Low-Complexity Brain Simulation Methods, Devices, Equipment, and Media

Abstract

The present invention relates to the field of computers, and provides a low-complexity brain simulation method, device, equipment and medium. The method includes: splitting the pulse neural network to be modeled into multiple neuron groups; constructing each synapse between any two neuron groups in each neuron group, where each synapse is provided with the same time delay; determining the synaptic variables of each synapse after the time delay; approximating the weight matrix of each synapse to the product of multiple small matrices in a low-rank manner; and performing brain simulation according to each neuron group, each synapse, each synaptic variable and its corresponding multiple small matrices. The low-complexity brain simulation method provided by the embodiments of the present invention significantly reduces the computational complexity of the synapses, thereby reducing the complexity of brain simulation.

Description

Technical Field

[0001] The present invention relates to the field of computers, and in particular, to a low-complexity brain simulation method, device, equipment and medium. Background Art

[0002] In a typical spiking neural network, if N neurons are used, there will be pN 2 synapses in the network, where p is the connection probability between neurons.

[0003] Therefore, based on the current spiking neural network simulation method, the computational complexity and memory requirements of brain dynamics simulation increase in an O(n 2 ) squared manner as the number of neurons grows. The complexity of O(n 2 ) squared makes it difficult to process network simulations of the order of one hundred thousand neurons on a single general-purpose computing device. If simulations of the order of one million neurons need to be processed, a supercomputer cluster needs to be used. Theoretically, if a dynamic model of the number of human brain neurons needs to be simulated, the squared complexity makes it impossible to perform whole-brain simulation even with all the supercomputers in the world combined. Summary of the Invention

[0004] The present invention provides a low-complexity brain simulation method, device, equipment and medium, aiming to reduce the O(n 2 ) complexity of brain simulation to O(n).

[0005] In a first aspect, the present invention provides a low-complexity brain simulation method, including:

[0006] Splitting the spiking neural network to be modeled into multiple neuron groups;

[0007] Constructing each synapse between any two neuron groups in each of the neuron groups, where each synapse is provided with the same time delay;

[0008] Determining the synaptic variables of each synapse after the time delay;

[0009] Low-rank approximating the weight matrix of each synapse as a product of multiple small matrices;

[0010] Performing brain simulation according to each of the neuron groups, each of the synapses, each synaptic variable and its corresponding multiple small matrices.

[0011] In one embodiment, the low-rank approximating the weight matrix of each synapse as a product of multiple small matrices includes:

[0012] If the weight connection information in the weight matrix is heterogeneous and known, each weight matrix is low-rank approximated as the product of a corresponding first matrix and a second matrix by means of matrix dimensionality reduction or matrix factorization methods.

[0013] The low-rank approximation of the weight matrix of each of the synapses as the product of a plurality of small matrices includes:

[0014] If the weight connection information in the weight matrix is heterogeneous and unknown, a third matrix and a fourth matrix are randomly generated, where the statistical properties of the third matrix and the fourth matrix are the same as those of the weight matrix without approximation.

[0015] Each weight matrix is low-rank approximated as the product of a corresponding one of the third matrix and the fourth matrix.

[0016] The number of the synaptic variables is the same as the number of neurons in the presynaptic neuron population, where the presynaptic neuron population is any one of a plurality of neuron populations.

[0017] After determining the synaptic variables of each of the synapses after the time delay, it further includes:

[0018] If the weight connections in each of the weight matrices are homogeneous, the connection weights are set to a scalar.

[0019] The splitting of the spiking neural network to be modeled into a plurality of neuron populations is performed based on a preset rule; the preset rule includes:

[0020] The first rule is that the neurons within each of the presynaptic neuron populations have the same firing pattern;

[0021] Or / and, the second rule is that the neurons within each of the presynaptic neuron populations have the same neurotransmitter receptors;

[0022] Or / and, the third rule is that the neurons within each of the presynaptic neuron populations have the same projection source;

[0023] Or / and, the fourth rule is that the maximum physical distance between the neurons within each of the presynaptic neuron populations does not exceed a preset distance threshold.

[0024] The brain simulation according to each of the neuron populations, each of the synapses, each synaptic variable and its corresponding plurality of small matrices includes:

[0025] At each moment, the state variables of each synapse are updated according to the states of the presynaptic neurons after the delay.

[0026] Multiply the state variable of each synapse by the scalar weight or multiple small matrix weights corresponding to each synapse to calculate the synaptic conductance of each synapse;

[0027] Generate the synaptic conductance of each synapse into a current acting on the postsynaptic membrane through a current generation function;

[0028] Accumulate the calculated postsynaptic membrane current onto the postsynaptic membrane;

[0029] Traverse all neuron groups and update the state variables of each neuron group.

[0030] In a second aspect, the present invention provides a low-complexity brain simulation device, including:

[0031] A splitting module, configured to split the spiking neural network to be modeled into multiple neuron groups;

[0032] A construction module, configured to construct each synapse between any two of the neuron groups, wherein the same time delay is set for each synapse;

[0033] A determination module, configured to determine the synaptic variables of each synapse after the time delay;

[0034] A low-rank approximation module, configured to low-rank approximate the weight matrix of each synapse as a product of multiple small matrices;

[0035] A simulation module, configured to perform brain simulation according to each neuron group, each synapse, each synaptic variable and its corresponding multiple small matrices.

[0036] In a third aspect, the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the low-complexity brain simulation method described in the first aspect is implemented.

[0037] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium, where the non-transitory computer-readable storage medium includes a computer program, and when the computer program is executed by the processor, the low-complexity brain simulation method described in the first aspect is implemented.

[0038] In a fifth aspect, the present invention further provides a computer program product, where the computer program product includes a computer program, and when the computer program is executed by the processor, the low-complexity brain simulation method described in the first aspect is implemented.

[0039] The low-complexity brain simulation method, device, equipment and medium provided by the present invention split the spiking neural network to be modeled into multiple neuron groups; construct each synapse between any two neuron groups in each neuron group, wherein each synapse is provided with the same time delay; determine the synaptic variables of each synapse after the time delay; approximate the weight matrix of each synapse to the product of multiple small matrices in a low-rank manner; perform brain simulation according to each neuron group, each synapse, each synaptic variable and its corresponding multiple small matrices.

[0040] During the brain simulation process of the present invention, a spiking neural network with N neurons is split into multiple neuron groups, and only N synaptic variables need to be stored and calculated. At the same time, only m*N synaptic weights need to be stored, where m is the size of the rank. Compared with the traditional neural dynamics modeling method, for a network with N neurons, N 2 synaptic variables need to be stored and calculated, and N 2 synaptic weights need to be stored and calculated. Further, it can be understood that the low-complexity brain simulation method provided by the present invention effectively reduces the brain dynamics simulation from O(n 2 ) squared complexity to O(n) linear complexity, significantly reducing the computational complexity of synapses and the memory requirements of computing devices, thereby reducing the simulation complexity of brain simulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the present invention, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0042] Figure 1 is a flowchart of the low-complexity brain simulation method provided by the present invention;

[0043] Figure 2 is a synaptic calculation flowchart provided by the present invention;

[0044] Figure 3 is a structural diagram of the low-complexity brain simulation device provided by the present invention;

[0045] Figure 4 is a schematic structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] To make the objectives, technical solutions and advantages of the present invention more clear, the following will, in conjunction with the accompanying drawings in the present invention, clearly and completely describe the technical solutions in the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts fall within the scope of protection of the present invention.

[0047] For the sake of brevity and intuitiveness in description, the following will elaborate on the solution of the present invention by describing several representative embodiments. A large number of details in the embodiments are only used to help understand the solution of the present invention. However, it is obvious that the implementation of the technical solution of the present invention may not be limited to these details. To avoid unnecessarily obscuring the solution of the present invention, some embodiments are not described in detail but only the framework is given. Hereinafter, "including" means "including but not limited to", and "according to..." means "at least according to..., but not limited to only according to...". Due to the language habits of Chinese, when the quantity of a component is not specifically indicated hereinafter, it means that the component can be one or more, or can be understood as at least one.

[0048] Further, in combination with Figures 1 to 4 describe the low-complexity brain simulation method, device, equipment and medium provided by the present invention. Figure 1 is the flowchart of the low-complexity brain simulation method provided by the present invention; Figure 2 is the synaptic calculation flowchart provided by the present invention; Figure 3 is the structural diagram of the low-complexity brain simulation device provided by the present invention; Figure 4 is the structural schematic diagram of the electronic device provided by the present invention.

[0049] The embodiments of the present invention provide an embodiment of the low-complexity brain simulation method. It should be noted that although the logical order is shown in the flowchart, under certain data, the steps shown or described can be completed in a different order from here.

[0050] The embodiments of the present invention take an electronic device as the execution subject for example. The embodiments of the present invention take the brain simulation system as one of the manifestations of the electronic device without limitation.

[0051] Referring to Figure 1 , Figure 1 is the flowchart of the low-complexity brain simulation method provided by the present invention. The low-complexity brain simulation method provided by the embodiments of the present invention includes:

[0052] Step 101, splitting the pulse neural network to be modeled into multiple neuron groups.

[0053] It should be noted that the spiking neural network to be modeled in the embodiments of the present invention is the object to be modeled. The spiking neural network to be modeled is a large-scale spiking neural network, that is, there are multiple neurons in the spiking neural network to be modeled. Therefore, first, the spiking neural network to be modeled needs to be split into multiple small neuron groups to obtain multiple neuron groups. It can be understood that each neuron group includes multiple neurons.

[0054] Furthermore, it can be understood that when splitting the spiking neural network to be modeled, the splitting rules of the neural network need to be involved. In one embodiment, the splitting rule can be that the neurons within each neuron group have the same firing pattern. Therefore, it can be understood that when splitting the spiking neural network to be modeled, the neurons with the same firing pattern are classified into the same neuron group.

[0055] Step 102, construct each synapse between any two neuron groups among the respective neuron groups, where each synapse is provided with the same time delay;

[0056] Step 103, determine the synaptic variables of each synapse after the time delay.

[0057] Furthermore, when constructing each synapse between any two neuron groups among the respective neuron groups, it should be noted that after constructing the synapses, the neuron groups can be divided into N i presynaptic neuron groups and N j postsynaptic neuron groups. Among them, a presynaptic neuron group includes multiple presynaptic neurons, a postsynaptic neuron group includes multiple postsynaptic neurons. A presynaptic neuron is the neuron in front of the synapse, and a postsynaptic neuron is the neuron behind the synapse. A synapse refers to the structure of mutual contact where the impulse of one neuron is transmitted to another neuron or to another cell.

[0058] At the same time, it is necessary to set that the time delays for each presynaptic neuron group to send neurotransmitter signals to multiple postsynaptic neuron groups are the same, that is, the time delay is t d milliseconds.

[0059] Furthermore, determining the synaptic variables of each synapse after the time delay is specifically: determining the synaptic connection state between the i-th presynaptic neuron group and the j-th postsynaptic neuron group after a time t d delay, the pulse state of the i-th presynaptic neuron group the synaptic dynamics h between the i-th presynaptic neuron group and the j-th neuron group ij , and the dynamics equation of each synapse:

[0060]

[0061] Step 104: Low-rank approximate the weight matrix of each of the synapses as a product of multiple small matrices.

[0062] Step 105: Perform brain simulation according to each of the neuron groups, each of the synapses, each synaptic variable, and its corresponding multiple small matrices.

[0063] If the synaptic weights of each of the synapses are weight matrices, then low-rank approximate the weight matrix of each of the synapses as a product of multiple small matrices. Among them, the small matrix can be understood as a kind of low-rank matrix. Therefore, it can be understood that if the synaptic weights of each of the synapses are weight matrices, then low-rank approximate the weight matrix of each of the synapses as a product of multiple low-rank matrices.

[0064] Furthermore, multiply each synaptic variable by multiple low-rank matrices to obtain the synaptic conductance of each of the synapses.

[0065] Furthermore, process the synaptic conductance of each of the synapses through the current generation function acting on the postsynaptic membrane of the synapse to generate the current acting on the postsynaptic membrane of each of the synapses.

[0066] Furthermore, perform brain simulation through the current acting on the postsynaptic membrane of each of the synapses and each postsynaptic neuron group.

[0067] The low-complexity brain simulation method provided by the present invention splits the spiking neural network to be modeled into multiple neuron groups; constructs each synapse between any two neuron groups in each neuron group, where each synapse is provided with the same time delay; determines the synaptic variable of each synapse after the time delay; low-rank approximate the weight matrix of each of the synapses as a product of multiple small matrices; perform brain simulation according to each neuron group, each synapse, each synaptic variable, and its corresponding multiple small matrices.

[0068] In the process of brain simulation of the present invention, a spiking neural network with N neurons is split into multiple neuron groups, and only N synaptic variables need to be stored and calculated. At the same time, only m*N synaptic weights need to be stored, where m is the size of the rank. Compared with the traditional neural dynamics modeling method, for a network with N neurons, N 2 synaptic variables need to be stored and calculated, and N 2 synaptic weights need to be stored and calculated. Further, it can be understood that the low-complexity brain simulation method provided by the present invention effectively reduces the brain dynamics simulation from O(n 2 ) square complexity to O(n) linear complexity, significantly reducing the computational complexity of the synapses and the memory requirements of the computing device, thereby reducing the simulation complexity of the brain simulation.

[0069] Furthermore, splitting the spiking neural network to be modeled in step 101 is performed based on a preset rule. It can be understood that neurons with the same properties in the spiking neural network to be modeled are classified through the preset rule to obtain multiple neuron groups.

[0070] Specifically, the spiking neural network to be modeled is split according to the preset rule to obtain multiple neuron groups. When splitting the spiking neural network to be modeled, the preset rule needs to comply with the following first rule, second rule, third rule, and fourth rule.

[0071] The first rule is that neurons within each neuron group have the same firing pattern. The second rule is that neurons within each neuron group have the same neurotransmitter receptors. The third rule is that neurons within each neuron group have the same projection source. The fourth rule is that the maximum physical distance between neurons within each neuron group does not exceed a preset distance threshold, where the preset distance threshold is set according to the actual situation. In the embodiment of the present invention, for example, if the preset distance threshold is set to 0.56 millimeters, that is, the fourth rule is that the maximum physical distance between neurons within each neuron group does not exceed 0.56 millimeters, which also means that the maximum area of each neuron group is 1 square millimeter.

[0072] That is to say, in the embodiment of the present invention, the spiking neural network to be modeled is split through the first rule, second rule, third rule, and fourth rule of the preset rule to obtain multiple neuron groups.

[0073] Furthermore, each presynaptic neuron group has N i neurons. Therefore, its dynamic equation is set as required:

[0074]

[0075] where the state of the i-th presynaptic neuron group is modeled as a vector f i is the evolution equation of the i-th presynaptic neuron group, is the spike firing state of the i-th presynaptic neuron group, is the input received by the i-th presynaptic neuron group, and t is time.

[0076] In the embodiment of the present invention, the spiking neural network is split through the first rule, second rule, third rule, and fourth rule of the preset rule, and neurons with the same properties are accurately classified into each neuron group.

[0077] Further, when performing brain simulation according to each synaptic variable and its corresponding multiple low-rank matrices as described in step 105, set the current equation for each synapse acting on the postsynaptic membrane as needed:

[0078]

[0079] Among them, the vector is the connection state of the synapse between the i-th presynaptic neuron group and the j-th postsynaptic neuron group, is the pulse state of the i-th presynaptic neuron group after a time delay t d , h ij is the synaptic dynamics between the i-th presynaptic neuron group and the j-th postsynaptic neuron group, o ij is the current generation function of the synapse acting on the postsynaptic membrane, is the conductance of the postsynaptic membrane neuron.

[0080] Further, approximating the weight matrix of each of the synapses to the product of multiple small matrices as described in step 104 can be specifically discussed in two cases.

[0081] The first case is that the weight connection information in the weight matrix is heterogeneous and known, and the second case is that the weight connection information is heterogeneous and unknown, that is, the weight matrix only has weight connections and no connection information between the weight connections.

[0082] For the first case, that is, the weight connection information is heterogeneous and known, specifically:

[0083] If the weight connection information in the weight matrix is heterogeneous and known, then each weight matrix is low-rank approximated to the product of a corresponding first matrix and a second matrix through a matrix dimensionality reduction method or a matrix factorization method.

[0084] For the weight connection information in each weight matrix W ij is heterogeneous and known, then each weight matrix is directly low-rank approximated to the product of a corresponding first matrix and a second matrix through a matrix dimensionality reduction method and a matrix factorization method, where the first matrix and the second matrix can be a type of low-rank matrix. Therefore, it can be understood that for the weight connection information in each weight matrix W ij is heterogeneous and known, then each weight matrix is directly low-rank approximated to the product of a corresponding first low-rank matrix U and a second low-rank matrix V through a matrix dimensionality reduction method and a matrix factorization method, that is, the synaptic weight W ij is low-rank approximated to UV, that is, W ij ≈UV. Among them, m is the size of the rank, which is set according to experience and needs, and generally m is much smaller than N i and Nj , N i is the number of neurons in the i-th presynaptic neuron group, and N j is the number of neurons in the j-th postsynaptic neuron group, that is, it can be expressed as:

[0085]

[0086] It should be noted that the matrix dimensionality reduction methods include but are not limited to Singular Value Decomposition, Random Projection, Principal Component Analysis, Nonnegative Matrix Factorization, Isomap, and Locally Linear Embedding; the matrix factorization methods include matrix factorization methods.

[0087] In the embodiments of the present invention, when it is determined that the weight connection information is heterogeneous and known, each weight matrix is directly low-rank approximated as the product of a first low-rank matrix and a second low-rank matrix through a matrix dimensionality reduction method or a matrix factorization method, providing a data basis for accurately calculating the synaptic conductances of each synapse.

[0088] For the second case, that is, the weight connection information is heterogeneous and unknown, specifically:

[0089] If the weight connection information in the weight matrix is heterogeneous and unknown, a third matrix and a fourth matrix are randomly generated, where the statistical properties of the third matrix and the fourth matrix are the same as those of the weight matrix without approximation;

[0090] Each weight matrix is low-rank approximated as the product of a corresponding said third matrix and a said fourth matrix.

[0091] For the weight connection information in each synaptic weight being heterogeneous and unknown, a third matrix and a fourth matrix with the same statistical properties as the original matrix W of each synapse ij are randomly generated, that is, the statistical properties of the third matrix and the fourth matrix are the same as those of the weight matrix without approximation, where the third matrix and the fourth matrix can be a kind of low-rank matrix. Therefore, it can be understood that for the weight connection information in each synaptic weight being heterogeneous and unknown, a third low-rank matrix U and a fourth low-rank matrix V with the same statistical properties as the original matrix of each synapse are randomly generated, that is, the statistical properties of each third low-rank matrix and the statistical properties of each fourth low-rank matrix are the same as those of the original matrix of each synapse. ij ​

[0092] Further, each weight matrix is low-rank approximated as the product of a corresponding third low-rank matrix and a fourth low-rank matrix.

[0093] In one embodiment, if the original matrix W ij obeys the Gaussian distribution N(μ,σ 2 ), then a third low-rank matrix U obeying the Gaussian distribution and a fourth low-rank matrix V obeying the uniform distribution on [0,1] can be generated such that the matrix UV after low-rank approximation of the third low-rank matrix U and the fourth low-rank matrix V also obeys the Gaussian distribution N(μ,σ 2 )

[0094] When it is determined in the embodiment of the present invention that the weight connection information in the weight matrix is heterogeneous and unknown, the product of a third low-rank matrix and a fourth low-rank matrix with the same statistical properties as the original matrix of each synapse is randomly generated, providing a data basis for accurately calculating the synaptic conductance of each synapse.

[0095] Further, after determining the synaptic variables of each synapse after the time delay described in step 103, it is also necessary to discuss the case where the synaptic weight of each synapse is not a weight matrix, which is specifically described as follows:

[0096] If the weight connections in each of the weight matrices are homogeneous, the connection weight is set to a scalar.

[0097] When the synaptic strength W ij between the i-th presynaptic neuron group and the j-th postsynaptic neuron group is homogeneous, that is, the synaptic weight of each synapse is a scalar, the synaptic conductance of each synapse is calculated by combining the synaptic weight and connection state of each synapse and any element of the connection state of each synapse through a preset formula. The preset formula is:

[0098]

[0099] The conductance g ij from the i-th presynaptic neuron group to the j-th postsynaptic neuron group at time t is obtained, where R is the synaptic variable and k is the k-th element of the synaptic state of the synapse.

[0100] Further, according to the current equation of each synapse acting on the postsynaptic membrane:

[0101]

[0102] The synaptic conductance of each synapse is processed to generate the current of each synapse acting on the postsynaptic membrane, so as to complete the brain simulation

[0103] When the synaptic weights of each synapse are scalars, the embodiments of the present invention accurately calculate the synaptic conductance of each synapse by combining the synaptic weights and connection states of each synapse and any element of the connection state through a preset formula, thereby ensuring the accuracy of brain simulation.

[0104] It should be noted that in actual calculation, the synaptic variables of the synapses between any two neuron groups among each presynaptic neuron group and multiple postsynaptic neuron groups The number is set to be the same as the number of neurons N in the i-th presynaptic neuron group i That is, the number of synaptic variables is the same as the number of neurons in the i-th presynaptic neuron group, so that the state of the presynaptic neuron group after delay can be used To update the state of the synapse one by one

[0105] Furthermore, the brain simulation according to each of the neuron groups, each of the synapses, each synaptic variable and its corresponding multiple small matrices described in step 105 is specifically as follows:

[0106] At each moment, update the state variables of each synapse according to the state of the presynaptic neuron after delay;

[0107] Multiply the state variable of each synapse by the scalar weight or multiple small matrix weights corresponding to each synapse to calculate the synaptic conductance of each synapse;

[0108] Generate the synaptic conductance of each synapse into the current acting on the postsynaptic membrane through a current generation function;

[0109] Accumulate the calculated postsynaptic membrane current onto the postsynaptic membrane;

[0110] Traverse all neuron groups and update the state variables of each neuron group.

[0111] When the synaptic strength between the i-th presynaptic neuron group and the j-th postsynaptic neuron group Is heterogeneous, that is, when the synaptic weights of each synapse are matrices, each weight matrix is low-rank approximated as the product of multiple low-rank matrices, that is, the synaptic weights of each synapse are low-rank approximated by multiple low-rank matrices. Furthermore, according to the connection state of each synapse and the multiple low-rank matrices corresponding to each synapse, calculate the synaptic conductance of each synapse.

[0112] Calculate the synaptic conductance of each synapse according to the connection state of each synapse in combination with each first low-rank matrix U and each second low-rank matrix V. The specific formula can be expressed as Furthermore, according to Update the conductance of the i-th presynaptic neuron population acting on the j-th postsynaptic neuron population

[0113] Further, according to the current equation of each synapse acting on the postsynaptic membrane:

[0114]

[0115] Process the synaptic conductance of each synapse to generate the current of each synapse acting on the postsynaptic membrane, so as to complete the brain simulation.

[0116] In the embodiment of the present invention, when the synaptic weights of each synapse are a weight matrix, the synaptic conductance of each synapse is accurately calculated through the connection state of each synapse and the plurality of low-rank matrices, thereby ensuring the accuracy of the brain simulation.

[0117] In one embodiment, referring to Figure 2 , Figure 2 is the synaptic calculation flow chart provided by the present invention, which can be understood as: after a time delay t d , the synaptic variables of each synapse are obtained Further, multiply each synaptic variable by the synaptic weight of each synapse, that is to obtain the synaptic conductance of each synapse

[0118] Further, process the synaptic conductance of each synapse through the current generation function of the synapse acting on the postsynaptic membrane , that is to generate the current of each synapse acting on the postsynaptic membrane. Finally, realize the brain simulation through the current of each synapse acting on the postsynaptic membrane and each postsynaptic neuron population.

[0119] Further, the low-complexity brain simulation device provided by the present invention corresponds to and refers to the low-complexity brain simulation method provided by the present invention.

[0120] Figure 3 As shown Figure 3 is the structural diagram of the low-complexity brain simulation device provided by the present invention. The low-complexity brain simulation device includes:

[0121] A splitting module 301, configured to split the pulse neural network to be modeled into a plurality of neuron populations;

[0122] A construction module 302, configured to construct each synapse between any two neuron populations in each of the neuron populations, wherein the same time delay is set for each of the synapses;

[0123] A determination module 303, configured to determine the synaptic variables of each of the synapses after the time delay;

[0124] A low-rank approximation module 304, configured to low-rank approximate the weight matrix of each of the synapses as a product of multiple small matrices;

[0125] A simulation module 305, configured to perform a brain simulation according to each of the neuron groups, each of the synapses, each synaptic variable, and its corresponding multiple small matrices.

[0126] Further, the low-rank approximation module 304 is further configured to:

[0127] If the weight connection information in the weight matrix is heterogeneous and known, each weight matrix is low-rank approximated as a product of a corresponding first matrix and a second matrix by a matrix dimensionality reduction method or a matrix factorization method.

[0128] Further, the low-rank approximation module 304 is further configured to:

[0129] If the weight connection information in the weight matrix is heterogeneous and unknown, a third matrix and a fourth matrix are randomly generated, where the statistical properties of the third matrix and the fourth matrix are the same as the statistical properties of the weight matrix without approximation;

[0130] Each weight matrix is low-rank approximated as a product of a corresponding one of the third matrix and a corresponding one of the fourth matrix.

[0131] Further, the simulation module 305 is further configured to:

[0132] At each moment, update the state variable of each synapse according to the state of the presynaptic neuron after the delay;

[0133] Multiply the state variable of each synapse by the scalar weight or the multiple small matrix weights corresponding to each synapse to calculate the synaptic conductance of each synapse;

[0134] Generate the synaptic conductance of each synapse as a current acting on the postsynaptic membrane through a current generation function;

[0135] Accumulate the calculated postsynaptic membrane current on the postsynaptic membrane;

[0136] Traverse all neuron groups and update the state variables of each neuron group.

[0137] The specific embodiments of the low-complexity brain simulation device provided by the present invention are basically the same as the embodiments of the above-mentioned low-complexity brain simulation method, and will not be elaborated here.

[0138] Figure 4 Illustrates a schematic physical structure diagram of an electronic device, such asFigure 4 As shown in Figure 4 , the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440. Among them, the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 may call the logical instructions in the memory 430 to execute a low-complexity brain simulation method, which includes:

[0139] Splitting the spiking neural network to be modeled into multiple neuron groups;

[0140] Constructing each synapse between any two neuron groups in each of the neuron groups, where each synapse is provided with the same time delay;

[0141] Determining the synaptic variables of each synapse after the time delay;

[0142] Approximating the weight matrix of each synapse to the product of multiple small matrices with low rank;

[0143] Performing brain simulation according to each of the neuron groups, each of the synapses, each synaptic variable, and its corresponding multiple small matrices.

[0144] In addition, when the logical instructions in the above-mentioned memory 430 are implemented in the form of software functional units and sold or used as an independent product, they may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0145] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the low-complexity brain simulation method provided by the above-mentioned various methods, and the method includes:

[0146] Splitting the spiking neural network to be modeled into multiple neuron groups;

[0147] Construct each synapse between any two neuron groups among the various neuron groups, where the same time delay is set for each synapse;

[0148] Determine the synaptic variables of each synapse after the time delay;

[0149] Approximate the weight matrix of each synapse to the product of multiple small matrices with low rank;

[0150] Perform brain simulation according to each neuron group, each synapse, each synaptic variable and its corresponding multiple small matrices.

[0151] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is used to execute the low-complexity brain simulation method provided above. The method includes:

[0152] Split the spiking neural network to be modeled into multiple neuron groups;

[0153] Construct each synapse between any two neuron groups among the various neuron groups, where the same time delay is set for each synapse;

[0154] Determine the synaptic variables of each synapse after the time delay;

[0155] Approximate the weight matrix of each synapse to the product of multiple small matrices with low rank;

[0156] Perform brain simulation according to each neuron group, each synapse, each synaptic variable and its corresponding multiple small matrices;

[0157] Perform brain simulation according to each synaptic variable and its corresponding multiple low-rank matrices.

[0158] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative labor.

[0159] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A low-complexity brain simulation method, characterized in that, Comprising: Splitting the spiking neural network to be modeled into multiple neuron groups; Constructing each synapse between any two neuron groups in each of the neuron groups, wherein the same time delay is set for each synapse; Determining the synaptic variables of each synapse after the time delay; Low-rank approximating the weight matrix of each synapse as a product of multiple small matrices, including: if the weight connection information in the weight matrix is heterogeneous and known, then low-rank approximating each weight matrix as a product of a corresponding first matrix and a second matrix through a matrix dimensionality reduction method or a matrix factorization method, or if the weight connection information in the weight matrix is heterogeneous and unknown, then randomly generating a third matrix and a fourth matrix, wherein the statistical properties of the third matrix and the fourth matrix are the same as those of the weight matrix without approximation; Low-rank approximating each weight matrix as a product of a corresponding one of the third matrix and the fourth matrix; Performing brain simulation according to each neuron group, each synapse, each synaptic variable and its corresponding multiple small matrices, including: At each moment, updating the state variable of each synapse according to the state of the presynaptic neuron after the delay; Multiplying the state variable of each synapse by the scalar weight or multiple small matrix weights corresponding to each synapse to calculate the synaptic conductance of each synapse; Generating the synaptic conductance of each synapse into a current acting on the postsynaptic membrane through a current generation function; Accumulating the calculated postsynaptic membrane current onto the postsynaptic membrane; Traversing all neuron groups and updating the state variables of each neuron group.

2. The low-complexity brain simulation method according to claim 1, wherein The number of the synaptic variables is the same as the number of neurons in the presynaptic neuron group, where the presynaptic neuron group is any one of the multiple neuron groups.

3. The low-complexity brain simulation method according to any one of claims 1 to 2, characterized in that, After determining the synaptic variables of each synapse after the time delay, it further includes: If the weight connections in each of the weight matrices are homogeneous, setting the connection weight to a scalar.

4. The low-complexity brain simulation method according to claim 1, characterized in that The splitting of the spiking neural network to be modeled into multiple neuron groups is performed based on a preset rule; the preset rule includes: The first rule is that the neurons in each of the presynaptic neuron groups have the same firing pattern; Or / and, the second rule is that the neurons in each of the presynaptic neuron groups have the same neurotransmitter receptors; Or / and, the third rule is that the neurons in each of the presynaptic neuron groups have the same projection source; Or / and, the fourth rule is that the maximum physical distance between the neurons in each of the presynaptic neuron groups does not exceed a preset distance threshold.

5. A low-complexity brain simulation device, characterized in that, Comprising: A splitting module, configured to split the spiking neural network to be modeled into multiple neuron groups; A constructing module, configured to construct each synapse between any two neuron groups in each of the neuron groups, wherein the same time delay is set for each synapse; A determining module, configured to determine the synaptic variables of each synapse after the time delay; A low-rank approximation module for approximating the weight matrix of each of the synapses as a product of multiple small matrices, including: if the weight connection information in the weight matrix is heterogeneous and known, each weight matrix is low-rank approximated as a product of a corresponding first matrix and a second matrix through a matrix dimensionality reduction method or a matrix factorization method, or if the weight connection information in the weight matrix is heterogeneous and unknown, a third matrix and a fourth matrix are randomly generated, wherein the statistical properties of the third matrix and the fourth matrix are the same as those of the weight matrix without approximation; each weight matrix is low-rank approximated as a product of a corresponding one of the third matrix and the fourth matrix; A simulation module for performing brain simulation according to each of the neuron groups, each of the synapses, each synaptic variable, and its corresponding multiple small matrices; including: At each moment, update the state variable of each synapse according to the state of the presynaptic neuron after delay; Multiply the state variable of each synapse by the scalar weight or multiple small matrix weights corresponding to each synapse to calculate the synaptic conductance of each synapse; Generate the synaptic conductance of each synapse into a current acting on the postsynaptic membrane through a current generation function; Accumulate the calculated postsynaptic membrane current onto the postsynaptic membrane; Traverse all neuron groups and update the state variables of each neuron group.

6. An electronic device, the electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the low-complexity brain simulation method according to any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium, the non-transitory computer-readable storage medium comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the low-complexity brain simulation method according to any one of claims 1 to 4.

Images