Fast memory encoding method and device based on multi-synaptic plasticity spiking neural network

Abstract

The present invention proposes a fast memory encoding method based on a multi-synaptic plasticity pulse neural network, including: step one: converting external stimuli into an input pulse sequence based on a hierarchical encoding strategy; step two: after the pulse neural network receives the input pulse, the membrane potential of the output layer neurons is updated based on the improved SRM model; step three: using the supervised group Tempotron to update the synaptic weights between the input and the output layer, and activating the output layer neurons to memorize the input; step four: after the output layer neurons are activated, using unsupervised STDP to update the synaptic weights between the activated neurons in the layer to form an enhanced recurrent subnetwork to store memory; step five: while executing step four, using unsupervised inhibition of synaptic plasticity to update the synaptic weights between the inhibition layer and the output layer, and inhibiting feedback to ensure the separation of the firing time of neural groups with different inputs. The present invention also proposes a fast memory encoding device based on a multi-synaptic plasticity pulse neural network. The present invention effectively improves the encoding speed and stability of memory.

Description

Technical Field

[0001] The present invention belongs to the field of brain-like intelligence and artificial intelligence, and relates to a method and device for fast memory encoding based on a multi-synaptic plasticity pulse neural network. Background Art

[0002] The Spiking Neural Network (SNN) simulates the information processing mechanism of the biological nervous system. It drives the calculation through discrete pulse events. Compared with traditional artificial neural networks, it has advantages such as low power consumption and stronger information expression ability. It is a powerful tool for analyzing and simulating brain cognitive functions. Memory, as a cognitive ability to recognize, retain and reproduce things experienced, is one of the most core components of brain intelligence.

[0003] Synaptic plasticity has long been considered the basis of learning and memory. Neuroscience research shows that the biological brain relies on the coordination of multiple synaptic plasticities to ensure the reliable implementation of cognitive tasks. However, due to the lack of understanding of the interaction between different forms of synaptic plasticity rules and the complexity of the structure and dynamics of neural networks, most of the existing spike memory models only use unsupervised synaptic plasticity rules (such as spike-timing-dependent-plasticity (STDP)) to train recurrent networks. By adjusting the synaptic weights between activated neurons in the recurrent network, a recurrent sub-network with enhanced connections is formed to store memory. The memory of external input patterns is encoded by the collective co-emission activity of the neural population. However, the learning process of the memory model based on unsupervised plasticity rules is slow, and it requires repeated input of perceptual stimuli to generate neural populations to express memory, and the memory efficiency is low. In addition, the memory model based on unsupervised plasticity rules cannot guarantee the stable generation of neural populations. During the learning process, interference may occur between neural populations that respond to different input patterns, thereby destroying memory. Summary of the invention

[0004] In order to solve the problem of low memory encoding efficiency and instability of the current pulse memory model based on unsupervised plasticity in the prior art, the present invention proposes a method and device for fast memory encoding based on a multi-synaptic plasticity pulse neural network, and its specific technical solution is as follows:

[0005] A spiking neural network fast memory encoding method based on multi-synaptic plasticity comprises the following steps:

[0006] Step 1: Convert external stimuli into input pulse trains based on a hierarchical coding strategy;

[0007] Step 2: After the spiking neural network receives the input pulse, it updates the membrane potential of the output layer neurons based on the improved SRM model;

[0008] Step 3: Use the supervised group Tempotron to update the synaptic weights between the input and output layers and activate the output layer neurons to memorize the input;

[0009] Step 4: After the output layer neurons are activated, the synaptic weights between the activated neurons in the update layer are updated using unsupervised STDP to form an enhanced recurrent sub-network storage memory;

[0010] Step 5: While executing step 4, use unsupervised inhibitory synaptic plasticity to update the synaptic weights from the inhibitory layer to the output layer, and inhibit feedback to ensure the separation of the firing times of neural groups that remember different inputs.

[0011] Preferably, the step 1 is specifically:

[0012] The external stimulus is seven images randomly selected from the seven categories of '0', '1', '2', '3', '4', '5', and '6' of the MNIST handwritten digit set. The size of each image is 28 pixels * 28 pixels. Each image is evenly divided to obtain 49 receiving fields of 4 pixels * 4 pixels.

[0013] The hierarchical coding strategy uses a two-layer structure consisting of a simple cell S layer and a complex cell C layer to extract features of external stimuli, and then encodes the extracted features through a delayed coding method, where:

[0014] S layer: extracts the edge direction features of the image. The filter kernel of the S layer consists of four directions: , , , The Gabor filters in the four directions simulate the receptive field of the visual cortex to perform filtering operations on the image blocks corresponding to each receiving field, and obtain 49*4 feature maps with a size of 4pixel *4 pixel after filtering;

[0015] C layer: first perform an addition operation on each feature map output by the S layer to obtain a feature map of size 49 pixel * 4 pixel, perform linear normalization on the feature map to distribute the data to the interval [0, 1], and then perform a maximum competition operation on a row basis. The maximum competition operation on a row basis includes: retaining the strongest feature in each row as the most matching directional feature at that position, and setting the feature values ​​of other directions to 0, to obtain a sparse expression of a feature map of size 4 pixel 9*4 pixel;

[0016] The obtained features are converted into input spike trains through delay coding and calculated according to the following formula:

[0017]

[0018] in, is the firing time of the ith neuron, T is the length of the encoding window, is the intensity value of the ith feature.

[0019] Preferably, the step 2 is specifically as follows:

[0020] The improved SRM model introduces an afterdepolarization potential in the refractory nucleus and The oscillation is injected into the output layer neurons as an external input current. In the pulse neural network, when one of the neurons in the output layer emits a pulse, the pulse will be transmitted to other neurons in the same layer and the inhibitory layer neurons. The membrane potential of the output layer neurons in the model receives excitation input from the input layer and other neurons in the same layer and inhibitory feedback from the inhibitory neurons. The membrane potential of the output layer neurons in the model is calculated according to the following formula:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] in, is the membrane potential of neuron i, is the total input of the neurons in the input layer, is the total input of neurons in the same layer, is the inhibitory feedback, is the neuron i in The afterdepolarization process of the membrane potential after the pulse is released. yes Oscillation source, , , They are the synaptic weights from input layer neurons to output layer neurons, neurons within the output layer, and inhibitory neurons to output layer neurons. is the normalization factor, , , They are inhibitory feedback, afterdepolarization, The amplitude of the oscillation, , are the pulse emission times of presynaptic neuron j and postsynaptic neuron i, respectively. , , , is the time constant, and f is The frequency of the oscillation, t is the time variable, is the initial phase.

[0028] Preferably, the supervisory group Tempotron in step 3 updates the synaptic weights between the input layer and the output layer, and the updating method is:

[0029]

[0030] in, is the weight of the connection from the input layer to the output layer, is the learning rate of the inter-layer weights, d is the desired output label including 0 or 1, is the pulse firing time of presynaptic neuron j, K is the number of neurons that need to be activated to reach the neural population size, is the moment corresponding to the maximum membrane potential of the kth most active postsynaptic neuron.

[0031] Preferably, the unsupervised STDP in step 4 updates the synaptic weights of activated neurons in the layer, and the updating method is:

[0032]

[0033] in, is the update amount of the connection weights in the output layer, and represent the amplitude of enhancement and attenuation respectively, and represent the pulse emission time of the presynaptic neuron j and the postsynaptic neuron i, respectively. , is the time constant, represents the decay time constant of NMDA receptor.

[0034] Preferably, the updating method for inhibiting synaptic plasticity in step 5 is:

[0035]

[0036] in, is the update amount of the connection weights from the suppression layer to the output layer, represents the pulse firing time of postsynaptic neuron i.

[0037] A fast memory encoding device based on a multi-synaptic plasticity pulse neural network includes one or more processors for implementing the fast memory encoding method based on a multi-synaptic plasticity pulse neural network.

[0038] A computer-readable storage medium stores a program, which, when executed by a processor, implements the fast memory encoding method based on a multi-synaptic plasticity pulse neural network.

[0039] Beneficial effects of the present invention:

[0040] The synergy of supervised group Tempotron, unsupervised STDP and unsupervised inhibitory plasticity proposed in this invention provides a feasible and efficient way for spiking neural networks to quickly memorize external stimuli. Compared with existing spiking memory models, the synergy of supervised and unsupervised plasticity is more biologically interpretable and effectively improves the encoding speed and stability of memory. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of a computational framework of a fast memory encoding method based on a multi-synaptic plasticity spiking neural network according to an embodiment of the present invention;

[0042] Figure 2 It is a flow chart of a method for fast memory encoding based on a multi-synaptic plasticity spiking neural network according to an embodiment of the present invention;

[0043] Figure 3 It is a schematic diagram of the external stimulus encoding process;

[0044] Figure 4 It is a schematic diagram of the group Tempotron learning strategy;

[0045] FIG5 (a) is a diagram of an embodiment of the present invention, in which the input pulse sequence after encoding each image in the spiking neural network is respectively in the order of '0', '1', '2', '3', '4', '5', and '6'. Schematic diagram of the input network at the trough of different periods of oscillation;

[0046] 5(b) is a schematic diagram of activities corresponding to '0', '1', '2', '3', '4', '5', '6', '0', '01', '012', '0123', '01234', '012345', '0123456' of an embodiment of the present invention;

[0047] Figure 6 (a) is a graph showing the change in synaptic weights between layers; Figure 6 (b) is a graph showing the change in synaptic weights within a layer;

[0048] Figure 7 It is a schematic diagram of the convergence speed of inter-layer synaptic weights;

[0049] Figure 8 It is a structural block diagram of a fast memory encoding device based on a multi-synaptic plasticity pulse neural network provided by the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and technical effect of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0051] In order to solve / alleviate the above-mentioned technical problems, the present invention constructs a spiking neural network model with full connections between layers and recurrent connections within the layers based on supervised group Tempotron and unsupervised STDP and inhibitory synaptic plasticity, and verifies the model function on a digital sequence memory task. The results show that the invention can effectively improve the encoding speed and stability of memory.

[0052] like Figure 1 As shown in the figure, it is a computational framework of a fast memory encoding method of a spiking neural network based on multi-synaptic plasticity, which includes an input layer, an output layer, and an inhibitory layer. The neurons in the input layer are fully connected to the neurons in the output layer, and the input pulses are transmitted to the output layer via the input layer. The synaptic weights between the input layer and the output layer are updated by the supervised group Tempotron; there are recurrent connections in the output layer, and the synaptic weights in the output layer are updated by unsupervised STDP; there are bidirectional connections between the inhibitory layer and the output layer, in which the output layer and the inhibitory layer are connected in a one-to-one manner, and the synaptic weights of this part of the connection are fixed and do not participate in weight training, while the inhibitory layer and the output layer are fully connected in a diagonal non-connected manner, and the synaptic weights are updated by the unsupervised inhibitory plasticity rule.

[0053] The fast memory encoding method based on multi-synaptic plasticity pulse neural network is as follows: Figure 2 As shown, the specific steps include:

[0054] Step 1: Convert external stimuli into input pulse trains based on a hierarchical coding strategy;

[0055] The real-valued data of external stimulation is first encoded into a pulse pattern before being input into the spiking neural network. The external stimulation of this embodiment is seven images randomly selected from the seven categories of '0', '1', '2', '3', '4', '5', and '6' in the MNIST handwritten digit set. The image size is 28 pixel * 28 pixel, which is divided into 49 receiving fields of 4 pixel * 4 pixel. Figure 3The encoding process of the number '0' is shown. First, its features are extracted through a two-layer structure consisting of a simple cell, the S layer, and a complex cell, the C layer. Then, the extracted features are encoded using a delayed encoding method, where:

[0056] S layer: extracts the edge direction features of the image. The filter kernels of the S layer are divided into two directions: , , , It is composed of 4 Gabor filters, which simulate the receptive field of the visual cortex to perform filtering operations on the receptive field image, and obtain a 49*4 feature map with a size of 4 pixel *4 pixel after filtering;

[0057] C layer: First, perform an addition operation on each feature map output by the S layer to obtain a feature map of size 49*4, perform linear normalization on the feature map to distribute the data to the interval [0, 1], and then perform a maximum competition operation on each row, that is, only the strongest feature is retained as the most matching directional feature at that position in each row, and the feature values ​​in other directions are set to 0, thus obtaining a sparse expression of a feature map of size 49*4.

[0058] The obtained features are converted into input spike trains through delay coding and calculated according to the following formula:

[0059]

[0060] in, is the firing time of the ith neuron, T=20e-3 is the length of the encoding window, is the intensity value of the ith feature.

[0061] The pulse sequence of each image is 49*4=196. Different images are represented by different pulse sequences. The encoded pulse sequence enters the network through the input layer. The total number of neurons in the input layer is 7*49*4=1372.

[0062] Step 2: After receiving the input pulse, the membrane potential of the output layer neurons is updated based on the improved SRM model;

[0063] In this embodiment, the neurons of the spiking neural network adopt an improved SRM model, which provides a concise description of the dynamics of spiking neurons through continuous kernel functions, which is conducive to clearly showing the contribution of multiple input sources to neuronal firing. Compared with the original SRM model, the improved SRM model takes into account more physiological details and introduces the post-depolarization potential in the insensitive nucleus to describe the process of the membrane potential slowly rising after the neuron fires a pulse:

[0064]

[0065] is the neuron i in The afterdepolarization process of the membrane potential after the pulse is released. is the amplitude of the afterdepolarization, is the pulse firing time of neuron i, is the time constant.

[0066] Introduced The important brain wave oscillating this synchronized neural activity, injected as external input current into the output neurons, is modeled as a cosine wave.

[0067]

[0068] in, yes Oscillation source, yes The amplitude of the oscillation, yes The frequency of oscillation, is the initial phase. In the spiking neural network, the input pulse sequence after each image encoding is in the order of '0', '1', '2', '3', '4', '5', and '6'. The troughs of different oscillation periods are input into the network, as shown in Figure 5(a). The excitatory input received by the output layer neurons in the network from the input layer is expressed as follows:

[0069]

[0070] in, is the total input of the neurons in the input layer, is the synaptic weight between input layer neuron j and output layer neuron i, is the normalization factor, is the pulse firing time of presynaptic neuron j, , is the time constant.

[0071] When the output layer neuron emits a pulse, the pulse is transmitted to other neurons in the same layer, as well as neurons in the inhibitory layer. The excitation input received by the output layer neuron from other neurons in the same layer is expressed as follows:

[0072]

[0073] is the total input of neurons in the same layer, is the synaptic weight between the presynaptic neuron j and the postsynaptic neuron i in the output layer.

[0074] The inhibitory feedback received from the inhibitory neuron is expressed as follows:

[0075]

[0076] is the inhibitory feedback, is the synaptic weight between the inhibitory neuron and the output layer neuron, is the amplitude of the inhibitory feedback, is the pulse emission time of the output layer neuron, is the time constant.

[0077] In summary, the membrane potential of the output layer neurons in the model is calculated according to the following formula:

[0078]

[0079] In this embodiment, 3. , , ,, , , , ,Simulation time interval , the duration of one simulation iteration , the number of iterations is set to 20.

[0080] Step 3: Use the supervised group Tempotron to update the synaptic weights between the input and output layers, and quickly activate a group of output neurons to memorize the input;

[0081] like Figure 4 The figure shows a schematic diagram of the group Tempotron learning strategy. In each iterative learning process, for each input pulse pattern, if the number of neurons in the output layer that respond to the input pattern is still short of K to reach the preset neural population size, the synaptic weights between the spiking neurons in the input layer and the unspiking neurons corresponding to the first K maximum membrane potentials in the output layer will be strengthened. The inter-layer weight update formula is as follows:

[0082]

[0083] in, is the update amount of the connection weights from the input layer to the output layer, is the learning rate of the inter-layer weights, d is the desired output label including: 0 or 1, is the time when the presynaptic neuron j fires a pulse, K is the number of neurons that need to be activated to reach the size of the neural population, is the moment corresponding to the maximum membrane potential of the kth most active postsynaptic neuron.

[0084] In this embodiment, , the number of output neurons is 100, and the number of neurons in the neural population is 12. Before learning, the initial weights between layers are randomly initialized according to the normal distribution, with a mean of 0.001 and a standard deviation of 0.002. As shown in Figure 6 (a), after learning, for each input pattern, the weights between the firing neurons of the input layer and the response neural population of the output layer are enhanced. Due to the enhancement of the inter-layer synaptic weights, near the time when the input pulse pattern is presented, the output layer presents the neural population firing activity encoding the input, as shown in the activities corresponding to '0', '1', '2', '3', '4', '5', and '6' in Figure 5 (b).

[0085] Step 4: After the output layer neurons are activated, the synaptic weights between the activated neurons in the update layer are updated using unsupervised STDP to form an enhanced recurrent sub-network storage memory;

[0086] After the output layer neuron is activated, when the firing time difference between the presynaptic neuron and the postsynaptic neuron is [ ] interval, the synaptic weight will be updated. When the presynaptic neuron fires before the postsynaptic neuron, the synaptic weight is enhanced; when the presynaptic neuron fires later than the postsynaptic neuron, the synaptic weight is weakened. The formula for updating the weight within the layer is as follows:

[0087]

[0088] in, is the update amount of the connection weights in the output layer, and represent the amplitude of enhancement and attenuation respectively, and represent the pulse emission time of the presynaptic neuron j and the postsynaptic neuron i, respectively. , is the time constant, represents the decay time constant of NMDA receptor.

[0089] In this embodiment, , Before learning, the weights of the recurrent connections within the layer are randomly initialized, and the initial weights are in the interval [0, 1e-5]. After learning, the internal connections of the neuron population that responds to each input pattern in the input layer are enhanced to form a recurrent subnetwork, as shown in Figure 6 (b). The autoassociative memory is stored in the enhanced recurrent connection weights within the neural population.

[0090] Step 5: While executing step 4, use unsupervised inhibitory plasticity to update the synaptic weights from the inhibitory layer to the output layer. Inhibitory feedback ensures the separation of the firing times of neural groups that remember different inputs.

[0091] The output layer is connected to the inhibitory layer in a one-to-one manner. The synaptic weights of this part of the connection are fixed and do not participate in weight training. The inhibitory layer is connected to the output layer in a full connection manner without diagonal connections. The values ​​on the diagonal of the weight matrix are 0, and the rest of the values ​​are 1. The time difference between the current time t and the last firing of the output neuron is [ ], the inhibitory weights of neurons in the output layer that have input to inhibitory neurons will be reduced, while the inhibitory weights of neurons that have no input to inhibitory neurons remain unchanged. The update formula for inhibiting synaptic plasticity is as follows:

[0092]

[0093] in, is the update amount of the connection weights from the suppression layer to the output layer, represents the pulse emission time of postsynaptic neuron i. After the output layer neural population is stimulated to emit pulses, the excitatory input from the neural population in the same layer, the depolarizing current after discharge, and the external Oscillatory input enables the neural population to respond in the absence of subsequent stimulatory input. After the output layer neurons fire, the inhibitory feedback from the inhibitory layer neurons prevents the neural population from firing continuously after firing, thus avoiding mutual interference between different memory items, as shown in the activities corresponding to '0', '01', '012', '0123', '01234', '012345', '0123456' in Figure 5 (b). Near the peak of the oscillation cycle, neural populations responding to different input stimuli fire in sequence at different times, and inhibitory feedback ensures the temporal separation of firing of neural populations that remember different inputs.

[0094] In this embodiment, the number of inhibitory neurons is 100, which is the same as the number of output neurons. .

[0095] like Figure 7 The figure shows the convergence speed of the inter-layer synaptic weights. Under the group Tempotron learning strategy, a group of output neurons are quickly activated to encode the input memory, and the inter-layer synaptic weights converge after only 4 learnings. The experimental results show that the fast memory encoding method of the pulse neural network based on multi-synaptic plasticity proposed in the present invention can effectively improve the encoding speed and stability of memory.

[0096] Corresponding to the aforementioned embodiment of the method for rapid memory encoding based on a multi-synaptic plasticity spiking neural network, the present invention also provides an embodiment of a device for rapid memory encoding based on a multi-synaptic plasticity spiking neural network.

[0097] See also Figure 8 An embodiment of the present invention provides a fast memory encoding device based on a multi-synaptic plasticity pulse neural network, which includes one or more processors for implementing the fast memory encoding method based on a multi-synaptic plasticity pulse neural network in the above embodiment.

[0098] The embodiments of the fast memory encoding device based on the multi-synaptic plasticity pulse neural network of the present invention can be applied to any device with data processing capability, and the device with data processing capability can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capability in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, if Figure 8 As shown, it is a hardware structure diagram of any device with data processing capability based on the multi-synaptic plasticity pulse neural network fast memory encoding device of the present invention, except Figure 8 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiments is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0099] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0100] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0101] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for fast memory encoding based on a multi-synaptic plasticity pulse neural network in the above embodiment is implemented.

[0102] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of a wind turbine, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0103] The above is only a preferred implementation case of the present invention and does not limit the present invention in any form. Although the implementation process of the present invention is described in detail above, for those familiar with the art, they can still modify the technical solutions recorded in the above examples, or replace some of the technical features therein with equivalents. All modifications, equivalent replacements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A fast memory encoding method based on a multi-synaptic plasticity pulse neural network, characterized in that: The following steps are involved: Step 1: Convert external stimuli into input pulse trains based on a hierarchical coding strategy; Step 2: After the spiking neural network receives the input pulse, it updates the membrane potential of the output layer neurons based on the improved SRM model; Step 3: Use the supervised group Tempotron to update the synaptic weights between the input and output layers and activate the output layer neurons to memorize the input; Step 4: After the output layer neurons are activated, the synaptic weights between the activated neurons in the update layer are updated using unsupervised STDP to form an enhanced recurrent sub-network storage memory; Step 5: While executing step 4, use unsupervised inhibition of synaptic plasticity to update the synaptic weights from the inhibition layer to the output layer, and inhibit feedback to ensure the separation of the firing time of the neural groups that remember different inputs; The step 1 is specifically as follows: The external stimulus is seven images randomly selected from the seven categories of '0', '1', '2', '3', '4', '5', and '6' of the MNIST handwritten digit set. The size of each image is 28 pixels * 28 pixels. Each image is evenly divided to obtain 49 receiving fields of 4 pixels * 4 pixels. The hierarchical coding strategy uses a two-layer structure consisting of a simple cell S layer and a complex cell C layer to extract features of external stimuli, and then encodes the extracted features through a delayed coding method, where: S layer: extracts the edge direction features of the image. The filter kernel of the S layer consists of four directions: , , , The Gabor filters in the four directions simulate the receptive field of the visual cortex to perform filtering operations on the image blocks corresponding to each receiving field, and obtain 49*4 feature maps with a size of 4pixel *4 pixel after filtering; C layer: first perform an addition operation on each feature map output by the S layer to obtain a feature map of size 49 pixel * 4 pixel, perform linear normalization on the feature map to distribute the data to the interval [0, 1], and then perform a maximum competition operation on a row basis. The maximum competition operation on a row basis includes: retaining the strongest feature in each row as the most matching directional feature at that position, and setting the feature values ​​of other directions to 0, to obtain a sparse expression of a feature map of size 4 pixel 9*4 pixel; The obtained features are converted into input spike trains through delay coding and calculated according to the following formula: ， in, is the firing time of the ith neuron, T is the length of the encoding window, is the intensity value of the i-th feature; The step 2 is specifically as follows: The improved SRM model introduces an afterdepolarization potential in the refractory nucleus and The oscillation is injected into the output layer neurons as an external input current. In the pulse neural network, when one of the neurons in the output layer emits a pulse, the pulse will be transmitted to other neurons in the same layer and the inhibitory layer neurons. The membrane potential of the output layer neurons in the model receives excitation input from the input layer and other neurons in the same layer and inhibitory feedback from the inhibitory neurons. The membrane potential of the output layer neurons in the model is calculated according to the following formula: ， in, is the membrane potential of neuron i, is the total input of the neurons in the input layer, is the total input of neurons in the same layer, is the inhibitory feedback, is the neuron i in The afterdepolarization process of the membrane potential after the pulse is released. yes Oscillation source, , , They are the synaptic weights from input layer neurons to output layer neurons, neurons within the output layer, and inhibitory neurons to output layer neurons. is the normalization factor, , , They are inhibitory feedback, afterdepolarization, The amplitude of the oscillation, , are the pulse emission times of presynaptic neuron j and postsynaptic neuron i, respectively. , , , is the time constant, and f is The frequency of the oscillation, t is the time variable, is the initial phase.

2. The fast memory encoding method based on multi-synaptic plasticity pulse neural network according to claim 1 is characterized in that: The supervisory group Tempotron in step 3 updates the synaptic weights between the input layer and the output layer, and the updating method is: ， in, is the weight of the connection from the input layer to the output layer, is the learning rate of the inter-layer weights, d is the desired output label including 0 or 1, is the pulse firing time of presynaptic neuron j, K is the number of neurons that need to be activated to reach the neural population size, is the moment corresponding to the maximum membrane potential of the kth most active postsynaptic neuron.

3. The fast memory encoding method based on multi-synaptic plasticity pulse neural network according to claim 1 is characterized in that: The unsupervised STDP in step 4 updates the synaptic weights of activated neurons in the layer, and the updating method is: ， in, is the update amount of the connection weights in the output layer, and represent the amplitude of enhancement and attenuation respectively, and represent the pulse emission time of the presynaptic neuron j and the postsynaptic neuron i, respectively. , is the time constant, represents the decay time constant of NMDA receptor.

4. The fast memory encoding method based on multi-synaptic plasticity pulse neural network according to claim 1 is characterized in that: The updating method for inhibiting synaptic plasticity in step 5 is: ， in, is the update amount of the connection weights from the suppression layer to the output layer, represents the pulse firing time of postsynaptic neuron i.

5. A fast memory encoding device based on a multi-synaptic plasticity pulse neural network, characterized in that: It comprises one or more processors for implementing the fast memory encoding method based on multi-synaptic plasticity pulse neural network according to any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the fast memory encoding method based on a multi-synaptic plasticity pulse neural network described in any one of claims 1 to 4 is implemented.

Images