A method and system for supporting learning of a brain-like simulation model

By refining the management of neuromorphic simulation model parameters and designing an approximate function library, simulating gradient calculations for non-differentiable parameters, and combining this with the BP algorithm, the limitations of the STDP learning rule in neuromorphic simulation models are overcome, achieving efficient learning of complex tasks.

CN122072756APending Publication Date: 2026-05-22CHINA NANHU ACAD OF ELECTRONICS & INFORMATION TECH
0 Cites 0 Cited by

Patent Information

Application Number
CN202411678324.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-22
Publication Date
2026-05-22

Smart Images

  • Figure CN122072756A_ABST
    Figure CN122072756A_ABST
Patent Text Reader

Abstract

The application discloses a method and system for supporting brain-like simulation model learning, and the method comprises the following steps: managing all parameters in the brain-like simulation model, and classifying the parameters into fixed parameters, variable parameters, non-differentiable parameters and learnable parameters according to calculation characteristics; designing an approximate function operator library for the non-differentiable parameters, and integrating the library into the system; designing usage rules of the approximate functions, so that the original calculation function of the non-differentiable parameters is used in the forward calculation process of the model, and the approximate function of the non-differentiable parameters is called to perform differential calculation in the gradient solving stage of the back propagation; designing an approximate function calling interface of the brain-like simulation model, replacing the non-differentiable calculation part in the model; and constructing a learnable brain-like simulation model network to perform task learning and training. Through the fine parameter management and the approximate function simulation gradient calculation, the application realizes efficient learning and training of the brain-like simulation model, and is successfully applied to complex classification tasks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of brain-like simulation technology, specifically relating to a method and system for supporting the learning of brain-like simulation models. Background Technology

[0002] With the rapid development of artificial intelligence technology, brain-like simulation models, as a method of simulating the human brain's neural network, have demonstrated their unique advantages and enormous potential in multiple fields. These models mimic the structure and function of biological neural networks, utilizing the complex computations of neurons and their synaptic connections to simulate the way the brain processes information, aiming to improve computational efficiency and learning capabilities.

[0003] In brain-inspired simulation models, Spike-Timing-Dependent Plasticity (STDP) is a common learning rule that adjusts synaptic weights based on the relative spike firing times between neurons, thereby simulating the synaptic plasticity of biological neural networks. While STDP has advantages in simulating biological neural networks, as an unsupervised learning mechanism, it lacks the ability to utilize task-related feedback information to optimize network performance. STDP primarily focuses on local synaptic changes, has limited ability to learn global structure, and is weak in learning multi-level, multi-scale features independent of time sequence, which limits its application in complex tasks such as multi-object detection and classification.

[0004] On the other hand, the backpropagation (BP) algorithm, as a core algorithm in traditional neural networks, can handle nonlinear relationships and approximate arbitrarily complex function mappings. The BP algorithm uses gradient information to guide the search process in the parameter space, gradually approaching the minimum value of the loss function by updating parameters along the negative gradient direction, thus solving the optimization problem of complex task classification. However, due to the existence of discontinuous and non-differentiable parameters in neuromorphic simulation models, such as the spike firing function of the Leaky Integrate-and-Fire (LIF) neuron model, the BP algorithm cannot be directly applied for gradient calculation. This poses a significant challenge to the application of this algorithm in neuromorphic simulation models.

[0005] Therefore, how to maintain the biological characteristics of brain-like simulation models while integrating the efficiency of the backpropagation (BP) algorithm has become a key technical challenge in current research. This invention addresses this challenge by proposing an innovative method that achieves an efficient combination of brain-like simulation models and the BP algorithm through refined management of the parameters of the brain-like simulation model and the use of approximate functions to simulate gradient calculations for non-differentiable parameters. This method provides a novel technical solution for the learning and training of brain-like simulation models, and is expected to overcome the limitations of existing technologies and promote the development of brain-like simulation technology. Summary of the Invention

[0006] One of the objectives of this invention is to provide a method for supporting the learning of brain-like simulation models. By finely managing the model parameters and using approximate functions to simulate the gradient of non-differentiable parameters, the method achieves efficient integration of brain-like simulation models and the BP algorithm, thereby improving the performance and applicability of brain-like simulation models in handling complex classification tasks.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0009] A method for supporting learning of a brain-like simulation model, the method comprising:

[0010] All parameters in the brain-inspired simulation model are managed and classified according to their computational characteristics into fixed parameters, variable parameters, non-differentiable parameters, and learnable parameters.

[0011] A library of approximate function operators for non-differentiable parameters was designed and integrated into a system that supports learning of neuromorphic simulation models.

[0012] The design rules for using approximation functions are such that the original computation function with non-differentiable parameters is used in the forward computation of the model, and the approximation function with non-differentiable parameters is called for differential calculation in the backpropagation gradient solving stage.

[0013] Design an approximate function call interface for a brain-inspired simulation model to replace the non-differentiable computational parts in the model;

[0014] We construct a learnable brain-like simulation model network for task learning and training.

[0015] Preferably, the brain-like simulation model includes a biological neuron model and a synapse model, and manages all parameters in the brain-like simulation model, including initializing and classifying each parameter in the brain-like simulation model. According to the computational characteristics of the parameters in the brain-like simulation model, the parameters are stored in four different dictionaries: fixed parameters, variable parameters, non-differentiable parameters, and learnable parameters. Each parameter has a unique identifier and an initial value.

[0016] Preferably, the fixed parameters include parameters such as the connection structure and delay of the synaptic model that remain unchanged during the calculation process; the variable parameters include parameters such as the pulses and voltages of the neuron model that are iteratively updated with simulation time during the forward calculation process; the non-differentiable parameters are binary and discontinuous neuron model pulse parameters; and the learnable parameters are the weight parameters of the synaptic model, which are used in the brain-like simulation model to simulate the connection strength between neurons and are key factors affecting the information transmission of the neural network and the overall performance of the model.

[0017] Preferably, the library of approximation function operators for non-differentiable parameters includes a variety of different approximation functions, including but not limited to Gaussian functions, logistic functions, and polynomial functions, which can approximate the behavior of the original non-differentiable parameters in a differentiable manner.

[0018] Preferably, the approximation function call interface can handle different types of non-differentiable parameter computation functions and can return outputs that conform to the approximation function usage rules according to the current computation stage (forward or backward). The interface is designed to be extensible, allowing different approximation functions to be used by modifying the approximation function name parameter without modifying the interface itself.

[0019] Preferably, the learning and training of the learnable brain-like simulation model network specifically includes the following steps:

[0020] Construct a learnable brain-like simulation model network and load the task dataset;

[0021] The network receives data input and performs forward computation, updating the parameters in the variable parameter dictionary during the process.

[0022] Design a loss function to calculate the difference between the forward output of the model and the true value;

[0023] The gradient is calculated using the backpropagation method based on the loss function value.

[0024] The parameters in the learnable parameter dictionary are updated based on the calculated gradient.

[0025] Preferably, the learnable brain-like simulation model network is a network composed of neuron models and synapse models that can perform simulation and learning, wherein the neuron models and synapse models are interconnected according to specific connection rules, and the input and output neurons in the network are specified.

[0026] Preferably, the difference between the forward output of the brain-like simulation and the true value can be calculated using a loss function designed according to the task type, including but not limited to using mean squared error loss, cross-entropy loss or other appropriate loss functions to quantify the difference between the simulation output and the true value; the forward output can be designed according to the model structure and task type, including but not limited to the spiking rate of all or specified neurons, the voltage value at the last moment, and the spiking calculated value according to a given calculation formula.

[0027] Preferably, the gradient descent optimization method includes, but is not limited to, stochastic gradient descent, batch gradient descent, or adaptive gradient algorithm.

[0028] The present invention provides a method for supporting learning of neuromorphic simulation models, which has the following advantages compared with the prior art:

[0029] 1) This invention integrates a pre-designed approximation function library into a system that supports learning of neuromorphic simulation models, and formulates the usage rules and calling interfaces of the approximation functions. This enables the use of the original functions in the forward computation stage of the model and the use of the corresponding approximation functions to calculate the gradient in the backpropagation stage. Without interfering with the normal simulation function of the neuromorphic simulation model, the model can use the backpropagation algorithm to learn, thereby enhancing the model's ability to handle complex tasks.

[0030] 2) This invention employs a refined parameter management strategy to meticulously classify the parameters in the brain-like simulation model into four types: fixed parameters, variable parameters, non-differentiable parameters, and learnable parameters. This classification method enables more efficient execution of corresponding computational operations for different types of parameters, allowing each type of parameter to fully play its role in model computation, thereby significantly improving the computational efficiency of the model and enhancing its adaptability in learning and simulation tasks.

[0031] The second objective of this invention is to provide a brain-like simulation system that supports learning. By integrating methods that support the learning of brain-like simulation models, the system enables efficient learning and training of brain-like simulation models, thereby enabling the application of brain-like simulation models in classification tasks.

[0032] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0033] A brain-like simulation system supporting learning includes a processor unit, a memory unit, simulation software, and an input / output unit. The processor unit executes computation and learning algorithms for the brain-like simulation model; the memory unit stores a parameter dictionary and an approximation function operator library from the brain-like simulation method; the simulation software integrates the brain-like simulation method, including modules for parameter management, gradient calculation, loss function definition, backpropagation, and gradient descent optimization, as well as a computation module for the brain-like simulation model; and the input / output unit receives external data and outputs simulation results. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method for supporting brain-like simulation model learning in Embodiment 1 of the present invention;

[0035] Figure 2 This is a flowchart of the brain-like simulation model learning and training method in Embodiment 2 of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0038] This embodiment aims to achieve the following in a neuromorphic simulation model: First, it differentiates model parameters through parameter management. Second, it designs approximation functions for non-differentiable parameters and establishes rules for their use. The original functions are used in forward computation, while the approximation functions are used in backward computation. Then, it designs an interface for calling these approximation functions to replace the non-differentiable computational parts. Finally, it constructs a learnable neuromorphic simulation model network for task training. Through this design, this embodiment achieves the learning task of a neuromorphic simulation model without interfering with its normal simulation functionality.

[0039] Example 1

[0040] like Figure 1 As shown in the figure, a method for supporting brain-like simulation model learning in this embodiment includes the following steps:

[0041] S1 manages all parameters in the brain-like simulation model and classifies these parameters according to their computational characteristics into fixed parameters, variable parameters, non-differentiable parameters, and learnable parameters.

[0042] In this embodiment, the brain-like simulation model includes a biological neuron model and a synapse model. It manages all parameters in the brain-like simulation model, including initializing and classifying each parameter in the brain-like simulation model. Based on the computational characteristics of the parameters in the brain-like simulation model, the parameters are stored in four different dictionaries: fixed parameters, variable parameters, non-differentiable parameters, and learnable parameters. Each parameter has a unique identifier and an initial value.

[0043] In this embodiment, the fixed parameters include parameters such as the connection structure and delay of the synaptic model that remain unchanged during the calculation process; the variable parameters include parameters such as the pulses and voltages of the neuron model that are iteratively updated with simulation time during the forward calculation process; the non-differentiable parameters are the binary and discontinuous pulse parameters of the neuron model; and the learnable parameters are the weight parameters of the synaptic model, which are used in the brain-like simulation model to simulate the connection strength between neurons and are key factors affecting the information transmission of the neural network and the overall performance of the model.

[0044] S2. Design a library of approximate function operators for non-differentiable parameters and integrate the library into a system that supports learning of brain-like simulation models.

[0045] In this embodiment, the non-differentiable parameter in the parameter dictionary of the brain-like simulation model is the pulse parameter. The pulse firing function of a neuron is usually a step function. This function jumps from 0 to 1 when the membrane potential exceeds the firing threshold, indicating that the neuron has fired a pulse. This function is mathematically non-differentiable because it has a discontinuity at the threshold point.

[0046] To fit the approximate calculation of non-differentiable parameters such as impulses, the approximation function operator library designed in this embodiment contains a variety of different approximation functions, including but not limited to Gaussian functions, logistic functions, and polynomial functions. These functions can approximate the behavior of the original non-differentiable parameters in a differentiable manner.

[0047] S3. Design the rules for using approximation functions so that the original calculation function with non-differentiable parameters is used in the forward calculation of the model, and the approximation function with non-differentiable parameters is called for differential calculation in the backpropagation gradient solving stage.

[0048] To ensure the continuity and differentiability of the model, this embodiment designs rules for the use of approximation functions, enabling the model to learn using the backpropagation algorithm without interfering with the normal simulation function of the brain-like simulation model.

[0049] S4. Design an approximate function call interface for the brain-like simulation model to replace the non-differentiable computational parts in the model.

[0050] The approximation function call interface in this embodiment can handle different types of non-differentiable parameter computation functions and can return outputs that conform to the approximation function usage rules according to the current computation stage (forward or backward). The interface is designed to be extensible, allowing different approximation functions to be used by modifying the approximation function name parameter without modifying the interface itself.

[0051] S5. Construct a learnable brain-like simulation model network for task learning and training.

[0052] This embodiment describes the learning and training of a learnable brain-like simulation model network, such as... Figure 2 As shown, the specific steps include:

[0053] A1. Construct a learnable brain-like simulation model network and load the task dataset;

[0054] A2. The network receives data input and performs forward calculations, updating the parameters in the variable parameter dictionary during the process;

[0055] A3. Design a loss function to calculate the difference between the forward output of the model and the true value;

[0056] A4. Calculate the gradient using the backpropagation method based on the loss function value;

[0057] A5. Update the parameters in the learnable parameter dictionary based on the calculated gradient.

[0058] The learnable brain-like simulation model network in this embodiment is a network composed of neuron models and synapse models that can perform simulation and learning. The neuron models and synapse models are interconnected according to specific connection rules, and the input and output neurons in the network are specified.

[0059] In this embodiment, forward computation refers to the computation process of the neuromorphic simulation model within a specific simulation time. During this process, the model determines the number of simulation steps to be executed based on a preset simulation duration and step size. Each simulation step involves updating the variable parameters in the model, and these updates reflect the model's computational state at each step.

[0060] In this embodiment, the difference between the forward output of the brain-like simulation and the true value can be calculated using a loss function designed according to the task type. This includes, but is not limited to, using mean squared error loss, cross-entropy loss, or other appropriate loss functions to quantify the difference between the simulation output and the true value. The forward output can be designed according to the model structure and task type, including, but not limited to, the firing rate of all or specified neurons, the voltage value at the last moment, and the pulse calculation value according to a given calculation formula.

[0061] The gradient descent optimization methods in this embodiment include, but are not limited to, stochastic gradient descent, batch gradient descent, or adaptive gradient algorithms.

[0062] This scheme supports the learnable function of brain-like simulation models by managing the parameters of the brain-like simulation model and designing approximate functions for non-differentiable parameters. The BP algorithm is used to complete the learnable training on the classification task. The following experiments further illustrate the feasibility of the method.

[0063] Experiment Description: This experiment, based on the JAX library, designed and implemented a brain-like simulation system that incorporates refined parameter management and approximate function substitution methods. Different brain-like simulation models were constructed and trained for the MNIST handwritten digit classification task to realize the application of brain-like simulation models in classification tasks.

[0064] The brain-like simulation model comprises two neuron clusters and three sets of experiments. The first set consists of two fully connected Exponential Leaky Integrate-and-Fire (ExpLIF) neuron models, the second set consists of two fully connected Generalized Leaky Integrate-and-Fire with Post-Synaptic Currents (GLIFPSC) neuron models, and the third set consists of two fully connected Izhikevich neuron models. In each set, the first neuron model has 784 neurons, and the second neuron model has 10 neurons. The first neuron model is specified to receive the processed dataset as input, and the second neuron model is specified to output the number of pulses fired by each neuron within 10 ms. After forward computation, the cross-entropy loss function is used to calculate the error between the output value and the true value in the dataset. Then, an adaptive gradient algorithm is used as the gradient descent optimization method to update the synaptic weight parameters in the model. After 10 epochs of training iterations, the classification accuracy is calculated. The training and testing results are shown in Table 1.

[0065] Table 1. Training and testing results of different neuron models

[0066] Model type Training accuracy Test accuracy ExpLIF 0.914 0.906 GLIFPSC 0.918 0.920 Izhikevich 0.907 0.911

[0067] The neuromorphic simulation model in the table has not undergone parameter tuning, but the experimental results show that after 10 training cycles, it has achieved high classification accuracy in both training and testing. The experiments demonstrate that the neuromorphic simulation model, through parameter management and the design of approximate functions for non-differentiable parameters, can be learned and trained using the BP algorithm, enabling its application in handwritten digit classification tasks.

[0068] Example 2

[0069] A system for supporting learning of a neuromorphic simulation model is provided, comprising a processor unit, a memory unit, simulation software, and an input / output unit. The processor unit executes computation and learning algorithms for the neuromorphic simulation model; the memory unit stores the parameter dictionary and approximation function operator library of the neuromorphic simulation methods described in claims 1 to 9; the simulation software integrates the neuromorphic simulation methods described in claims 1 to 9, including modules for parameter management, gradient calculation, loss function definition, backpropagation, and gradient descent optimization, as well as a computation module for the neuromorphic simulation model; and the input / output unit receives external data and outputs simulation results.

[0070] For specific limitations on systems that support learning from neuromorphic simulation models, please refer to the limitations on methods that support learning from neuromorphic simulation models mentioned above, which will not be repeated here.

[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0072] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for supporting learning in neuromorphic simulation models, characterized in that, The method for supporting learning of neuromorphic simulation models includes: All parameters in the brain-inspired simulation model are managed and classified according to their computational characteristics into fixed parameters, variable parameters, non-differentiable parameters, and learnable parameters. A library of approximate function operators for non-differentiable parameters was designed and integrated into a system that supports learning of neuromorphic simulation models. The design rules for using approximation functions are such that the original computation function with non-differentiable parameters is used in the forward computation of the model, and the approximation function with non-differentiable parameters is called for differential calculation in the backpropagation gradient solving stage. Design an approximate function call interface for a brain-inspired simulation model to replace the non-differentiable computational parts in the model; We construct a learnable brain-like simulation model network for task learning and training.

2. The method for supporting brain-like simulation model learning as described in claim 1, characterized in that, The brain-like simulation model includes a biological neuron model and a synapse model. It manages all parameters in the brain-like simulation model, including initializing and classifying each parameter in the brain-like simulation model. Based on the computational characteristics of the parameters in the brain-like simulation model, the parameters are stored in four different dictionaries: fixed parameters, variable parameters, non-differentiable parameters, and learnable parameters. Each parameter has a unique identifier and an initial value.

3. The method for supporting learning of neuromorphic simulation models as described in claim 2, characterized in that, The fixed parameters include parameters such as the connection structure and delay of the synaptic model that remain unchanged during the calculation; the variable parameters include parameters such as the impulses and voltages of the neuron model that are iteratively updated with simulation time during the forward calculation. Non-differentiable parameters are binary and discontinuous impulse parameters of neuron models; Learnable parameters are the weight parameters of the synaptic model. These parameters are used in brain-inspired simulation models to simulate the connection strength between neurons and are key factors affecting the information transmission of neural networks and the overall performance of the model.

4. The method for supporting brain-like simulation model learning as described in claim 1, characterized in that, The aforementioned library of approximation function operators for non-differentiable parameters contains a variety of different approximation functions, including but not limited to Gaussian functions, logistic functions, and polynomial functions, which can approximate the behavior of the original non-differentiable parameters in a differentiable manner.

5. The method for supporting learning of neuromorphic simulation models as described in claim 1, characterized in that, The approximation function call interface can handle different types of non-differentiable parameter computation functions and can return outputs that conform to the approximation function usage rules according to the current computation stage (forward or backward). The interface is designed to be extensible, allowing different approximation functions to be used by modifying the approximation function name parameter without modifying the interface itself.

6. The method for supporting brain-like simulation model learning as described in claim 1, characterized in that, The learning and training of the aforementioned learnable brain-like simulation model network specifically includes the following steps: Construct a learnable brain-like simulation model network and load the task dataset; The network receives data input and performs forward computation, updating the parameters in the variable parameter dictionary during the process. Design a loss function to calculate the difference between the forward output of the model and the true value; The gradient is calculated using the backpropagation method based on the loss function value. The parameters in the learnable parameter dictionary are updated based on the calculated gradient.

7. The method for supporting brain-like simulation model learning as described in claim 6, characterized in that, The learnable brain-like simulation model network is a network composed of neuron models and synapse models that can perform simulation and learning. The neuron models and synapse models are interconnected according to specific connection rules, and the input and output neurons in the network are specified.

8. The method for supporting brain-like simulation model learning as described in claim 6, characterized in that, The difference between the forward output of the brain-like simulation and the true value can be calculated using a loss function designed according to the task type, including but not limited to using mean squared error loss, cross-entropy loss or other appropriate loss functions to quantify the difference between the simulation output and the true value; the forward output can be designed according to the model structure and task type, including but not limited to the spiking rate of all or specified neurons, the voltage value at the last moment, and the spiking calculated value according to a given calculation formula.

9. The method for supporting brain-like simulation model learning as described in claim 6, characterized in that, The gradient descent optimization methods mentioned include, but are not limited to, stochastic gradient descent, batch gradient descent, or adaptive gradient algorithms.

10. A system for supporting learning of neuromorphic simulation models, comprising a processor unit, a memory unit, simulation software, and an input / output unit, characterized in that, The processor unit is used to execute the computation and learning algorithms of the brain-like simulation model. The memory unit is used to store the parameter dictionary and approximation function operator library of the brain-like simulation method according to claims 1 to 9. The simulation software integrates the brain-like simulation method according to claims 1 to 9, including modules such as parameter management, gradient calculation, loss function definition, backpropagation and gradient descent optimization, as well as the computation module of the brain-like simulation model. The input / output unit is used to receive external data and output simulation results.