Cortical spike signal motion decoding method and decoding device based on recurrent spiking neural network

Abstract

The present invention discloses a cortical spike signal motion decoding method and decoding device based on a recurrent spiking neural network, the decoding method comprising: obtaining a cortical spike sequence signal; constructing a recurrent spiking neural network, the recurrent spiking neural network comprising an input layer, a hidden layer and an output layer connected in sequence; the hidden layer comprising: one or more serially connected recurrent hidden layers, each recurrent hidden layer comprising: a plurality of standard leaky integral discharge neurons connected in a recurrent topology; training the recurrent spiking neural network; using the trained recurrent spiking neural network to perform motion decoding on the required cortical spike sequence signal. The present invention reduces the power consumption of the decoding device by introducing activity regularization technology, adopting an iterative synaptic pruning strategy, converting the model accuracy and other means. The present invention provides a competitive solution for the practical application of a fully implantable brain-computer interface.

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram (EEG) data analysis, and in particular relates to a cortical peak potential signal motion decoding method and decoding device based on a recurrent pulse neural network. Background Art

[0002] In recent years, brain-computer interface technology has made significant progress in the fields of motion control, rehabilitation therapy, and neuroscience research. Invasive brain-computer interfaces collect cortical nerve signals to achieve precise control of external devices, providing patients with movement disorders with a new way to improve their quality of life. However, there is a risk of infection caused by externally installed components in the existing technology, which has prompted research to turn to fully implantable systems. Since implantable devices need to work inside the human body for a long time and their energy supply is relatively limited, it is necessary to reduce power consumption as much as possible while ensuring decoding performance to extend the service life of the device. At the same time, in order to achieve real-time and smooth human-computer interaction, the system delay must also be controlled at an extremely low level to ensure that the commands issued by the brain can be quickly converted into actions of external devices to avoid obvious lags.

[0003] Spiking Neural Network (SNN) communicates between neurons through spikes, naturally adapts to cortical spike signals, and has significant advantages of low power consumption and low latency. These characteristics make SNN a potential ideal decoding solution for fully implantable brain-computer interfaces. However, most current feedforward SNN decoders only show moderate decoding performance in the task of decoding cortical spike signal trajectories, while traditional artificial neural networks (ANN) have high decoding capabilities but high energy consumption requirements, which is inconsistent with the requirements of implantable devices with limited energy supply. In contrast, recurrent spiking neural networks (RSNN) significantly enhance the ability to capture spatiotemporal dependent features by introducing recurrent connections between neurons. It not only inherits the low power consumption and low latency characteristics of SNN, but also shows obvious potential advantages in decoding performance. However, the application of RSNN in the field of cortical spike motion trajectory decoding is still not widespread, and there is a lack of optimized design for actual fully implantable systems. Summary of the invention

[0004] In view of the shortcomings of the existing technology in balancing motion decoding performance and resource efficiency, the present invention proposes a cortical peak potential signal motion decoding method and decoding device based on recurrent spiking neural network (RSNN). This method aims to solve the contradiction between accuracy and energy efficiency that is prevalent in traditional decoding technology, especially for fully implantable brain-computer interface (BCI) systems, which require efficient and accurate decoding under extremely limited power consumption and computing resources.

[0005] The method of the present invention utilizes the cortical spike sequence signals collected by the implantable brain-computer interface, and realizes efficient prediction of the target motion trajectory through preprocessing, neural network modeling and training. In order to reduce energy consumption, the present invention introduces an activity regularization method to control the average firing rate of neurons, reduces network redundant connections through synaptic pruning, and changes the model parameters to half-precision floating point numbers after the model training is completed, further reducing power consumption and storage size. This method has the characteristics of superior decoding performance and energy efficiency optimization, and is particularly suitable for low-power fully implantable brain-computer interface systems.

[0006] A method for cortical spike signal motion decoding based on a recurrent spiking neural network comprises the following steps: S1: obtaining a cortical spike sequence signal; S2: constructing a recurrent spiking neural network, wherein the recurrent spiking neural network comprises an input layer, a hidden layer and an output layer connected in sequence; the hidden layer comprises: one or more recurrent hidden layers connected in series, each recurrent hidden layer comprises: a plurality of standard leaky integral discharge neurons connected in a recurrent topology; S3: using a proxy gradient method, an activity regularization method, an iterative synaptic pruning method and the cortical spike sequence signal obtained in step S1 to train the recurrent spiking neural network, to obtain a trained recurrent spiking neural network; S4: using the trained recurrent spiking neural network to perform motion decoding on the required cortical spike sequence signal.

[0007] In step S1, a cortical peak potential sequence signal is obtained, which specifically includes: collecting original neural signals from the cerebral cortex and performing preprocessing to obtain a cortical peak potential sequence signal.

[0008] In step S1, the preprocessing includes denoising, filtering and peak potential detection.

[0009] In step S2, the input layer has the same number of channels as the cortical peak potential train signal.

[0010] In step S2, the multiple cyclic topologically connected standard leaky integral discharge neurons specifically include: connecting the multiple standard leaky integral discharge neurons to each other.

[0011] In step S2, the synaptic activities of each connection in the input layer, hidden layer and output layer decay over time according to an exponential function.

[0012] In step S2, the output layer includes: non-pulse leaky integrating neurons, and the membrane potential of the non-pulse leaky integrating neurons is output by motion decoding of a recurrent spiking neural network.

[0013] In step S3, the recurrent spiking neural network is trained with the goal of minimizing the error between the decoding result and the true motion trajectory.

[0014] In step S3, the recurrent spiking neural network is trained using the proxy gradient method, the activity regularization method, the iterative synaptic pruning method, and the cortical spike potential sequence signal obtained in step S1, including: inputting the cortical spike potential sequence signal obtained in step S1 into the recurrent spiking neural network, using the proxy gradient method to calculate the error gradient of the recurrent spiking neural network to optimize the recurrent spiking neural network parameters, and using the activity regularization method to limit the standard leaky integral discharge neuron activity intensity of the hidden layer in the recurrent spiking neural network. After training the recurrent spiking neural network, using the iterative synaptic pruning method to gradually remove the synaptic connections with the smallest absolute value of weights (i.e., low connection strength) between the input layer, the hidden layer, and the output layer, and in the hidden layer of the recurrent spiking neural network, and finally converting the parameters of the recurrent spiking neural network into half precision to obtain a trained recurrent spiking neural network.

[0015] A decoding device, comprising:

[0016] one or more processors;

[0017] A memory for storing one or more programs;

[0018] When the one or more programs are executed by the one or more processors, the one or more processors implement the cortical peak potential signal motion decoding method based on recurrent pulse neural network.

[0019] Furthermore, the cortical peak potential signal motion decoding method based on recurrent spiking neural network of the present invention comprises the following steps:

[0020] (1) Signal acquisition and preprocessing: Cortical spike trains (CST) signals are collected and preprocessed, including filtering, denoising, and spike detection, to preserve the fine time series characteristics of the signal and improve signal quality;

[0021] (2) Construct a recurrent spiking neural network (RSNN), which mainly consists of three parts:

[0022] Input layer: matches the number of channels of cortical spike potential sequence signals and is used to receive multi-channel neural signals;

[0023] Hidden layers, including: one or more cascaded recurrent hidden layers, the recurrent hidden layers are composed of standard Leaky Integrate-and-Fire (LIF) neurons, and the neurons have recurrent topological connections to capture the dynamic spatiotemporal dependency characteristics of the signal;

[0024] Output layer: composed of non-pulse leaky integrator (LI) neurons, used to generate motion decoding results;

[0025] (3) Network training and optimization:

[0026] Through the end-to-end training process, the network parameters of the RSNN model are optimized. The training goal is to minimize the error between the decoding result output by the network and the actual motion trajectory, thereby improving the decoding accuracy.

[0027] (4) Motion decoding:

[0028] The trained recurrent spiking neural network is used to decode the cortical spike sequence signal and predict the target motion trajectory.

[0029] In order to improve the decoding accuracy of the recurrent spiking neural network RSNN, the connection weights of the recurrent spiking neural network and the heterogeneity time constant of the neurons in the network can be trained, and the connection synaptic activity of the network Be dynamic:

[0030]

[0031] in, is the synaptic activity attenuation coefficient, is the weighted upstream input received by the synapse, is the synaptic activity of the neuron at the previous moment (t-1 moment), is the synaptic activity of the neuron at this moment (time t).

[0032] Furthermore, in order to adapt to the requirements of fully implantable brain-computer interfaces for low power consumption and resource efficiency, the present invention proposes a series of optimization strategies. It aims to optimize the system's energy management, computing efficiency and decoding performance to ensure that an efficient and stable decoding model can be achieved under limited hardware resources and power supply conditions. These strategies include multi-faceted optimization measures from neuronal activity control, network structure simplification to hardware adaptation. Specifically, the present invention proposes the following optimization measures: First, introduce an activity regularization method (i.e., activity regularization technology) to constrain the average peak potential firing rate of neurons; second, adopt an iterative synaptic pruning method (i.e., iterative synaptic pruning strategy) to remove unnecessary synaptic connections in the neural network; third, convert model parameters and intermediate calculation results to half precision to save space and reduce power consumption.

[0033] Each optimization strategy is designed to meet the key performance requirements of a fully implantable brain-computer interface and is implemented as follows:

[0034] Introducing activity regularization technology: Activity regularization technology is a technology that constrains the average peak potential firing rate of neurons to avoid over-activation of neurons, thereby reducing overall energy consumption. During the training process, this technology dynamically adjusts the firing frequency of neurons by adding a regularization loss term to keep its activity within a reasonable range, which can not only reduce power waste but also maintain the stability of decoding accuracy. Specifically, the following regularization term is introduced in the loss function :

[0035]

[0036] in, is the number of neurons, is the firing rate of the ith neuron, is the upper threshold of the neuron firing rate, is the regularization strength. It is a modified linear unit activation function, which is one of the most commonly used activation functions in modern deep learning. Its mathematical expression is:

[0037]

[0038] This regularization strategy significantly reduces energy consumption by suppressing activities above a set threshold while avoiding negative impact on the network decoding performance.

[0039] Adopt iterative synaptic pruning strategy: Synaptic pruning is a technique that reduces network complexity and computational overhead by removing redundant or unnecessary synaptic connections in a neural network. The present invention analyzes the contribution of network weights and gradually removes synaptic connections of low importance, thereby reducing computational complexity and optimizing resource efficiency. The specific implementation steps are as follows:

[0040] S3.1: Calculate the importance score of each synaptic connection based on the absolute value of the weight of the current state of the network;

[0041] S3.2: Sort by importance score from low to high, and remove the one with the lowest importance score. connection;

[0042] S3.3: Retrain the pruned network to restore decoding performance;

[0043] S3.4: Repeat steps S3.1-S3.3, pruning and retraining one by one, until the decoding performance of the network degrades beyond the set tolerance.

[0044] This iterative synaptic pruning strategy gradually adjusts the network structure to avoid severe performance degradation that may be caused by one-time large-scale pruning. Experiments show that this strategy significantly reduces network complexity while maintaining decoding accuracy, making it suitable for resource-constrained implantable hardware.

[0045] Convert model parameters and intermediate calculation results to half-precision: In traditional neural networks, it is common to use floating-point precision for calculations, but this results in higher power consumption and computational costs. By converting 32-bit floating-point numbers to 16-bit half-precision floating-point numbers, storage and computational requirements are significantly reduced, thereby adapting to low-power hardware systems. After training is completed, model parameters are stored in half-precision format and uniformly represented using half-precision during the inference phase. This strategy not only reduces memory usage, but also reduces the computational power consumption of the model during the inference phase, showing obvious advantages in embedded devices.

[0046] It should also be noted that technical personnel in the relevant field can clearly understand that for the convenience and simplicity of description, the division of steps or modules in the above-described systems and methods is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] The RSNN model (trained recurrent spiking neural network) proposed in the present invention has broad application potential for decoding tasks of multi-dimensional motion target trajectories, especially in decoding fine movements such as finger motion trajectories. Experimental results show that the model not only surpasses traditional feedforward spiking neural networks (SNNs) and classical artificial neural networks (ANNs) in terms of decoding accuracy, but also shows significant optimization effects in terms of resource efficiency. It is particularly suitable for resource-constrained fully implantable brain-computer interface systems. Specifically, in terms of decoding accuracy, in the test of the cortical peak potential dataset of two macaques, the RSNN small model of the present invention achieved an average determination coefficient of , indicating that it has a high decoding performance for finger motion trajectories. Compared with the baseline feedforward SNN and ANN models of similar scale, the decoding accuracy is improved by about 13.60%. In addition, the RSNN model of the present invention performs outstandingly in optimizing resource utilization, with a connection sparsity of 45% and an activation sparsity of up to 98.36%. Compared with traditional SNN and ANN models, the energy consumption is only 70.92% of the SNN model of similar scale and 0.20% of the ANN model of similar scale. The above results show that the RSNN model proposed in the present invention can achieve efficient decoding performance with low power consumption in resource-constrained scenarios, providing a new technical solution for the practical application of fully implantable brain-computer interfaces. In addition, with the advancement of hardware technology and further optimization of algorithms, the computational efficiency and energy efficiency of the RSNN model still have a large room for improvement. The future fully implantable brain-computer interface system can further achieve the goals of high efficiency, low power consumption, and real-time decoding on the basis of this technology, bringing more possibilities for the practical application of brain-computer interfaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic flow chart of the cortical peak potential signal motion decoding method based on recurrent spiking neural network of the present invention.

[0050] Figure 2 This is the RSNN model architecture proposed in the present invention.

[0051] Figure 3 This is the RSNN model training flow chart proposed in the present invention.

[0052] Figure 4 The comparison results of the accuracy of the RSNN model proposed in this invention in decoding the target motion trajectory and other models.

[0053] Figure 5 The power consumption of the RSNN model proposed in this invention for decoding target trajectories is compared with other models. DETAILED DESCRIPTION

[0054] In order to describe the present invention in more detail, the technical scheme of the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0055] like Figure 1As shown, a cortical spike potential signal motion decoding method based on a recurrent spiking neural network includes the following steps: S1: obtaining a cortical spike potential sequence signal; S2: constructing a recurrent spiking neural network, the recurrent spiking neural network includes an input layer, a hidden layer and an output layer connected in sequence; the hidden layer includes: one or more recurrent hidden layers connected in series, each recurrent hidden layer includes: a plurality of standard leaky integral discharge neurons connected in a recurrent topology; S3: using a proxy gradient method, an activity regularization method, an iterative synaptic pruning method and the cortical spike potential sequence signal obtained in step S1 to train the recurrent spiking neural network, to obtain a trained recurrent spiking neural network; S4: using the trained recurrent spiking neural network to perform motion decoding on the required cortical spike potential sequence signal.

[0056] Cortical spike data collection and preprocessing. Signal data were selected from the public dataset "Nonhuman Primate Reaching with Multichannel Sensorimotor CortexElectrophysiology". In the experiment, the macaques sat in front of a screen that displayed a grid of target points, and the macaques could move their arms and fingers at their own pace. During the experiment, a certain point in the grid would light up, which was the target point. The macaques needed to move their arms to touch the target point on the screen within a specified time. The target point was randomly refreshed after the macaque's finger touched the screen, and the next target point was displayed immediately without an interval waiting time. During this process, the cortical spike signals were recorded by electrode arrays implanted in the primary motor cortex (M1) and primary sensory cortex (S1), and the finger movement trajectory was recorded synchronously. The recorded cortical spike signals were filtered, denoised, and spike detected to preserve the fine time series of the signal.

[0057] The present invention selected experimental trials of two test macaques. In the selected experimental trials: for the first macaque (Loco), the experiment recorded signals from both the M1 and S1 electrode arrays (192 channels); while for the second macaque (Indy), the experiment only recorded signals from the M1 electrode array (96 channels). The optical motion capture system records the motion trajectory (x, y) and target point position of the macaque's fingertips on a two-dimensional plane, with a sampling frequency of 250 Hz. The recorded spike timestamps are converted into discretized time steps (4 milliseconds), and each time step stores the number of spike events in the time window, which is generally 0 or 1. The discretized spike event number matrix represents the multi-channel cortical spike sequence CST signal as the input of the model. In each experiment, the data is divided into a training set (65%), a validation set (10%), and a test set (25%) to ensure the independence of model training, validation, and testing. It should be noted that the purpose of using public data sets is to use standard data to test the performance of this model. In practical applications, CST signals can be spike sequence signals collected from clinical or non-human animals using electrode arrays. In addition to finger motion trajectories, decoded variables can also be other multi-dimensional motion target trajectories.

[0058] Recurrent Spiking Neural Network (RSNN) Model Construction. The present invention designs a decoding model based on recurrent spiking neural network (RSNN) to predict the target motion trajectory from the cortical peak potential signal. The structure of the RSNN model is as follows: Figure 2 As shown in the figure: The RSNN model consists of input layer 1, hidden layer 2 and output layer 3. Hidden layer 2 contains one or more recurrent hidden layers. In addition to receiving the output of the previous layer, the recurrent hidden layer also receives the weighted output of the previous layer to enhance the ability to process time series signals. Then the input layer 1, hidden layer 2 and output layer 3 are connected in sequence to form the RSNN model. It should be understood that Figure 2 In the network structure, hidden layer 2 includes only one cyclic hidden layer, loop 4 represents the cyclic hidden layer receiving the output of the previous moment of this layer. In the specific experiment of the present invention, only one cyclic hidden layer structure is used.

[0059] The input layer 1 mainly provides an interface between a neural network and external data. The input layer 1 sends out the cortical spike potential sequence signal collected in the experiment. The number of channels is consistent with the number of electrodes for collecting experimental signals. In the experiment of the present invention, the input layer 1 uses 192 channels (the first macaque) or 96 channels (the second macaque);

[0060] The recurrent hidden layer consists of 64 standard leaky integrate-and-fire (LIF) neurons5, each of which dynamically captures the time dependency of the input signal. The characteristic of standard leaky integrate-and-fire neurons is that the output of the neuron is also a pulse train signal similar to CST. The membrane potential activity of the LIF neuron at this moment The dynamics obey the following differential equation:

[0061]

[0062] in, is the membrane potential activity of the LIF neuron at the previous moment (t-1 moment), is the membrane potential activity attenuation coefficient, is the synaptic activity of the neuron at the previous moment, is the neuron pulse firing activity at this moment (time t), which is 1 when a pulse is fired and 0 when it is not. The firing depends on the membrane potential activity of the LIF neuron at the previous moment. and threshold :

[0063]

[0064] Neuronal synaptic activity at this moment Satisfies the following differential equation:

[0065]

[0066] in, is the synaptic activity attenuation coefficient, is the weighted upstream input received by the synapse. is defined as follows:

[0067]

[0068] in, is the upstream input nerve impulse at this moment, It is the neural pulse emitted by the current layer at the previous moment. and Corresponding to The feedforward synaptic connection weights and correspond to The recurrent hidden layer forms a recurrent loop by receiving the weighted sum of the neural pulses emitted at the previous moment, thereby enhancing the processing capability for time series signals.

[0069] The membrane potential activity decay coefficient and the synaptic activity decay coefficient Respectively meet:

[0070]

[0071]

[0072] in is the time step, are the time constants of membrane potential and synaptic activity, respectively. To capture the spatiotemporal dependency of input signals, neurons in the hidden layer have recursive connections. LIF neurons encode and transmit information through pulse trains, which can reduce energy consumption.

[0073] The output layer 3 consists of two non-spiking leaky integrator (LI) neurons 6, which are used to transform the spike activity of hidden layer 2 Converted into continuous value output, the two-dimensional velocity (X and Y directions) of the target motion trajectory is predicted respectively. The dynamic equation of membrane potential activity is:

[0074]

[0075]

[0076]

[0077] in, is the weighted upstream input received by the LI neuron, is the output corresponding to the recurrent hidden layer The feedforward synaptic connection matrix of . Except that the input of output layer 3 depends only on the output of the recurrent hidden layer, the meaning of each variable is consistent with the relevant variables in the dynamic equation of LIF neuron membrane potential activity. is the membrane potential activity of the LI neuron at this moment, is the membrane potential activity of LI neuron at the previous moment, and are the membrane potential activity attenuation coefficient and synaptic activity attenuation coefficient of LI neurons, and They are the synaptic activities of LI neuron at this moment and last moment respectively. The synaptic activities of LI neuron satisfy the dynamic equation consistent with that of LIF neuron.

[0078] Model training. The recurrent spiking neural network (RSNN) model of the present invention adopts an end-to-end training strategy to optimize the model parameters by minimizing the root mean square error (RMSE) between the output layer prediction value and the true target trajectory. The loss function is defined as:

[0079]

[0080] in is the total loss function, Represents the error between the predicted value and the target value label, is an activity regularization term, which is used to constrain the average firing rate of neurons, keep their activity within a reasonable range, avoid over-activation, and reduce energy consumption:

[0081]

[0082] in, is the number of neurons, is the firing rate of the ith neuron, is the upper threshold of the neuron firing rate, is the regularization strength, is the rectified linear unit activation function.

[0083] The connection weights of the recurrent spiking neural network model and the heterogeneous time constants of the neurons in the network can all be trained. Model training is divided into three stages: pre-training, fine-tuning, and pruning. Figure 3 As shown. In the pre-training stage, the data from all experiments in the public dataset of each macaque were mixed and trained for 100 cycles to better initialize the network weights. Then, the experimental data of each monkey was fine-tuned for 200 cycles to obtain a model adapted to the experimental conditions of the monkey. In the network, a considerable number of connections occupy computing resources but have little impact on network performance. Therefore, in the pruning stage, an iterative pruning strategy is adopted to remove the connections with the lowest importance according to the absolute value score of the weight. , and then retrain for 100 cycles after pruning until the model decoding performance drops by more than 2%. In traditional neural networks, the activation function of neurons is usually continuously differentiable. However, the neuron model in RSNN has the discontinuity of pulse emission, and conventional gradient optimization methods are not applicable. In order to solve the problem of non-differentiable pulse emission mechanism, the surrogate gradient descent method is used to optimize the RSNN network parameters. The core idea of ​​the surrogate gradient descent method is to introduce a continuously differentiable surrogate function to approximate the true gradient. The training is combined with the cosine annealing learning rate scheduler to dynamically adjust the learning rate to accelerate stable learning. At the same time, Dropout (drop rate of 0.3) is introduced in the hidden layer and input layer to alleviate overfitting. Model training mainly uses the surrogate gradient descent method to adjust the weight matrix [ , , ] and time constant [ ] for adjustment. Low-power hardware systems have higher requirements for memory and computing resources, and the performance improvement of 32-bit floating-point numbers on general-purpose computers is not significant compared to 16-bit half-precision floating-point numbers. Using 16-bit half-precision floating-point numbers can reduce memory usage and power consumption. Therefore, after training, the RSNN model will be converted to half-precision, and half-precision representation will be uniformly used in the inference phase. Through the above training process, the RSNN model can achieve high decoding accuracy while further improving sparsity and energy efficiency, providing an efficient decoding solution for resource-constrained scenarios of fully implantable brain-computer interfaces.

[0084] The target motion trajectory is predicted in the model inference stage. The present invention successfully decodes the finger motion trajectory from the cortical peak potential sequence signal through the recurrent spike neural network RSNN. The experimental results show that Figure 4 As shown in Figure 2, RSNN is significantly better than the classic algorithms Kalman Filter (KF), feedforward spike neural network (SNN) and traditional artificial neural network (ANN) in terms of prediction accuracy; at the same time, Figure 5 As shown in the figure, the RSNN model of the present invention has obvious advantages in energy efficiency compared with feedforward spike neural networks (SNN) and especially traditional artificial neural networks (ANN). Specifically, in the test of two macaque cortical spike potential datasets, the RSNN model of the present invention achieved an average determination coefficient of , indicating that its decoding performance of finger motion trajectory is high, comparable to the baseline feed-forward SNN of similar scale ( ) or ANN ( ) model, the decoding accuracy is improved by about 13.60%. The average determination coefficient of the classic algorithm KF =0.18, average determination coefficient of RSNN The significant improvement is greater than 200%, which shows that RSNN has a greater advantage in decoding performance. In addition, the RSNN model of the present invention performs outstandingly in optimizing resource utilization, with a connection sparsity of 45% and an activation sparsity of up to 98.36%. In synaptic operations, the ANN model requires 4967 multiplication and addition operations, the SNN model requires 414 accumulation operations, and the RSNN model, although more complex, only requires 304 accumulation operations. According to the energy consumption of a multiplication and addition operation is about 31 times that of an accumulation operation, the synaptic operation energy efficiency ratio shows that the energy consumption between RSNN synapses is about 1 / 1.41 (70.92%) of the SNN model; in particular, it is about 1 / 506.5 (0.20%) of the ANN model, and the energy consumption differs by more than 500 times. The above results show that compared with other models, the RSNN model proposed in the present invention can achieve high-precision decoding, and at the same time shows significant advantages in energy consumption, which meets the requirements of low power consumption for fully implanted brain-computer interfaces, and provides a competitive solution for the practical application of fully implanted brain-computer interfaces.

[0085] The above-described embodiments are only some preferred solutions of the present invention, but are not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A cortical spike signal motion decoding method based on recurrent spiking neural network, characterized in that: The following steps are involved: S1: obtain cortical spike train signals; S2: constructing a recurrent spiking neural network, wherein the recurrent spiking neural network comprises an input layer, a hidden layer, and an output layer connected in sequence; The hidden layer comprises: one or more cyclic hidden layers connected in series, each cyclic hidden layer comprises: a plurality of standard leaky integral discharge neurons connected in cyclic topology; S3: training the recurrent spiking neural network using the proxy gradient method, the activity regularization method, the iterative synaptic pruning method, and the cortical spike potential sequence signal obtained in step S1 to obtain a trained recurrent spiking neural network, specifically including: The cortical spike sequence signal obtained in step S1 is input into the recurrent spiking neural network, and the error gradient of the recurrent spiking neural network is calculated using the proxy gradient method to optimize the parameters of the recurrent spiking neural network. At the same time, the activity regularization method is used to limit the activity intensity of the standard leaky integral discharge neurons in the hidden layer of the recurrent spiking neural network. After the recurrent spiking neural network is trained, the synaptic connections with the smallest absolute weights between the input layer, the hidden layer and the output layer, and in the hidden layer of the recurrent spiking neural network are gradually removed using the iterative synaptic pruning method. Finally, the parameters of the recurrent spiking neural network are converted to half precision to obtain a trained recurrent spiking neural network. S4: Use the trained recurrent spiking neural network to perform motion decoding on the required cortical spike potential sequence signals.

2. The method for cortical peak potential signal motion decoding based on recurrent spiking neural network according to claim 1, characterized in that: In step S1, obtaining a cortical spike potential sequence signal specifically includes: The original neural signals of the cerebral cortex are collected and preprocessed to obtain the cortical peak potential sequence signals.

3. The method for cortical peak potential signal motion decoding based on recurrent spiking neural network according to claim 2, characterized in that: In step S1, the preprocessing includes denoising, filtering and peak potential detection.

4. The method for cortical peak potential signal motion decoding based on recurrent spiking neural network according to claim 1, characterized in that: In step S2, the input layer has the same number of channels as the cortical peak potential train signal.

5. The method for cortical peak potential signal motion decoding based on recurrent spiking neural network according to claim 1, characterized in that: In step S2, a plurality of standard leaky integrate-and-discharge neurons connected in a cyclic topology specifically include: Connect multiple standard leaky integrate-and-discharge neurons to each other.

6. The method for cortical peak potential signal motion decoding based on recurrent spiking neural network according to claim 1, characterized in that: In step S2, the synaptic activities of each connection in the input layer, hidden layer and output layer decay over time according to an exponential function.

7. The cortical peak potential signal motion decoding method based on recurrent spiking neural network according to claim 1, characterized in that: In step S2, the output layer includes: non-pulse leaky integrating neurons, and the membrane potential of the non-pulse leaky integrating neurons is output by motion decoding of a recurrent spiking neural network.

8. The method for cortical peak potential signal motion decoding based on recurrent spiking neural network according to claim 1, characterized in that: In step S3, the recurrent spiking neural network is trained with the goal of minimizing the error between the decoding result and the true motion trajectory.

9. A decoding device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the cortical peak potential signal motion decoding method based on recurrent pulse neural network as described in any one of claims 1-8.

Images