG signal samples corresponding to each training task into an initial brain activity state classification model constructed based on a spiking neural network; training the brain activity state classification model according to a target rule, in which, in a forward propagation step in the target rule, Hebbian information corresponding to each synapse in the brain activity state classification model is updated according to the spike trains corresponding to each training task, and is determined based on the synaptic co-discharge frequency and used to represent the degree of association between the training task and the synapse; and in a back propagation step in the target rule, determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and the back propagation result.

[0078] Furthermore, the logic instructions in the memory 830 may be stored in a computer-readable storage medium when they are realized in the form of a software functional unit and sold or used as an independent product. Based on this understanding, the essential part of the technical solution of the present invention or the part contributing to the prior art, or a part of the technical solution, may be embodied in the form of a software product, and the computer software product is stored in a storage medium and includes some instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method according to each embodiment of the present invention. The aforementioned storage medium includes various media capable of storing program codes, such as a USB memory, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0079] In another aspect, the present invention further provides a computer program product, the computer program product including a computer program stored in a non-transitory computer-readable storage medium, the computer program including program instructions, when the program instructions are executed by a computer, the computer can execute a method for training a brain activity state classification model according to each of the above methods, the method including: acquiring spike trains of EEG signal samples corresponding to a plurality of brain activity state classification training tasks; inputting the spike trains of the EEG signal samples corresponding to each training task into an initial brain activity state classification model constructed based on a spiking neural network; and training the brain activity state classification model based on a target rule, in which, in a forward propagation step in the target rule, Hebbian information corresponding to each synapse in the brain activity state classification model is updated according to the spike trains corresponding to each training task, the Hebbian information being determined based on the synaptic co-discharge frequency and being used to represent the degree of association between the training task and the synapse; and in a back propagation step in the target rule, determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and a back propagation result.

[0080] In yet another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, realizes and executes the above-mentioned respective training methods for a brain activity state classification model, the method including: obtaining spike trains of EEG signal samples corresponding to a plurality of brain activity state classification training tasks; inputting the spike trains of the EEG signal samples corresponding to each training task into an initial brain activity state classification model constructed based on a spiking neural network; training the brain activity state classification model based on a target rule, in which, in a forward propagation step in the target rule, Hebbian information corresponding to each synapse in the brain activity state classification model is updated according to the spike trains corresponding to each training task, the Hebbian information is determined based on the synaptic co-discharge frequency and is used to represent the degree of association between the training task and the synapse, and in a back propagation step in the target rule, determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and the back propagation result.

[0081] The above-described device embodiments are merely schematic, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., located in one place or distributed among multiple network units. Some or all of these modules can be selected according to actual needs to achieve the objectives of the aspects of the present embodiment. Those skilled in the art can understand and implement the present embodiment without creative labor.

[0082] From the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be realized in the form of software plus a necessary general-purpose hardware platform, and of course, can also be realized by hardware. Based on this understanding, the essential part of the above technical solution or the part contributing to the prior art can be embodied in the form of a software product, which may be stored in a storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes some instructions for making a computer device (which may be a personal computer, a server, or a network device, etc.) execute the method described in each embodiment or any part of each embodiment.

[0083] Finally, it should be noted that the above embodiments are merely for illustrating the technical solutions of the present invention, and are not intended to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art may still modify the technical solutions described in the above embodiments or equally replace some of the technical features therein, and it should be understood that even if such modifications or replacements are made, the essence of the corresponding technical solutions will not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. obtaining spike trains of electroencephalographic signal samples corresponding to a plurality of brain activity state classification training tasks; inputting spike trains of electroencephalogram signal samples corresponding to each of the training tasks into an initial brain activity state classification model constructed based on a spiking neural network, training the brain activity state classification model based on a target rule, updating Hebbian information corresponding to each synapse in the brain activity state classification model according to the spike trains corresponding to each of the training tasks, the Hebbian information being determined based on a synaptic co-discharge frequency and used to represent a degree of association between the training task and the synapse in a forward propagation step in the target rule, and determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and a back propagation result in a back propagation step in the target rule. A method for training a brain activity state classification model, comprising:

2. The method of updating the Hebbian information corresponding to each synapse in the brain activity state classification model according to the spike train corresponding to each training task includes: Updating the Hebbian information corresponding to each synapse in the brain activity state classification model using the following formula: represents the Hebbian information of the i-th synapse before the j-th task in the spike train, represents the Hebbian information of the i-th synapse after the j-th task in the spike train, ω represents a preset update rate, represents the co-discharge frequency of the i-th synapse in the brain activity state classification model corresponding to the j-th task in the spike train, represents a goal list that stores the Hebbian information of synapses corresponding to each training task, represents the Hebbian information of the i-th synapse corresponding to the j-th task stored in the target list.

2. The method for training a brain activity state classification model according to claim 1 .

3. The above-mentioned Hebbian information corresponding to each synapse in the brain activity state classification model is updated. Updating synaptic Hebbian information based on synaptic co-discharge states in a single time window; and / or updating synaptic Hebbian information based on average discharge rates within a plurality of time windows; 3. The method for training a brain activity state classification model according to claim 2.

4. In the back propagation step of the target rule, determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and the back propagation result; In the back propagation step, for any one synapse of the brain activity state classification model, if the Hebbian information of the synapse is greater than a first threshold, the synapse is determined to be relevant to the task, and a weight value of the synapse in the brain activity state classification model is locked; if the Hebbian information of the synapse is not greater than the first threshold, the weight value of the synapse is modified according to the back propagation result.

4. The method for training a brain activity state classification model according to claim 3.

5. an acquisition module for acquiring spike trains of electroencephalogram signal samples corresponding to a training task of a plurality of brain activity state classifications; a training module for inputting spike trains of electroencephalogram signal samples corresponding to each of the training tasks into an initial brain activity state classification model constructed based on a spiking neural network, training the brain activity state classification model based on a target rule, updating Hebbian information corresponding to each synapse in the brain activity state classification model according to the spike trains corresponding to each of the training tasks in a forward propagation step in the target rule, the Hebbian information being determined based on a synaptic co-discharge frequency and used to represent a degree of association between the training task and the synapse, and determining a weight value of the synapse in the brain activity state classification model according to the Hebbian information corresponding to each synapse and a back propagation result in a back propagation step in the target rule. A training device for a brain activity state classification model, comprising:

6. The training module is used to update the Hebbian information corresponding to each synapse in the brain activity state classification model by the following formula: represents the Hebbian information of the i-th synapse before the j-th task in the spike train, represents the Hebbian information of the i-th synapse after the j-th task in the spike train, ω represents a preset update rate, represents the co-discharge frequency of the i-th synapse in the brain activity state classification model corresponding to the j-th task in the spike train, represents a goal list that stores the Hebbian information of synapses corresponding to each training task, represents the Hebbian information of the i-th synapse corresponding to the j-th task stored in the target list.

6. The training device for a brain activity state classification model according to claim 5.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and operable by the processor, When the processor executes the program, the method for training a brain activity state classification model according to any one of claims 1 to 4 is realized.

1. An electronic device comprising:

8. A non-transitory computer-readable storage medium having a computer program stored thereon, When the computer program is executed by a processor, the computer program realizes the method for training a brain activity state classification model according to any one of claims 1 to 4. A non-transitory computer-readable storage medium comprising:

Images