Frequency adaptive lif neuron and control method, method for simulating biological adaptation
By introducing vacancy modulation of the resistance state change rate of the memristor in LIF neurons, a frequency-adaptive LIF neuron was designed, which solved the problem of difficulty in simulating biological adaptation in existing technologies, achieved stable adaptation of neuron output frequency, and improved the response accuracy of neural networks.
Patent Information
- Application Number
- CN202210828987.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-14
- Publication Date
- 2025-11-25
- Estimated Expiration
- 2042-07-14
AI Technical Summary
Existing hardware neural networks struggle to accurately simulate the adaptive process of organisms to environmental stimuli, especially in special scenarios such as strong light stimulation, where existing neuron simulation methods cannot effectively simulate the gradual adaptation process of biological responses.
A frequency-adaptive LIF neuron is designed. By introducing vacancies into a threshold-switching memristor and adjusting the vacancy concentration, the rate of resistance change of the memristor slows down as the state switching cycle increases. Combined with the change in the charging speed of the charging capacitor, the output firing pulse frequency gradually decreases until it stabilizes, simulating a biological adaptive process.
It achieves a gradual decrease in the output frequency of neurons until it stabilizes, effectively simulating the adaptive process of organisms under environmental stimuli and improving the accuracy of neural network response to environmental stimuli.
Smart Images
Figure CN115329945B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of neural networks, and more particularly relates to a frequency adaptive LIF neuron and a control method and a method for simulating biological adaptation. BACKGROUND
[0002] Neuromorphic computing based on spiking neural networks aims to simulate the cognitive behavior of the brain and ultimately achieve human-level machine intelligence through the implementation of hardware neural networks. It has attracted widespread attention due to its low energy consumption and high similarity to the biological nervous system. Memristors have great potential in logic operations and brain-like computing due to their high speed, low power consumption, simple structure, easy integration, and compatibility with CMOS technology. The neuron based on threshold transition memristors only needs to design a simple auxiliary circuit, and the firing behavior and reset process of the threshold-driven neuron can be easily simulated by the volatile switching characteristics of the threshold transition memristor.
[0003] Currently, when building a hardware neural network, people often pay more attention to how to reduce the delay caused by the device to improve the calculation speed and reduce the power consumption. However, in some special scenarios, such as simulating the process of accepting environmental stimuli (such as strong light stimulation) in life, it is difficult to obtain accurate results using the current neural network, therefore, how to use neurons to simulate the adaptive process of biological adaptation to environmental stimuli is also one of the current research focuses. SUMMARY
[0004] In view of the above defects or improvement needs of the prior art, the present application provides a frequency adaptive LIF neuron and a control method and a method for simulating biological adaptation, which aims to simulate the adaptive process of biological adaptation using neurons.
[0005] To achieve the above-mentioned purpose, according to one aspect of the present application, a frequency adaptive LIF neuron is provided, which comprises a resistor R S , a resistor R o , a threshold transition memristor and a charging capacitor C M , the first end of the resistor R S is used to receive a control pulse, the second end is grounded through the charging capacitor C M , the first end of the threshold transition memristor is connected with the second end of the resistor R S , the second end of the threshold transition memristor is grounded through the resistor R o and the second end outputs a firing pulse.
[0006] The threshold transition memristor includes a resistive switching layer, a bottom electrode and a top electrode respectively located on both sides of the resistive switching layer, the top electrode has an active electrode material, the resistive switching layer has vacancies, the off-state resistance change speed of the threshold transition memristor from a low resistance state to a high resistance state slows down with the increase of the state switching period, and the off-state resistance change speed of the threshold transition memristor changes with the adjustment of the vacancy concentration, the charging speed of the charging capacitor C M The charging speed of the charging capacitor C M becomes slower with the slowing down of the off-state resistance change speed of the threshold transition memristor, so that the frequency of the output firing pulse gradually decreases until it tends to be stable.
[0007] In one embodiment, the top electrode is a stack composed of an active material layer and a titanium layer, the active electrode material includes silver or copper, the resistive switching layer includes gallium selenide or perfluorosulfonic acid resin, and the bottom electrode is a single-layer or multi-layer electrode structure formed by one or more of platinum, chromium, gold, and titanium.
[0008] In one embodiment, the thickness of the resistive switching layer is 5-10 nm.
[0009] In one embodiment, the vacancies are vacancies of the resistive switching layer material itself or vacancies caused by plasma treatment.
[0010] According to another aspect of the present application, a frequency adaptive control method of LIF neurons is provided, comprising:
[0011] The frequency adaptive LIF neuron is built, and the vacancy concentration of the threshold transition memristor is adjusted according to the change requirement of the off-state resistance change speed;
[0012] An input control pulse is input to charge the charging capacitor C M , and the state switching of the threshold transition memristor is controlled, the off-state resistance change speed of the threshold transition memristor from a low resistance state to a high resistance state slows down with the increase of the state switching period, and the charging speed of the charging capacitor C M becomes slower with the slowing down of the off-state resistance change speed of the threshold transition memristor, so that the frequency of the output firing pulse gradually decreases until it tends to be stable, realizing the frequency adaptation of LIF neurons.
[0013] In one embodiment, the resistive switching layer is subjected to plasma treatment to obtain vacancies, and the power range of the plasma treatment is (20 W, 80 W) and the time range is (100 s, 180 s).
[0014] In one embodiment, the plasma is an inert gas plasma.
[0015] According to another aspect of the present application, there is provided a method for simulating biological adaptation, which uses the frequency adaptive LIF neuron to build a neural network, uses environmental stimulation as input of the neural network to control state switching of the LIF neuron, and slows down output frequency of the neural network over time until it tends to be stable, so as to simulate the process that the biological reaction slows down and finally adapts to the environment when the biological is stimulated by the environment.
[0016] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:
[0017] The present application analyzes biological reaction characteristics of gradual adaptation of the biological under stimulation based on careful research on the adaptation process of the biological under environmental stimulation. Based on this, a frequency adaptive LIF neuron simulation circuit based on a memristor is designed, wherein the charging capacitor C M and the threshold transition memristor interact to finally output a firing pulse. In the designed LIF neuron simulation circuit, the threshold transition memristor is very critical. According to the conventional thinking of those skilled in the art, in order to reduce device power consumption and maintain stability of device operation, it is required that the resistance state transition speed of the memristor is faster and more balanced. Therefore, the resistance state transition speed of the conventional threshold transition memristor is relatively fast and the resistance state change speed is relatively balanced. However, using the conventional threshold transition memristor to build the above LIF neuron simulation circuit cannot realize the biological adaptation process.
[0018] In the present application, the research and development team found that when there is an active electrode material in the top electrode of the memristor, that is, the resistance state change of the memristor is realized by controlling the state of the conductive wire formed by the active electrode material in the functional layer, vacancies can be formed in the functional layer of the memristor and the concentration of the vacancies can be adjusted to make the resistance state change speed of the memristor slow down over time and eventually tend to be stable. Based on the above analysis, the concentration of the vacancies can be controlled as required to prepare a threshold transition memristor with a resistance state change speed that meets the requirements, and then the above LIF neuron simulation circuit is built using the memristor. When the off resistance state change speed of the threshold transition memristor from a low resistance state to a high resistance state slows down over time, the charging speed of the charging capacitor C M also slows down over time as the off resistance state change speed of the threshold transition memristor slows down, and finally the frequency of the output firing pulse gradually decreases until it tends to be stable. The change rule of the firing pulse is exactly adapted to the process that the biological reaction gradually slows down and finally adapts to the environment under stimulation, and thus the adaptation process of the biological under environmental stimulation can be simulated. BRIEF DESCRIPTION OF DRAWINGS
[0019] Figure 1This is a schematic diagram of a LIF neuron simulation circuit based on a threshold switching memristor, according to one embodiment.
[0020] Figure 2 This is a schematic diagram of the structure of a threshold switching memristor according to one embodiment;
[0021] Figure 3(a) is an IV curve of a gallium selenide threshold switching memristor after argon plasma treatment according to an embodiment.
[0022] Figure 3(b) is an IV curve of an untreated gallium selenide threshold-switching memristor according to an embodiment;
[0023] Figure 4(a) shows the current response curve captured by the threshold switching memristor formed in Example 1 under a triangular wave pulse signal with a pulse amplitude of 0.7V;
[0024] Figure 4(b) shows the performance of the threshold switching memristor formed in Example 1 over 10 cycles at four different pulse amplitudes;
[0025] Figure 5(a) is a graph showing the adaptive change of firing pulses with membrane potential in a LIF neuron model using a gallium selenide threshold switching memristor according to an embodiment.
[0026] Figure 5(b) shows the firing pulse frequency versus time curve of a LIF neuron model using a gallium selenide threshold switching memristor in one embodiment;
[0027] Figure 6(a) is a graph showing the adaptive change of firing pulses with membrane potential in a LIF neuron model using a perfluorosulfonic acid resin threshold switching memristor according to an embodiment.
[0028] Figure 6(b) shows the firing pulse frequency versus time curve of a LIF neuron model using a perfluorosulfonic acid resin threshold switching memristor in one embodiment. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0030] like Figure 1 The diagram shown is a schematic of an analog circuit for a frequency-adaptive LIF neuron in one embodiment, which includes a resistor R. S Resistance R o Threshold switching memristor (TSM) and charging capacitor C M Among them, the resistance R SThe first end of the threshold transition memristor TSM is used to receive a control pulse Input, and the second end of the threshold transition memristor TSM is connected to the ground through a charging capacitor C M The first end of the threshold transition memristor TSM is connected to the resistance R S The second end of the threshold transition memristor TSM is connected to the ground through a resistance R o The second end of the threshold transition memristor TSM outputs a firing pulse Output.
[0031] The threshold transition memristor TSM is not a conventional memristor, but a memristor with vacancies in the functional layer, and the vacancy concentration of the memristor is different, and the off-state transition speed of the memristor is slow with the increase of the state switching period. The top electrode of the memristor needs to have an active electrode material, that is, the memristor realizes the resistance state transition by controlling the state of the active electrode material to form a conductive wire in the functional layer. When there is no vacancy in the functional layer, the memristor also has the characteristic of resistance state transition, and the vacancy modulation is added to make the memristor eventually have the characteristic that the off-state transition speed is slow with the increase of the state switching period.
[0032] The above-mentioned threshold transition memristor TSM is selected to build the above-mentioned LIF neuron simulation circuit, and the interaction relationship between the threshold transition memristor TSM and the charging capacitor C M is combined, and the charging capacitor C M is charged by the input control pulse Input. When the voltage of the charging capacitor C M rises to the opening threshold voltage of the threshold transition memristor TSM, the threshold transition memristor TSM changes from high resistance to low resistance and turns on, and the charging capacitor C M is discharged through the opened threshold transition memristor TSM and outputs a firing pulse. When the voltage of the charging capacitor C M falls to the off threshold voltage of the threshold transition memristor TSM, the threshold transition memristor TSM changes from low resistance to high resistance and turns off, and the charging capacitor C M is charged again by the control pulse Input, and the cycle continues. Among them, with the increase of the state switching period of the threshold transition memristor TSM, the off speed gradually slows down, so that the charging speed of the charging capacitor C M gradually slows down, the charging time to the opening threshold voltage of the threshold transition memristor TSM becomes longer and longer, and finally the frequency of the output firing pulse gradually becomes smaller and tends to be stable, thereby simulating the adaptive process of the biological environment stimulation.
[0033] In an embodiment, as Figure 2As shown, the top electrode of the threshold switching memristor TSM is a stack of an active material layer and a titanium layer, and the active electrode material includes Ag or Cu. The titanium layer can prevent excessive injection of silver ions and copper ions, while adhering the resistive switching layer and the active material layer. In the initial study, it was found that the silver conductive filaments formed would overgrow after multiple applied biases, and the generated silver conductive filaments could not spontaneously melt after the bias was removed, which affected the subsequent device resistance state switching process. Based on this finding, the research and development team improved the device performance, using a resistive switching layer with vacancies, which can limit the migration of silver ions and prevent the excessive injection of silver ions. During this period, it was found that when using a resistive switching layer with vacancies, the resistance state change of the memristor showed a special change rule, that is, the turn-off resistance state change speed of the memristor from a low resistance state to a high resistance state became slower with the increase of the state switching period, and it was found that the adjusted vacancy concentration could adjust the turn-off resistance state change speed of the memristor. Based on this finding, the above neuron simulation circuit was built to simulate the adaptive process of living beings.
[0034] The material of the resistance change layer is not limited, and the means for introducing vacancies into the resistance change layer is not limited, as long as the memristor after introducing the vacancies presents the special change rule described above. Specifically, the resistance change layer can be gallium selenide or perfluorosulfonic acid resin. Specifically, the way of introducing vacancies can be plasma treatment or the material itself has vacancies. Taking the treatment of gallium selenide by argon plasma to form a resistance change layer with vacancies as an example. Argon plasma treatment will introduce gallium vacancies on the surface of the gallium selenide sheet, which can limit the migration of silver ions and prevent the excessive injection of silver ions, thereby affecting the growth and breaking process of the silver conductive filament. However, a suitable vacancy concentration has a great influence on the performance of the device, so by using a suitable argon plasma power and time, a suitable gallium vacancy concentration can be introduced on the surface of the gallium selenide. By guiding the migration of silver ions and adjusting the growth of silver conductive filaments through the introduced gallium vacancies, the excessive growth of silver conductive filaments can be prevented, thereby improving the performance of the device and further effectively regulating the time delay of the device in each resistance state switching process. In an embodiment, the power range of the plasma treatment is (20 W, 80 W), and the time range is (100 s, 180 s). If the power is too high, the structure of the resistance change layer will be damaged, and if the power is too low, the vacancies cannot be introduced or the vacancy concentration is too low to achieve the regulation effect. In an embodiment, the thickness of the resistance change layer is 5 nm to 10 nm. A too thick resistance change layer is not conducive to the migration of silver ions, which affects the growth of silver conductive filaments, thereby greatly increasing the opening time of the device. A too thin resistance change layer is easy to cause excessive growth of silver conductive filaments, which may even cause the device to be directly conductive after applying a few biases, thereby reducing the cycle stability of the device. In addition, the oxidation process of a too thin resistance change layer is more likely to occur, thereby affecting the performance of the device. Therefore, a reasonable thickness of the resistance change layer can prevent excessive oxidation on the one hand, and can more effectively regulate the growth of silver conductive filaments on the other hand. In an embodiment, the bottom electrode is a multi-layer electrode formed by one or two metals such as platinum, chromium, gold, and titanium.
[0035] The application also relates to a frequency adaptive control method of a LIF neuron.
[0036] First, the vacancy concentration of the threshold transition memristor is adjusted according to the change requirement of the off resistance state change speed, a suitable threshold transition memristor is prepared, and the threshold transition memristor is used to build the frequency adaptive LIF neuron in the above.
[0037] An input control pulse Input is input to charge the capacitor C M , and the state switching of the threshold transition memristor is controlled. The off resistance state change speed of the threshold transition memristor from a low resistance state to a high resistance state slows down with the increase of the state switching period. The capacitor C MThe charging speed of the threshold switching memristor is slowed down with the change of the off-state resistance of the threshold switching memristor, so that the frequency of the output firing pulse gradually decreases until it tends to be stable, realizing the frequency adaptation of the LIF neuron.
[0038] The frequency adaptation process of the LIF neuron has been introduced above, and will not be repeated here.
[0039] In the above method, the preparation of the threshold switching memristor is particularly crucial. The following is a specific example to illustrate the preparation process and its performance.
[0040] In this embodiment, the substrate is a silicon substrate; the bottom electrode adopts a double-layer electrode structure composed of platinum and titanium, the thickness of platinum is 20 nm, and the thickness of titanium is 10 nm, wherein the titanium contacts the silicon substrate as an adhesion layer, and the platinum is grown on the titanium; the top electrode adopts silver and titanium, the thickness of silver is 50 nm, and the thickness of titanium is 3 nm, the titanium contacts the resistance switching layer as a barrier layer to prevent excessive silver ion injection, and also plays a role in adhering the silver electrode and the resistance switching layer, and the silver is on the upper layer of the titanium; the resistance switching layer is a two-dimensional gallium selenide flake treated by argon plasma, and the thickness is 5-10 nm.
[0041] In this embodiment, a threshold switching memristor based on gallium selenide with adjustable resistance state switching delay is prepared by photolithography, magnetron sputtering, two-dimensional material directional transfer and argon plasma treatment, which specifically includes the following steps:
[0042] (1) The silicon substrate is sequentially cleaned with acetone, ethanol and deionized water for 15 minutes each time, and the cleaning is repeated 3 times. After taking out, the cleaned silicon substrate is blown dry with nitrogen.
[0043] (2) The photoresist is spin-coated on the silicon substrate, and then the photoresist is uniformly coated using a spin coater. After that, the bottom electrode pattern is obtained by using ultraviolet lithography technology and developing. The line width of the bottom electrode is 5 um. Then, the Pt / Ti electrode with a thickness of 20 / 10 nm is deposited by using a magnetron sputtering system. After deposition, the excess photoresist is removed by lift-off treatment using acetone to obtain a patterned platinum / titanium bottom electrode.
[0044] (3) A two-dimensional directional transfer method is used to dry transfer the mechanically exfoliated gallium selenide flake to the bottom electrode on a two-dimensional material transfer platform.
[0045] (4) After the transfer is completed, the device is placed in the vacuum chamber of the reactive ion etching machine for argon plasma treatment. The argon flow is set to 50 sccm, the power is 20 W, and the treatment time is 2 min.
[0046] (5) Again spin-coat uniform glue and use ultraviolet lithography to carry out top electrode over-etching treatment, so that the top electrode pattern and the bottom electrode with the resistance change layer are in cross structure, and the top electrode pattern is obtained after development, and the top electrode line width is 5 um; then, a 50 / 3 nm Ag / Ti metal stack structure is deposited as a top electrode by using a magnetron sputtering system, titanium is deposited under the silver lower layer, and after soaking in acetone, cleaning and drying, a patterned stack top electrode is obtained.
[0047] (6) In the above preparation steps, the step (4), i.e. argon plasma treatment, is crucial for regulating the switching time delay of the device, and by using a suitable argon plasma treatment power and time, a suitable gallium vacancy concentration can be introduced on the surface of the two-dimensional gallium selenide sheet. In order to further verify the reliability of the method provided in the present application, the threshold transition memristor prepared in Example 1 and Comparative Example 1 without argon plasma treatment are compared.
[0048] Comparative Example 1:
[0049] In the present comparative example, except that the resistance change layer material is not subjected to argon plasma treatment, the remaining steps are consistent with Example 1.
[0050] The Keithley 4200SCS semiconductor analysis tester is used to test the electrical performance of the threshold transition memristor corresponding to Example 1 and Comparative Example 1, respectively, and the specific steps are as follows:
[0051] A bias voltage is applied to the top electrode of Example 1 and Comparative Example 1, and the bottom electrode is grounded, and direct current I-V characteristic scanning is performed, respectively.
[0052] In Example 1, the voltage scanning range is set to 0V-1.4V, the limiting current is 1uA, and 100 times of continuous cycle direct current I-V characteristic scanning is performed, and the I-V characteristic curve is as shown in Fig. 3(a).
[0053] In Comparative Example 1, the voltage scanning range is set to 0V-1V, the limiting current is 1uA, and the device can only perform 10 times of cycle direct current I-V characteristic scanning, and the I-V characteristic curve is as shown in Fig. 3(b). In the subsequent scanning cycle, the device is completely turned on, which indicates that the conductive wire cannot spontaneously melt at this time, and the device cannot be turned off.
[0054] According to the comparison of Fig. 3(a) and Fig. 3(b), it can be seen that compared with Comparative Example 1, the 100 times of cycle direct current I-V curve results of Example 1 show that:
[0055] Example 1 has a larger on / off ratio (ratio of on-state current to off-state current), and the state switching is more obvious;
[0056] The high resistance state current of embodiment 1 is more stable and does not gradually increase with the increase of voltage;
[0057] The device cycle stability of embodiment 1 (>100 times) is greatly improved compared with comparative example 1 (<10 times);
[0058] Meanwhile, the voltage distribution of embodiment 1 is more random, and can more conform to the random firing process in the LIF neuron simulation.
[0059] That is, the key performance parameters of the selector in embodiment 1 are optimized, and are more suitable for the simulation and implementation of the frequency adaptive LIF neuron model.
[0060] Meanwhile, the application provides simulation and emulation of the frequency adaptive function of the LIF neuron based on the threshold transition memristor. Due to the space regulation, the resistance state change speed changes correspondingly in the resistance state switching process of the device each time. In the LIF neuron, the speed of the resistance state change of the threshold transition device changes the charging speed of the membrane potential of the neuron, thereby changing the frequency change of the neuron each time.
[0061] Further, the embodiment 1 is applied with different pulse amplitude triangular wave pulse signals of 0.7V, 0.9V, 1.0V and 1.2V for capturing the current response signals. In FIG. 4(a), it is observed that there is an obvious time delay in the device current on-response under the 0.7V triangular wave pulse signal, i.e. the time delay of the device resistance state switching. The time delay distribution is counted by applying the embodiment 1 with 10 pulse signal cycles respectively under the triangular wave pulse signals of 0.7V, 0.9V, 1.0V and 1.2V, and FIG. 4(b) is obtained by fitting. It is observed that the resistance state switching delay of the device will adaptively change with the number of applied cycles. In FIG. 4(b), the frequency of the triangular wave pulse signal is fixed, each cycle includes a time period t1 and a time period t2, the conductive wire is formed in the resistance variable layer and the memristor is turned on to a low resistance state in t1, and the conductive wire is melted and the memristor is in a high resistance state in t2. For the same triangular wave of the same pulse amplitude, the on-delay becomes smaller and smaller with the increase of the switching cycles. It means that with the increase of the switching cycles, it is more and more difficult to melt the conductive wire in t2, and there are more residual conductive wires, so the next cycle is easier to open and the on-delay is smaller. Therefore, FIG. 4(b) reflects the change rule of the slower and slower off-speed from the smaller and smaller on-delay. For example, in the first pulse cycle, the on-delay of the high resistance to the low resistance in the time period t1 is 400ns, and the melting speed of the conductive wire is fast in the time period t2, so the low resistance state can be changed to a high resistance state of 800MΩ; in the second pulse cycle, the on-delay of the high resistance to the low resistance in the time period t1 is 350ns, and the melting speed of the conductive wire is slow in the time period t2, so the low resistance state can be changed to a high resistance state of 600MΩ; in the third pulse cycle, since the resistance value of the high resistance state is smaller, it is easier to change from the high resistance state to the low resistance state, and the delay is smaller; and so on. With the slower and slower off-speed, the resistance value of the high resistance state reached in the same off-time is smaller, the next cycle is easier to open, and the on-delay is smaller.
[0062] Corresponding to the LIF neuron in the present application, with the increase of the switching cycles, the off-speed of the memristor is slower and slower, and during the off period, the charging capacitor C M begins to charge, the off-speed of the memristor is slower and slower, the charging speed of the charging capacitor C M is slower and slower, and the time for charging to the on threshold voltage of the memristor is longer and longer, so that the frequency of the output firing pulse gradually becomes smaller and smaller until it tends to be stable. Moreover, the delay caused by different pulse amplitudes is not the same, so that the neuron can also simulate the biological response process under different stimulations by converting the strength of the external stimulation into the pulse amplitude of the control pulse.
[0063] The LIF neuron circuit is simulated by using the time delay of Fig. 4(b) through Verilog. In the constructed LIF neuron circuit, the parameters of Example 1 and the characteristics of the resistance state switching time delay are adjusted, and the frequency adaptive function of the LIF neuron is successfully simulated. Fig. 5(a) is the change of the firing pulse of the LIF neuron within 10 ms with the charging process. Due to the vacancy regulating the speed of the resistance state change of the device, there is a time delay in the process of the resistance state switching. With the increase of the cycle number, the resistance state switching speed of the device becomes slow, thereby causing the charging speed of the LIF neuron to become slow in the subsequent firing process, and the charging voltage V M The frequency becomes smaller, thereby causing the firing pulse V out The frequency becomes smaller and smaller. Fig. 5(b) is the change of the frequency of the firing pulse with time, which more directly shows the process of the frequency adaptive change of the firing pulse of the LIF neuron by counting the time change of the firing pulse of Fig. 5(a). The material of the resistance variable layer in Example 1 is replaced by Nafion (perfluorosulfonic acid resin), and other processing steps are the same, and good simulation results can also be obtained. As shown in Fig. 6(a), the change of the firing pulse of the LIF neuron within 10 ms with the charging process, and Fig. 6(b) is the change of the frequency of the firing pulse with time, which also shows the process of the frequency adaptive change of the firing pulse of the LIF neuron. It should be noted that the present scheme does not limit the specific material and processing technology of the memristor, as long as the memristor finally has the above-mentioned delay characteristics.
[0064] Based on the obvious performance optimization obtained in the above examples, and the implementation and simulation of the frequency adaptive LIF neuron, the feasibility of the method for regulating the resistance state switching time delay of the device by the vacancy in the present application, and then realizing the frequency adaptation in the LIF neuron is verified.
[0065] Correspondingly, the present application also relates to a method for simulating biological adaptation, which uses the above-mentioned frequency adaptive LIF neuron to build a neural network, uses the environmental stimulus as the input of the neural network to control the state switching of the LIF neuron, and the output frequency of the neural network becomes slow with the increase of time until it tends to be stable, so as to simulate the process that the biological reaction becomes slow and finally adapts to the environment when the biological is stimulated by the environment.
[0066] Those skilled in the art can easily understand that the above description is only the preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A frequency-adaptive LIF neuron circuit, characterized in that, Including resistors ,resistance Threshold switching memristors and charging capacitors ,resistance The first terminal is used to receive control pulses, and the second terminal is connected to a charging capacitor. Grounded, the first terminal of the threshold switching memristor is connected to the resistor. The second terminal of the threshold switching memristor is connected via a resistor. Grounded and the second terminal outputs a pulse; The threshold switching memristor includes a resistive switching layer and a bottom electrode and a top electrode located on both sides of the resistive switching layer. The top electrode has an active electrode material, and the resistive switching layer has vacancies. The rate of change of the turn-off resistance state of the threshold switching memristor from a low resistance state to a high resistance state slows down with the increase of the state switching cycle, and the rate of change of the turn-off resistance state of the threshold switching memristor changes with the adjustment of the vacancy concentration. The charging capacitor... The charging speed slows down as the rate of change of the threshold switching memristor's off-resistance state decreases, so that the frequency of the output firing pulse gradually decreases until it stabilizes. The threshold switching memristor uses a silicon substrate as its substrate. The bottom electrode is a double-layer electrode structure composed of platinum and titanium, with a platinum thickness of 20 nm and a titanium thickness of 10 nm. The titanium contacts the silicon substrate as an adhesion layer, and the platinum is grown on top of the titanium. The top electrode is composed of silver and titanium, with a silver thickness of 50 nm and a titanium thickness of 3 nm. The titanium is in contact with the resistive switching layer in the lower layer to act as a barrier layer to prevent excessive silver ion implantation, and at the same time serves to adhere the silver electrode and the resistive switching layer. The silver is on top of the titanium. The resistive switching layer is a two-dimensional gallium selenide thin film treated with argon plasma, with a thickness of 5 nm to 10 nm.
2. The frequency-adaptive LIF neuron circuit as described in claim 1, characterized in that, The vacancies are either vacancies inherent in the resistive switching layer material itself or vacancies resulting from plasma treatment.
3. A frequency adaptive control method for LIF neurons, characterized in that, include: Construct a frequency-adaptive LIF neuron circuit as described in any one of claims 1 to 2, and adjust the vacancy concentration of the threshold switching memristor according to the change in the rate of change of the turn-off resistance state. Input control pulses to charge the capacitor Charging is performed, controlling the state switching of the threshold switching memristor. The rate of change of the threshold switching memristor from a low resistance state to a high resistance state (off-state) slows down as the state switching cycle increases. The charging capacitor... The charging speed slows down as the rate of change of the turn-off resistance state of the threshold switching memristor slows down, so that the frequency of the output firing pulse gradually decreases until it stabilizes, thereby realizing the frequency adaptation of the LIF neuron.
4. The frequency adaptive control method as described in claim 3, characterized in that, Plasma treatment was performed on the resistive switching layer to obtain vacancies. The power range of the plasma treatment was (20 W, 80 W) and the time range was (100 s, 180 s).
5. The frequency adaptive control method as described in claim 4, characterized in that, The plasma is an inert gas plasma.
6. A method for simulating biological adaptation, characterized in that, A neural network is constructed using the frequency-adaptive LIF neuron circuit described in any one of claims 1 to 2. Environmental stimuli are used as inputs to control the state switching of LIF neurons. The output frequency of the neural network slows down over time until it stabilizes, thus simulating the process by which an organism slows down its response to environmental stimuli and eventually adapts to the environment.