A stochastic neuron circuit and a discharge mode extraction method thereof

Abstract

The present application relates to a kind of random neuron circuit and its discharge mode extraction method, a kind of neuron circuit is constructed based on threshold transition device, resistance and capacitor, the pulse emission characteristics of neuron is realized based on the circuit.Compared with other neuron circuits based on threshold transition device, the neuron circuit can realize the conversion of different discharge modes under fixed parameters, and due to the randomness of neuron, the mode switching of circuit output has the probability transition behavior, which is beneficial to construct high-order intelligent brain-like system.At the same time, the present application uses the integral and resonance characteristics of H-H neuron itself to extract the mode of the probability output of the previous neuron, uses the resonance characteristics of unimodal detection neuron circuit to extract the unimodal output of intermediate neuron circuit, and uses the integral characteristics of bimodal detection neuron circuit to extract the bimodal output of intermediate neuron circuit, so as to complete the calculation process of multiplexing encoding and decoding of information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain-like bionics, and particularly to a random neuron circuit and a discharge mode extraction method thereof. BACKGROUND

[0002] Biological neurons can generate 23 different discharge behaviors, and different discharge types have different effects. The diversity of neuron discharge types endows them with the ability to complete high-level behaviors such as encoding input features, network synchronous oscillation, and decision-making. To build a high-order brain-like intelligent system, artificial neuron circuits need to be able to generate different pulse patterns like biological neurons. Currently, neuron circuits based on threshold transition devices have been reported, but most of them can only generate one discharge mode, cannot faithfully simulate the multiple discharge behaviors of biological neurons, and have complex circuit parameters that are not easy to adjust, which is not conducive to the construction of a high-order intelligent brain-like system.

[0003] Among various neuron models, the Hodgkin-Huxley (H-H) model has high biological fidelity and almost simulates all discharge modes of biological neurons. H-H neuron circuits based on CMOS technology have been widely studied and implemented. However, the complexity of the circuits hinders their practical application. In recent years, high-dynamic memory resistors have provided a promising platform for building compact H-H neurons. However, these circuits usually show a complex parameter tuning process to achieve different neuron discharge modes, which makes them hardware-unfriendly. Therefore, it is a problem worth paying attention to to achieve different discharge modes in H-H neuron circuits with fixed circuit element parameters. SUMMARY

[0004] The present application is to overcome the defects of the prior art and provide a random neuron circuit and a discharge mode extraction method thereof.

[0005] The object of the present application can be achieved by the following technical solutions:

[0006] A random neuron circuit, comprising a first branch, a second branch, a resistor R1, a resistor R2 and a capacitor;

[0007] The first branch comprises a direct current biased threshold transition memristor TS1, and the second branch comprises a direct current biased threshold transition memristor TS2;

[0008] One end of the resistor R1 is connected to the first branch, and the other end serves as an input end;

[0009] One end of the resistor R2 is connected to the first branch, and the other end is connected to the second branch;

[0010] The one end of the first branch not connected with the resistor R1 is grounded; the one end of the second branch connected with the resistor R2 is used as an output end, and the other end is grounded.

[0011] Further, the threshold transition memristor TS1 and the threshold transition memristor TS2 are respectively connected with an external capacitor in parallel or use a parasitic capacitor of the device itself as an energy storage capacitor.

[0012] Further, the threshold transition memristor TS1 and the threshold transition memristor TS2 both include a top electrode, a threshold transition layer and a bottom electrode.

[0013] Further, the preparation of the threshold transition memristor TS1 and the threshold transition memristor TS2 includes the following steps:

[0014] forming a SiO2 layer by oxidizing on a silicon wafer;

[0015] depositing a bottom electrode on the SiO2 layer;

[0016] depositing a threshold transition layer on the bottom electrode;

[0017] depositing a top electrode on the threshold transition layer.

[0018] Further, the thickness of the bottom electrode is 10nm-100nm, and the bottom electrode includes a bottom electrode made of TiN, Poly-Si, Pd, W, Pt or Au.

[0019] Further, the thickness of the threshold transition layer is 5nm-50nm, and the threshold transition layer is deposited by using a material with a volatile threshold transition characteristic.

[0020] Further, the thickness of the top electrode is 10nm-100nm, and the top electrode includes a top electrode made of TiN, Poly-Si, Pd, Pt, W, Cu, Ag or Au.

[0021] A discharge mode extraction method of a random neuron circuit, based on a random neuron circuit as described above, includes the following steps:

[0022] constructing a discharge mode extraction circuit, which includes an intermediate neuron circuit, a single-peak detection neuron circuit and a double-peak detection neuron circuit, the output end of the intermediate neuron circuit is connected to a comparator, and then connected to the input ends of the single-peak detection neuron circuit and the double-peak detection neuron circuit respectively;

[0023] extracting the single-peak output of the intermediate neuron circuit by using the resonance characteristic of the single-peak detection neuron circuit, and filtering out the double-peak output;

[0024] The double-peak detection neuron circuit extracts the double-peak output of the intermediate neuron circuit by using the integral characteristic of the double-peak detection neuron circuit, and filters out the single-peak output.

[0025] Further, the intermediate neuron circuit, the single-peak detection neuron circuit and the double-peak detection neuron circuit in the discharge mode extraction circuit have the same structure, and each circuit comprises a first branch, a second branch, a resistor R1, a resistor R2 and a capacitor.

[0026] The first branch comprises a direct current biased threshold transition memristor TS1, and the second branch comprises a direct current biased threshold transition memristor TS2.

[0027] One end of the resistor R1 is connected with the first branch, and the other end serves as an input end.

[0028] One end of the resistor R2 is connected with the first branch, and the other end is connected with the second branch.

[0029] The end of the first branch not connected with the resistor R1 is grounded, and the end of the second branch connected with the resistor R2 serves as an output end, and the end not connected with the resistor R2 is grounded.

[0030] Further, the capacitor and the resistor parameters of each circuit in the intermediate neuron circuit, the single-peak detection neuron circuit and the double-peak detection neuron circuit in the discharge mode extraction circuit are different.

[0031] Compared with the prior art, the present application has the following beneficial effects:

[0032] 1. The present application constructs a neuron circuit based on a threshold transition device, a resistor and a capacitor, and realizes the pulse emission characteristic of the neuron based on the circuit. Compared with other neuron circuits based on threshold transition devices, the neuron circuit can realize the conversion of different discharge modes under fixed parameters, and due to the randomness of the neuron, the mode switching of the circuit output has a probabilistic transition behavior, which is beneficial to construct a high-order intelligent brain-like system.

[0033] 2. The present application uses the integral and resonance characteristics of the H-H neuron itself to extract the mode of the probabilistic output of the previous neuron, proposes a new type of neuron calculation method, uses the resonance characteristic of the single-peak detection neuron circuit to extract the single-peak output of the intermediate neuron circuit, and filters out the double-peak output; uses the integral characteristic of the double-peak detection neuron circuit to extract the double-peak output of the intermediate neuron circuit, and filters out the single-peak output, thereby completing the calculation process of multiplexing encoding and decoding of information. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 It is a neuron circuit principle diagram constituted by a TS device and a resistor, an external capacitor in the embodiment of the present application;

[0035] Figure 2 This is a schematic diagram of a neuron circuit composed of the parasitic capacitance and resistance of the TS device itself in an embodiment of the present invention;

[0036] Figure 3 This is a flowchart illustrating the fabrication process of the threshold switching memristor device in an embodiment of the present invention.

[0037] Figure 4 This is an IV characteristic diagram of the threshold switching memristor device in an embodiment of the present invention;

[0038] Figure 5 The threshold switching memristor device V in this embodiment of the invention th V hold Characteristic diagram;

[0039] Figure 6 Schematic diagram of different output modes of neuron circuits under different input conditions;

[0040] Figure 7 This is a schematic diagram showing the output change of a neuron circuit under triangular wave input.

[0041] Figure 8 A schematic diagram illustrating the probabilistic pattern transition behavior caused by the threshold randomness of the device;

[0042] Figure 9 The graph shows the probability of two firing modes of a neuron circuit as a function of input voltage.

[0043] Figure 10 The circuit schematic for extracting firing patterns from the probabilistic output of a preneuron;

[0044] Figure 11 This is a schematic diagram of the output of each loop in the discharge mode extraction circuit. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0046] To address the shortcomings of existing technologies, this invention constructs a neuron circuit based on a threshold switching device, resistors, and capacitors. This circuit realizes the pulse firing characteristics of neurons. Compared to other neuron circuits based on threshold switching devices, this circuit can achieve switching between different firing modes under fixed parameters. Furthermore, due to the randomness of neurons, the mode switching of the circuit output exhibits probabilistic switching behavior. This invention also relates to a novel neuron computation method based on this circuit.

[0047] like Figures 1-2As shown, the neuron circuit proposed by the application is composed of two similar parts, each part is composed of a DC biased threshold transition memristor (TS1 or TS2) in parallel with a capacitor (C1 or C2) (which can also be the parasitic capacitor of the device itself) and finally in series with a load resistor (R1 or R2).

[0048] Figure 1 The schematic diagram of the neuron circuit composed of a threshold transition device (TS) and a resistor and a capacitor; Figure 2 The schematic diagram of the neuron circuit composed of a threshold transition device (TS), a parasitic capacitor and a resistor.

[0049] When there is an input signal at the input end, C1 is charged, and the voltage across TS1 gradually increases. When the threshold of TS1 is reached, TS1 opens, C1 discharges, C2 charges, and the voltage across TS2 gradually rises. When the threshold of TS2 is exceeded, TS2 opens, C2 discharges, and the output voltage rises, generating a single spike (tonic spike) output. When the input voltage increases, the forward charging of C1 becomes faster, and the reverse charging becomes slower (the effective voltage decreases), so TS1 will open faster, but the opening time will increase. Since TS1 opens faster, the increase in input voltage has less effect on the discharge rate of C2, so the opening time of TS2 gradually advances, and the closing time of TS1 gradually delays. When the opening time of TS2 advances to a certain extent, the voltage of the constant voltage source charges C1 through the coupled R2, causing TS1 to fail to close normally. As the voltage across TS2 gradually decreases, TS2 closes, C1 normally reverse charges, and V K The voltage is reduced through R2 coupling, and TS1 closes. After TS1 closes, C1 discharges, and V K Continued reduction eventually leads to the opening and closing of TS2 again, generating a burst of spike output.

[0050] The basic structure of the threshold transition memristor device includes an upper electrode, an intermediate layer, and a lower electrode. As shown in the figure, Figure 3 As shown, the preparation scheme of the threshold transition memristor device includes the following steps:

[0051] Step 1: Oxidize on a silicon wafer to form a SiO2 layer, the thickness of SiO2 is 100-300 nm, and the thickness of the oxide layer can also be reduced or increased according to actual process conditions.

[0052] Step 2: Deposit a lower electrode (BE) on the silicon oxide, the thickness of the lower electrode is 10-100 nm. The lower electrode material can be TiN, Poly-Si, Pd, Pt, W, or Au, etc. inert conductive material.

[0053] Step 3: Deposit a threshold switching layer (TS) on the bottom electrode, with a thickness of 5nm-50nm. The TS material can be NbO2, VO2, SiTe, SiO2:Ag, a-Si:Cu, a-Si:Ag, or AM4Q8(A=Ga, Ge; M=V, Nb, Ta, Mo; Q=S, Se) mixed material. The functional layer material includes but is not limited to the above materials, and any material with a volatile threshold switching property can be applied.

[0054] Step 6: Deposit a top electrode (TE) on the functional layer, with a thickness of 10nm-100nm. The intermediate electrode material can be TiN, Poly-Si, Pd, Pt, W, Cu, Ag, or Au, etc.

[0055] The prepared threshold switching memristive device should have an I-V characteristic as shown in Figures 4-5 under voltage scanning.

[0056] When the voltage applied on the threshold switching layer TE exceeds a certain voltage value (V th ), the TS device is switched from a high resistance state to a low resistance state. During the voltage sweep back, when the voltage is less than a certain voltage value (V hold ), the device returns to the high resistance state from the low resistance state due to the insufficient voltage on the TE. The V th , V hold of the device have randomness, satisfying a Gaussian distribution.

[0057] For a neuron circuit built by two threshold switching devices, two capacitors, and two resistors as shown in Figure 1 , the different output modes of the neuron under different input conditions are shown in Figure 6 . When the input is small, the neuron fires in a single-peak tonic mode, and when the input increases, the neuron fires in a double-peak burst mode. The change in the neuron output under a triangular wave input is shown in Figure 7 . As the input gradually increases, the output mode of the neuron changes from a single-peak mode to a double-peak mode, and the firing frequency gradually increases.

[0058] The probabilistic mode transition behavior caused by the threshold randomness of the device is shown in Figure 8 . Under pulse input, the mode of the neuron output is a random single-peak or double-peak mode. The curve of the probability of the two firing modes of the neuron changing with the input voltage is shown in Figure 9 . The greater the input, the greater the probability of the neuron output in the burst mode and the smaller the probability of the neuron output in the tonic mode under the same pulse width.

[0059] The H-H neuron itself is used to extract the pattern of the probability output of the pre-neuron by using the integral and resonance characteristics of the H-H neuron, and the circuit principle is as shown in Figure 10 The circuit includes a relay neuron circuit, a tonic detection neuron circuit and a burst detection neuron circuit, and the output end of the relay neuron circuit is connected to the input end of the tonic detection neuron circuit and the burst detection neuron circuit after being connected to a comparator.

[0060] The structures of the three circuits are the same, and are consistent with the neuron circuit structure mentioned in the above embodiment. The difference between the three circuits is only the difference in the capacitance and resistance parameters.

[0061] The relay neuron circuit encodes the input voltage into a pulse output with single and double peaks alternating, and the proportion of the double peaks increases with the increase of the input voltage, so it can be used to represent the input intensity. The principle of generating such single and double peak alternation is that the V th , V hold has randomness.

[0062] The tonic detection neuron has a resonance characteristic, and only a specific frequency input can make it emit a spike. Therefore, it can be used to extract the single peak (tonic spiking) in the single and double peak alternation pattern emitted by the relay neuron circuit.

[0063] The burst detection neuron has an integral characteristic. For an input with a large pulse interval, the capacitor will leak during the interval, so it will not produce a spike output. For an input with a small pulse interval, it can effectively integrate the input voltage to reach the threshold voltage of the circuit, thereby generating a spike. Therefore, it can be used to extract the double peak (bursting) in the single and double peak alternation pattern emitted by the relay neuron.

[0064] The capacitance and resistance parameters of the three neuron circuits are different, so they can realize their respective functions. As a preferred embodiment, the parameters of each circuit are set as follows:

[0065] In the relay neuron circuit, the V th of the TS device (threshold switching memristor) is 1.65V, the V hold is 1.3V, R1 is 23KΩ, R2 is 4KΩ, C1 is 1nF, C2 is 330pF, E1 is -1.55V, and E2 is 1.55V. In this embodiment, the resistance and capacitance values can be adjusted according to the actual situation.

[0066] In the single-peak detection neuron circuit, R1 = 7KΩ, R2 = 4KΩ, C1 = 2nF, C2 = 500pF, E1 = -1.55V, E2 = 1.55V;

[0067] In the dual-peak detection neuron circuit, R1 = 25KΩ, R2 = 4KΩ, C1 = 2nF, C2 = 330pF, E1 = -1.55V, E2 = 1.55V;

[0068] It is worth noting that the values ​​of capacitors and resistors in each circuit are set according to the V of the selected TS device. th V hold The parameter settings in this embodiment are for illustrative purposes only. In actual implementation, the settings should be adjusted according to the type of TS device selected.

[0069] like Figure 11 As shown, the resonance characteristic of a single-peak detection neuron can be used to extract the single-peak output of the preceding neuron and filter out the double-peak output; similarly, the integral characteristic of a double-peak detection neuron can be used to extract the double-peak output of the preceding neuron and filter out the single-peak output. This completes the computational process of multiplexing encoding and decoding information.

[0070] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A stochastic neuron circuit, characterized by, comprising a first branch, a second branch, a resistor , a resistor and a capacitor; The first branch includes a direct current biased threshold switching memristor The second branch includes a direct current biased threshold switching memristor ​ the resistance one end is connected with the first branch, and the other end serves as an input end; the resistance one end is connected with the first branch, and the other end is connected with the second branch; The first branch is grounded at one end not connected to the resistor The second branch is connected to the resistor at one end as an output end, and grounded at one end not connected to the resistor The second branch is connected to the resistor at one end as an output end, and grounded at one end not connected to the resistor The second branch is connected to the resistor at one end as an output end, and grounded at one end not connected to the resistor The threshold switching memristor And preparation of threshold switching memristor The preparation includes the following steps: forming a SiO2 layer on a silicon wafer by oxidation; depositing a bottom electrode on the SiO2 layer; depositing a threshold transition layer on the bottom electrode; depositing a top electrode on the threshold transition layer; the thickness of the bottom electrode is 10nm-100nm, and the bottom electrode is made of TiN, Poly-Si, Pd, W, Pt or Au; the thickness of the threshold transition layer is 5nm-50nm, and the threshold transition layer is deposited by using a material with volatile threshold transition characteristics; the thickness of the top electrode is 10nm-100nm, and the top electrode is made of TiN, Poly-Si, Pd, Pt, W, Cu, Ag or Au.

2. A stochastic neuron circuit according to claim 1, wherein, The threshold-transition memristor And threshold-transition memristor Respectively parallel external capacitor or take the parasitic capacitor of the device itself as the energy storage capacitor.

3. A stochastic neuron circuit according to claim 1, wherein, The threshold transition memristor And threshold transition memristor Both include a top electrode, a threshold transition layer, and a bottom electrode.

4. A method of spike pattern extraction of a stochastic neuron circuit, characterized by, Based on the random neuron circuit implementation as claimed in any one of claims 1-3, comprising the following steps: constructing a discharge pattern extraction circuit, which comprises an intermediate neuron circuit, a single-peak detection neuron circuit and a double-peak detection neuron circuit, the output end of the intermediate neuron circuit is connected to a comparator, and then connected to the input end of the single-peak detection neuron circuit and the double-peak detection neuron circuit respectively; extracting the single-peak output of the intermediate neuron circuit by using the resonance characteristics of the single-peak detection neuron circuit, and filtering out the double-peak output; extracting the double-peak output of the intermediate neuron circuit by using the integration characteristics of the double-peak detection neuron circuit, and filtering out the single-peak output.

5. The discharge pattern extraction method of a stochastic neuron circuit according to claim 4, wherein The intermediate neuron circuit, the single-peak detection neuron circuit and the double-peak detection neuron circuit in the discharge mode extraction circuit are of the same structure, each circuit includes a first branch, a second branch, a resistor , a resistor and a capacitor; The first branch includes a direct current biased threshold switching memristor The second branch includes a direct current biased threshold switching memristor ; the resistance one end is connected with the first branch, and the other end serves as an input end; the resistance one end connected to the first branch and the other end connected to the second branch; The first branch is grounded at one end not connected to the resistor The second branch is connected to the resistor at one end as an output end, and grounded at one end not connected to the resistor The second branch is connected to the resistor at one end as an output end, and grounded at one end not connected to the resistor ​ 6. The discharge pattern extraction method of a stochastic neuron circuit according to claim 5, wherein In the intermediate neuron circuit, the single-peak detection neuron circuit and the double-peak detection neuron circuit in the discharge pattern extraction circuit, the capacitance and resistance parameters of each circuit are different.

Images