Spintronic neuron devices and related methods, devices, and electronic devices

Abstract

This application proposes a spintronic neuron device, modeling method, and spiking neural network training method. The spintronic neuron device comprises two electrodes at two ends, a transmission component composed of a multilayer film located between the two electrodes, and multiple output components deployed at different locations within the transmission component. This allows different output components to be selected for operation, changing the transmission path length of magnetic domain wall motion and automatically adjusting the emission threshold of the spintronic neuron device. By modeling the biomimetic behavior of the device's characteristics and utilizing the resulting neuron models with different performance and mechanisms, a spiking neural network for recognition tasks is constructed. The performance of the spiking neural network can be optimized by varying the number of neuron models, the threshold adjustment rate of the emission threshold, and / or the synaptic weights. This approach eliminates the need for peripheral circuitry, achieves homeostasis within the spiking neural network, and improves its stability and robustness to varying inputs.

Description

Technical Field

[0001] The present application mainly relates to the field of artificial intelligence, and in particular to a spintronic neuron device and related methods, devices and electronic equipment. Background Art

[0002] Spiking neural networks (SNs) achieve the information processing and computational behavior of biological SNs by interconnecting artificial neurons. They convert input samples into pulse sequences with varying delays and carrying the sample's characteristics, transmitting them to multiple layers of neurons connected by synapses. Finally, by observing the neurons' behavior, the processing results for the input samples are obtained. This information processing approach makes it well-suited for implementation in neuromorphic hardware, providing sparse yet powerful computing power to meet task processing requirements.

[0003] For the spiking neural network hardware circuit, artificial neuron devices with different mechanisms are used as computing nodes in the spiking neural network to simulate the information encoding and processing process of the human brain and train the spiking neural network for recognition tasks. In this training process, in order to avoid overfitting, it is necessary to achieve homeostasis of the spiking neural network to ensure that the response frequency of each neuron is within a specific range, thereby preventing some neurons from responding too much or too little, and ensuring that neurons can respond quickly to changes in input signals.

[0004] Current artificial neuron devices are typically implemented with peripheral circuits or software algorithms, which increases overall network power consumption and latency. Furthermore, the resulting hardware spiking neural networks have long training cycles and high optimization costs. Summary of the Invention

[0005] In order to solve the above problems, this application proposes the following technical solutions:

[0006] This application also proposes a spintronic neuron device, comprising:

[0007] a first electrode;

[0008] a second electrode;

[0009] a transmission component connected to the first electrode and the second electrode respectively, wherein the transmission component is composed of a multilayer film;

[0010] a plurality of output components disposed at different locations within the transmission component;

[0011] Wherein, when different output components are in working states, the transmission path lengths of the magnetic domain walls injected through the first electrode or the second electrode moving in the transmission component are different, so that the spintronic neuron device has different emission thresholds.

[0012] Optionally, the plurality of output components are distributed in an array on the transmission component, and the relative positions between the plurality of output components are adjustable;

[0013] Wherein, when the magnetic domain wall moves to the position of the output component in the working state in the transmission component, it is detected and stops moving to the next output component distributed in the array.

[0014] This application also proposes a spintronic neuron device modeling method, including:

[0015] Obtaining a selection instruction for a plurality of output components in the spintronic neuron device as described above, and determining a target output component to enter a working state; different output components correspond to different magnetic domain wall transmission path lengths;

[0016] Performing electrical signal input control on the spintronic neuron device to trigger the generated magnetic domain wall motion and obtain corresponding neuron physical properties;

[0017] The spintronic neuron device is modeled according to the different neuron physical properties of the multiple output components to obtain a neuron model of a corresponding type.

[0018] Optionally, the electrical signal input control of the spintronic neuron device to trigger the generated magnetic domain wall motion and obtain corresponding neuron physical properties includes:

[0019] Controlling the spintronic neuron device with an electrical signal input triggers the generated magnetic domain wall forward and reverse motions, thereby mimicking the accumulation and leakage processes of neuron membrane potentials and obtaining corresponding accumulation and leakage rates;

[0020] Determining that the magnetic domain wall moves to the position of the target output component, detecting the neuron membrane potential when the target output component transmits a pulse signal, obtaining an emission threshold for the target output component, and acquiring refractory period information of the spintronic neuron device; the refractory period information can indicate that after detecting that the magnetic domain wall moves to the position of the target output component, it stops responding to the input electrical signal and moves in the opposite direction to the starting position;

[0021] The step of constructing a neuron model of a corresponding type for the spintronic neuron device based on different neuron physical properties of the plurality of output components includes:

[0022] According to the accumulation rate, the leakage rate, the emission threshold and the refractory period information of the multiple output components, corresponding characteristic bionic behaviors of the spintronic neuron device are modeled respectively to obtain corresponding types of neuron models.

[0023] This application also proposes a pulse neural network training method, which includes:

[0024] Obtaining a neural network to be trained and a sample data set for a recognition task; the spiking neural network is constructed from multiple different types of neuron models to achieve the recognition task, and the neuron models are obtained using the spintronic neuron device modeling method described above;

[0025] Training the spiking neural network based on the plurality of sample data included in the sample data set to optimize network parameters of the spiking neural network; wherein the network parameters include one or more of the number of neuron models, a threshold adjustment rate for a firing threshold of at least one neuron model, and synaptic weights between different network layers in the spiking neural network; the threshold adjustment rate is used to indicate adjustment of the firing threshold of the corresponding neuron model;

[0026] It is determined that the recognition accuracy of the spiking neural network with optimized network parameters for the sample data meets the training termination condition, and the spiking neural network finally obtained by training is determined as the spiking neural network model for implementing the recognition task.

[0027] Optionally, the spiking neural network includes an input layer, an excitatory neuron layer, an inhibitory neuron layer, and an output layer, wherein:

[0028] The excitatory neuron layer is connected to the input layer and the output layer by artificial synapses in a fully connected manner, and the excitatory neuron layer is connected to the inhibitory neuron layer by a complementary fully connected manner; the complementary fully connected manner means that in the excitatory neuron layer and the inhibitory neuron layer, two neuron models with corresponding relationships are connected point-to-point, and two neuron models without corresponding relationships are connected by a fully connected manner;

[0029] A corresponding synaptic weight is configured for each artificial synapse so that the synaptic weight can be adjusted through a pulse time-dependent plasticity (STDP) learning mechanism during the spiking neural network training process; wherein the synaptic weight of the artificial synapse between the inhibitory neuron layer and the excitatory neuron layer is less than zero; the synaptic weights of the artificial synapses between the excitatory neuron layer and the inhibitory neuron layer, between the input layer and the excitatory neuron layer, and between the excitatory neuron layer and the output layer are all greater than zero.

[0030] Optional, where:

[0031] In the case where some neuron models in the excitatory neuron layer transmit pulses, randomly selecting a pulse transmitted by one neuron model as a valid pulse;

[0032] transmitting the effective pulse to the inhibitory neuron layer to obtain an inhibitory response signal;

[0033] Feeding back the inhibitory response signal to another part of the neuron model in the excitatory neuron layer to inhibit the accumulation of the pulse neuron membrane potential of the neuron model;

[0034] If, before selecting the effective pulse, the pulse emitted by the neuron model that has already emitted a pulse is determined to be a valid pulse, the emission threshold of the neuron model is adjusted by selecting different output ports of the spintronic neuron device used to construct the neuron model to achieve the homeostasis of the pulse neural network.

[0035] The present application also proposes a neuron modeling device, comprising:

[0036] a target output component determination module, configured to obtain a selection instruction for a plurality of output components in the spintronic neuron device and determine a target output component to be put into operation; different output components correspond to different magnetic domain wall transmission path lengths;

[0037] a neuron physical property acquisition module, configured to control the electrical signal input to the spintronic neuron device, trigger the generated magnetic domain wall motion, and obtain corresponding neuron physical properties;

[0038] The neuron model construction module is used to construct a neuron model of a corresponding type for the spintronic neuron device according to different neuron physical properties for the multiple output components.

[0039] This application also proposes a pulse neural network training device, which includes:

[0040] An acquisition module, configured to acquire a neural network to be trained and a sample data set for a recognition task; the spiking neural network is constructed from a plurality of different types of neuron models to implement the recognition task, the neuron models being obtained using the spintronic neuron device modeling method described above;

[0041] A training module, configured to train the spiking neural network based on the plurality of sample data included in the sample data set to optimize the network parameters of the spiking neural network; wherein the network parameters include the number of the neuron models, a threshold adjustment rate for the emission threshold of at least one of the neuron models, and one or more of the synaptic weights between different network layers in the spiking neural network; the threshold adjustment rate is used to indicate the adjustment of the emission threshold of the corresponding neuron model.

[0042] The pulse neural network model determination module is used to determine whether the recognition accuracy of the pulse neural network with optimized network parameters for the sample data meets the training termination condition, and determine the pulse neural network finally trained as the pulse neural network model for implementing the recognition task.

[0043] The present application also proposes an electronic device, comprising:

[0044] Communication module;

[0045] A memory for storing a first program for implementing the above-mentioned spintronic neuron device modeling method and a second program for implementing the above-mentioned spiking neural network training method;

[0046] The processor is configured to load and execute the first program to implement the above-mentioned spintronic neuron device modeling method, or load and execute the second program to implement the above-mentioned pulse neural network training method. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.

[0048] Figure 1 This is a schematic structural diagram of an optional embodiment of the spintronic neuron device proposed in this application;

[0049] Figure 2a A schematic diagram of an optional working state of the spintronic neuron device proposed in this application;

[0050] Figure 2b This is a schematic diagram of another optional working state of the spintronic neuron device proposed in this application;

[0051] Figure 3 A schematic flow chart of an optional embodiment of the spintronic neuron device modeling method proposed in this application;

[0052] Figure 4 A schematic diagram of the biomimetic behavior of the spintronic neuron device proposed in this application;

[0053] Figure 5 Schematic diagram of LIF characteristics obtained based on the biomimetic behavior of the spintronic neuron device proposed in this application;

[0054] Figure 6 This is a flow chart of an optional embodiment of the spiking neural network training method proposed in this application;

[0055] Figure 7 Schematic diagram of the network structure of a spiking neural network suitable for the spiking neural network training method proposed in this application;

[0056] Figure 8 This is a flow chart of another optional embodiment of the spiking neural network training method proposed in this application;

[0057] Figure 9 A synaptic weight distribution diagram of each synaptic connection between the input layer and the excitatory neuron layer obtained after one training in the spiking neural network training method proposed in this application;

[0058] Figure 10 A graph showing the relationship between the recognition accuracy and the number of training times of the spiking neural network on the test set of the MNIST dataset, applicable to the spiking neural network training method proposed in this application;

[0059] Figure 11 A graph showing the relationship between recognition accuracy and training times on the test set of the MNIST dataset for a spiking neural network with different numbers of neurons in the excitatory neuron layer, suitable for the spiking neural network training method proposed in this application;

[0060] Figure 12 A graph showing the relationship between the recognition accuracy and the number of training times of multiple trainings of a spiking neural network under different threshold adjustment rates in the spiking neural network training method proposed in this application;

[0061] Figure 13 This is a schematic structural diagram of an optional embodiment of the spintronic neuron device modeling apparatus proposed in this application;

[0062] Figure 14 This is a schematic structural diagram of an optional embodiment of the spiking neural network training device proposed in this application;

[0063] Figure 15 The figure is a schematic diagram of the hardware structure of an optional embodiment of an electronic device suitable for the spintronic neuron device modeling method and the pulse neural network training method proposed in this application. DETAILED DESCRIPTION

[0064] In response to the issues described in the background technology section, this application aims to enable the training of spiking neural networks composed of artificial neurons with diverse performance and mechanisms, and to optimize spiking neural networks from multiple perspectives. To this end, this application proposes a spintronic neuron device with adaptively adjustable emission thresholds, constructs a spiking neuron array, and then constructs a hardware spiking neural network based on this. This spintronic neuron device can achieve adaptive adjustment of the emission threshold without the need for peripheral circuitry, achieving the internal homeostasis characteristics of the spiking neural network, and training a highly efficient, stable, and low-power spiking neural network.

[0065] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. For the convenience of description, only the parts related to the relevant inventions are shown in the drawings. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. That is to say, based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0066] Reference Figure 1 , which is a schematic structural diagram of an optional embodiment of the spintronic neuron device proposed in this application, as shown Figure 1 As shown, the spintronic neuron device includes two-terminal electrodes, namely a first electrode 11 and a second electrode 12, a transmission component 13 connecting the first electrode 11 and the second electrode 12, and a plurality of output components 14 disposed at different positions in the transmission component 13, wherein:

[0067] The first electrode 11 and the second electrode 12 are located at the two ends of the spin electronic neuron device, and are used to receive input signals and transmit pulses respectively, thereby realizing signal transmission between different spin electronic neuron devices. This application does not describe in detail the structure of the electrodes at the two ends of the spin electronic neuron device and their working principles.

[0068] The transmission component 13 can be composed of a multilayer film to support the movement of the domain wall (Domain Wall, DW for short, i.e., the interface between different magnetic domains). This application is used to simulate the basic behaviors of biological neurons and synapses, such as the leaky integral ignition model (LIF) and the spike-timing-dependent plasticity model (STDP). The constituent materials and structures of the transmission component 13 are not restricted and can be determined according to the circumstances.

[0069] Because different output components 14 are deployed at different positions in the transmission component 13, the lengths of the transmission component 13 between these multiple output components 14 and the same electrode are different. In this way, when different output components 14 are selected to be in the working state, the path lengths of the magnetic domain walls injected through the first electrode or the second electrode moving in the transmission component 13 are different, so that the spintronic neuron device has different emission thresholds. The emission threshold refers to the value accumulated to the membrane potential of the neuron when an electrical signal triggers a nerve impulse to transmit the electrical signal to other neurons, that is, when the neuron emits a pulse signal. This application does not limit the value of the emission threshold corresponding to each output component.

[0070] It can be seen from this that in a spin electronic neuron device proposed in the present application, different output components 14 are selected and used, and the transmission path lengths of the magnetic domain wall injected from one end electrode on the self-transmission component 13 are different. In this way, the magnetic domain wall moves on the transmission component 13 to be received by the different selected output components 14, and the emission threshold for completing the corresponding neuron pulse emission will be different. Therefore, the present application can control the selection and use of different output components 14 of the spin electronic neuron device and change the length of the magnetic domain wall transmission path, so as to flexibly realize the adjustment of the emission threshold of the spin electronic neuron device.

[0071] Therefore, in order to ensure that the response frequency of neurons is within a specific range, avoid excessive or insufficient response of neurons, and ensure that neurons can respond quickly to changes in input signals, and avoid overfitting during the training process of the hardware neural network constructed using the spin electronic neuron device proposed in this application, this application only needs to select and use appropriate output components for the spin electronic neuron device, without the help of peripheral circuits, to achieve internal homeostasis of the neural network, thereby improving the stability and robustness of the neural network to different inputs.

[0072] In some embodiments, as Figure 1 As shown, the plurality of output components 14 are distributed in an array on the transmission component 13. When preparing the spintronic neuron device, the relative positions of the plurality of output components 14 (such as the distance and / or position relationship between two adjacent output components 14) are adjustable and are not limited to Figure 1 The array arrangement shown has the same distance between adjacent output assemblies 14. Thus, during the device fabrication phase, the resolution of the multiple emission threshold variations of the device can be adjusted by changing the relative positions of the multiple output assemblies 14. The adjustment relationship between the two can be determined through testing.

[0073] For example, Figure 2aAs shown, if the first output component on the first electrode side is selected to enter the working state, the magnetic domain wall injected from the first electrode will be received when it moves to the position of the output component, and will stop moving to the next output component. The emission threshold corresponding to the output component at this time can be determined as the initial emission threshold of the spintronic neuron device, and its value can be determined by detection.

[0074] Similarly, if Figure 2b As shown, if the sixth output component on the first electrode side is selected to enter the operating state, the magnetic domain wall injected from the first electrode will only be received when it moves to the location of this output component. Here, the emission threshold of the transmitted pulse signal is greater than the aforementioned initial emission threshold. It can be seen that the longer the transmission path of the magnetic domain wall, the greater the emission threshold of the neuron's pulse signal when the corresponding output component is selected for operation. The adjustment relationship between the two can be determined during the device preparation stage. In this way, during the operation phase of the spintronic neuron device, different output components 14 can be selected to achieve adaptive adjustment of the emission threshold of the spintronic neuron device to the pulse signal.

[0075] Reference Figure 3 , is a flow chart of an optional embodiment of the spintronic neuron device modeling method proposed in this application, which can be applied to electronic devices such as Figure 3 As shown, the method may include:

[0076] Step S31, obtaining a selection instruction for multiple output components in the spintronic neuron device, and determining a target output component to enter a working state;

[0077] In combination with the above description of the spin electronic neuron device proposed in this application, the spin electronic neuron device itself has multiple emission thresholds for the emitted pulse signals. By selecting different output components to work, the transmission path length of the magnetic domain wall movement is changed, thereby achieving adaptive adjustment of the emission threshold. In this way, after the spin electronic neuron device inputs an electrical signal, the transmission time of the magnetic domain wall movement in the spin electronic neuron device can be changed. In other words, the response speed of the spin electronic neuron device to the input signal can be adaptively adjusted without the help of peripheral circuits, supporting the realization of internal homeostasis within the constructed neural network.

[0078] This application models spintronic neuron devices, using them as the basic unit of neural networks (i.e., neurons) to construct spiking neural networks. The physical properties of neurons, such as LIF, can be realized based on the motion of magnetic domain walls. Because spintronic neuron devices utilize different output components, the transmission path lengths of magnetic domain wall motion vary, resulting in different response speeds to input electrical signals.

[0079] Therefore, in order to accurately understand the operating behavior of the different output components of the spintronic neuron device and the LIF characteristics achieved thereby, each output component can be selected to enter an operating state and then the simulated behavior of the spintronic neuron device characteristics can be detected, thereby obtaining the LIF characteristics mimicked by the spintronic neuron device. This application does not limit the implementation method for selecting the operation of any output component in the spintronic neuron device.

[0080] Step S32, performing electrical signal input control on the spintronic neuron device to trigger the generated magnetic domain wall motion and obtain corresponding neuron physical properties;

[0081] In practical applications of the present application, an electrical signal can be input into a spin electronic neuron device. When the spin electronic neuron device receives the electrical signal, i.e., a current pulse, the magnetic domain wall is triggered to start moving in the forward direction, i.e., moving along the transmission component toward the selected target output component, i.e., forward motion. When the electrical signal is stopped from being input into the spin electronic neuron device, i.e., the current pulse is removed, the magnetic domain wall will start moving in the reverse direction.

[0082] It can be seen from this that the present application can control the forward and reverse movement of the magnetic domain wall in the spin electronic neuron device by inputting an electrical signal into the spin electronic neuron device to correspond to the accumulation process and leakage process of the bionic neuron membrane potential, and detect this process when the magnetic domain wall moves to the end point of the spin electronic neuron device (that is, the location of the target output component of the selected work) to simulate the discharge behavior of the neuron.

[0083] Among them, a pulse signal will be emitted when the magnetic domain wall moves forward to the end point. At the same time, the magnetic domain wall is detected at the target output component, which can trigger the neuron to enter the refractory period state. In this state, the spin electronic neuron device will no longer respond to the external input electrical signal, and the magnetic domain wall will maintain the reverse motion state, moving backward from the target output component to the starting position, such as the end of the first electrode where the electrical signal was previously input.

[0084] Based on the above analysis, in order to obtain the neuronal physical properties of the spintronic neuron device, such as the LIF characteristics, the spintronic neuron device can be controlled by electrical signal input to trigger the forward and reverse motion of the generated magnetic domain wall, thereby mimicking the accumulation and leakage processes of the neuron membrane potential and obtaining the corresponding accumulation and leakage rates. From this, we can understand the response of the spintronic neuron device to the input electrical signal by selecting different output components to work.

[0085] In addition, as described above, the movement of the magnetic domain wall to the position of the target output component is determined, the neuronal membrane potential when the target output component emits a pulse signal is detected, the emission threshold for the target output component is obtained, and the refractory period information of the spin electronic neuron device is obtained; the refractory period information can characterize that after the magnetic domain wall is detected to move to the position of the target output component, it stops responding to the input electrical signal and moves in the opposite direction to the starting position, thereby understanding the situation in which the spin electronic neuron device is in a refractory period state. This application does not limit the content of the refractory period information and its representation method.

[0086] Thus, this application can select each output component and perform characteristic simulation according to the above method to simulate the LIF characteristics of the spintronic neuron device in its current operating state. Therefore, the accumulation rate, leakage rate, emission threshold, and refractory period information for different output components can be used to construct the neuron physical properties of the spintronic neuron device. It should be noted that other neuron physical properties can also be simulated according to the method described above, which is not detailed in this application.

[0087] Step S33 : Modeling the spintronic neuron device according to the different neuron physical properties of the multiple output components to obtain a neuron model of a corresponding type.

[0088] According to the above method, the spintronic neuron device is used to simulate the LIF characteristics of biological neurons. After obtaining the actual neuronal physical characteristics of the spintronic neuron device, different neuronal physical characteristics can be modeled based on the mathematical equations of the spintronic neuron device, that is, the equivalent mathematical model, to obtain neuron models with different performance and mechanisms that can serve as basic units of neural networks. This application does not elaborate on this method of modeling artificial neuron devices.

[0089] Reference Figure 4 The schematic diagram of the bionic behavior of the spintronic neuron device is shown. The magnetic field H of the spintronic neuron device is set. build_in , and the size of the DM interaction (Dzyaloshinsky-Moriya interaction, DMI), under the pre-set normal working conditions of the spin electronic neuron device, the relative position change of the magnetic domain wall DW of the spin electronic neuron device compared to its movement starting point is detected, that is, as the test time (time unit: nanoseconds ns), the position Position (length unit: nanometers nm) of the magnetic domain wall DW of the spin electronic neuron device relative to the starting point is detected to form the following Figure 4 The curve shown in FIG5 is used to obtain the physical properties of each neuron in the spintronic neuron device.

[0090] Based on this, the above-mentioned step S33 may include: modeling the corresponding characteristic bionic behaviors of the spin electronic neuron device (such as the forward accumulation process and reverse leakage process of the neuron membrane potential, the pulse emission process and refractory period of the pulse signal emitted by the target output component, etc.) according to the accumulation rate, leakage rate, emission threshold and refractory period information of multiple output components, and obtaining the corresponding type of neuron model, that is, obtaining a neuron model with different neuron physical properties, so as to subsequently construct neurons with different functions in the corresponding network layer of the pulse neural network. The implementation process is not described in detail in this application.

[0091] In combination with the structure of the spintronic neuron device proposed in this application, it can realize the adaptive adjustment of the emission threshold of the spintronic neuron device to the pulse signal through the multiple output components it has, without the need for peripheral circuits, thus reducing hardware costs and improving the convenience of adjusting the emission threshold. Figure 5 The bionic schematic diagram of the LIF characteristics of the spintronic neuron device shown is a simulation diagram obtained by simulating the above neuron model. By inputting electrical signals to the spintronic neuron device, the movement of the magnetic domain wall (such as Figure 5 The forward accumulation process (shown by the A-to-B curve, and the reverse leakage process (shown by the B-to-C curve)) of this neuron is simulated by modeling its LIF characteristics, or the physical properties of neurons. This allows for the development of a neuron model with properties corresponding to those of a spintronic neuron device. This model can then be used to construct the various layers of a spiking neural network to meet information processing requirements. Because this neuron model can flexibly adjust its firing threshold and control its response speed to input electrical signals, it can achieve homeostasis in a spiking neural network without the need for peripheral circuitry or software algorithms.

[0092] Reference Figure 6 , is a flow chart of an optional embodiment of the pulse neural network training method proposed in this application, which can be applied to electronic devices, such as servers or terminal devices with data processing capabilities, such as Figure 6 As shown, the method may include:

[0093] Step S61, obtaining a spiking neural network to be trained and a sample data set for a recognition task;

[0094] In practical applications, information transmission in spiking neural networks relies on synapses, which can be divided into excitatory and inhibitory synapses. Excitatory synapses increase membrane potential by outputting certain electrical signals (i.e., current); inhibitory synapses, on the other hand, decrease membrane potential. As can be seen, spiking neural networks include excitatory neurons, which are modeled by the forward accumulation of neuronal membrane potential, inspired by the forward motion of the magnetic domain walls of spintronic neuron devices, and inhibitory neurons, which are modeled by the reverse leakage of neuronal membrane potential, inspired by the reverse motion of the magnetic domain walls of spintronic neuron devices.

[0095] Thus, the present application, in accordance with the method described above, constructs corresponding neuron models for the different biomimetic behaviors of spintronic neuron devices, giving them corresponding neuron physical properties. Subsequently, based on the network architecture of a spiking neural network, multiple different types of neuron models are utilized to construct the spiking neural network. It should be understood that by modeling the different behaviors of spintronic neuron devices, neuron models with different performance and mechanisms can be obtained. The implementation process is similar and can be referred to the corresponding description of the above embodiment. This embodiment will not be described in detail here.

[0096] It should be noted that this application does not impose any restrictions on the network structure of the initially constructed pulse neural network (which includes the number of neuron models and the LIF characteristics such as the current emission threshold of each neuron model). The network structure can be adaptively adjusted according to the task processing requirements, or a pulse neural network with a general structure can be initially constructed, and then the pulse neural network model for the recognition task can be trained in a targeted manner.

[0097] After determining the current information processing task, i.e., a recognition task, such as object recognition or classification, a sample dataset for the recognition task can be obtained, which can include a large amount of sample data (such as positive and negative training samples, test samples, verification samples, etc.). For example, in an image processing scenario, the sample dataset can include the MNIST dataset (Modified National Institute of Standards and Technology database), a handwritten digit dataset used to train various image processing systems and binary image datasets for machine learning training and testing. Of course, other types of datasets can also be included. This application does not limit the content and source of the sample dataset, which can be determined as appropriate.

[0098] Step S62: Training the spiking neural network based on the multiple sample data included in the sample data set to optimize network parameters of the spiking neural network; the network parameters include the number of neuron models, a threshold adjustment rate for a firing threshold of at least one neuron model, and one or more of synaptic weights between different network layers in the spiking neural network; the threshold adjustment rate is used to indicate adjustment of the firing threshold of the corresponding neuron model;

[0099] Step S63: determine whether the recognition accuracy of the spiking neural network for the sample data meets the training termination condition, and determine the spiking neural network finally obtained by training as the spiking neural network model for implementing the recognition task.

[0100] Following the above analysis, the spiking neural network can include an input layer, an excitatory neuron layer, an inhibitory neuron layer, and an output layer. In the embodiment of the present application, the first three network layers are composed of one or more neuron models, while the final output layer does not contain an actual neuron model. It can be used to output the task recognition results obtained based on the response of each neuron model in the excitatory neuron layer. Among them, the neuron model constituting the input layer can be recorded as input neurons, the neuron model constituting the excitatory neuron layer can be recorded as excitatory neurons, and the neuron model constituting the inhibitory neuron layer can be recorded as inhibitory neurons.

[0101] Take the scenario of building a pulse neural network for image processing tasks as an example to illustrate, combined with Figure 7 The network structure diagram of the pulse neural network shown in the figure shows that the input layer processes the sample data into a pulse sequence suitable for the pulse neural network, and inputs it as input excitation to the excitatory neuron layer. Since the neuron model constituting the excitatory neuron layer has the LIF characteristics of the spin electronic neuron device, it responds to the input excitation, and the emitted pulse signal will be input into the subsequent neuron layer.

[0102] Among them, because the input of the inhibitory neuron layer comes from the excitatory neuron layer, and the output response is fed back to the excitatory neuron layer, it inhibits the response behavior of the neuron model in the excitatory neuron layer, that is, it has an inhibitory influence on the behavior of the excitatory neuron layer. After this signal processing, the excitatory neuron layer can be processed according to the "Winner Take All (WTA)" principle to obtain the final output of the spiking neural network. That is, based on the response of the excitatory neuron layer, the task recognition result of the input sample data can be determined, and the task recognition result is output through the output layer, such as selecting a certain number in the input sample image.

[0103] Based on this, we take the digit recognition task of the sample images in the MNIST dataset as an example to illustrate. Figure 8As shown in the flow chart, the sample image is input into the initially constructed pulse neural network, and the sample image is encoded in Poisson sequence through the neuron model of the input layer to obtain a pulse sequence corresponding to the pulse neural network, that is, according to the different pixel intensities of different pixels in the sample image, each pixel is encoded in Poisson sequence and converted into a set of pulse sequences, so that the information in the sample image is carried by the delay of the pulse in the pulse sequence.

[0104] Afterwards, the pulse sequence can be used as input stimulation and input into the excitatory neuron layer. According to the connection relationship between the excitatory neuron layer and the inhibitory neuron layer, the corresponding neuron model will be caused to undergo membrane potential accumulation, pulse reflection, and membrane potential leakage. The output layer selects a number in the sample image as the task recognition result output based on the response of the excitatory neuron layer, such as Figure 7 As shown in the figure, the task of recognizing digits in images of the MNIST dataset is implemented.

[0105] After each training of the spiking neural network in the above manner, the recognition loss of the task recognition result can be obtained, which represents the recognition accuracy of the spiking neural network for the input sample data. If the training termination condition is not met, such as the recognition accuracy does not reach the accuracy threshold, or does not converge to the maximum recognition accuracy, or the number of training times does not reach the preset training threshold, etc., the network parameters of the current spiking neural network are optimized based on the recognition loss, which includes but is not limited to the number of neuron models, the emission threshold of at least one neuron model, the threshold adjustment rate for the emission threshold of different neuron models, and one or more of the synaptic weights between different network layers in the spiking neural network, which can be determined as appropriate. Afterwards, the sample image can be continuously input into the spiking neural network with optimized network parameters for training. After multiple trainings, it is determined that the training termination condition is met, and the spiking neural network finally trained is determined as the spiking neural network model for implementing the recognition task, such as the spiking neural network model for recognizing image digits in the MNIST dataset. It should be noted that the training implementation process for spiking neural network models for implementing other tasks is similar and will not be described in detail in this application. Generally, the network parameters of spiking neural network models for implementing different tasks will be different and will not be described in detail in this application.

[0106] In the practical application of the present application, the different network layers in the pulse neural network constructed above can be connected through artificial synapses. In this embodiment, the excitatory neuron layer of the pulse neural network can be connected to the input layer and the output layer by a full connection method for artificial synaptic connection, while the excitatory neuron layer and the inhibitory neuron layer can be connected to the artificial synapse by a complementary full connection method. The complementary full connection method can refer to the point-to-point connection between the excitatory neurons in the excitatory neuron layer and the inhibitory neurons corresponding to the excitatory neurons in the inhibitory neuron layer. The excitatory neurons will generate excitatory inputs to the inhibitory neurons. At the same time, the inhibitory neurons and other neuron models other than their corresponding excitatory neurons adopt a full connection method, and the inhibitory neurons generate feedback inhibitory inputs to such excitatory neurons. It can be seen that between the excitatory neuron layer and the inhibitory neuron layer in the pulse neural network, two neuron models with a corresponding relationship can be connected point-to-point, and two neuron models without a corresponding relationship can be connected by a full connection method. The present application does not elaborate on the process of implementing this complementary connection.

[0107] It should be noted that this application will usually configure corresponding synaptic weights for each artificial synapse, so that during the training process of the spiking neural network, the synaptic weights can be adjusted through the pulse time dependent plasticity STDP learning mechanism to optimize the performance of the spiking neural network. Among them, in this STDP learning mechanism, if the time when the presynaptic pulse stimulus arrives at the synapse is earlier than the time when the postsynaptic pulse stimulus arrives at the synapse, the synaptic weight will be enhanced; conversely, if the time when the presynaptic pulse stimulus arrives at the synapse is later than the time when the postsynaptic pulse stimulus arrives at the synapse, the synaptic weight will be weakened and enhanced. This application does not elaborate on the principle of the STDP learning mechanism.

[0108] In addition, in the pulse neural network constructed in the present application, the synaptic weight of the artificial synapse between the inhibitory neuron layer and the excitatory neuron layer is less than zero, that is, it is a negative number; the other synaptic weights are all greater than zero, such as the synaptic weights of the artificial synapses between the excitatory neuron layer and the inhibitory neuron layer, between the input layer and the excitatory neuron layer, and between the excitatory neuron layer and the output layer are all greater than zero, that is, they are positive numbers.

[0109] In summary, in the embodiments of the present application, since the present application proposes that the spin electronic neuron device has multiple output components, it can adaptively adjust the transmission path length of the magnetic domain wall and change the emission threshold of the device. By modeling the behavior of the spin electronic neuron device, neuron models with different performances and mechanisms are obtained, which are used as basic units to construct a pulse neural network, so that the neurons in the pulse neural network have the LIF characteristics of the spin electronic neuron device. In the process of training it using a sample data set, it can be optimized from multiple angles, such as adaptively adjusting the number of neuron models that constitute the pulse neural network, the emission threshold of at least one neuron model and its threshold adjustment rate, and adjusting the synaptic weights between the two connected neuron models to change the degree of influence of the previous neuron on the next neuron, so as to optimize the performance of the pulse neuron, thereby providing a reference for the optimization of subsequent hardware neural networks and realizing an efficient, stable, and low-power pulse neural network.

[0110] During the training process of the aforementioned spiking neural network, the neuron model can be modified, such as by changing the mathematical method of the input neuron model to change the properties of the neuron model. This is suitable for training and optimizing various spiking neural networks, thereby improving the compatibility of spiking neural network training methods. Furthermore, the present application can achieve adaptive adjustment of the emission threshold of the spintronic neuron device without the need for peripheral circuitry, ensuring that the response time of the spiking neural network to external input is controlled within a preset range, achieving the internal homeostasis characteristics of the spiking neural network, avoiding overfitting, and reducing the power consumption of the spiking neural network.

[0111] Based on the spiking neural network training method described in the above embodiment, after the spiking neural network to be trained is trained once using multiple sample images included in the MNIST data set, assuming that the excitatory neuron layer of the spiking neural network includes 400 excitatory neurons, and the excitatory neuron layer is constructed by 400 neuron models with corresponding LIF characteristics, the synaptic weight distribution diagram of each synaptic connection between the input layer and the excitatory neuron layer obtained after one training can be as follows Figure 9 As shown in Figure 2, the synaptic weight is a property of the synapse between two neurons. The larger the synaptic weight, the greater the influence of the preceding neuron on the succeeding neuron.

[0112] In the process of the pulse neural network processing the input sample image, the input layer converts the sample image into a pulse sequence. Figure 9The synaptic weight distribution diagram shown in the figure shows a weighted summation of the pulse sequence and the synaptic weight matrix. The resulting input stimulus is then transmitted to the excitatory neuron layer. Simultaneously, based on the STDP algorithm, the pulse delays between the input layer and the excitatory neuron layer are compared, and the synaptic weights are adjusted to optimize the performance of the spiking neural network. Because different synaptic weights cause different neurons to respond differently to the same sample image, these differences can be used to label the excitatory neurons in the excitatory neuron layer, such as the 10 digits 0 to 9 in the sample image, so that the trained spiking neural network model can perform recognition and classification tasks.

[0113] It should be noted that the number of excitatory neurons contained in the excitatory neuron layer in the pulse neural network, including but not limited to 400 in the above example, can be determined based on actual conditions, and the number of excitatory neurons can also be adaptively adjusted during the model training process. The implementation process is not described in detail in this application.

[0114] In conjunction with the description of the spintronic neuron device proposed in this application, different output components correspond to different emission thresholds. In other words, the emission threshold of a spintronic neuron device is related to the position of the output component it selects for operation, and the difference between different emission thresholds is related to the relative positions of the different output components. When preparing this spintronic neuron device, the relative positions of the different output components can be reasonably determined, thereby determining the threshold adjustment ratio between the different emission thresholds. This application does not impose any restrictions on this value, allowing the neuron model constructed based on this threshold adjustment ratio to achieve adaptive adjustment of the emission threshold.

[0115] It can be seen that in practical applications, by configuring a reasonable threshold adjustment rate, the difficulty of neurons emitting pulses can be increased, thereby reducing the frequency of their emitted pulses. In this way, in a spiking neural network, when a neuron switches from an excited state to a resting state, the membrane potential change of the neuron in a short period of time can be suppressed by changing the ion concentration difference inside and outside the neuron membrane, thereby suppressing the neuron from being excited multiple times in a short period of time. This behavior can be achieved by setting the threshold adjustment rate of the neuron in the bionic behavior of the spintronic neuron device. Accordingly, the spintronic neuron device proposed in this application can adjust its emission threshold by changing the transmission path length of the magnetic domain wall motion, without the need for peripheral circuits or software algorithms, thereby improving the performance of the spiking neural network constructed thereby.

[0116] Based on the above analysis, in some embodiments, it is assumed that the excitatory neuron layer in the pulse neural network is composed of 1,600 neuron models, the initial emission threshold of the neuron model is 100 nm, and its threshold adjustment rate is +7 nm each time a pulse is emitted, that is, each time the neuron emits a pulse, the emission threshold of the neuron in the subsequent working cycle can be increased by 7 nm. In this way, during the model training process, the emission threshold of the neuron model can be adjusted according to the threshold adjustment rate to achieve adaptive adjustment of the neuron emission threshold.

[0117] In addition, assuming that the STDP learning mechanism has a pre-stage learning rate of 0.005 and a post-stage learning rate of 0.1, refer to Figure 10 The relationship between the recognition accuracy and the number of training times on the test set of the MNIST dataset is shown in the figure. After completing two training sessions on the entire MNIST dataset as described above, the recognition accuracy of the spiking neural network on the test set of the MNIST dataset has exceeded 90%. After completing 5 training sessions, the recognition accuracy reaches 92.38%.

[0118] like Figure 10 The curve showing the relationship between the accuracy rate and the number of training cycles on the MNIST dataset shows that, from a macro perspective, the recognition accuracy of the spiking neural network on the test set of the MNIST dataset increases with the number of training cycles, especially in the first three training cycles. This optimization process, based on the STDP learning mechanism, adjusts the synaptic weights between the input layer and the excitatory neuron layer, resulting in different neurons responding differently to the pixel values of different pixels in the input sample image. Furthermore, the pixels at the same pixel in different sample images differ from each other. This difference, after weighted summation using the synaptic weight matrix, is reflected in the neuron's firing behavior, thereby enabling digit recognition in the sample image. Based on this, the neurons that fire can be assigned recognized labels, establishing a correspondence between neurons and recognition results.

[0119] After the number of training times of the pulse neural network reaches saturation, such as Figure 10 Continuing to train the spiking neural network during the last three training sessions will cause the correspondence between neurons and recognition results to change. It is possible that the original correct correspondence may be modified into an incorrect correspondence by an accidental sample image, causing fluctuations in the recognition accuracy of the spiking neural network. To avoid this situation, the present application can configure a maximum number of training times, i.e., a training threshold, and determine the actual number of training times reaching the maximum number of training times as a training termination condition, but is not limited to this.

[0120] Combined with the above analysis, the number of neuron models can also be adjusted during the spiking neural network training process. Let’s take the above example where the emission threshold of the neuron device model is 100nm, the threshold adjustment rate is 7nm, and the learning rates of the front and back stages in the STDP learning algorithm are 0.005 / 0.01 as an example. If the number of neuron models N in the excitatory neuron layer is changed, as shown in the following example: Figure 11 As shown in the figure, the number of excitatory neurons was reduced from 1600 to 900, and then to 400, and the training was repeated five times. The recognition accuracy of the spiking neural network obtained after each training on the test set of the MNIST dataset decreased. When the number of excitatory neurons was 1600, the accuracy of the spiking neural network reached 92.38%. Based on this, in order to improve the performance of the spiking neural network, the number of excitatory neurons can be increased.

[0121] In addition, during the training of the pulse neural network, the present application can also change the threshold adjustment rate of the emission threshold of the neuron model. The present application takes the scenario where the pulse neural network contains 900 excitatory neurons, the emission threshold of the neuron device model is 100nm, and the pre-stage and post-stage learning rates in the STDP learning algorithm correspond to 0.005 / 0.01 as an example. Figure 12 The graph shows the relationship between the recognition accuracy of the pulse neural network and the number of training times under different threshold adjustment rates. The selected threshold adjustment rates △θ can be 2.5nm, 7nm and 12.5nm respectively. After multiple trainings, it can be seen that the increase in the threshold adjustment rate at any time can improve the recognition accuracy of the pulse neural network.

[0122] In combination with the above analysis, the present application can increase the number of neuron models, increase the threshold adjustment rate of the neuron model emission threshold, and the synaptic weight between the input layer and the excitatory neuron layer, so as to achieve the training purpose of improving the performance of the pulse neural network.

[0123] Reference Figure 13 , which is a structural diagram of an optional embodiment of the spintronic neuron device modeling device proposed in this application, such as Figure 13 As shown, the modeling device may include:

[0124] a target output component determination module 131, configured to obtain a selection instruction for a plurality of output components in the spintronic neuron device according to claim 1 or 2, and determine a target output component to enter a working state; different output components correspond to different magnetic domain wall transmission path lengths;

[0125] a neuron physical property acquisition module 132 for controlling the electrical signal input to the spintronic neuron device, triggering the generated magnetic domain wall motion, and obtaining corresponding neuron physical properties;

[0126] The modeling module 133 is configured to model the spintronic neuron device according to the different neuron physical properties of the multiple output components to obtain a neuron model of a corresponding type.

[0127] Optionally, the neuron physical property acquisition module 132 may include:

[0128] A first obtaining unit is configured to control the spintronic neuron device by inputting an electrical signal to trigger the generated magnetic domain wall to generate forward and reverse motions, thereby mimicking the accumulation and leakage processes of neuron membrane potential and obtaining corresponding accumulation and leakage rates;

[0129] a second obtaining unit, configured to determine that the magnetic domain wall moves to the position of the target output component, detect the neuron membrane potential when the target output component transmits a pulse signal, obtain an emission threshold for the target output component, and acquire refractory period information of the spintronic neuron device; the refractory period information can indicate that after detecting that the magnetic domain wall moves to the position of the target output component, it stops responding to the input electrical signal and moves in the opposite direction to the starting position;

[0130] Based on this, the modeling module 133 may include:

[0131] A modeling unit is configured to model the corresponding characteristic bionic behaviors of the spintronic neuron device according to the accumulation rate, the leakage rate, the emission threshold, and the refractory period information of the multiple output components, so as to obtain a neuron model of a corresponding type.

[0132] Reference Figure 14 , is a structural diagram of an optional embodiment of the pulse neural network training device proposed in this application, such as Figure 14 As shown, the training device may include:

[0133] An acquisition module 141 is configured to acquire a neural network to be trained and a sample data set for a recognition task; the spiking neural network is constructed from a plurality of different types of neuron models to implement the recognition task, and the neuron models are obtained using the spintronic neuron device modeling method described above;

[0134] A training module 142 is configured to train the spiking neural network based on the plurality of sample data included in the sample data set to optimize network parameters of the spiking neural network;

[0135] Among them, the network parameters include the number of neuron models, the threshold adjustment rate for the emission threshold of at least one neuron model, and one or more of the synaptic weights between different network layers in the pulse neural network; the threshold adjustment rate is used to indicate the adjustment of the emission threshold corresponding to the neuron model.

[0136] The pulse neural network model determination module 143 is used to determine whether the recognition accuracy of the pulse neural network with optimized network parameters for the sample data meets the training termination condition, and determine the pulse neural network finally trained as the pulse neural network model for implementing the recognition task.

[0137] Optionally, the above-mentioned pulse neural network includes an input layer, an excitatory neuron layer, an inhibitory neuron layer and an output layer. The connection relationship between each network layer can refer to the description of the corresponding part of the above embodiment, which will not be described in detail in this embodiment.

[0138] Based on this, in practical applications, the training module 142 may include a synaptic weight adjustment unit for adjusting the synaptic weights during the spiking neural network training process through a spike time-dependent plasticity (STDP) learning mechanism. The synaptic weights of the artificial synapses between the inhibitory neuron layer and the excitatory neuron layer are less than zero; the synaptic weights of the artificial synapses between the excitatory neuron layer and the inhibitory neuron layer, between the input layer and the excitatory neuron layer, and between the excitatory neuron layer and the output layer are all greater than zero.

[0139] Optionally, the training device may further include:

[0140] An effective pulse selection module, configured to randomly select a pulse emitted by a neuron model as a valid pulse when some neuron models in the excitatory neuron layer emit pulses;

[0141] an inhibitory response signal obtaining module, configured to transmit the effective pulse to the inhibitory neuron layer to obtain an inhibitory response signal;

[0142] an inhibition accumulation module, configured to feed back the inhibitory response signal to another part of the neuron model in the excitatory neuron layer to inhibit the accumulation of the pulse neuron membrane potential of the neuron model;

[0143] The emission threshold adjustment module is used to determine the pulses emitted by the neuron model that has already emitted pulses as valid pulses before selecting the valid pulses, and to adjust the emission threshold of the neuron model by selecting different output ports of the spintronic neuron device used to construct the neuron model, thereby achieving the homeostasis of the spiking neural network. It should be noted that the various modules, units, etc. in the above-mentioned device embodiments can all be stored in the memory as program modules, and the processor executes the above-mentioned program modules stored in the memory to achieve the corresponding functions. Regarding the functions achieved by each program module and its combination, as well as the technical effects achieved, reference can be made to the description of the corresponding parts of the above-mentioned method embodiment, and this embodiment will not be repeated here.

[0144] The present application also provides a computer-readable storage medium on which a computer program can be stored. The computer program can be called and loaded by a processing device to implement the various steps of the spintronic neuron device modeling method and the pulse neural network training method described in the above embodiments. The specific implementation process can refer to the description of the corresponding parts of the above embodiments and will not be repeated in this embodiment.

[0145] Reference Figure 15 , is a hardware structure diagram of an optional embodiment of an electronic device suitable for the above-mentioned spintronic neuron device modeling method and / or pulse neural network training method proposed in this application. The electronic device can be a server or a terminal device with data processing capabilities. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a cloud server capable of cloud computing, etc. The terminal device can include but is not limited to smart phones, tablet computers, laptops, desktop computers, robots, etc. Take the electronic device as an example to illustrate, such as Figure 15 As shown, the electronic device may be a communication module 151, a memory 152 and a processor 153, wherein:

[0146] The number of each of the communication module 151, memory 152, and processor 153 can be at least one, and they can be connected to each other via a bus. That is, the various components of the electronic device can be connected to the bus to achieve data exchange between different components. The implementation process is not described in detail in this application. Among them, the bus can include a data bus, an address bus, etc., and the connection relationship between each component and different buses can be determined based on processing requirements. This application does not impose any restrictions on this.

[0147] Memory 152 can be used to store a first program for implementing the aforementioned spintronic neuron device modeling method and a second program, i.e., computer instruction code, for implementing the aforementioned spiking neural network training method. These two program codes can be stored in the same or different memories, depending on the circumstances. Processor 153 can load and execute the first program stored in the memory to implement the various steps of the spintronic neuron device modeling method described in the corresponding method embodiment, or load and execute the second program to implement the spiking neural network training method described in the above embodiment. The specific implementation process can be referred to the description of the corresponding parts of the above embodiment and is not repeated here in this embodiment.

[0148] In the embodiment of the present application, the memory 152 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device or other volatile solid-state storage device. The processor 153 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices. The present application does not limit the structure and model of the above-mentioned memory 152 and processor 153, and they can be flexibly adjusted according to actual needs.

[0149] It should be understood that Figure 15 The structure of the electronic device shown does not constitute a limitation on the electronic device in the embodiment of the present application. In actual applications, the electronic device may include Figure 15 More components shown, or a combination of certain components, such as when the electronic device is a terminal device, may also generally include at least one input component such as a touch sensing unit for sensing touch events on a touch display panel, a keyboard, a mouse, a camera, a microphone, etc.; at least one output component such as a display, a speaker, a vibration mechanism, a lamp, etc.; an antenna; a sensor module; a power supply module, etc., which can be determined based on the product type and function of the terminal device, and this application does not list them one by one here.

[0150] Finally, it should be noted that the terms "system," "device," "unit," and / or "module" used in this application are a method for distinguishing different components, elements, parts, portions, or assemblies at different levels. However, if other terms can achieve the same purpose, they can be replaced by other expressions.

[0151] As used in this application and the claims, unless the context clearly indicates an exception, the terms "a," "an," "an," and / or "the" are not intended to refer to the singular and may include the plural, unless the context clearly indicates otherwise. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements. The phrase "comprises a..." does not preclude the presence of additional identical elements in the process, method, product, or apparatus that includes the elements.

[0152] Among them, in the description of the embodiments of the present application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present application, "multiple" refers to two or more than two. The following terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0153] In addition, the various embodiments in this specification are described in a progressive or parallel manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between the various embodiments can be referred to in conjunction with each other. For the devices and electronic devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and relevant parts can be referred to the method description.

[0154] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A spintronic neuron device, characterized in that: include: a first electrode; a second electrode; a transmission component connected to the first electrode and the second electrode respectively, wherein the transmission component is composed of a multilayer film; a plurality of output components disposed at different locations within the transmission component; Wherein, when different output components are in working states, the transmission path lengths of the magnetic domain walls injected through the first electrode or the second electrode moving in the transmission component are different, so that the spintronic neuron device has different emission thresholds.

2. The spintronic neuron device according to claim 1, characterized in that The plurality of output components are distributed in an array on the transmission component, and the relative positions between the plurality of output components are adjustable; Wherein, when the magnetic domain wall moves to the position of the output component in the working state in the transmission component, it is detected and stops moving to the next output component distributed in the array.

3. A spintronic neuron device modeling method, characterized in that: include: Obtaining a selection instruction for a plurality of output components in the spintronic neuron device according to claim 1 or 2, and determining a target output component to enter a working state; different output components correspond to different magnetic domain wall transmission path lengths; Performing electrical signal input control on the spintronic neuron device to trigger the generated magnetic domain wall motion and obtain corresponding neuron physical properties; The spintronic neuron device is modeled according to the different neuron physical properties of the multiple output components to obtain a neuron model of a corresponding type.

4. The method according to claim 3, characterized in that The electrical signal input control of the spintronic neuron device to trigger the generated magnetic domain wall motion and obtain corresponding neuron physical properties includes: Controlling the spintronic neuron device with an electrical signal input triggers the generated magnetic domain wall forward and reverse motions, thereby mimicking the accumulation and leakage processes of neuron membrane potentials and obtaining corresponding accumulation and leakage rates; Determining that the magnetic domain wall moves to the position of the target output component, detecting the neuron membrane potential when the target output component transmits a pulse signal, obtaining an emission threshold for the target output component, and acquiring refractory period information of the spintronic neuron device; the refractory period information can indicate that after detecting that the magnetic domain wall moves to the position of the target output component, it stops responding to the input electrical signal and moves in the opposite direction to the starting position; The step of modeling the spintronic neuron device based on the different neuron physical properties of the multiple output components to obtain a corresponding type of neuron model includes: According to the accumulation rate, the leakage rate, the emission threshold and the refractory period information of the multiple output components, corresponding characteristic bionic behaviors of the spintronic neuron device are modeled respectively to obtain corresponding types of neuron models.

5. A pulse neural network training method, characterized in that: The method comprises: Obtaining a neural network to be trained and a sample data set for a recognition task; the spiking neural network is constructed by multiple different types of neuron models to achieve the recognition task, and the neuron models are obtained by the spintronic neuron device modeling method according to claim 3 or 4; Training the spiking neural network based on the plurality of sample data included in the sample data set to optimize network parameters of the spiking neural network; wherein the network parameters include one or more of the number of neuron models, a threshold adjustment rate for a firing threshold of at least one neuron model, and synaptic weights between different network layers in the spiking neural network; the threshold adjustment rate is used to indicate adjustment of the firing threshold of the corresponding neuron model; It is determined that the recognition accuracy of the spiking neural network with optimized network parameters for the sample data meets the training termination condition, and the spiking neural network finally obtained by training is determined as the spiking neural network model for implementing the recognition task.

6. The method according to claim 5, characterized in that The spiking neural network includes an input layer, an excitatory neuron layer, an inhibitory neuron layer, and an output layer, wherein: The excitatory neuron layer is connected to the input layer and the output layer by artificial synapses in a fully connected manner, and the excitatory neuron layer is connected to the inhibitory neuron layer by a complementary fully connected manner; the complementary fully connected manner means that in the excitatory neuron layer and the inhibitory neuron layer, two neuron models with corresponding relationships are connected point-to-point, and two neuron models without corresponding relationships are connected by a fully connected manner; A corresponding synaptic weight is configured for each artificial synapse so that the synaptic weight can be adjusted through a pulse time-dependent plasticity (STDP) learning mechanism during the spiking neural network training process; wherein the synaptic weight of the artificial synapse between the inhibitory neuron layer and the excitatory neuron layer is less than zero; the synaptic weights of the artificial synapses between the excitatory neuron layer and the inhibitory neuron layer, between the input layer and the excitatory neuron layer, and between the excitatory neuron layer and the output layer are all greater than zero.

7. The method according to claim 6, characterized in that in: In the case where some neuron models in the excitatory neuron layer transmit pulses, randomly selecting a pulse transmitted by one neuron model as a valid pulse; transmitting the effective pulse to the inhibitory neuron layer to obtain an inhibitory response signal; Feeding back the inhibitory response signal to another part of the neuron model in the excitatory neuron layer to inhibit the accumulation of the pulse neuron membrane potential of the neuron model; If, before selecting the effective pulse, the pulse emitted by the neuron model that has already emitted a pulse is determined to be a valid pulse, the emission threshold of the neuron model is adjusted by selecting different output ports of the spintronic neuron device used to construct the neuron model to achieve the homeostasis of the pulse neural network.

8. A spintronic neuron device modeling device, characterized in that: The device comprises: a target output component determination module, configured to obtain a selection instruction for a plurality of output components in the spintronic neuron device according to claim 1 or 2, and determine a target output component to enter a working state; different output components correspond to different magnetic domain wall transmission path lengths; a neuron physical property acquisition module, configured to control the electrical signal input to the spintronic neuron device, trigger the generated magnetic domain wall motion, and obtain corresponding neuron physical properties; A modeling module is used to model the spintronic neuron device according to the different neuron physical properties of the multiple output components to obtain a neuron model of a corresponding type.

9. A pulse neural network training device, characterized in that: The device comprises: An acquisition module, configured to acquire a neural network to be trained and a sample data set for a recognition task; the spiking neural network is constructed from a plurality of different types of neuron models to implement the recognition task, the neuron models being obtained by the spintronic neuron device modeling method according to claim 3 or 4; a training module, configured to train the spiking neural network based on the plurality of sample data included in the sample data set to optimize network parameters of the spiking neural network; wherein the network parameters include one or more of the number of neuron models, a threshold adjustment rate for a firing threshold of at least one neuron model, and synaptic weights between different network layers in the spiking neural network; the threshold adjustment rate is used to indicate adjustment of the firing threshold of the corresponding neuron model; The pulse neural network model determination module is used to determine whether the recognition accuracy of the pulse neural network with optimized network parameters for the sample data meets the training termination condition, and determine the pulse neural network finally obtained by training as the pulse neural network model for implementing the recognition task.

10. An electronic device, characterized in that: The electronic device comprises: Communication module; A memory for storing a first program for implementing the spintronic neuron device modeling method according to claim 3 or 4, and a second program for implementing the pulse neural network training method according to any one of claims 5 to 7; A processor is configured to load and execute the first program to implement the spintronic neuron device modeling method according to claim 3 or 4, or to load and execute the second program to implement the pulse neural network training method according to any one of claims 5 to 7.

Images