A dynamic wake-up neuron circuit and an event-driven neural network system

Abstract

The application relates to a dynamic wake-up neuron circuit and an event-driven neural network system, wherein the dynamic wake-up neuron circuit comprises an adjustable resistor R1, a capacitor C1, a bidirectional threshold value conversion device TS and a capacitor C2, the adjustable resistor R1 is connected with the capacitor C1 in series, the bidirectional threshold value conversion device TS is connected with the capacitor C2 in parallel, the series connection of the adjustable resistor R1 and the capacitor C1 is connected with the parallel connection of the bidirectional threshold value conversion device TS and the capacitor C2 in series, one end of the bidirectional threshold value conversion device TS is connected with the capacitor C1, and the other end is grounded, one end of the adjustable resistor R1 is connected with an excitation signal, a connection point of the bidirectional threshold value conversion device TS and the capacitor C1 is a voltage output end, and the other end of the bidirectional threshold value conversion device TS grounded is a current output end, that is, an event signal output end. Compared with the prior art, the application has the advantages of simple circuit structure, high-density and large-scale integration, small area, low energy consumption and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing and dynamic intelligent sensing technology, in particular to a dynamic wake-up neuron circuit and an event-driven neural network system. BACKGROUND

[0002] With the development of new technologies such as artificial intelligence, Internet of Things, and cloud computing, human society is developing from informationization to intelligentization. The construction of an intelligent society puts higher demands on the development of information technology, and systems need to be able to acquire, process, and make decisions in real time on external environmental information. Due to the increasing amount of interaction data, the conventional data acquisition and processing mode based on "frames" poses a huge challenge to traditional information systems that separate sensing and computing. One solution to alleviate this dilemma is to accelerate hardware and continuously improve the computing power of the system, which is the mainstream solution in the current industry and academia. However, since data growth is endless, this solution will eventually encounter its bottleneck. Another solution is to draw on the event-driven information processing mode of biological sensing systems to alleviate the problem of large data volume from the source. Event-driven cameras are a typical example of this information processing mode. Unlike traditional frame cameras, they do not capture images at a fixed rate, but asynchronously measure the brightness changes of each pixel and output a series of events that encode time, location, and brightness changes. This camera only outputs information on the location of the pixels in the receptive field that have changed, and there is no information output for pixels that have not changed, avoiding repeated collection of static data and reducing the amount of information. In addition, compared with traditional cameras, event cameras offer attractive features: high temporal resolution, high dynamic range, low power consumption, high pixel bandwidth, and reduced motion object trailing in "frame" mode. Therefore, event cameras have great application potential in challenging scenarios such as low latency, high speed, and high dynamic range of traditional cameras. Spiking neural networks are a neural network algorithm inspired by the human brain, with event-driven characteristics that can further process event information, and are one of the ideal algorithms for building low-power neuromorphic hardware systems.

[0003] Implementing the above dynamic signal detection and event output at the data source requires a dynamic signal detection circuit for each pixel, which only outputs event responses to changing signals. Positive and negative events can be given for signals that increase or decrease, respectively. The dynamic signal detection circuit in the prior art is shown in Figure 1 and Figure 2 As can be seen from the figure, the existing dynamic signal detection circuit has a complex structure, which is not conducive to high-density and large-scale integration. SUMMARY

[0004] The present application aims at providing a dynamic wake neuron circuit and an event-driven neural network system with simple circuit structure, high-density and large-scale integration and small area.

[0005] The present application can be realized by the following technical solutions.

[0006] The dynamic wake neuron circuit comprises an adjustable resistor R1, a capacitor C1, a bidirectional threshold switching device TS and a capacitor C2.

[0007] Preferably, the adjustable resistor R1 is an adjustable non-volatile resistive switching device.

[0008] Preferably, the bidirectional threshold switching device TS is a vertical bidirectional threshold switching device or a planar bidirectional threshold switching device.

[0009] Preferably, the bidirectional threshold switching device TS has a symmetrical current-voltage characteristic curve.

[0010] More preferably, the vertical bidirectional threshold switching device is obtained by the following method.

[0011] Step 1: oxidizing a silicon wafer to form a SiO2 layer;

[0012] Step 2: depositing a lower electrode BE on the SiO2 layer;

[0013] Step 3: depositing a SiO2 isolation layer on the lower electrode BE;

[0014] Step 4: etching a hole through the SiO2 isolation layer and the lower electrode BE on the deposited thin film by etching process, and finally etching a certain thickness on the SiO2 layer to ensure that the cross section of the lower electrode BE is completely exposed;

[0015] Step 5: depositing a functional layer FL on the etched hole;

[0016] Step 6: depositing a first intermediate electrode ME1 on the deposited functional layer FL;

[0017] Step 7: depositing a capacitor dielectric thin film on the first intermediate electrode ME1;

[0018] Step 8: depositing a second intermediate electrode ME2 on the capacitive dielectric film;

[0019] Step 9: depositing a resistive film on the second intermediate electrode ME2;

[0020] Step 10: depositing a top electrode TE on the resistive film.

[0021] More preferably, the acquisition method of the planar bidirectional threshold transition device is as follows:

[0022] Step 1: oxidizing on a silicon wafer to form a SiO2 layer;

[0023] Step 2: depositing a bottom electrode BE on the SiO2 layer;

[0024] Step 3: depositing a functional layer FL on the bottom electrode BE;

[0025] Step 4: depositing a first intermediate electrode ME1 on the functional layer DL;

[0026] Step 5: depositing a capacitive dielectric film on the first intermediate electrode ME1;

[0027] Step 6: depositing a second intermediate electrode ME2 on the capacitive dielectric film;

[0028] Step 7: depositing a resistive film on the second intermediate electrode ME2;

[0029] Step 8: depositing a top electrode TE on the resistive film.

[0030] Preferably, the capacitance C2 is a parasitic capacitance or an external capacitance.

[0031] An event-driven neural network system, comprising a DVS pixel array, an address event representation module and a neuromorphic core connected in sequence; the DVS pixel array is composed of an array of DVS pixel point circuits, used for detecting moving objects; the neuromorphic core is used for identifying moving objects; the DVS pixel point circuit comprises the dynamic wake-up neuron circuit as described above.

[0032] Preferably, the dynamic wake-up neuron circuit is provided with a signal amplification module and a photoelectric detector; one end of the signal amplification module is connected with an adjustable resistor R1, and the other end is connected with the photoelectric detector; the other end of the photoelectric detector is grounded; the excitation signal is input into the dynamic wake-up neuron circuit through the signal amplification module; the photoelectric detector is used for turning on the dynamic wake-up neuron circuit when the brightness of the pixel point changes.

[0033] More preferably, the DVS pixel circuit is provided with an intensity detection circuit, comprising a trigger switch, a bidirectional threshold switching device and a capacitor; the bidirectional threshold switching device is connected in parallel with the capacitor; the trigger switch is connected in series with the parallel connection of the bidirectional threshold switching device and the capacitor; the excitation signal is input into the dynamic wake-up neuron circuit and the intensity detection circuit through a mirror current source, respectively.

[0034] Preferably, the neuromorphic core comprises:

[0035] a synaptic weight array storing a weight matrix composed of neural network weights, for performing multiplication and addition calculation of the weight matrix and an input vector;

[0036] a neuron array composed of a plurality of neuron circuits, each neuron circuit comprising a capacitor for performing integration operation and a bidirectional threshold switching device for determining the firing of an action potential as a threshold switch;

[0037] The neuron circuit integrates the weighted current or voltage signal from the synapse, and fires an action potential signal when the integrated potential reaches the threshold of the neuron, and the mode of the input signal can be determined according to the firing frequency of the action potential of the neurons in the neuron array, so that the object detected by the DVS pixel array can be identified.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] I. Simple circuit structure, conducive to high-density and large-scale integration: The dynamic wake-up neuron circuit and the circuit structure of the event-driven neural network system in the present application are simple, capable of simultaneously sensing the dynamic change of light and outputting the real-time light intensity, and conducive to high-density and large-scale integration.

[0040] II. Small area: The event-driven neural network system in the present application only uses one photodetector to sense the change of light, and additionally uses a signal amplifier module to amplify the change of signal, so that the circuit structure is simple and the area is small, which is conducive to integration.

[0041] III. Low energy consumption and large dynamic range: The dynamic wake-up neuron circuit in the present application uses a novel bidirectional threshold switching device, which has the advantages of fast switching speed and low switching energy consumption, which is conducive to reducing the energy consumption of dynamic signal detection and improving the response accuracy.

[0042] IV. High configurability: The dynamic wake-up neuron circuit in the present application uses a variable resistor (e.g., a memristor) as the resistor R1, and the resistance value of the resistor R1 can adjust the response accuracy of the neuron circuit, which is conducive to the application requirements of different dynamic scenes.

[0043] V. Delay low: the event-driven neural network system in the application adopts a computing architecture integrating storage and computing, can perform in-situ calculation on input signals from the AER circuit, does not need to perform reciprocating transmission between the computing unit and the storage unit, and is beneficial to improving the calculation energy efficiency and the real-time performance of information processing. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 It is a structural schematic diagram of a dynamic signal detection circuit in the prior art;

[0045] Figure 2 It is another structural schematic diagram of a dynamic signal detection circuit in the prior art;

[0046] Figure 3 It is a structural schematic diagram of a dynamic wake-up neuron circuit in the application;

[0047] Figure 4 It is a structural schematic diagram of a vertical bidirectional threshold transition device in the application;

[0048] Figure 5 It is a structural schematic diagram of a planar bidirectional threshold transition device in the application;

[0049] Figure 6 It is an I-V characteristic diagram of a bidirectional threshold transition device in the application;

[0050] Figure 7 It is an output characteristic schematic diagram of a dynamic wake-up neuron in the application under a dynamic input signal;

[0051] Figure 8 It is a time output characteristic schematic diagram of a dynamic wake-up neuron in the application under a sinusoidal voltage input signal;

[0052] Figure 9 It is a relationship diagram between event output event intervals of a dynamic wake-up neuron circuit in the application and input signal change rates;

[0053] Figure 10 It is a structural schematic diagram of a strength detection neuron circuit in the application;

[0054] Figure 11 It is a circuit output characteristic diagram of a strength detection neuron circuit in the application under a sinusoidal voltage input signal;

[0055] Figure 12 It is a relationship diagram between event output frequencies of a strength detection neuron circuit in the application and input signal strengths;

[0056] Figure 13 It is a structural schematic diagram of an event-driven neural network system in the application;

[0057] Figure 14 This is a schematic diagram of the DVS pixel circuit in this invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0059] This embodiment relates to a circuit for dynamically waking up neurons, the structure of which is as follows: Figure 3 As shown, the device includes an adjustable resistor R1, a capacitor C1, a bidirectional threshold converter TS, and a capacitor C2. The adjustable resistor R1 and capacitor C1 are connected in series, and the bidirectional threshold converter TS and capacitor C2 are connected in parallel. The series connection of the adjustable resistor R1 and capacitor C1 is connected in series with the parallel connection of the bidirectional threshold converter TS and capacitor C2. One end of the bidirectional threshold converter TS is connected to capacitor C1, and the other end is grounded. One end of the adjustable resistor R1 is connected to the excitation signal. The connection point between the bidirectional threshold converter TS and capacitor C1 is the voltage output terminal, and the grounded end of the bidirectional threshold converter TS is the current output terminal, i.e., the event signal output terminal.

[0060] In this embodiment, the adjustable resistor R1 is an adjustable non-volatile resistive switching device, whose basic structure includes an upper electrode, an intermediate layer, and a lower electrode. The resistance value of this device is adjustable, affecting the charging and discharging speed of capacitor C1 when the input signal changes, and thus affecting the output frequency of the event. The larger the resistance value, the higher the frequency of the output event. When the input signal is a fixed value, the circuit has no output event signal.

[0061] The input signal of the dynamically activated neuron circuit is a voltage signal, and the output signal is an oscillating frequency signal. The oscillation frequency of the output signal is related to the rate of change of the input signal and the resistance R1. The event output signal of the TS device is a current signal. The frequency of the event output signal is related to the rate of change of the input signal and the resistance R1. The TS device is a novel bidirectional threshold switching device, and its basic structure includes an upper electrode, an intermediate layer, and a lower electrode. The value of capacitor C1 is greater than the value of capacitor C2.

[0062] When the input terminal has an input signal, the circuit detects the change of the signal through the charging and discharging of C1 and C2. When the voltage between the two ends of C2 exceeds the transition voltage of the TS device, the TS device changes to a low resistance state, and C2 discharges through the TS device. When the voltage between the two ends of C2 decreases to the holding voltage of the TS device, the TS device changes to a high resistance state, and a first event signal is generated. At the same time, the TS device outputs a voltage oscillation signal at the connection end of C1. When the input signal continuously changes, the circuit generates continuous output event pulse signals. Since the output event signal is a current signal, the polarity of the signal is related to the change direction of the input signal. When the input signal increases, the current event signal is positive; when the input signal decreases, the current event signal is negative.

[0063] The bidirectional threshold transition device TS in the embodiment is a vertical bidirectional threshold transition device or a planar bidirectional threshold transition device, the structures of the two devices are shown in Figure 4 and Figure 5 .

[0064] The obtaining method of the vertical bidirectional threshold transition device is as follows:

[0065] Step 1: Form a SiO2 layer on a silicon wafer, the thickness of SiO2 is 100 nm to 300 nm, and the thickness of the oxide layer can be reduced or increased according to actual process conditions.

[0066] Step 2: Deposit a lower electrode BE on the silicon oxide, the thickness of the lower electrode is 10 nm to 100 nm. The lower electrode material can be TiN, Poly-Si, Pd, Pt, W or Au and the like inert conductive material.

[0067] Step 3: Deposit a SiO2 isolation layer on the lower electrode, the thickness of SiO2 is 100 nm to 300 nm, and the thickness of the oxide layer can be reduced or increased according to actual process conditions.

[0068] Step 4: Etch a hole in the deposited film by using an etching process. The etching depth directly reaches the bottommost SiO2, and the bottommost SiO2 is etched to a certain depth to ensure that the cross section of the lower electrode is completely exposed. However, the bottommost SiO2 should have a sufficient thickness after etching.

[0069] Step 5: Deposit a functional layer FL on the etched hole, the thickness of the functional layer is 5 nm to 50 nm. The functional layer 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 the above materials but is not limited to the above materials, and materials with volatile threshold transition characteristics can be applied.

[0070] Step 6: Deposit a first intermediate electrode ME1 on the functional layer, the thickness of the intermediate electrode is 10nm-100nm. The material of the intermediate electrode can be TiN, Poly-Si, Pd, Pt, W, Cu, Ag or Au and other conductive materials.

[0071] Step 7: Deposit a capacitive dielectric film on the first intermediate electrode ME1, the capacitance of the deposited capacitive film is 1fF-100pF. The capacitance of the capacitive film is greater than the parasitic capacitance of the FL functional layer.

[0072] Step 8: Deposit a second intermediate electrode BE2 on the capacitive film, the electrode material is not limited.

[0073] Step 9: Deposit a resistive film on the second intermediate electrode ME2, the resistance of the deposited resistive film is 1Ω-1MΩ.

[0074] Step 10: Deposit an upper electrode TE on the resistive film, the material of the upper electrode is not limited.

[0075] The obtaining method of the planar bidirectional threshold transition device is:

[0076] Step 1: Form a SiO2 layer on a silicon wafer by oxidation, the thickness of the SiO2 is 100nm-300nm, and the thickness of the oxide layer can also be reduced or increased according to the actual process conditions.

[0077] Step 2: Deposit a lower electrode BE on the silicon oxide, the thickness of the lower electrode is 10nm-100nm. The material of the lower electrode can be TiN, Poly-Si, Pd, Pt, W or Au and other inert conductive materials.

[0078] Step 3: Deposit a functional layer FL on the lower electrode, the thickness of the functional layer is 5nm-50nm. The material of the functional layer 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 the above materials but is not limited to the above materials, and materials with volatile threshold transition characteristics can be applied.

[0079] Step 4: Deposit a first intermediate electrode ME1 on the functional layer, the thickness of the intermediate electrode is 10nm-100nm. The material of the intermediate electrode can be TiN, Poly-Si, Pd, Pt, W, Cu, Ag or Au and other conductive materials.

[0080] Step 5: Deposit a capacitive dielectric film on the first intermediate electrode ME1, the capacitance of the deposited capacitive film is 1fF-100pF. The capacitance of the capacitive film is greater than the parasitic capacitance of the FL functional layer.

[0081] Step 6: deposit a second intermediate electrode ME2 on the capacitor film, the electrode material is not limited.

[0082] Step 7: deposit a resistance film on the second intermediate electrode ME2, the resistance value of the resistance film is 1Ω-1MΩ.

[0083] Step 8: deposit a top electrode TE on the resistance film, the top electrode material is not limited.

[0084] The bidirectional threshold transition device TS should have an I-V characteristic as shown in Figure 6 When the voltage applied on ME1 exceeds a certain voltage value (VTH+ or VTH-), the TS device is converted from a high resistance state to a low resistance state. During the voltage back sweep, when the voltage is less than a certain voltage value (VHOLD+ or VHOLD-), the TS device is restored from a low resistance state to a high resistance state. It shows a bidirectional threshold transition characteristic.

[0085] The output characteristic of the dynamic wake-up neuron circuit under the dynamic input signal is as shown in Figure 7 The output characteristic of the dynamic wake-up neuron circuit under the sinusoidal voltage input signal is as shown in Figure 8 It can be seen that the dynamic wake-up neuron circuit in this embodiment can effectively output the event signal. The relationship between the event output interval of the dynamic wake-up neuron circuit and the input signal change rate is as shown in Figure 9 The greater the input signal change rate, the shorter the time interval between the output events. It is illustrated that the faster the input signal changes, the greater the output event frequency.

[0086] In order to meet the needs of intensity detection, this embodiment proposes an intensity detection circuit as shown in Figure 10 The intensity detection circuit includes a resistance and a TS device. The resistance R2 is connected in series with the parallel structure of the bidirectional threshold transition device TS and the capacitor C3. The capacitor C3 can be a parasitic capacitor of the TS device or an external capacitor. One end of the TS device is grounded, and the other end is connected to one end of the resistance R2. The other end of the resistance R2 serves as an input end for receiving an excitation signal. The connection point of the TS device connected to the resistance R2 serves as a voltage output end, which outputs a shock frequency signal related to the intensity. The end of the TS device grounded serves as a current signal detection port, which outputs an event frequency signal.

[0087] The circuit output characteristic of the intensity detection neuron circuit under the sinusoidal voltage input signal is as shown in Figure 11 It can be seen that the above intensity detection neuron circuit can effectively output the intensity signal. The relationship between the event output frequency of the intensity detection neuron circuit and the input signal intensity is as shown in Figure 12As shown, the greater the input signal strength, the higher the output event frequency. Furthermore, the relationship exhibits a quasi-linearity, indicating that the output frequency accurately reflects the strength of the input signal.

[0088] This embodiment also relates to an event-driven neural network system, the structure of which is as follows: Figure 13 As shown, it includes a DVS pixel array, an address event expression module, and a neuromorphic core connected in sequence. The DVS pixel array is an array composed of several DVS pixel point circuits, used to detect moving objects. The neuromorphic core is used to identify moving objects. The DVS pixel point circuits include the aforementioned dynamic wake-up neuron circuits.

[0089] DVS pixel circuitry Figure 14 As shown, a signal amplification module and a photodetector are also provided. One end of the signal amplification module is connected to an adjustable resistor R1, and the other end is connected to the photodetector. The other end of the photodetector is grounded. The excitation signal is input to the dynamic wake-up neuron circuit through the signal amplification module. The photodetector is used to turn on the dynamic wake-up neuron circuit when the brightness of the pixel changes.

[0090] The DVS pixel circuit includes an intensity detection circuit, such as Figure 14 As shown, it includes a trigger switch, a bidirectional threshold switching device and a capacitor. The bidirectional threshold switching device and the capacitor are connected in parallel. The trigger switch is connected in series with the parallel structure of the bidirectional threshold switching device and the capacitor. The excitation signal is input to the dynamic wake-up neuron circuit and the intensity detection circuit through a mirror current source.

[0091] The neuromorphic core in this embodiment includes:

[0092] Synaptic weight array, which stores a weight matrix composed of neural network weights, is used to perform multiplication and addition calculations between the weight matrix and the input vector;

[0093] A neuron array consists of several neuron circuits, each including a capacitor and a bidirectional threshold switching device. The capacitor is used to perform integration operations, and the bidirectional threshold switching device is used as a threshold switch to determine the release of the action potential.

[0094] During operation, the DVS pixel array detects moving objects and outputs parallel event signals. The AER module receives these event signals, processes them by converting parallel signals to serial signals and serial signals to parallel signals, and then transmits them to the subsequent neuromorphic core module. The output of the AER module serves as the input to the neuromorphic core module.

[0095] The neuromorphic core is composed of a synaptic weight array and a neuron array, and the synaptic weight array stores the weights of the neural network. When the signal from the AER is added to the input end of the core, the weight array performs the multiplication and addition calculation of the weight matrix and the input vector. The weight array has the characteristics of storage and calculation, and the synaptic weight unit at the cross point can be a memory device such as a memristor, SRAM, DRAM, flash, etc.

[0096] The output of the synaptic weight array is the input of the neuron array. The neuron circuit in the neuron array is composed of a capacitor and a threshold conversion device. The capacitor performs the integration function, and the threshold conversion device acts as a threshold switch to determine the firing of the action potential. The neuron circuit integrates the weighted current or voltage signal from the synapse, and when the integrated potential reaches the threshold of the neuron, the action potential signal is fired. The firing frequency of the neuron circuit action potential is positively correlated with the strength of the input, and the pattern of the input signal can be judged according to the firing frequency of the neuron action potential in the neuron array, so that the object detected by the DVS pixel array can be identified.

[0097] The neuromorphic core has the function of online learning, and its weights can be dynamically adjusted according to the stimulation of the external environment to realize the recognition of different input images and to cope with the changing external environment.

[0098] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An event-driven neural network system, characterized in that, The event-driven neural network system comprises a DVS pixel array, an address event expression module, and a neuromorphic core connected in sequence. The DVS pixel array is an array of several DVS pixel point circuits used for detecting moving objects. The neuromorphic core is used for recognizing moving objects. Each DVS pixel point circuit includes a dynamically activated neuron circuit, which includes adjustable resistors. ,capacitance Bidirectional threshold switching device and capacitor The adjustable resistor mentioned above With capacitor Series connection; the aforementioned bidirectional threshold switching device With capacitor Parallel connection; the aforementioned adjustable resistor With capacitor Series structure and bidirectional threshold switching device With capacitor The parallel structure is connected in series; the bidirectional threshold switching device One end is connected to the capacitor Connected at one end, with the other end grounded; the adjustable resistor mentioned One end of the device is connected to an excitation signal; the bidirectional threshold switching device With capacitor The connection point is the voltage output terminal, a bidirectional threshold converter. The grounded end is the current output end, which is also the event signal output end; The dynamic wake-up neuron circuit includes a signal amplification module and a photodetector; one end of the signal amplification module is connected to an adjustable resistor. One end is connected to the other end, and the other end is connected to the photodetector; the other end of the photodetector is grounded; the excitation signal is input to the dynamic wake-up neuron circuit through the signal amplification module; the photodetector is used to cause the dynamic wake-up neuron circuit to output an event signal when the brightness of the pixel changes. The DVS pixel circuit includes an intensity detection circuit, comprising a trigger switch, a bidirectional threshold switching device, and a capacitor; the bidirectional threshold switching device is connected in parallel with the capacitor; the trigger switch is connected in series with the parallel structure of the bidirectional threshold switching device and the capacitor; the excitation signal is input to the dynamic wake-up neuron circuit and the intensity detection circuit respectively through a mirror current source; the trigger switch is triggered by the output of the dynamic wake-up neuron circuit.

2. The event-driven neural network system according to claim 1, characterized in that, The adjustable resistor It is an adjustable non-volatile resistive switching device.

3. The event-driven neural network system according to claim 1, characterized in that, The bidirectional threshold switching device It can be a vertical bidirectional threshold switching device or a planar bidirectional threshold switching device.

4. The event-driven neural network system according to claim 3, characterized in that, The method for obtaining the vertical bidirectional threshold switching device is as follows: Step 1: Oxidize the silicon wafer to form a SiO2 layer; Step 2: Deposit the lower electrode BE on the SiO2 layer; Step 3: Deposit a SiO2 isolation layer on the lower electrode BE; Step 4: Using an etching process, holes are etched into the deposited thin film. The holes penetrate the SiO2 isolation layer and the lower electrode BE. Finally, a certain thickness is etched into the SiO2 layer to ensure that the cross-section of the lower electrode BE is fully exposed. Step 5: Deposit the functional layer FL on the etched holes; Step 6: Deposit the first intermediate electrode ME1 on the functional layer FL; Step 7: Deposit a capacitor dielectric film on the first intermediate electrode ME1; Step 8: Deposit the second intermediate electrode ME2 on the capacitor film; Step 9: Deposit a resistive thin film on the second intermediate electrode ME2; Step 10: Deposit electrode TE on the resistive thin film.

5. The event-driven neural network system according to claim 3, characterized in that, The method for obtaining the planar bidirectional threshold switching device is as follows: Step 1: Oxidize the silicon wafer to form a SiO2 layer; Step 2: Deposit the lower electrode BE on the SiO2 layer; Step 3: Deposit the functional layer FL on the lower electrode BE; Step 4: Deposit the first intermediate electrode ME1 on the functional layer FL; Step 5: Deposit a capacitor dielectric film on the first intermediate electrode ME1; Step 6: Deposit the second intermediate electrode ME2 on the capacitor dielectric film; Step 7: Deposit a resistive thin film on the second intermediate electrode ME2; Step 8: Deposit electrode TE on the resistive thin film.

6. The event-driven neural network system according to claim 1, characterized in that, The capacitor These are parasitic capacitances or external capacitances.

7. The event-driven neural network system according to claim 1, characterized in that, The neuromorphic core includes: Synaptic weight array, which stores a weight matrix composed of neural network weights, is used to perform multiplication and addition calculations between the weight matrix and the input vector; A neuron array consists of several neuron circuits, each including a capacitor and a bidirectional threshold switching device. The capacitor is used to perform integration operations, and the bidirectional threshold switching device is used as a threshold switch to determine the release of action potentials. The neuron circuit integrates the weighted current or voltage signal from the synapse. When the integrated potential reaches the threshold of the neuron, it emits an action potential signal. The pattern of the input signal can be determined based on the firing frequency of the action potentials of the neurons in the neuron array, thereby enabling the identification of objects detected by the DVS pixel array.

Images