A node scale adaptive neuromorphic computing neuron classification method

Abstract

The application discloses a node scale adaptive neural quasi-state computing neuron classification method, and belongs to the field of electroencephalogram spike classification decoding. The application constructs a pulse neural network framework containing a two-layer structure pulse neural network and a attention neuron node, combines prior knowledge of a spike waveform, automatically guides addition and deletion of network nodes, optimizes allocation of computing resources according to specific requirements, and maximally reduces waste of hardware resources. The method has the characteristics of low hardware overhead, fast computing speed and high result consistency on different data sets, can improve the speed of a spike classification process, and is expected to provide full-automatic neuron classification technical support for a wireless implantable electroencephalogram signal acquisition device.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal spike potential classification and decoding, and in particular to a neuromorphic computation method for neuron classification that is adaptive to node size. Background Technology

[0002] Spike sequencing plays a crucial role in neuroscience research by accurately classifying spike events generated by individual neurons through signal channels. By precisely identifying and classifying these neurons, spike sequencing provides valuable insights into the functional characteristics, connectivity patterns, and temporal dynamics of specific neuronal types. This information is essential for understanding neural circuits and unraveling the complexities of brain function.

[0003] Traditional spike sequencing methods primarily rely on manual inspection or semi-automated processes. However, recent advancements in microelectronics and nanoscale structures have enabled neural recordings to be performed on a massive scale using thousands of channels, making manual spike sequencing virtually impossible. Large-scale recording also presents critical challenges in computation and signal transmission. The increased data volume leads to higher signal processing and transmission costs, making brain implants nearly unfeasible.

[0004] For applications that primarily require spike potential training signals, chip-based spike potential sequencing is considered a potential solution for reducing transmission bandwidth. As the name suggests, these sequencing methods reside within the brain implant device, transmitting only spike potential events to downstream tasks. In this way, the scale of data transmitted can be significantly reduced, potentially enabling wireless transmission. Therefore, the key issue is finding a low-cost, automated spike potential sequencing method suitable for large-scale recording for brain implant computation.

[0005] Although seemingly simple, fully automating spike potential sorting at low cost remains a challenging task. Several methods have been employed to transfer whole or partial sorting results. These methods range from analyzing individual action potentials to attempting spike potential extraction and classification. For example, Chinese patent document CN115844422A discloses a method for classifying neuronal spike potentials, including acquiring raw spike potential signals, preprocessing the acquired spike potential signals; performing spike potential detection on the preprocessed dataset using a heuristic adaptive threshold to obtain a spike potential signal dataset; reducing the dimensionality of the spike potential data using principal component analysis to obtain eigenvalues ​​and eigenvectors, and mapping the spike potential points to a feature space constructed from the eigenvectors; using the eigenvalues ​​and spike potential data as input for K-means clustering, iterating continuously to ensure that the cluster centers remain unchanged while guaranteeing that each sample is closest to its corresponding cluster center, thus obtaining the classification result.

[0006] Recent advances in large-scale spike sequencing techniques have demonstrated high accuracy while operating in a fully autonomous manner. However, these methods often face the challenge of high computational complexity, making it difficult to achieve rapid channel scaling within the power constraints of implanted devices.

[0007] Currently, these methods face two major obstacles. First, implementing algorithms for complex tasks incurs significant hardware costs and requires meeting higher computational resource demands within the limited space of the device. Second, the temperature sensitivity and rejection response of nerve cells necessitate low power consumption and small size in implanted devices. Achieving a suitable balance between power consumption and circuit area is a key area that requires further exploration in this field. Summary of the Invention

[0008] This invention provides a neuromorphic computation neuron classification method with adaptive node size. It improves upon the problems of slow manual classification, inconsistent classification results from different experts, and long time consumption in the classification of spike potential signals. This method improves the speed of the spike potential classification process to a certain extent and maintains high classification consistency on different datasets. In addition, this method also helps in the deployment of implanted chips.

[0009] A neuromorphic computational neuron classification method with adaptive node size includes the following steps:

[0010] (1) Obtain the original multi-channel neural signals, remove low-frequency local field potentials by bandpass filtering, and perform whitening preprocessing and artifact removal on the neural signals of each channel;

[0011] (2) Detect and align candidate peak potentials on each signal channel. Specifically, use a nonlinear energy operator to calculate the energy intensity at each position of the discrete signal. A time window that exceeds the threshold is determined to be a candidate peak potential. Then, align the candidate peak potentials according to the peak position.

[0012] (3) Construct a spiking neural network framework, which includes a two-layer spiking neural network and an attention neuron node; wherein, the first layer of the spiking neural network is a perceptual layer and the second layer is a cognitive layer, and the neurons between the two layers are connected in a fully connected manner and the connection synapses are dynamically updated; the attention neuron controls all neurons in the network through unidirectional connection;

[0013] (4) Input the aligned candidate spike potentials into the spiking neural network framework. The perceptual layer of the spiking neural network performs pulse encoding on the candidate spike potentials and maps the discrete signals of each time point of the candidate spike potentials to a set of pulse sequences in a Gaussian receptive field encoding manner.

[0014] Neurons in the cognitive layer respond to different pulse sequence inputs and update the synapses connecting the activated neurons with the corresponding neurons in the sensory layer through a winner-takes-all mechanism; when the accumulated voltage of the cognitive layer neurons exceeds the voltage threshold, the pulse sequence is output as a timestamp sequence in response to different cellular action potentials;

[0015] Attention neurons respond to prior knowledge of the waveform of the input candidate spike potential, modify the waveform of the input candidate spike potential, and control the threshold changes of the strategy of adding, deleting and merging nodes in the perceptual layer.

[0016] (5) For the original neural signal corresponding to the timestamp sequence, the cognitive layer nodes dynamically update the threshold to divide the spike potential and noise according to the prior knowledge of the pulse waveform, and each channel reconstructs the waveform emitted from different cells according to the timestamp sequence output by the spiking neural network.

[0017] Furthermore, in step (1), the bandpass filter uses a third-order Butterworth filter with a bandpass frequency of 300–3000 Hz.

[0018] In step (2), the energy intensity at each location of the discrete signal is calculated using a nonlinear energy operator, as shown in the formula:

[0019] ψ[x(n)]=x 2 (n)-x(n+1)·x(n-1)

[0020] Where x(n) are the sampling points of the waveform at time n. Here, the threshold is set to 0.05. The first 25 and the last 38 energy operators exceeding the threshold, a total of 64 time points, are truncated and regarded as candidate peak potentials.

[0021] In step (2), when aligning candidate peak potentials according to the peak position, the peak maximum value position is first upsampled and interpolated, and after realignment, the waveform is downsampled to the original length.

[0022] In step (4), the Gaussian receptive field encoding method is as follows:

[0023]

[0024] Where μ is the center position of the receptive field neuron, θ is the width of the receptive field neuron, and S t For a signal sequence at time t, J (t,m,n) The pulse firing of the sensory layer neuron (m,n) at time t. The Poisson process for Gaussian receptive fields.

[0025] In step (4), the winner-takes-all mechanism is as follows: when a neuron is activated, other neurons are inhibited and not updated; only the weight of the synapse connecting the activated neuron and the corresponding sensory layer neuron is updated.

[0026] When updating the synapses connecting activated neurons to their corresponding sensory layer neurons, the neuron selection method is as follows:

[0027]

[0028] in, For the selected cognitive layer neuron to perform the update, z(∈,t) is the voltage value of the cognitive layer neuron at time t;

[0029] The weights of the synapses connecting the two layers are updated as follows:

[0030]

[0031] in, For the updated synaptic weights at time t+1, ω t For the synaptic weights at time t before the update, τ + Let τ be the post-burst constant. - is the pre-burst emission constant.

[0032] In step (4), the attention neurons respond to the prior knowledge of the waveform of the input candidate spike potential and modify the waveform of the input candidate spike potential. When a candidate spike potential waveform is input into the network, waveform masking is performed based on the prior knowledge of the waveform. The generation method is as follows:

[0033]

[0034] Where S t Given a signal sequence at time t, G t For waveform modification and masking; waveform modification and masking G t The generation method is as follows:

[0035]

[0036] Among them, l AP t is the action peak potential width. A1 and t A2 Let l represent the time points of the pre-hyperpolarization peak and the post-hyperpolarization peak, respectively. TP This represents the duration from the trough to the peak of the high-potential wave.

[0037] In step (4), when the strategy of adding nodes in the perception layer is implemented, if the difference between the waveform after the input spike potential is covered and the waveform stored in the network is less than the similarity threshold Th simIf a new node is added to the perception layer, the method for updating and comparing nodes in the perception layer is as follows:

[0038]

[0039] in, For the masked waveform, ω ∈ Φ represents the connection weights between the selected node ∈ in the perception layer and the previous layer, where Φ is a uniform matrix with the same size as the weights.

[0040] In step (4), when implementing the perceptual layer node reduction and merging strategy, if the waveform difference stored in the network is less than the similarity threshold Th sim Then, after merging these two nodes, the comparison method for updating the perception layer nodes is as follows:

[0041]

[0042] Where, ω i With ω j These represent the connection weights corresponding to different nodes in the perception layer.

[0043] In step (4), when adjusting the policy of the perception layer nodes, the similarity threshold Th in the policy is... sim The updates are as follows:

[0044]

[0045] Where α is the scaling control coefficient, β is the waveform number control coefficient, and K is the input peak potential waveform cycle;

[0046] Output threshold Th in the strategy output The update is as follows:

[0047]

[0048] in, z is the threshold for the next round of pruning strategy after the update. (t,ζ) Let ζ be the voltage value of the sensing layer node at time t, β be the waveform number control coefficient, and K be the input peak potential waveform cycle.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. This invention, through a network addition and deletion strategy, has the effect of adaptive node scale and can achieve the minimum hardware overhead under the same computing power.

[0051] 2. By integrating prior neural knowledge, this invention has a strong recognition versatility and can adapt to EEG signal acquisition scenarios with more signal channels.

[0052] 3. By deploying on a biomimetic neural chip, this invention achieves ultra-low power consumption and hardware overhead, and is expected to realize online closed-loop brain-computer interface spike potential classification. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the paradigm of a neuromorphic computational neuron classification method with adaptive node size according to the present invention.

[0054] Figure 2 This is a diagram showing the effect of the method of the present invention on a simulation dataset. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0056] This example uses a simulated dataset. Two different peak potentials are taken from a real peak potential library and emitted at different frequencies. Mechanical noise and far-end peak potential noise are added to simulate the real dataset.

[0057] Although the network can automatically learn the occurrence and transition of spike potentials, some hyperparameters need to be set to adjust the network structure strategy before operation. A reasonable set of parameters allows the entire network to quickly add or remove nodes from different waveform inputs without over-clustering or misclassification. Therefore, the following parameters were chosen: α: scaling control coefficient 0.04; β: waveform number control coefficient 10; I max : Feeling the upper realm of the wild 200; I min : Lower bound of receptive field -200; γ: Shape factor 2 of field neuron; d r Average distance between adjacent receptive fields: 13; τ + Short-term plasticity time presynaptic constant: 0.2; τ - Short-term plasticity time postsynaptic constant: 0.1.

[0058] like Figure 1 As shown, a neuromorphic computational neuron classification method with adaptive node size includes the following steps:

[0059] 1. EEG signal preprocessing: The raw EEG signal is acquired and bandpass filtered to remove low-frequency local field potentials. A third-order Butterworth filter is used here with a bandpass frequency of [300Hz, 3000Hz]. Artifacts are then removed. The signal is finally normalized to the range [-1, 1].

[0060] 2. Candidate Peak Potential Detection: Candidate peak potentials are detected on each signal channel. This invention uses a nonlinear energy operator (NEO) ψ to detect the energy intensity at each location of the discrete signal. The formula is as follows:

[0061] ψ[x(n)]=x 2 (n)-x(n+1)·x(n-1)

[0062] Here, x(n) represents the sampling points of the waveform at time n. The threshold is set to 0.05, and the first 25 and last 38 energy points exceeding the threshold, totaling 64 time points, are truncated and considered as candidate peak potentials.

[0063] 3. Peak Alignment: Peak alignment aligns each peak potential with its point of maximum amplitude (in some cases, the maximum peak potential may be the minimum waveform value). Since the peak value of a wave only reaches its peak for a very short time and typically lies between signal sampling points, the peak value of the waveform cannot be accurately measured. To avoid peak potential misalignment due to low sampling, the peak potential maximum position is determined by interpolating the waveform using cubic splines. After realignment, for ease of subsequent calculations, a 256-bit length signal is truncated with the peak position set to 100, and all points are scaled by a factor of 100 as network input.

[0064] 4. Prior Knowledge Modification: The spike potential waveforms at different time points carry unique physiological significance, and certain characteristic points often exhibit statistical regularities within the waveform. Introducing the pre- and post-hyperpolarization peak times and action potential width as key features for assessing neural activity, this modifies the input spike potential waveform, thus masking G. t The formula is:

[0065]

[0066] Among them l AP t represents the action peak potential width, and t represents the time length of the waveform zero-crossing value between the preceding and following hyperpolarization peaks. A1 and t A2 Let l represent the time points of the pre-hyperpolarization peak and the post-hyperpolarization peak, respectively. TP This represents the time from the trough to the peak of the potential wave.

[0067] 5. Input Signal Field Encoding: Neural signals received within a sliding time window prior to the current time point are encoded into pulse signals and passed to subsequent spiking neural networks. This network layer is called the "receptive layer," which maps continuous input signals to a sequence of pulses using Gaussian receptive fields. This specific neural encoding technique is an extension of a sorting encoder, allowing vectors of real-valued elements to be mapped to a series of pulse sequences. The receptive field allows for the encoding of continuous values ​​using an ensemble of neurons with overlapping sensitivity profiles.

[0068] The Gaussian receptive field encoding method is as follows:

[0069]

[0070] Where μ is the center position of the receptive field neuron, and after normalizing the input, the upper and lower bounds of the receptive field are [-200, 200], so the range of m is [1, 22]; θ is the width of the receptive field neuron, set to 10; S t The signal sequence at time t has a length of 64; J (t,m,n) The pulse firing of the sensory layer neuron (m,n) at time t. The Poisson process for Gaussian receptive fields.

[0071] 6. Spiking Neural Network Classification: This class consists of two layers. The first layer is the perceptual layer, with a size of 22×256 neurons. The second layer is the cognitive layer, which is fully connected to the neurons in the first layer and dynamically updated. Here, the size of the cognitive layer is initialized to 0. The cognitive layer follows a winner-take-all (WTA) mechanism: when one neuron is activated, other neurons are inhibited and not updated.

[0072] Neuron ∈ is an integrally fired (IF) neuron. The selection of neurons in the cognitive layer based on the winner-takes-all mechanism is:

[0073]

[0074] here It is the selected neuron that performs the update.

[0075] Initially, all weights are initialized to 0. Each neuron in the cognitive layer is fully connected to the perceptual layer. Each time an output neuron fires beyond a threshold, the Hebbian learning rule is applied to its incoming synapse. Using this rule, the synaptic weights connecting the two layers are... Using the constant value τ + or τ - renew:

[0076]

[0077] Where, τ +=0.2, τ - =0.1.

[0078] 7. Node addition / deletion strategy update: such as Figure 1 As shown by the addition and pruning / merging strategies, when the perceptron layer node addition strategy is implemented, if the difference between the waveform after input spike potential masking and the waveform stored in the network is less than the similarity threshold Th... sim If a new node is added to the perception layer, then when implementing the perception layer node reduction and merging strategy, if the waveform difference stored in the network is less than the similarity threshold Th... sim Then, the two nodes will be merged.

[0079] When adjusting the policy of the perception layer nodes, the similarity threshold Th in the policy is... sim The updates are as follows:

[0080]

[0081] Where α is the scaling control coefficient of 0.04; β is the waveform number control coefficient of 10; and K is the input peak potential waveform cycle.

[0082] Output threshold Th in the strategy output The update is as follows:

[0083]

[0084] in, z is the threshold for the next round of pruning strategy after the update. (t,ζ) Let t be the voltage value of the sensing layer node ζ, β be the waveform number control coefficient of 10, and K be the input peak potential waveform round.

[0085] 8. Comparison Results: Compare the obtained timestamp sequence with the real labels to verify the performance of the method.

[0086] To demonstrate that this method can achieve a minimal computing hardware size while completing the computational task, we examined the changes in weights and nodes within the network. For example... Figure 2 In this simulation dataset, blue spikes were emitted continuously throughout the entire process, while orange spikes were emitted only for a limited time during data acquisition. The changes in connection weights corresponding to nodes demonstrate that this method can automatically add nodes as spikes appear; furthermore, the SHAP values ​​of the spike waveforms conform to prior neuroscience knowledge, indicating that this method can track changes in neurons and alter network size.

[0087] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A neuromorphic computational neuron classification method with adaptive node size, characterized in that, Includes the following steps: (1) Obtain the original multi-channel neural signals, remove low-frequency local field potentials by bandpass filtering, and perform whitening preprocessing and artifact removal on the neural signals of each channel; (2) Detect and align candidate peak potentials on each signal channel. Specifically, use a nonlinear energy operator to calculate the energy intensity at each position of the discrete signal. A time window that exceeds the threshold is determined to be a candidate peak potential. Then, align the candidate peak potentials according to the peak position. (3) Construct a spiking neural network framework, which includes a two-layer spiking neural network and an attention neuron node; wherein, the first layer of the spiking neural network is a perceptual layer, the second layer is a cognitive layer, and the neurons between the two layers are connected in a fully connected manner and the connection synapses are dynamically updated; the attention neuron controls all neurons in the network through unidirectional connection. (4) The aligned candidate spike potentials are input into the spiking neural network framework. The perceptual layer of the spiking neural network performs pulse encoding on the candidate spike potentials and maps the discrete signals of each time point of the candidate spike potentials to a set of pulse sequences in a Gaussian receptive field encoding manner. Neurons in the cognitive layer respond to different pulse sequence inputs and update the synapses connecting the activated neurons with the corresponding neurons in the sensory layer through a winner-takes-all mechanism; when the accumulated voltage of the cognitive layer neurons exceeds the voltage threshold, the pulse sequence is output as a timestamp sequence in response to different cellular action potentials; The attention neuron node responds to the waveform prior knowledge of the input candidate spike potential, modifies the waveform of the input candidate spike potential, and controls the threshold change of the addition, deletion and merging strategy of the perception layer node; Specifically, the attention neurons respond to prior knowledge of the waveform of the input candidate spike potential, modifying the waveform of the input candidate spike potential, and performing waveform masking. The generation method is as follows: in In time The signal sequence, Waveform modification and masking; Waveform modification and masking The generation method is as follows: in, The width of the action peak potential. and These represent the time points of the pre-hyperpolarization peak and the post-hyperpolarization peak, respectively. The duration from the trough to the peak of the high-potential wave; When implementing the strategy of adding nodes in the perception layer, if the difference between the waveform after the input spike potential is masked and the waveform stored in the network is less than the similarity threshold... If a new node is added to the perception layer, the method for updating and comparing nodes in the perception layer is as follows: in, The waveform after masking. Select a node on the perception layer Connection weights with the previous layer, It is a completely uniform matrix with the same size and weights. When implementing the perceptron layer node reduction and merging strategy, if the waveform differences stored in the network are less than the similarity threshold... Then, these two nodes are merged. The comparison method for updating the perception layer nodes is as follows: in, and These are the connection weights corresponding to different nodes in the perception layer; (5) For the original neural signal corresponding to the timestamp sequence, the cognitive layer nodes dynamically update the threshold to divide the spike potential and noise according to the prior knowledge of the pulse waveform. Each channel reconstructs the waveform emitted from different cells according to the timestamp sequence output by the spiking neural network.

2. The neuromorphic computational neuron classification method with node-scale adaptive computation according to claim 1, characterized in that, In step (1), a third-order Butterworth filter is used for bandpass filtering, with a bandpass frequency of 300~3000Hz.

3. The neuromorphic computational neuron classification method with node-scale adaptive computation according to claim 1, characterized in that, In step (2), the energy intensity at each location of the discrete signal is calculated using a nonlinear energy operator, as shown in the formula: in, It is the first Sampling points of a time waveform.

4. The neuromorphic computational neuron classification method with node-scale adaptive computation according to claim 1, characterized in that, In step (2), when aligning candidate spike potentials according to the spike position, the position of the maximum spike potential is first upsampled and interpolated, and after re-alignment, the waveform is downsampled to the original length.

5. The neuromorphic computational neuron classification method with node-scale adaptive computation according to claim 1, characterized in that, In step (4), the Gaussian receptive field encoding method is as follows: in, For the central location of the receptive field neuron, To the width of the receptive field neurons, In time The signal sequence, For the neurons of the sensory layer In time Pulse emission, The Poisson process for Gaussian receptive fields.

6. The neuromorphic computational neuron classification method with node-scale adaptive computation according to claim 1, characterized in that, In step (4), the winner-takes-all mechanism is as follows: when a neuron is activated, other neurons are inhibited and not updated; only the weights of the synapses connecting the activated neuron and the corresponding sensory layer neuron are updated. When updating the synapses connecting activated neurons to their corresponding sensory layer neurons, the neuron selection method is as follows: in, For the selected cognitive layer neurons to perform the update, For cognitive layer neurons in Voltage value over time; The weights of the synapses connecting the two layers are updated as follows: in, For the updated Temporal synaptic weights Before the update Temporal synaptic weights For the post-burst emission constant, is the pre-burst emission constant.

7. The neuromorphic computational neuron classification method with node-scale adaptive computation according to claim 1, characterized in that, In step (4), when adjusting the strategy for the perception layer nodes, the similarity threshold in the strategy is... The updates are as follows: in, This is the scaling control factor. This is the waveform number control coefficient. The input peak potential waveform cycle; Output threshold in the strategy The update is as follows: in, This serves as the threshold for the next round of pruning strategies after the update. for Time-aware layer nodes voltage value, This is the waveform number control coefficient. This refers to the number of cycles of the input spike potential waveform.

Images