CN113869505A - Pulse neural network optimization method and system based on brain glial cells - Google Patents
Pulse neural network optimization method and system based on brain glial cells Download PDFInfo
- Publication number
- CN113869505A CN113869505A CN202111471537.2A CN202111471537A CN113869505A CN 113869505 A CN113869505 A CN 113869505A CN 202111471537 A CN202111471537 A CN 202111471537A CN 113869505 A CN113869505 A CN 113869505A
- Authority
- CN
- China
- Prior art keywords
- neuron
- glial cell
- glial
- connection
- event
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Theoretical Computer Science (AREA)
- Molecular Biology (AREA)
- Neurology (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Feedback Control In General (AREA)
Abstract
The invention relates to a method and a system for optimizing a pulse neural network based on brain glial cells, wherein the method comprises the following steps: constructing a pulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron; determining various types of connection parameters; and optimizing the impulse neural network according to the various types of connection parameters to obtain the optimized impulse neural network. The spiking neural network fusing the glial cells of the brain established by the invention not only can realize the weight adjustment of synapses meeting plasticity rules, but also can realize the adjustment of transmission delay and signal types of the synapses, and can also add synapses and delete synapses, thereby improving the learning capacity of the SNN network.
Description
Technical Field
The invention relates to the field of neural network training, in particular to a method and a system for optimizing a pulse neural network based on brain glial cells.
Background
Glial cells refer to non-neuronal cells within the brain, and early studies thought that glial cells were used only to immobilize and support neurons. While the fact is not so simple, recent studies have found that glial cells can participate in directing neuronal development, precisely controlling the differentiation location and growth direction of neurons, regulating neuronal activity, regulating synapse formation and maintaining their stability, etc. This shows that glial cells have the function of regulating brain and play an important role in brain evolution, brain repair, and memory reasoning. Therefore, the research and study of glial cells on the regulatory function of neural networks is very urgent and valuable.
The traditional impulse neural network is composed of neurons and synaptic connections thereof, and the SNN network has the characteristic that a state function and an error function are not microminiature due to the fact that information is transmitted in the form of impulse time series, so that a supervision learning algorithm of the traditional neural network cannot be used. Meanwhile, existing plasticity researches such as pulse timing dependent plasticity (STDP), Long Term Potentiation (LTP), long term inhibition (LTD), short term oscillation (STF), Short Term and (STD) are local plasticity methods, and global plasticity is lacked. In addition, the existing local plasticity methods can only perform weight optimization on synapses of the spiking neural network, namely, only network weight parameters can be optimized, state optimization of network nodes (neuron models) cannot be realized, and an effective supervision training method suitable for SNN is not available. Therefore, at present, only the pulse neural network learning algorithm and the unsupervised training method of local plasticity are relied on, so that the learning capability of the SNN network is difficult to be effectively improved, and an effective supervised training method is lacked.
Disclosure of Invention
The invention aims to provide a method and a system for optimizing a spiking neural network based on brain glial cells so as to improve the learning capacity of an SNN network.
In order to achieve the above object, the present invention provides a method for optimizing a spiking neural network based on brain glial cells, the method comprising:
step S1: constructing a pulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron;
step S2: determining various types of connection parameters;
step S3: and optimizing the impulse neural network according to the various types of connection parameters to obtain the optimized impulse neural network.
Optionally, the optimizing the spiking neural network according to the various types of connection parameters to obtain an optimized spiking neural network specifically includes:
step S31: judging whether an event occurs or not; if an event occurs, "step S32" is executed; if no event occurs, continuously judging whether the event occurs or not until a set condition is met, and outputting a pulse neural network fused with the glial cells;
step S32: determining the event type; when the event type is a neuron impulse event, executing step S33; when the event type is a glial cell communication event, then "step S35" is performed; when the event type is the glial cell conditioning event, then "step S36" is performed;
step S33: looking up an event source neuron related connection table, and determining a neuron fan-out connection type; when the neuron fan-out connection type is neuron-neuron, reading neuron-neuron connection parameters; when the type of the neuron fan-out connection is neuron-glial cell, reading the connection parameters of the neuron-glial cell;
step S34: taking the neuron-neuron connection parameters as neuron input, updating the target neuron state parameters and output, and returning to the step S31; inputting the neuron-glial cell connection parameters into the glial cell kinetic model, updating the target glial cell state parameters and output, and returning to step S31;
step S35: looking up an event source glial cell fan-out connection table, and reading a glial cell-glial cell connection parameter; inputting the glial cell-glial cell connection parameter into the glial cell kinetic model, updating the target glial cell state parameter, and returning to step S31;
step S36: looking up an event source glial cell fan-out connection table, and determining a glial cell fan-out connection type; when the type of the glial cell fan-out connection is a glial cell-neuron synapse, reading a connecting parameter of the glial cell-neuron synapse; when the type of the glial cell fan-out connection is glial cell-neuron, reading a glial cell-neuron connection parameter;
step S37: updating the target synapse state parameter according to the glial cell-neuronal synapse connection parameter, and returning to step S31; the target neuron state parameter is updated according to the glial cell-neuron connection parameter, and the process returns to step S31.
Optionally, the glial cell kinetic model comprises:
the system comprises a glial communication event processing module, a neuron pulse event processing module and a glial cell state machine; the glial cell state machine is respectively connected with the glial communication event processing module and the neuron impulse event processing module.
Optionally, the specific calculation formula of the colloid communication event processing module is as follows:
wherein the content of the first and second substances,
indicating the strength of the period of the gum communication event,
represents a weight parameter in the glial cell-glial cell line parameter,
representing the statistical period of glial communication events by glial cells,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
Optionally, the specific calculation formula of the neuron pulse event processing module is as follows:
wherein the content of the first and second substances,
represents the intensity of the period of the neuron impulse event,
represents a weight parameter in the neuron-glial cell junction parameter,
representing the statistical period of glial cell versus neuronal impulse events,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
The invention also provides a pulsed neural network optimization system based on brain glial cells, the system comprising:
the impulse neural network construction module is used for constructing an impulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron;
the connection parameter determining module is used for determining various types of connection parameters;
and the optimization module is used for optimizing the impulse neural network according to the various types of connection parameters to obtain the optimized impulse neural network.
Optionally, the optimization module specifically includes:
the judging unit is used for judging whether an event occurs or not; if an event occurs, executing an event type determining unit; if no event occurs, continuously judging whether the event occurs or not until a set condition is met, and outputting a pulse neural network fused with the glial cells;
an event type determining unit for determining an event type; when the event type is a neuron pulse event, executing a first connection parameter determining unit; when the event type is a glial cell communication event, then a "second update unit" is performed; when the event type is a glial cell regulating event, executing a second connection parameter determining unit;
the first connection parameter determining unit is used for searching the event source neuron related connection table and determining the neuron fan-out connection type; when the neuron fan-out connection type is neuron-neuron, reading neuron-neuron connection parameters; when the type of the neuron fan-out connection is neuron-glial cell, reading the connection parameters of the neuron-glial cell;
the first updating unit is used for taking the neuron-neuron connection parameters as neuron input, updating the state parameters and output of the target neuron and returning to the judging unit; inputting the neuron-glial cell connection parameters into a glial cell kinetic model, updating the state parameters and output of target glial cells, and returning to a judgment unit;
the second updating unit is used for checking the glial cell fan-out connection table of the event source and reading the glial cell-glial cell connection parameter; inputting connecting parameters of the glial cells into the glial cell kinetic model, updating the state parameters of the target glial cells, and returning to a judgment unit;
the second connection parameter determining unit is used for checking an event source glial cell fan-out connection table and determining the type of the glial cell fan-out connection; when the type of the glial cell fan-out connection is a glial cell-neuron synapse, reading a connecting parameter of the glial cell-neuron synapse; when the type of the glial cell fan-out connection is glial cell-neuron, reading a glial cell-neuron connection parameter;
a third updating unit, configured to update the target synapse state parameter according to the glial cell-neuronal synapse connection parameter, and return to the "determining unit"; and updating the state parameter of the target neuron according to the connecting parameter of the glial cell-neuron, and returning to the judgment unit.
Optionally, the glial cell kinetic model comprises:
the system comprises a glial communication event processing module, a neuron pulse event processing module and a glial cell state machine; the glial cell state machine is respectively connected with the glial communication event processing module and the neuron impulse event processing module.
Optionally, the specific calculation formula of the colloid communication event processing module is as follows:
wherein the content of the first and second substances,
indicating the strength of the period of the gum communication event,
represents a weight parameter in the glial cell-glial cell line parameter,
system for representing glial communication event by glial cellsThe counting period is counted, and the time period is counted,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
Optionally, the specific calculation formula of the neuron pulse event processing module is as follows:
wherein the content of the first and second substances,
represents the intensity of the period of the neuron impulse event,
represents a weight parameter in the neuron-glial cell junction parameter,
representing the statistical period of glial cell versus neuronal impulse events,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention relates to a method and a system for optimizing a pulse neural network based on brain glial cells, wherein the method comprises the following steps: constructing a pulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron; determining various types of connection parameters; and optimizing the impulse neural network according to the various types of connection parameters to obtain the optimized impulse neural network. The spiking neural network fusing the glial cells of the brain established by the invention not only can realize the weight adjustment of synapses meeting plasticity rules, but also can realize the adjustment of transmission delay and signal types of the synapses, and can also add synapses and delete synapses, thereby improving the learning capacity of the SNN network.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
FIG. 1 is a flow chart of a method for optimizing a spiking neural network based on brain glial cells according to the present invention;
FIG. 2 is a diagram illustrating the process of realizing the plasticity of the spiking neural network of glial cells according to the present invention;
FIG. 3 is a schematic diagram of a spiking neural network of a fusion glial cell of the present invention;
FIG. 4 is a schematic representation of the expression of the binding relationship between neurons and glial cells according to the invention;
FIG. 5 is a schematic diagram of the parameters of the neuron-glial cell junction of the present invention;
FIG. 6 is a schematic representation of the expression of glial cell-neuronal synaptic junction relationship according to the invention;
FIG. 7 is a schematic diagram of glial cell-neuronal synaptic connection parameters of the present invention;
FIG. 8 is a schematic representation of the expression of glial cell-glial cell gap junction relation according to the invention;
FIG. 9 is a schematic representation of glial cell-glial cell line-connecting parameters of the present invention;
FIG. 10 is a schematic diagram illustrating an event-driven optimal update process of an impulse neural network according to the present invention;
FIG. 11 is a schematic representation of a glial cell kinetic model of the present invention;
FIG. 12 is a diagram illustrating a processing event by the colloidal communication event handler of the present invention;
FIG. 13 is a schematic diagram of a neuron impulse event processing module for processing events according to the present invention;
FIG. 14 is a diagram of a system for optimizing a spiking neural network based on glial cells of the brain according to the present invention;
FIG. 15 is a diagram of the plastic functional classification of the present invention;
fig. 16 shows an SNN training method provided in the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a method and a system for optimizing a spiking neural network based on brain glial cells so as to improve the learning capacity of an SNN network.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
The Spiking Neural Networks (SNNs) are currently the most bioanalytically artificial neural networks, and aim to mimic the process of information transfer in the human brain, unlike conventional neural networks, a Spiking neural network does not transmit information every cycle, only when the accumulation of membrane potential reaches a threshold, the neural network excites a neuron to generate a signal, i.e., a pulse, and transmits the pulse to other neurons through synapses, and the neuron receiving the pulse increases or decreases its own membrane potential according to the pulse. Although SNNs are theoretically more powerful than second generation artificial neural networks, the main problem with the application of SNNs is the training of SNNs. Although we have unsupervised biological learning methods such as Hebbian learning (Hebbian learning) and STDP, there is no effective supervised training method suitable for SNN that can provide superior performance over second generation networks. Since impulse training is not trivial, we cannot train SNNs using gradient descent without losing accurate time information. Therefore, in order to correctly use SNN to solve real world tasks, we need to develop an efficient supervised learning approach. This is a difficult task as it involves determining how the human brain learns given the biological realism of these networks.
In order to solve the existing technical problem, namely to solve the problem of insufficient plasticity of the existing SNN network, the invention is influenced by the high-level function research result of biological glial cells, and provides a pulse neural network plasticity optimization method and system based on the glial cells in a heuristic manner. The method is a universal pulse neural network plasticity method, and can be realized by software, such as an existing SNN simulator (e.g. NEURON); it may also be implemented in hardware, such as SpiNNaker.
The pulse neural network plasticity means that parameters in the pulse neural network can be adaptively modified according to external input pulses. The plasticity of the neural network comprises neuron synapse plasticity and neuron state plasticity. The neuron synaptic plasticity refers to that synaptic connection weight parameters, synaptic connection delay parameters and synaptic connection type parameters can be modified in real time according to corresponding plasticity rules; neuron state plasticity means that kinetic parameters of neurons can be modified in real time according to corresponding plasticity rules. The plasticity rules here are embodied in the glial cell kinetic model and the glial cell network connection.
Example 1
As shown in fig. 1, the invention discloses a method for optimizing a spiking neural network based on brain glial cells, which comprises the following steps:
step S1: constructing a pulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron.
Step S2: various types of connection parameters are determined.
Step S3: and optimizing the impulse neural network according to the various types of connection parameters to obtain the optimized impulse neural network.
As shown in figure 2, the first 2 steps of the invention are the steps of establishing a plastic impulse neural network and setting connection parameters, and the 3 rd step is an optimization step, namely, a training impulse sequence is input into the impulse neural network, and the glial cell network part in the constructed impulse neural network fusing the brain glial cells optimizes the impulse neural network according to the impulse activity generated by the impulse neural network part.
The individual steps are discussed in detail below:
step S1: constructing a pulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron.
Step S1: the method for constructing the pulse neural network fusing the brain glial cells based on the brain glial cells specifically comprises the following steps:
step S11: and correspondingly determining the required population number of the glial cells according to the population division of the impulse neural network to be regulated.
Step S12: and establishing a connection relation between the glial population and the neuron population, namely establishing a neuron-glial cell and glial cell-neuron network. In the method, a one-to-one connection is established between the glial population and the neuronal population, i.e., one glial cell is connected to only one neuronal cell.
Step S13: and establishing a connection relation between the glial cell population and the glial cell population, namely establishing the glial cell-glial cell. In the method, full connections are established between glial cells within a population of glia, i.e., one glial cell is connected to all other cells in the population.
Step S14: and establishing a connection relation between the glial cells and the synapses accessed by the neuron population, namely establishing a glial cell-neuron synapse network. In the method, glial cells are connected to all the afferent synapses of their accessory neurons.
At present, the connection relationship between glial cell-neuron and neuron-glial cell is basically recognized at present by conducting special chemical molecules or proteins bidirectionally through the fluid between the intercellular spaces, which is called virtual synapse; the glial cell-glial cell junction has been shown to be achieved by interconnecting glial gaps (a protein structure) between each other. The connection between glial cell-neuronal synapses is made by a structure known as the astrocyte process.
The linkage is a substance or information transmission pathway that mimics brain cells on organisms. As shown in fig. 3, the neural network fusing the glial cells to be established is a hierarchical network, and the neural network is formed by interconnecting one or more neuron populations (neuron sets), and similarly, the glial cell network is formed by interconnecting one or more glial cell populations (glial cell sets), and finally, the glial cell populations and the neuron populations are interconnected. From bottom to top, the first layer is an interconnection network formed by neuron populations, namely an impulse neural network, neurons among or inside the neuron populations are interconnected with each other to form a neuron-neuron network, and in addition, the neuron-neuron network is also interconnected with corresponding glial cell populations to form a neuron-glial network; the second layer is a glial cell population, and glial cells between or within the glial populations are interconnected to form a glial cell-glial cell network, and in addition, with a corresponding neuronal population, a glial cell-neuronal network is formed, and finally, the glial populations also form a glial cell-neuronal synaptic network with all of the connected synapses of the corresponding neuronal population.
To implement the present steps of the present invention in a computer, a two-dimensional matrix shown in table 1, i.e., a virtual synaptic connection relationship matrix, is assigned a unique number to each node (neuron or glial cell) in the network, a row index indicates the number of cells before connection (pre-neuron or pre-glial cell) as the starting point of connection, a column index indicates the number of cells after connection (post-glial cell or post-glial cell) as the ending point of connection, a "0" indicates no connection, and a "1" indicates a connection.
TABLE 1 virtual synaptic connection relationship description matrix
Similarly, for the expression of directional connection between glial cells and neuronal synapses, a two-dimensional matrix shown in table 2, i.e., a matrix of glial protrusion connection relationships, is indicated by a row index indicating the start of connection, i.e., the glial cell number before connection, and a column index indicating the end of connection, i.e., the synaptic number after connection, with "0" indicating no connection and "1" indicating connection.
TABLE 2. colloidal projection connection relation description matrix
Similarly, we can express the directional connection relationship between glial cells and glial cells as a two-dimensional matrix shown in table 3, i.e., a glial gap connection relationship matrix, where the row index indicates the starting point of connection, i.e., the glial cell number before connection, the column index indicates the end point of connection, i.e., the glial cell number after connection, "0" indicates no connection, and "1" indicates connection.
TABLE 3. description matrix of gap-to-gap connection relationship of gum
The linking parameter is a load that mimics the mass or information transfer of brain cells on an organism. To implement the present steps of the present invention in a computer, we generate a virtual synaptic connection parameter description matrix according to the virtual synaptic connection relationship description matrix generated in the above steps, as shown in table 4, a row index indicates the starting point of a connection, i.e., the number of cells before connection (pre-neuron or pre-glial cell), a column index indicates the end point of the connection, i.e., the number of cells after connection (post-glial cell or post-glial cell), the related connection parameters are filled according to the virtual synaptic connection relationship description matrix, i.e., the virtual synaptic connection parameter words (numerical description of the virtual synapse) corresponding to the filling of the connection are filled, and no data is filled in the non-connection.
TABLE 4 virtual synapse connection parameter description matrix
Similarly, we also generate a glia-process connection parameter description matrix according to the glia-process connection relationship description matrix of table 2, as shown in table 5, the row index indicates the starting point of connection, i.e., the glial cell number before connection, the column index indicates the end point of connection, i.e., the post-connection synaptic number, and the related connection parameters are filled according to the glia-process connection relationship description matrix, i.e., there is connected filling corresponding glia-process connection parameter word (numerical description of glia-process connection), and there is no connection filling any data.
TABLE 5 glial protrusion connection parameter description matrix
Similarly, a glial gap connectivity parameter description matrix is generated according to the glial gap connectivity description matrix in table 3, as shown in table 6, the row index indicates the starting point of the connection, i.e., the glial cell number before the connection, the column index indicates the end point of the connection, i.e., the glial cell number after the connection, and the related connection parameters are filled according to the glial gap connectivity description matrix, i.e., the corresponding glial gap connectivity parameter words (numerical description of the glial gap connectivity) are filled with the connection, and no data is filled in the connectionless connection.
TABLE 6. colloid gap connection parameter description matrix
Step S2: determining various types of connection parameters, and the specific steps comprise:
step S21: determining a neuron-glial cell junction parameter.
As shown in FIG. 4, the virtual synaptic parameters of the neuron-glial cell orientation of the present invention are divided into two parts: neuron-glial cell junction parameter, glial cell index; the virtual synapse parameters in glial cell-neuron direction of the present invention are divided into two parts: neuron parameter adjustment and neuron index. The pre-connection glial cell number and the post-connection neuron number are used to locate where the parameter needs to be stored in the network connection parameter matrix. Similarly, the pre-connection neuron number and the post-connection glial cell number are also used to locate where this parameter needs to be stored in the virtual synaptic connection parameter description matrix.
As shown in fig. 5, the virtual synapse parameters are suitable for LIF neuron model and glial cell engineering custom model, and the neuron tuning parameters are: membrane voltage adjustment, input current adjustment, and pulse delivery threshold adjustment. The neuron-glial cell parameters were: neuron impulse event intensity.
Step S22: determining a parameter of glial cell-neuronal synaptic connection.
As shown in FIG. 6, the expression of a gum protrusion of the present invention can be divided into two categories: synaptic parameter adjustment and synaptic index. The pre-synaptic glial cell number and synaptic index are used to locate the position where this parameter needs to be deposited in the glial process connection parameter description matrix.
As shown in fig. 7, the parameters of a colloidal process are the following parameters, and according to the neuron-neuron synaptic connection model, the synaptic parameter adjustment amount is: the amount of synaptic weight adjustment, the amount of synaptic delay adjustment, and whether synapse type is changed.
Step S23: determining glial cell-glial cell junction parameters.
As shown in FIG. 8, the expression of a single glial cell gap in this method is divided into two contexts: glial-cell affecting parameters, glial cell index. The pre-ligation glial cell number and the post-ligation glial cell number are used to locate where the parameter needs to be stored in the glial gap junction parameter description matrix.
As shown in fig. 9, the glial cell-glial cell line parameter is a collagen gap junction parameter, and the parameters are: the intensity of the gum communication event.
As shown in fig. 10, the updating and optimizing of the whole network is a loop iteration process, and the implementation in the computer is based on event driving. The events herein are mainly divided into two types: neuronal impulse events, glial cell activity events. The neuron pulse event refers to that neurons in the network generate pulses, then a target neuron or a target glial cell to which the pulses are transmitted is found according to fan-out connection of the neuron of an event source (the neuron generating the event is taken as a node, and the connection is transmitted outwards), then connection parameters are found, and the state of the target neuron or the glial cell is updated according to the connection parameters.
Among them, glial activity events are divided into two categories: glial cell communication events and glial cell regulatory events. The glial cell communication event refers to that glial cells generate inter-glial message transfer activity (used for feeding back the global activity state of different glial cell populations), then target glial cells to which communication messages are transferred are found according to event source glial cell fan-out connection (connection transferred outwards by taking the glial cells generating the event as nodes), then connection parameters are found in a connection relation description matrix and a connection parameter description matrix, and the state of the target glial cells is updated according to the connection parameters. The glial regulating event refers to that the glial regulating activity is generated by the glial cell (used for regulating the impulse neural network), then according to the event source glial cell fan-out connection, a target neuron cell or a target synapse to which a communication message is transmitted is found, then a connection parameter is found, and according to the connection parameter, the model parameter of the target neuron is regulated or the connection weight parameter of the target synapse is regulated. The steps are therefore summarized as follows:
step S3: the optimizing the spiking neural network according to the various types of connection parameters to obtain the optimized spiking neural network specifically includes:
step S31: judging whether an event occurs or not; if an event occurs, "step S32" is executed; if no event occurs, continuously judging whether the event occurs or not until a set condition is met, and outputting a pulse neural network fused with the glial cells;
step S32: determining the event type; when the event type is a neuron impulse event, executing step S33; when the event type is a glial cell communication event, then "step S35" is performed; when the event type is the glial cell conditioning event, then "step S36" is performed;
step S33: looking up an event source neuron related connection table, and determining a neuron fan-out connection type; when the neuron fan-out connection type is neuron-neuron, reading neuron-neuron connection parameters; when the type of the neuron fan-out connection is neuron-glial cell, reading the connection parameters of the neuron-glial cell;
step S34: taking the neuron-neuron connection parameters as neuron input, updating the target neuron state parameters and output, and returning to the step S31; inputting the neuron-glial cell connection parameters into the glial cell kinetic model, updating the target glial cell state parameters and output, and returning to step S31;
step S35: looking up an event source glial cell fan-out connection table, and reading a glial cell-glial cell connection parameter; inputting the glial cell-glial cell connection parameter into the glial cell kinetic model, updating the target glial cell state parameter, and returning to step S31;
step S36: looking up an event source glial cell fan-out connection table, and determining a glial cell fan-out connection type; when the type of the glial cell fan-out connection is a glial cell-neuron synapse, reading a connecting parameter of the glial cell-neuron synapse; when the type of the glial cell fan-out connection is glial cell-neuron, reading a glial cell-neuron connection parameter;
step S37: updating the target synapse state parameter according to the glial cell-neuronal synapse connection parameter, and returning to step S31; the target neuron state parameter is updated according to the glial cell-neuron connection parameter, and the process returns to step S31.
After the interconnected architecture of glial cells and neuronal cells is established, a kinetic model of glial cells needs to be established. At present, no general glial cell kinetic model exists for the glial cell with respect to the function of neural network plasticity. Therefore, according to the specific engineering application, the invention establishes the engineering self-defined model, namely the model meeting the engineering requirements. The self-defined model is considered from the engineering perspective, serves and meets specific engineering requirements, and the internal functions and logic of the model are artificially and subjectively defined. As shown in fig. 11, a custom model of colloidal cell engineering is divided into three modules: the system comprises a glial communication event processing module, a neuron pulse event processing module and a glial cell state machine.
The glial communication event processing module is mainly responsible for processing an incoming glial cell communication event and updating a glial communication state. After receiving communication events from other glial cells, the glial cells firstly remove the matched connection and read connection parameters, the connection parameters are used as input of event characteristic statistics, the characteristics of the event transmitted by the current connection are counted, then the glial communication state is updated, finally the state is transmitted to a glial cell state machine, and whether a glial cell communication event is generated or not is judged by the state machine. The processing flow of the neuron pulse event processing module is consistent with that of the glial communication event processing module, and is not described herein again.
As shown in fig. 11, the glial cell kinetic model constructed included: the system comprises a glial communication event processing module, a neuron pulse event processing module and a glial cell state machine; the glial cell state machine is respectively connected with the glial communication event processing module and the neuron impulse event processing module. The glial communication event processing module and the neuron impulse event processing module are shown in fig. 12 and fig. 13, respectively.
The specific calculation formula of the colloid communication event processing module is as follows:
wherein the content of the first and second substances,
indicating the strength of the period of the gum communication event,
represents a weight parameter in the glial cell-glial cell line parameter,
representing the statistical period of glial communication events by glial cells,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
The specific calculation formula of the neuron pulse event processing module is as follows:
wherein the content of the first and second substances,
represents the intensity of the period of the neuron impulse event,
represents a weight parameter in the neuron-glial cell junction parameter,
indicating glial cell versus neuronal vesselsThe statistical period of the flushing event is,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
The specific calculation formula of the glial cell state machine is as follows:
and configuring corresponding model parameters according to the selected dynamic model. The type of the model and the adjustment of the parameters thereof are the most critical and difficult steps in the invention, and the adjustment is related to the learning and memory effects of the whole impulse neural network. The glial cell kinetic model in the method needs to set the following parameters of the glial cell kinetic model:
。
example 2
As shown in fig. 14, the present invention also provides a pulsed neural network optimization system based on brain glial cells, the system comprising:
the impulse neural network construction module 141 is used for constructing an impulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron.
A connection parameter determination module 142 for determining various types of connection parameters.
And the optimizing module 143 is configured to optimize the spiking neural network according to the various types of connection parameters, so as to obtain an optimized spiking neural network.
As an optional implementation manner, the optimization module 143 of the present invention specifically includes:
the judging unit is used for judging whether an event occurs or not; if an event occurs, executing an event type determining unit; and if no event occurs, continuously judging whether the event occurs or not until a set condition is met, and outputting a pulse neural network fused with the glial cells.
An event type determining unit for determining an event type; when the event type is a neuron pulse event, executing a first connection parameter determining unit; when the event type is a glial cell communication event, then a "second update unit" is performed; when the event type is a glial cell regulating event, the "second connection parameter determining unit" is executed.
The first connection parameter determining unit is used for searching the event source neuron related connection table and determining the neuron fan-out connection type; when the neuron fan-out connection type is neuron-neuron, reading neuron-neuron connection parameters; when the neuron fan-out connection type is neuron-glial cell, reading the connection parameters of the neuron-glial cell.
The first updating unit is used for taking the neuron-neuron connection parameters as neuron input, updating the state parameters and output of the target neuron and returning to the judging unit; and inputting the connecting parameters of the neuron and the glial cell into the glial cell kinetic model, updating the state parameters and the output of the target glial cell, and returning to a judgment unit.
The second updating unit is used for checking the glial cell fan-out connection table of the event source and reading the glial cell-glial cell connection parameter; and inputting the connecting parameters of the glial cells into the glial cell kinetic model, updating the state parameters of the target glial cells, and returning to the judgment unit.
The second connection parameter determining unit is used for checking an event source glial cell fan-out connection table and determining the type of the glial cell fan-out connection; when the type of the glial cell fan-out connection is a glial cell-neuron synapse, reading a connecting parameter of the glial cell-neuron synapse; when the glial cell fan-out junction type is glial cell-neuron, then the glial cell-neuron junction parameters are read.
A third updating unit, configured to update the target synapse state parameter according to the glial cell-neuronal synapse connection parameter, and return to the "determining unit"; and updating the state parameter of the target neuron according to the connecting parameter of the glial cell-neuron, and returning to the judgment unit.
The same parts as those in embodiment 1 will not be described in detail.
The technical scheme disclosed by the invention has the following technical effects:
1. the network structure plasticity can be realized:
the glial-neuronal synaptic network established by the invention not only can realize the weight adjustment of synapses meeting plasticity rules, but also can realize the adjustment of transmission delay and signal types of the synapses, and can add synapses and delete synapses. The classification of the plasticity function is shown in fig. 15.
2. The network node plasticity, namely neuron plasticity can be realized:
the glial-neuron network established by the invention can realize the adjustment of the neuron state in accordance with plasticity rules, and the LIF neuron is taken as an example, and the current membrane potential, resting potential, time constant, threshold voltage and the like of the LIF neuron can be adjusted. In addition, the neurons which accord with plasticity rules can be directly deleted or new neurons can be generated.
3. Global feedback unsupervised learning based on global neural network activity can be realized:
due to the feedback processing node of the glial cell and the feedback loop of the neuron-glial network, not only can a plastic neural network based on feedback regulation of local neural network activity be designed, but also a plastic neural network based on feedback regulation of global neural network activity can be designed.
4. The supervised feedback learning of the pulse neural network can be realized:
because of the feedback processing node of the glial cell and the feedback loop of the neuron-glial network, the flexible construction mode of the feedback network and the customized glial cell dynamics model, supervised feedback learning can be designed (the expected value is defined as a group of pulse sequences with specific frequency in the SNN, and the characteristic frequency can be added into characteristic statistics of the glial cell). The SNN training method is shown in fig. 16.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.
The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.
Claims (10)
1. A method for brain glial cell-based impulse neural network optimization, the method comprising:
step S1: constructing a pulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron;
step S2: determining various types of connection parameters;
step S3: and optimizing the impulse neural network according to the various types of connection parameters to obtain the optimized impulse neural network.
2. The method according to claim 1, wherein the optimizing the spiking neural network according to the various types of connection parameters to obtain an optimized spiking neural network comprises:
step S31: judging whether an event occurs or not; if an event occurs, "step S32" is executed; if no event occurs, continuously judging whether the event occurs or not until a set condition is met, and outputting a pulse neural network fused with the glial cells;
step S32: determining the event type; when the event type is a neuron impulse event, executing step S33; when the event type is a glial cell communication event, then "step S35" is performed; when the event type is the glial cell conditioning event, then "step S36" is performed;
step S33: looking up an event source neuron related connection table, and determining a neuron fan-out connection type; when the neuron fan-out connection type is neuron-neuron, reading neuron-neuron connection parameters; when the type of the neuron fan-out connection is neuron-glial cell, reading the connection parameters of the neuron-glial cell;
step S34: taking the neuron-neuron connection parameters as neuron input, updating the target neuron state parameters and output, and returning to the step S31; inputting the neuron-glial cell connection parameters into the glial cell kinetic model, updating the target glial cell state parameters and output, and returning to step S31;
step S35: looking up an event source glial cell fan-out connection table, and reading a glial cell-glial cell connection parameter; inputting the glial cell-glial cell connection parameter into the glial cell kinetic model, updating the target glial cell state parameter, and returning to step S31;
step S36: looking up an event source glial cell fan-out connection table, and determining a glial cell fan-out connection type; when the type of the glial cell fan-out connection is a glial cell-neuron synapse, reading a connecting parameter of the glial cell-neuron synapse; when the type of the glial cell fan-out connection is glial cell-neuron, reading a glial cell-neuron connection parameter;
step S37: updating the target synapse state parameter according to the glial cell-neuronal synapse connection parameter, and returning to step S31; the target neuron state parameter is updated according to the glial cell-neuron connection parameter, and the process returns to step S31.
3. The method of brain glial cell-based impulse neural network optimization according to claim 2, wherein the glial cell kinetic model comprises:
the system comprises a glial communication event processing module, a neuron pulse event processing module and a glial cell state machine; the glial cell state machine is respectively connected with the glial communication event processing module and the neuron impulse event processing module.
4. The method for brain glial cell-based impulse neural network optimization according to claim 3, wherein the specific calculation formula of the glial communication event processing module is as follows:
wherein the content of the first and second substances,
indicating the strength of the period of the gum communication event,
represents a weight parameter in the glial cell-glial cell line parameter,
representing the statistical period of glial communication events by glial cells,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
5. The method for optimizing the spiking neural network based on the brain glial cells according to claim 4, wherein the specific calculation formula of the neuron spiking event processing module is as follows:
wherein the content of the first and second substances,
represents the intensity of the period of the neuron impulse event,
represents a weight parameter in the neuron-glial cell junction parameter,
representing the statistical period of glial cell versus neuronal impulse events,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
6. A brain glial cell-based spiking neural network optimization system, the system comprising:
the impulse neural network construction module is used for constructing an impulse neural network fusing the brain glial cells based on the brain glial cells; the spiking neural network fused with the brain glial cells comprises five types of connections, namely neuron-neuron, neuron-glial cell, glial cell-neuron synapse, glial cell-glial cell and glial cell-neuron;
the connection parameter determining module is used for determining various types of connection parameters;
and the optimization module is used for optimizing the impulse neural network according to the various types of connection parameters to obtain the optimized impulse neural network.
7. The brain glial cell-based spiking neural network optimization system according to claim 6, wherein the optimization module specifically comprises:
the judging unit is used for judging whether an event occurs or not; if an event occurs, executing an event type determining unit; if no event occurs, continuously judging whether the event occurs or not until a set condition is met, and outputting a pulse neural network fused with the glial cells;
an event type determining unit for determining an event type; when the event type is a neuron pulse event, executing a first connection parameter determining unit; when the event type is a glial cell communication event, then a "second update unit" is performed; when the event type is a glial cell regulating event, executing a second connection parameter determining unit;
the first connection parameter determining unit is used for searching the event source neuron related connection table and determining the neuron fan-out connection type; when the neuron fan-out connection type is neuron-neuron, reading neuron-neuron connection parameters; when the type of the neuron fan-out connection is neuron-glial cell, reading the connection parameters of the neuron-glial cell;
the first updating unit is used for taking the neuron-neuron connection parameters as neuron input, updating the state parameters and output of the target neuron and returning to the judging unit; inputting the neuron-glial cell connection parameters into a glial cell kinetic model, updating the state parameters and output of target glial cells, and returning to a judgment unit;
the second updating unit is used for checking the glial cell fan-out connection table of the event source and reading the glial cell-glial cell connection parameter; inputting connecting parameters of the glial cells into the glial cell kinetic model, updating the state parameters of the target glial cells, and returning to a judgment unit;
the second connection parameter determining unit is used for checking an event source glial cell fan-out connection table and determining the type of the glial cell fan-out connection; when the type of the glial cell fan-out connection is a glial cell-neuron synapse, reading a connecting parameter of the glial cell-neuron synapse; when the type of the glial cell fan-out connection is glial cell-neuron, reading a glial cell-neuron connection parameter;
a third updating unit, configured to update the target synapse state parameter according to the glial cell-neuronal synapse connection parameter, and return to the "determining unit"; and updating the state parameter of the target neuron according to the connecting parameter of the glial cell-neuron, and returning to the judgment unit.
8. The brain glial cell-based spiking neural network optimization system according to claim 7, wherein the glial cell kinetic model comprises:
the system comprises a glial communication event processing module, a neuron pulse event processing module and a glial cell state machine; the glial cell state machine is respectively connected with the glial communication event processing module and the neuron impulse event processing module.
9. The brain glial cell-based impulse neural network optimization system of claim 8, wherein the specific calculation formula of the glial communication event processing module is:
wherein the content of the first and second substances,
indicating the strength of the period of the gum communication event,
represents a weight parameter in the glial cell-glial cell line parameter,
representing the statistical period of glial communication events by glial cells,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
10. The glial cell line-based spiking neural network optimization system according to claim 9, wherein the specific calculation formula of the neuron spiking event processing module is:
wherein the content of the first and second substances,
represents the intensity of the period of the neuron impulse event,
represents a weight parameter in the neuron-glial cell junction parameter,
indicating glial cell versus neuronal vesselsThe statistical period of the flushing event is,
indicating the number of colloid communication events received by the connection during the current statistical period,
representing the intermediate parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111471537.2A CN113869505A (en) | 2021-12-06 | 2021-12-06 | Pulse neural network optimization method and system based on brain glial cells |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111471537.2A CN113869505A (en) | 2021-12-06 | 2021-12-06 | Pulse neural network optimization method and system based on brain glial cells |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113869505A true CN113869505A (en) | 2021-12-31 |
Family
ID=78985879
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111471537.2A Pending CN113869505A (en) | 2021-12-06 | 2021-12-06 | Pulse neural network optimization method and system based on brain glial cells |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113869505A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115310871A (en) * | 2022-10-09 | 2022-11-08 | 中科南京智能技术研究院 | Water resource allocation system |
| CN115547042A (en) * | 2022-09-20 | 2022-12-30 | 中科南京智能技术研究院 | Intelligent control system and method for large-scale urban road traffic lights |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109711531A (en) * | 2018-11-24 | 2019-05-03 | 浙江大学 | The artificial neural network optimization method of neuron interaction mechanism in a kind of consideration layer |
-
2021
- 2021-12-06 CN CN202111471537.2A patent/CN113869505A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109711531A (en) * | 2018-11-24 | 2019-05-03 | 浙江大学 | The artificial neural network optimization method of neuron interaction mechanism in a kind of consideration layer |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115547042A (en) * | 2022-09-20 | 2022-12-30 | 中科南京智能技术研究院 | Intelligent control system and method for large-scale urban road traffic lights |
| CN115310871A (en) * | 2022-10-09 | 2022-11-08 | 中科南京智能技术研究院 | Water resource allocation system |
| CN115310871B (en) * | 2022-10-09 | 2023-01-06 | 中科南京智能技术研究院 | Water resource allocation system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US9881251B2 (en) | Structural plasticity in spiking neural networks with symmetric dual of an electronic neuron | |
| US9697462B1 (en) | Synaptic time multiplexing | |
| CN113869505A (en) | Pulse neural network optimization method and system based on brain glial cells | |
| CN105719000B (en) | A kind of neuron hardware unit and the method with this unit simulation impulsive neural networks | |
| US8250011B2 (en) | Autonomous learning dynamic artificial neural computing device and brain inspired system | |
| US8914315B2 (en) | Multi-compartment neuron suitable for implementation in a distributed hardware model by reducing communication bandwidth | |
| WO2015020802A2 (en) | Computed synapses for neuromorphic systems | |
| KR20160125967A (en) | Method and apparatus for efficient implementation of common neuron models | |
| KR20160084401A (en) | Implementing synaptic learning using replay in spiking neural networks | |
| CN105701540B (en) | A kind of self-generating neutral net construction method | |
| US20190012597A1 (en) | Method and A System for Creating Dynamic Neural Function Libraries | |
| CN110232443A (en) | Realize the impulsive neural networks Digital Analog Hybrid Circuits system of liquid condition machine | |
| CN114202068A (en) | Self-learning implementation system for brain-like computing chip | |
| KR102535635B1 (en) | Neuromorphic computing device | |
| Vaario et al. | Toward evolutionary design of autonomous systems | |
| Vivekanand et al. | Robot Locomotion Control Using Central Pattern Generator with Non-linear Bio-mimetic Neurons | |
| Mercado et al. | Generation of virtual creatures under multidisciplinary biological premises | |
| Çakal | DynapSIM: A fast, optimizable, and mismatch aware mixed-signal neuromorphic chip simulator | |
| Fang et al. | Achieving swarm intelligence with spiking neural oscillators | |
| RU2484527C1 (en) | Simulator for self-forming networks of informal neurons | |
| CN115392460A (en) | Pulsed neural network simulation system and method fusing brain glial cells | |
| Venkatesh et al. | A high level Petri net model of olfactory bulb | |
| Suh | The context-aware learning model | |
| Vahie | Dynamic neuronal ensembles: neurobiologically inspired discrete event neural networks | |
| Hassan | On behavioral dynamics evaluation of glial cells role in comparison with brain neurons functions using artificial neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20211231 |
|
| RJ01 | Rejection of invention patent application after publication |