A dendritic platform potential mechanism-based pulse neuron model and identification method

Abstract

The application discloses a kind of impulse neuron model and identification method based on dendritic platform potential mechanism, it is applied to artificial intelligence, brain-like computing, impulse neural network, neuro-morphic computing and time series signal processing technical field, for the problems such as insufficient time series information retention, limited local feature expression ability and insufficient classification stability of existing model in impulse time series task;The present application builds soma input channel and dendritic input channel, introduces continuous platform state, platform positive feedback mechanism, slow recovery termination variable and soma coupling term, so that artificial impulse neuron can model local spatiotemporal aggregation input, while solving the problems of limited short-term memory capacity and insufficient nonlinear state expression within neuron, thereby improving the representation ability of impulse neural network for local spatiotemporal aggregation and cross-time dependence in pulse sequence.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, brain-like computing, spiking neural networks, neuromorphic computing and temporal signal processing, and particularly relates to a pulse temporal recognition technology based on a dendritic plateau potential spiking neuron model. Background Technology

[0002] Spiking neural networks (SNNs) transmit information over time via discrete pulses, offering advantages such as event-driven processing, sparse computation, and low-power deployment, making them suitable for temporal signal processing tasks like EEG and speech recognition. Existing pulse time series recognition models typically employ neuron models such as LIF, PLIF, and ALIF. These models primarily describe neuronal dynamics through membrane potential leakage integration, threshold firing, and post-firing reset, offering simple structures and ease of training and deployment. However, when processing pulse time series, traditional neuron models often rely on integrating a single membrane potential state, making it difficult to fully express the nonlinear amplification effect of locally spatiotemporally concentrated inputs. Furthermore, the post-firing reset of neurons limits their ability to preserve long-term temporal information states. Therefore, existing models still suffer from insufficient temporal information preservation, limited local feature representation capabilities, and inadequate classification stability in pulse time series tasks. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a spiking neuron model and identification method based on the dendritic plateau potential mechanism.

[0004] One of the technical solutions adopted in this invention is: a pulse timing recognition model based on dendritic plateau potential mechanism, comprising: an input module, a dendritic local integration module, a bistable plateau dynamics module, a cell body adaptive update module, and an output module;

[0005] The input module is used to expand the input pulse timing data by time steps to obtain a data tensor;

[0006] The dendrite local integration module updates the dendrite local potential based on the input data tensor using the dendrite leakage integral method;

[0007] The bistable platform dynamics module updates the platform state based on the updated dendritic local potential;

[0008] The cell body adaptive update module updates the cell body membrane potential based on the updated platform state;

[0009] The output module generates an output pulse based on the updated cell membrane potential.

[0010] The second technical solution adopted in this invention is: a method for identifying spiking neuron models based on dendritic plateau potential mechanism, comprising:

[0011] S1: Acquire pulse timing data and expand it by time step;

[0012] S2: Construct a spiking neural network based on the spiking neuron model described in any one of claims 1-7;

[0013] S3. Expand the pulse time series data obtained in step S1 into the constructed spiking neural network by time step, and calculate the cross-entropy loss based on the classification results output by the spiking neural network.

[0014] S4: Backpropagation and parameter update are performed based on cross-entropy loss. The trained spiking neural network is saved after training is completed.

[0015] S5: Expand the pulse timing data to be processed by time step and input it into the trained neural network model to obtain the corresponding recognition result.

[0016] The beneficial effects of this invention are as follows: By constructing cell body input pathways and dendritic input pathways, and introducing continuous plateau states, plateau positive feedback mechanisms, slow recovery termination variables, and cell body coupling terms, this invention enables artificial spiking neurons to model local spatiotemporal clustered inputs while addressing the limitations of short-term memory and insufficient expression of nonlinear states within neurons. This improves the ability of spiking neural networks to represent local spatiotemporal clusters and cross-temporal dependencies in spiking sequences. The method of this invention has the following advantages:

[0017] This invention explicitly introduces dendritic input pathways and plateau state variables, enabling spiking neurons to not only integrate based on cell body membrane potential, but also to form nonlinear responses to locally spatiotemporally concentrated pulse inputs.

[0018] This invention forms bistable or quasi-bistable temporal dynamics through positive feedback of the platform state and suppression of slow recovery variables, enabling neurons to exhibit dynamic characteristics of rapid triggering, sustained maintenance and recovery termination, thereby improving the persistence, temporal separability and short-term memory capacity of output pulses.

[0019] This invention significantly improves the performance of SNN in pulse timing recognition tasks: tests on speech datasets show that BDP-SNN achieves classification accuracy of 94.73% and 80.62% on SHD and SSC datasets, respectively. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of the BDP-LIF neuron model of the present invention.

[0021] Figure 2 This is a flowchart illustrating the state update process of the dendritic platform module of the present invention. Detailed Implementation

[0022] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0023] Figure 1 The diagram shows that the input sequence enters the cell body projection pathway and the dendritic projection pathway respectively, and is coupled to the cell body membrane potential by the dendritic plateau state to generate an output pulse.

[0024] Figure 2 The relationships between dendrite-driven, plateau positive feedback, recovery inhibition, sharp gating, plateau state update, and recovery variable accumulation are illustrated.

[0025] The specific steps are as follows:

[0026] 1) Constructing such Figure 1 The dendritic plateau potential spiking neuron model shown is (BDP-LIF).

[0027] In this embodiment, the dendritic plateau potential spiking neurons are configured with both a cell body input pathway and a dendritic input pathway. The cell body input pathway is used to form a conventional membrane potential integral, while the dendritic input pathway is used to capture local spatiotemporally concentrated inputs in the pulse sequence. The cell body input pathway is the backbone LIF integral path of a traditional SNN; the dendritic input pathway adds a local nonlinear slow-state branch to the input at the same layer to simulate the continuous modulation of the cell body membrane potential by the NMDA plateau. Both utilize conventional membrane potential integration.

[0028] Then, update the dendritic local potential according to the dendritic leakage integral method:

[0029]

[0030] in, This represents the local potential of the dendrites at time step t. Indicates the dendrite potential leakage coefficient. This represents the input current entering the dendrite path at time step t, initially set to 0. The dendrite local potential characterizes the local accumulation of the input pulse on the dendrite branches. The subscript 'd' stands for dendrite input path.

[0031] When the dendritic local potential exceeds the dendritic plateau trigger threshold, a dendritic plateau driving term is generated:

[0032]

[0033] in, Represents the dendrite-driven term. This indicates the dendritic plateau trigger threshold. This driving term reflects whether the input pulse at the current time step forms a sufficiently strong aggregation response locally in the dendrites.

[0034] Furthermore, the dendritic driving term, the positive feedback term of the platform state at the previous time step, the suppression term of the recovery variable, and the platform threshold bias are all input into the platform gating function to obtain the platform gating input:

[0035]

[0036] in, Indicates platform gating input, This indicates the platform status at the previous moment. This indicates the variable to be restored from the previous time step. Indicates dendrite-driven weights, Indicates the strength of positive feedback from the platform. Indicates restoration of inhibition strength. This indicates the platform threshold bias.

[0037] The platform gating output is represented as:

[0038]

[0039] in, This indicates the platform's gated output, and k represents the gate slope.

[0040] Update the platform status based on the platform gating output:

[0041]

[0042] in, This indicates the platform status at the current time step. This represents the plateau state retention coefficient. The plateau state describes the activity level of the dendritic plateau potential and is initially set to 0. When the pulse input forms a local cluster within a short time window, the plateau state is triggered; after triggering, the positive feedback term of the plateau maintains it for a period of time, thereby enhancing the expressive power of continuous phoneme segments or short acoustic events.

[0043] To prevent the platform state from remaining unchanged for too long and affecting subsequent input responses, this invention sets a slow recovery termination variable:

[0044]

[0045] in, Indicates the restored variable. Indicates the coefficient for maintaining the recovered variable. This represents the cumulative coefficient of the recovery variable. The recovery variable gradually accumulates as the platform state remains active, and through the recovery inhibition term, it generates negative feedback on the platform gating input, causing the platform state to gradually decay or terminate after a certain period of time.

[0046] Finally, the platform state is coupled to the cell membrane potential update process:

[0047]

[0048] in, Indicates cell membrane potential. Indicates the membrane potential leakage coefficient. This indicates the coupling strength from the plateau state to the cell membrane potential. This indicates the reset item after issuance. This represents the output pulse from the previous time step. The output pulse is generated by the hard threshold firing function.

[0049]

[0050] in, (·) represents the step function. This indicates the cell body emission threshold. When... Greater than When the condition is met, the neuron outputs a 1 pulse; otherwise, it outputs a 0 pulse. The output pulse here is the classification score vector.

[0051] 2) Spiking Neural Network Structure Design (BDP-SNN)

[0052] In this embodiment, the dendritic plateau potential spiking neurons are used to construct a spiking neural network. The network as a whole consists of two fully connected layers and one readout layer, which has a simple structure and facilitates the verification of the effectiveness of the neuron model in time series recognition tasks.

[0053] Instead of generating discrete impulses, the Readout_Layer accumulates or reads out the time-dimensional information from the previous layer's output to obtain the output value corresponding to each label. This output value can be understood as the classification score for different categories.

[0054] 3) Learning framework and loss function design

[0055] This invention constructs a unified end-to-end supervised learning framework to jointly train a two-layer fully connected spiking neural network and a ReadoutLayer. During training, it is necessary to simultaneously update the network connection weights and relevant dynamic parameters in the dendritic plateau potential spiking neurons, such as the dendritic plateau trigger threshold, plateau coupling strength, plateau positive feedback strength, recovery inhibition strength, membrane potential leakage coefficient, and firing threshold.

[0056] The dendritic platform trigger threshold, platform coupling strength, platform positive feedback strength, recovery inhibition strength, membrane potential leakage coefficient, and firing threshold are set as learnable parameters and automatically updated during training through error backpropagation and gradient optimization. Initially, the dendritic platform trigger threshold is set to 0.9, the platform coupling strength to 0.9, the platform positive feedback strength to 2.0, the recovery inhibition strength to 1.5, the membrane potential leakage coefficient to be uniformly and randomly initialized in the range of [exp(-1 / 5), exp(-1 / 30)], and the firing threshold to 1.0.

[0057] Since the impulse firing function H(·) is a step function, its true derivative is almost zero everywhere, making it difficult to train directly using the backpropagation algorithm. Therefore, this invention employs the SLAYER surrogate gradient method during the training phase to address the issue of the impulse firing function's non-differentiability.

[0058] Its derivative is approximated using the SLAYER exponential surrogate gradient function:

[0059]

[0060] in, This represents the difference between the membrane potential and the firing threshold. This represents the difference between the membrane potential at time step t and the firing threshold. Indicates the proxy gradient magnitude coefficient. This represents the surrogate gradient decay coefficient. It should be noted that... This is used only to describe the shape of the SLAYER surrogate gradient function and is different from the leakage coefficient in the membrane potential update equation. In one specific embodiment... Take 0.4, Take 5.0.

[0061] This implementation uses the cross-entropy loss function as the classification training objective. Let the classification score vector output by ReadoutLayer be... If the true class label is y, then the cross-entropy loss can be expressed as:

[0062]

[0063] in, This represents the output score corresponding to the true category. Indicates the first The output score corresponds to each category. During training, the cross-entropy loss is minimized to make the model's category predictions gradually approach the true labels.

[0064] 4) Implementation process as follows Figure 2 As shown, it includes the following steps:

[0065] S1: Acquire pulse timing data. In this embodiment, the pulse timing data processed is specifically voice data.

[0066] S2: Build a spiking neural network structure, construct the SNN network according to the task requirements, and initialize the parameters.

[0067] The pulse sequence is input into the network step by step according to time frames, and the network performs feature extraction and temporal evolution.

[0068] S3: Calculate the classification result and cross-entropy loss.

[0069] S4: Perform backpropagation and parameter updates, and save the final model after training is complete.

[0070] 5) Experimental verification and effect description

[0071] This embodiment further validates the performance on the pulse speech dataset. BDP-LIF achieves excellent performance on both the SHD and SSC datasets. The experimental results are shown in Table 1.

[0072] Table 1. Comparison of the accuracy of the model of this invention with existing models.

[0073]

[0074] Applying the technical solution of this invention to the processing of pulse sequences obtained by pulse encoding of continuous signals is also within the scope of protection of this invention.

[0075] Those skilled in the art will recognize that the embodiments described herein are for the purpose of helping to understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A pulse timing identification model based on dendritic plateau potential mechanism, characterized in that, include: The module includes an input module, a dendrite local integration module, a bistable platform dynamics module, a cell body adaptive update module, and an output module. The input module is used to expand the input pulse timing data by time steps to obtain a data tensor; The dendrite local integration module updates the dendrite local potential based on the input data tensor using the dendrite leakage integral method; The bistable platform dynamics module updates the platform state based on the updated dendritic local potential; The cell body adaptive update module updates the cell body membrane potential based on the updated platform state; The output module generates an output pulse based on the updated cell membrane potential.

2. The pulse timing identification model based on dendritic plateau potential mechanism according to claim 1, characterized in that, The expression for updating the dendritic local potential is: ； in, This represents the local potential of the dendrites at time step t. This represents the local potential of the dendrites at time step t-1. Indicates the dendrite potential leakage coefficient. This represents the input current entering the dendritic pathway at time step t.

3. The pulse timing identification model based on the dendritic plateau potential mechanism according to claim 2, characterized in that, The bistable platform dynamics module includes a dendritic platform driving unit, a platform gating unit, and a platform state update unit; The dendrite platform driving unit outputs a dendrite platform driving term when the local potential of the dendrite exceeds the dendrite platform trigger threshold. The input to the platform gating unit is the output obtained by inputting the dendrite driving term, the positive feedback term of the platform state at the previous time step, the recovery variable suppression term, and the platform threshold bias into the platform gating function. The inputs of the platform state update unit include the output of the platform gating unit and the platform state of the previous time step. The output of the platform state update unit is the platform state of the current time step.

4. The pulse timing identification model based on the dendritic plateau potential mechanism according to claim 3, characterized in that, The platform gating function is represented as follows: ； in, Indicates platform gating input, This indicates the platform state at time t-1. This indicates that the variable is restored at time t-1. Indicates dendrite-driven weights, Indicates the strength of positive feedback from the platform. Indicates restoration of inhibition strength. This indicates the platform threshold bias.

5. The pulse timing identification model based on the dendritic plateau potential mechanism according to claim 4, characterized in that, Let the recovery variable at time t be... , The expression is: ； in, Indicates the coefficient for maintaining the recovered variable. This represents the cumulative coefficient of the restored variable.

6. The pulse timing identification model based on dendritic plateau potential mechanism according to claim 5, characterized in that, The calculation formula for the cell body adaptive update module is: ； in, Indicates cell membrane potential. Indicates the membrane potential leakage coefficient. This indicates the coupling strength from the plateau state to the cell membrane potential. This indicates the reset item after issuance. This indicates the output pulse at the previous moment.

7. The pulse timing identification model based on dendritic plateau potential mechanism according to claim 5, characterized in that, The output pulse is generated by a hard threshold firing function.

8. A method for identifying spiking neuron models based on dendritic plateau potential mechanism, characterized in that, include: S1: Acquire pulse timing data and expand it by time step; S2: Construct a spiking neural network based on the spiking neuron model described in any one of claims 1-7; S3. Expand the pulse time series data obtained in step S1 into the constructed spiking neural network by time step, and calculate the cross-entropy loss based on the classification results output by the spiking neural network. S4: Backpropagation and parameter update are performed based on cross-entropy loss. The trained spiking neural network is saved after training is completed. S5: Expand the pulse timing data to be processed by time step and input it into the trained neural network model to obtain the corresponding recognition result.

9. The method for identifying a spiking neuron model based on a dendritic plateau potential mechanism according to claim 8, characterized in that, During training, the derivative of the hard threshold firing function is approximated by the SLAYER exponential surrogate gradient function.

10. The method for identifying a spiking neuron model based on a dendritic plateau potential mechanism according to claim 9, characterized in that, The spiking neural network comprises two fully connected layers and one readout layer, and uses the spiking neuron model described in any one of claims 1-7 as the neurons in the two fully connected layers.

Images