Artificial neuron

Abstract

The application discloses an artificial neuron, comprising a microwave unit, a device unit and a weight unit. The microwave unit is used for generating a microwave signal; the device unit is responsive to the microwave signal and outputs a direct current voltage signal conforming to a plurality of clock curves with different peak response points; and the weight unit assigns a weight value to adjust the strength of the direct current voltage signal. The artificial neuron based on group coding with high flexibility can realize a single neuron with a plurality of adjustable nonlinear activation characteristics and can adapt to the needs of different artificial intelligence computing scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain-like computing, and in particular to an artificial neuron based on population coding. BACKGROUND

[0002] At present, the vigorous development of artificial intelligence and big data technology puts forward higher requirements on the computing power and energy efficiency of computing devices. However, the mainstream chips are all based on the Von Neumann architecture, and there is a natural bottleneck in the improvement of computing power and energy efficiency. Therefore, people begin to simulate the working mechanism of the brain and explore the neural morphological computing with high computing power, high energy efficiency and brain-like.

[0003] The hardware system relied on by neural morphological computing mainly includes artificial neurons and artificial synapses, wherein the artificial neurons are used for nonlinear processing of input signals to realize the functions of software nonlinear activation functions such as ReLU, Sigmoid and Softplus from the hardware bottom. However, whether the artificial neurons are constructed by using traditional semiconductor CMOS technology or new materials such as resistive switching materials and phase change materials, at present, they can only realize the function of one kind of nonlinear activation function, lack flexibility and are difficult to adapt to the needs of different artificial intelligence computing scenarios. In contrast, the brain realizes population coding through neuron groups, which not only has high computing flexibility, but also realizes the optimal balance between the reliability and energy efficiency of computing. Therefore, it is of great potential to simulate the population coding mechanism by using new type of nanodevices to construct reconfigurable neurons with high flexibility.

[0004] The information disclosed in this BACKGROUND section is only for the purpose of increasing the understanding of the background of the present application and should not be taken as an acknowledgment or any form of suggestion that this information forms prior art with respect to the present application. SUMMARY

[0005] The present application aims to provide an artificial neuron based on population coding with high flexibility, which realizes a single neuron with multiple adjustable nonlinear activation characteristics.

[0006] To achieve the above-mentioned purpose, the embodiments of the present application provide an artificial neuron, which comprises a microwave unit, a device unit and a weight unit.

[0007] The microwave unit is used to generate a microwave signal; the device unit responds to the microwave signal and outputs a direct current voltage signal conforming to multiple clock curves with different peak response points; and the weight unit assigns a weight value to adjust the strength of the direct current voltage signal.

[0008] In one or more embodiments of the present application, the weight unit is configured to assign multiple weight values corresponding to the multiple clock curves output by the device unit to adjust the strength of the direct current voltage signal output by the device unit.

[0009] In one or more embodiments of the present application, the weight unit comprises a plurality of weight devices, wherein one of the weight devices provides a weight value for the direct current voltage signal conforming to one of the clock curves of the device unit.

[0010] In one or more embodiments of the present application, the weight device comprises a CMOS transistor or a memristor.

[0011] In one or more embodiments of the present application, the device unit comprises a spin-torque diode unit, which is capable of outputting a plurality of direct current voltage signals conforming to clock curves with different peak response points by virtue of different self-magnetic anisotropies.

[0012] In one or more embodiments of the present application, the spin-torque diode unit comprises a spin-torque diode; the artificial neuron further comprises a magnetic anisotropy adjusting device, which acts on the spin-torque diode and is used to regulate the magnetic anisotropy of the spin-torque diode according to the input control electrical signal acting thereon, so that the spin-torque diode outputs a plurality of direct current voltage signals conforming to clock curves with different peak response points.

[0013] In one or more embodiments of the present application, a plurality of the weight devices are arranged in parallel, and the spin-torque diode is selectively electrically connected to the weight devices in the weight unit through a connection switch.

[0014] In one or more embodiments of the present application, the spin-torque diode comprises a first electrode layer, a second electrode layer, and a magnetic tunnel junction. The second electrode layer is arranged opposite to the first electrode layer; the magnetic tunnel junction is arranged between the first electrode layer and the second electrode layer, and comprises a magnetic free layer, a barrier layer, a pinned layer, and an anti-ferromagnetic layer, which are sequentially stacked from the first electrode layer to the second electrode layer; wherein the magnetic free layer comprises an in-plane magnetic anisotropy layer or a perpendicular magnetic anisotropy layer, and the spin-torque diode outputs a plurality of direct current voltage signals conforming to clock curves with different peak response points by changing the magnetic anisotropy of the magnetic free layer; and the micro-unit is connected to the first electrode layer and / or the second electrode layer.

[0015] In one or more embodiments of the present application, the magnetic anisotropy adjusting device comprises a magnetic field generating device, which is used to generate a magnetic field according to the input control electrical signal acting thereon, so as to regulate the magnetic anisotropy of the magnetic free layer of the spin-torque diode.

[0016] In one or more embodiments of the present application, the magnetic anisotropy adjusting member comprises a magnetoelectric coupling device for generating a stress and transmitting a strain force to the magnetic free layer according to an input control electrical signal acting thereon, so as to regulate the magnetic anisotropy of the magnetic free layer.

[0017] In one or more embodiments of the present application, the magnetoelectric coupling device comprises a piezoelectric film disposed on the first electrode layer or the second electrode layer, the piezoelectric film being used for generating a strain force according to an input control electrical signal acting thereon.

[0018] In one or more embodiments of the present application, the magnetic anisotropy adjusting member comprises a memristive device for generating an electric field according to an input control electrical signal acting thereon, and maintaining the electric field after the control electrical signal is removed, so as to achieve a memristive control of the magnetic anisotropy of the magnetic free layer.

[0019] In one or more embodiments of the present application, the magnetic anisotropy adjusting member comprises a memristive device, the memristive device comprising a resistive switching film or a phase change film disposed on the first electrode layer or the second electrode layer or covering the magnetic tunnel junction.

[0020] In one or more embodiments of the present application, the artificial neuron further comprises an integration circuit connected to the spin-torque diode and the weight unit, the integration circuit being used for integrating and outputting the direct current voltage signals of the plurality of clock curves conforming to different peak response points generated by the spin-torque diode after the weight unit assigns weight values.

[0021] In one or more embodiments of the present application, the spin-torque diode unit comprises a plurality of spin-torque diodes, the plurality of spin-torque diodes having different magnetic anisotropies to output direct current voltage signals of a plurality of clock curves conforming to different peak response points; the input end of each spin-torque diode is electrically connected to the micro-unit, and the output end of each spin-torque diode is electrically connected to one weight device, each weight device providing a weight value for the direct current voltage signal of the clock curve output by the spin-torque diode connected thereto; wherein the direct current voltage signals of the plurality of clock curves conforming to different peak response points are collectively output after being assigned different weights, so as to activate a nonlinear objective function.

[0022] In one or more embodiments of the present application, the spin-torque diode comprises a first electrode layer, a second electrode layer, and a magnetic tunnel junction. The second electrode layer is disposed opposite to the first electrode layer; the magnetic tunnel junction is disposed between the first electrode layer and the second electrode layer, and comprises a magnetic free layer, a barrier layer, a pinned layer, and an anti-ferromagnetic layer, which are sequentially stacked from the first electrode layer to the second electrode layer; wherein the magnetic free layer comprises an in-plane magnetic anisotropy layer or a perpendicular magnetic anisotropy layer; and the microwave unit is connected to the first electrode layer and / or the second electrode layer.

[0023] In one or more embodiments of the present application, the plurality of spin-torque diodes have magnetic free layers with different thicknesses or sizes, thereby having different magnetic anisotropies to output a plurality of clock curve-shaped direct current voltage signals with different peak response points.

[0024] In one or more embodiments of the present application, the microwave unit comprises a microwave source or a microwave antenna for inputting a microwave signal.

[0025] In one or more embodiments of the present application, the material of the magnetic free layer comprises at least one of CoFeB, Py, Co, CoFe, FePt, Co / Ni.

[0026] In one or more embodiments of the present application, the thickness of the magnetic free layer is 0.1 nm to 10 nm.

[0027] In one or more embodiments of the present application, the thickness of the first electrode layer is 10 nm to 200 nm.

[0028] In one or more embodiments of the present application, the material of the first electrode layer comprises one of Pt, Au, Pb, Ta, and Cr.

[0029] In one or more embodiments of the present application, the thickness of the second electrode layer is 10 nm to 200 nm.

[0030] In one or more embodiments of the present application, the material of the second electrode layer comprises one of Pt, Au, Pb, Ta, and Cr.

[0031] In one or more embodiments of the present application, the thickness of the barrier layer is 0.1 nm to 10 nm, preferably 1 nm to 5 nm.

[0032] In one or more embodiments of the present application, the material of the barrier layer comprises one of MgO, Al2O3, and TiO2.

[0033] In one or more embodiments of the present application, the thickness of the pinning layer is 0.1 nm to 10 nm.

[0034] In one or more embodiments of the present application, the material of the pinning layer comprises one or more combinations of CoFeB, Py, Co, CoFe, FePt.

[0035] In one or more embodiments of the present application, the magnetization direction of the pinning layer is in-plane or perpendicular in-plane.

[0036] Compared with the prior art, the artificial neuron of the embodiment of the present application has the following advantages and beneficial effects:

[0037] I. High flexibility. The artificial neuron in the present application can realize various nonlinear input-output characteristics in the same neuron by adjusting the weight unit, thereby simulating different activation functions to adapt to the needs of different application scenarios.

[0038] II. High integration. The artificial neuron in the present application has a magnetic tunnel junction device in the device unit, which can be as small as a few nanometers, so high-density integration can be realized.

[0039] III. Low power consumption. The spin torque diode used in the artificial neuron of the present application has high microwave response sensitivity, and can respond to lower power microwave signals. In addition, by using the microwave energy collection characteristics of the spin torque diode, energy can also be captured from the environment to realize a self-driven artificial neuron. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is a schematic diagram of the principle of the artificial neuron of an embodiment of the present application.

[0041] Figure 2 is a schematic diagram of the signal output by the spin torque diode of an embodiment of the present application.

[0042] Figure 3 is a schematic diagram of the principle of the artificial neuron using time division multiplexing in embodiment one of the present application;

[0043] Figure 4 is a schematic diagram of the output signal of the artificial neuron using time division multiplexing in embodiment one of the present application under different magnetic fields;

[0044] Figure 5 is a schematic diagram of the output signal of the artificial neuron using time division multiplexing in embodiment one of the present application.

[0045] Figure 6 is a schematic diagram of the principle of the artificial neuron using time division multiplexing in embodiment two of the present application.

[0046] Figure 7 This is a schematic diagram illustrating the principle of the artificial neuron in Embodiment 3 of the present invention.

[0047] Figure 8 This is a schematic diagram of the principle of the artificial neuron in Embodiment 4 of the present invention. Detailed Implementation

[0048] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0049] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0050] like Figure 1 As shown, one embodiment of the present invention provides an artificial neuron, including a microwave unit 10, a device unit 20, and a weighting unit 30. The microwave unit 10 is used to generate microwave signals; the device unit 20 responds to the microwave signals and outputs DC voltage signals with multiple bell-shaped curves conforming to different peak response points; the weighting unit 30 assigns weight values ​​to adjust the intensity of the DC voltage signals.

[0051] In one embodiment, the weighting unit 30 is configured to assign multiple weight values ​​corresponding one-to-one with multiple bell-shaped curves output by the device unit 20 to adjust the strength of the DC voltage signal output by the device unit 20. The weighting unit 30 includes multiple weighting devices 31, wherein one weighting device 31 provides a weight value for the DC voltage signal of the device unit 20 conforming to one bell-shaped curve. In a specific embodiment, the weighting device 31 includes a CMOS transistor or a memristor; preferably, the weighting device 31 is a memristor.

[0052] In one embodiment, device unit 20 includes a spin-dipole unit. The spin-dipole unit, due to its varying magnetic anisotropy, can output DC voltage signals with multiple bell-shaped curves conforming to different peak response points. Specifically, the spin-dipole unit may include one or more spin-dipoles 21. In the case of a single spin-dipole, by adjusting the magnetic field, stress, voltage, etc., based on time-division multiplexing technology, multiple microwave response characteristic curves with similar microwave response frequencies and symmetrical characteristics can be obtained. In the case of multiple spin-dipoles, each spin-dipole has a similar microwave response frequency and symmetrical microwave response curve.

[0053] Each spin-torque diode 21 includes a first electrode layer 211, a second electrode layer 212, and a magnetic tunnel junction 213. The second electrode layer 212 is arranged opposite to the first electrode layer 211. The magnetic tunnel junction 213 is arranged between the first electrode layer 211 and the second electrode layer 212. The magnetic tunnel junction 213 includes a magnetic free layer 2131, a barrier layer 2132, a pinned layer 2133, and an anti-ferromagnetic layer 2134, which are sequentially stacked from the first electrode layer 211 to the second electrode layer 212. The magnetic free layer 2131 includes an in-plane magnetic anisotropy layer or a perpendicular magnetic anisotropy layer. The microwave unit 10 is configured to drive the magnetic moment precession of the magnetic free layer 2131 and cause the magnetic tunnel junction 213 to output a rectified voltage signal.

[0054] The working principle of the artificial neuron of the present application is briefly described as follows. First, the magnetic moment of the pinned layer 2133 is fixed by the anti-ferromagnetic layer 2134, so that only the magnetic moment of the magnetic free layer 2131 can move under the action of an external magnetic field or an electrical signal. When the microwave unit 10 injects a microwave with a frequency close to the intrinsic ferromagnetic resonance frequency of the magnetic free layer 2131 into the spin-torque diode 21, an electric current is generated between the magnetic free layer 2131 and the pinned layer 2133 due to the electron tunneling effect. Further, the magnetic moment of the magnetic free layer 2131 is driven to precess by the spin transfer torque effect, so that the magnetic moment arrangement state between the magnetic free layer 2131 and the pinned layer 2133 changes periodically, and the periodic resistance change is exhibited through the tunneling magnetoresistance effect. The periodic resistance change is superimposed with the periodic current change of the input microwave signal of the microwave unit 10, and a direct current voltage signal is output. When the magnetic free layer 2131 has a proper magnetic anisotropy, the size of the output direct current voltage satisfies the symmetric Lorentz curve with the frequency of the input microwave signal, as shown in FIG. 2. Figure 2

[0055] Under the condition of a single spin-torque diode, by using time division multiplexing technology, a plurality of microwave response curves with different characteristic response frequencies can be obtained by adjusting the ferromagnetic resonance frequency of the magnetic tunnel junction 213 of the single spin-torque diode, i.e., the tuning curve required for population coding in the brain can be simulated. By adjusting the output through the weight unit 30, and combining time division multiplexing and integration circuit, the nonlinear activation function of a plurality of artificial neurons can be realized on a single spin-torque diode.

[0056] ​With multiple spin-dual diodes, the magnetic anisotropy of the magnetic free layer 2131 is altered by changing the size of the magnetic tunnel junction 213 and the thickness of the magnetic free layer 2131, resulting in microwave response curves with different characteristic response frequencies for use as tuning curves for population coding. Furthermore, by adjusting the output of each spin-dual diode through the weighting unit 30, the nonlinear activation functions of various artificial neurons can be achieved.

[0057] It should be noted that the spin-dip diode unit is only one specific manifestation of device unit 20, but device unit 20 should not be limited to only being a spin-dip diode unit. Device unit 20 can also have other manifestations, as long as the input-output relationship of the device unit 20 in that manifestation satisfies a symmetrical and linear bell curve, such as... Figure 2 As shown.

[0058] The above is merely a simple description of the basic principle of the artificial neuron of the present invention. The specific technical principles of this application will be further elaborated below in conjunction with different embodiments.

[0059] Example 1:

[0060] like Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the principle of an artificial neuron based on a single spin diode. The artificial neuron in this embodiment includes a microwave source 10, a single spin diode 21, a magnetic anisotropy adjustment element 22, multiple weighting devices 31, and an integrating circuit. The microwave source 10 is connected to the first electrode layer 211 and the second electrode layer 212 of the single spin diode 21, and is used to input microwave signals. The magnetic anisotropy adjustment element 22 is connected to the single spin diode 21 and is used to adjust the magnetic anisotropy of the magnetic free layer 2131 of the spin diode 21 according to the input control electrical signal, so that the spin diode 21 responds to the microwave signal output from the microwave source 40 and outputs DC voltage signals with multiple bell-shaped curves conforming to different peak response points. Multiple weighting devices 31 are connected in parallel. The spin diode 21 can be electrically connected to one of the weighting devices 31 in the weighting unit 30 through the connection switch 40. The integrating circuit is connected to the spin diode 21 and the weighting unit 30. The integrating circuit is used to integrate and output the DC voltage signals generated by the spin diode 21 that conform to multiple bell curves with different peak response points after being assigned weight values ​​by the weighting unit 30.

[0061] In the embodiment, the thickness of the first electrode layer 211 of the spin-torque diode 21 is preferably 10-200 nm, and the material is preferably Pt or Au or Pb or Ta or Cr; the thickness of the second electrode layer 212 is preferably 10-200 nm, and the material is preferably Pt or Au or Pb or Ta or Cr; the thickness of the barrier layer 2132 is preferably 0.1-10 nm, and the material is preferably MgO or Al2O3 or TiO2, and the barrier layer 2132 allows the charge carriers to tunnel between the magnetic free layer 2131 and the pinned layer 2133; the thickness of the pinned layer 2133 is preferably 0.1-10 nm, and the material is preferably CoFeB or Py or Co or CoFe or FePt or a combination thereof. The magnetization direction of the pinned layer 2133 is in-plane or perpendicular in-plane. The in-plane orientation is specifically that the magnetic moment of the pinned layer 2133 is arranged along the plane in which the pinned layer 2133 is located, the magnetization direction is fixed by the anti-ferromagnetic layer 2134, and does not change significantly under the working conditions of the spin-torque diode 21. The perpendicular in-plane orientation is specifically a structure in which the magnetic moment of the pinned layer 2133 is arranged perpendicular to the plane in which the pinned layer 2133 is located. The thickness of the magnetic free layer 2131 is 0.1-10 nm, and the material is at least one of CoFeB or Py or Co or CoFe or FePt or a Co / Ni multilayer film. The magnetization direction of the magnetic free layer 2131 is in-plane or perpendicular in-plane.

[0062] Under the action of the microwave current, the magnetic moment of the magnetic free layer 2131 causes the resistance of the magnetic tunnel junction 213 to change periodically, thereby outputting a rectified voltage, and the rectified voltage satisfies the following relationship with the frequency of the injected microwave signal:

[0063]

[0064] wherein V dc is the output rectified voltage, fr is the intrinsic response frequency of the device, f is the frequency of the input microwave signal. Δf is the full width at half maximum of the microwave response of the device, V A and V S are the amplitudes of the antisymmetric and symmetric Lorentz components in the microwave response of the device, V C is the compensation voltage.

[0065] By selecting the aforementioned materials and controlling the thickness, the magnetic free layer 2131 has a suitable magnetic anisotropy, thereby suppressing the antisymmetric Lorentz component in the rectified voltage of the microwave response of the magnetic tunnel junction 213, and only retaining the symmetric Lorentz component, i.e. V A is 0, thereby simulating the symmetric bell-shaped curve required by the population coding, as shown in Figure 2 .

[0066] In the embodiment, the magnetic anisotropy adjusting member 22 is a magnetic field generating device, which is configured to generate a magnetic field according to an input control electrical signal acting thereon, so as to regulate the magnetic anisotropy of the magnetic free layer 2131, and further to cause the microwave characteristic response frequency of the magnetic tunnel junction 213 to be shifted, so as to obtain multiple clock-shaped tuning curves on a single spin-torque diode based on time-division multiplexing technology. As shown in FIG. 2, the dotted line is the experimental measured spin-torque diode response curve under different magnetic fields, and the solid line is the corresponding Lorentz function. Figure 4

[0067] Further, the output of the spin-torque diode 21 under different adjustment conditions is adjusted by the weight unit 30, wherein the number of weight calculation in the weight unit can be determined according to specific requirements. In the embodiment, nine weight devices are used. In combination with the integral circuit, the nonlinear input-output relationship of multiple software activation functions is realized. Figure 5 Two commonly used nonlinear activation functions, Sigomiod and ReLU activation functions, are obtained based on the data. Figure 4 The solid line is the target fitting function, and the black data points are the input-output relationship obtained based on the experimental data points.

[0068] It should be noted that the weight value can be obtained by the following derivation:

[0069] Firstly, the neuron population response curve can be expressed as:

[0070]

[0071] where w i represents the weight of the output of the i-th neuron, and r i (θ) is the response function of the neuron. Therefore, by changing the weight w i , different population response curves can be obtained to adapt to different task requirements.

[0072] ​Multiple neuron activation functions are implemented on a single spin-diode using spin-diode population encoding. This involves setting the population response curve H(θ) in the formula as a specific activation function and calculating the weight value corresponding to each curve based on all bell-shaped response curves, implemented using a weighting device. In single-spin-diode mode, the first step is to determine the required number of curves n, corresponding to the n states required by the spin-diode and weighting device. When an external signal is input, the spin-diode responds and the weighting device stores the output in an integrator circuit. Subsequently, a magnetic anisotropy adjustment element changes the state of the spin-diode, altering the characteristic response value of the microwave response curve. The weighting device then re-selects the weights to represent the response value, and the output in this state is stored in the integrator circuit. Repeating these steps, the outputs from the n states are integrated using the integrator circuit to obtain the output of the time-division multiplexed spin-diode population, thereby achieving nonlinear activation functions for multiple artificial neurons.

[0073] Example 2:

[0074] like Figure 6 As shown, the difference between this embodiment and Embodiment 1 lies only in that the magnetic anisotropy adjustment component 22 employs a magnetoelectric coupling device. This device generates stress based on the input control electrical signal acting upon it and transmits the stress to the magnetic free layer 2131 of the spin-torque diode 21 to regulate the magnetic anisotropy of the magnetic free layer 2131, thereby achieving regulation of the microwave response center frequency of the magnetic tunnel junction 213. In a preferred embodiment, the magnetoelectric coupling device includes a piezoelectric thin film disposed on the first electrode layer 211 or the second electrode layer 212 of the spin-torque diode 21. The piezoelectric thin film generates stress based on the input electrical signal acting upon it.

[0075] Specifically, the piezoelectric thin film can be made of piezoelectric or ferroelectric materials, with barium titanate (BaTiO3), lead magnesium niobate (PMN-PT), and lead zirconate titanate (PZT) being preferred ferroelectric materials. In this embodiment, the voltage-induced magnetic anisotropy is mainly controlled by the strain transfer mechanism. When a voltage is applied, an electric field is generated on the piezoelectric thin film. Due to the inverse piezoelectric effect of the ferroelectric or piezoelectric materials, the piezoelectric thin film will generate strain. This strain is transferred through the interface to the magnetic free layer 2131 of the magnetic tunnel junction 213, and then, through the magnetostrictive effect, changes the magnetic anisotropy of the magnetic free layer 2131. It should be noted that the stress mainly changes the magnetic anisotropy of the magnetic free layer 2131, thereby changing the microwave center response frequency of the spin-torque diode 21.

[0076] Furthermore, by inputting multiple different signals into the piezoelectric film, multiple tuning curves required for group coding can be obtained.

[0077] Example 3:

[0078] As shown in Figure 7 The difference between this embodiment and embodiments 1 and 2 is that a plurality of spin-torque diodes 21 are contained in the spin-torque diode unit 20 in this embodiment. The artificial neuron in this embodiment is realized by population coding of the plurality of spin-torque diodes 21, without a magnetic anisotropy adjusting member. The number of spin-torque diodes 21 is determined according to specific requirements, and the number selected in this embodiment is 5. The input end of each spin-torque diode 21 is electrically connected to the microwave unit 10, and the output end of each spin-torque diode 21 is electrically connected to a weight device 31, and each weight device 31 provides a weight value for the direct current voltage signal output by the spin-torque diode 21 connected thereto in accordance with the clock curve.

[0079] Specifically, by controlling the size of the magnetic tunnel junction 213 of each spin-torque diode 21, or the thickness of the magnetic free layer 2131, etc., all the spin-torque diodes 21 have symmetric Lorentz linear microwave response characteristic curves and slightly different microwave center response frequencies, so as to obtain a plurality of tuning curves required for population coding. Further, by controlling the corresponding weight device 31 of each spin-torque diode 21, the activation function of the artificial neuron is realized.

[0080] Compared with embodiments 1 and 2, the artificial neuron realized by population coding of a plurality of spin-torque diodes can effectively reduce the required time for calculation and improve the calculation speed.

[0081] Embodiment 4:

[0082] As shown in Figure 8 The difference between this embodiment and embodiment 3 is that the microwave unit 10 in this embodiment is a microwave antenna, which detects the microwaves in the environment and converts the microwaves into microwave electrical signals transmitted to the spin-torque diode 21, and further, the spin-torque diode 21 directly processes the microwave electrical signals and outputs a direct current signal.

[0083] Compared with the traditional microwave signal detection and processing system, the artificial neuron realized by population coding of the spin-torque diode 21 can directly process the target microwave information in the environment, without the need for traditional analog-digital conversion, which can effectively improve the energy efficiency and directly perform artificial neural network calculation.

[0084] Compared with the prior art, the artificial neuron of the present application has the following technical effects:

[0085] I. High flexibility. The artificial neuron in the present application can realize a plurality of nonlinear input-output characteristics in the same neuron by adjusting the weight unit, thereby simulating different activation functions to adapt to the needs of different application scenarios.

[0086] II. High integration. The artificial neuron of the present application has a device unit with a magnetic tunnel junction, and the size can be as low as several nanometers, so high-density integration can be achieved.

[0087] III. Low power consumption. The spin torque diode used in the artificial neuron of the present application has high microwave response sensitivity, and can respond to lower power microwave signals. In addition, using the microwave energy harvesting characteristics of the spin torque diode, energy can also be captured from the environment to achieve a self-driven artificial neuron.

[0088] The foregoing description of specific exemplary embodiments of the application has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the application to the precise forms disclosed, and obviously many modifications and variations are possible in light of the above teaching. The exemplary embodiments were chosen and described in order to explain the principles of the application and its practical application, thus enabling others skilled in the art to implement and utilize the application in various different exemplary embodiments and with various modifications as are suited to the particular use contemplated. The scope of the application is intended to be defined by the claims and their equivalents.

Claims

1. An artificial neuron, characterized in that, include: Microwave unit, used to generate microwave signals; The device unit responds to the microwave signal and outputs a DC voltage signal that conforms to multiple bell-shaped curves with different peak response points. The device unit includes a spin-torque diode unit. A weighting unit is used to assign weight values ​​to adjust the intensity of the DC voltage signal; the weighting unit includes multiple weighting devices. The spin-difference diode unit includes a spin-difference diode, and the artificial neuron further includes an integrator circuit connected to the spin-difference diode and the weighting unit. The integrator circuit is used to integrate and output the DC voltage signals generated by the spin-difference diode from multiple bell-shaped curves conforming to different peak response points, after the weighting unit assigns weight values. Alternatively... The spin-moment diode unit includes multiple spin-moment diodes. The input terminal of each spin-moment diode is electrically connected to the microwave unit, and the output terminal of each spin-moment diode is electrically connected to a weighting device. Each weighting device provides a weight value for the DC voltage signal that conforms to the bell curve output by the spin-moment diode connected to it. The DC voltage signals of multiple bell curves that conform to different peak response points are output as a group after being assigned different weights, so as to activate the nonlinear objective function.

2. The artificial neuron as described in claim 1, characterized in that, The weighting unit is configured to assign multiple weight values ​​corresponding one-to-one with multiple bell curves output by the device unit to adjust the intensity of the DC voltage signal output by the device unit.

3. The artificial neuron as described in claim 2, characterized in that, One of the weighting devices provides a weight value for the DC voltage signal of the device unit that conforms to a bell curve.

4. The artificial neuron as described in claim 3, characterized in that, The weighting devices include CMOS transistors or memristors.

5. The artificial neuron as described in claim 1, characterized in that, The spin-dipole unit can output DC voltage signals with multiple bell-shaped curves that conform to different peak response points due to its different magnetic anisotropy.

6. The artificial neuron as described in claim 5, characterized in that, The spin-moment diode unit includes a spin-moment diode; The artificial neuron also includes a magnetic anisotropy modulator, which acts on the spin-moment diode to regulate the magnetic anisotropy of the spin-moment diode according to the input control electrical signal acting on it, so that the spin-moment diode outputs a DC voltage signal that conforms to multiple bell-shaped curves with different peak response points.

7. The artificial neuron as described in claim 6, characterized in that, Multiple weighting devices are connected in parallel, and the spin-torque diode can be selectively electrically connected to one of the weighting devices in the weighting unit via a connection switch.

8. The artificial neuron as described in claim 5, characterized in that, The spin-torque diode unit includes multiple spin-torque diodes; The multiple spin-torque diodes have different magnetic anisotropies to output DC voltage signals with multiple bell-shaped curves that conform to different peak response points.

9. The artificial neuron as described in claim 1, characterized in that, The microwave unit includes a microwave source or a microwave antenna for inputting microwave signals.

Images