Self-adaptive sensitized artificial neuron as well as preparation method and application thereof

Abstract

The invention relates to the technical field of artificial neurons, in particular to a self-adaptive sensitized artificial neuron and a preparation method and application thereof. The invention relates to a self-adaptive sensitized artificial neuron, which comprises a perovskite nickelate thin film device. The self-adaptive sensitization artificial neuron is applied as the self-adaptive sensitization artificial neuron based on the self-adaptive sensitization performance of the perovskite nickelate thin-film device, so that when the single perovskite nickelate thin-film device is used as the artificial neuron, the self-adaptive sensitization characteristic is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial neurons, and particularly to an adaptive sensitized artificial neuron, a preparation method thereof, and an application thereof. Background Art

[0002] Traditional artificial neurons rely on a complex circuit composed of multiple electronic components or artificial adjustment of external inputs to change the neuron state. The implementation method of neuron state switching is very complex, and the device itself cannot self-regulate.

[0003] In addition, traditional artificial neurons cannot directly interact with the environment for information, and can only passively receive external signals. That is, under the same external conditions, the device can only exhibit a corresponding state, lacking different processing methods for the same external signal, that is, the device cannot autonomously interact with the environment for information. After the neuron device is manufactured, it can only receive signals within a fixed and limited range and cannot exceed the limit of its own reception range. In the face of changes in signal intensity under different environments, there are often problems that many signals cannot be received, resulting in information loss. In a neural network composed of the above neurons, in the face of input changes, it does not have self-adaptability and sufficient generalization ability, or requires complex algorithms to have self-adaptability. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems in the related art to some extent.

[0005] Based on this, in a first aspect, an adaptive sensitized artificial neuron is provided, including a perovskite nickelate thin film device.

[0006] According to the adaptive sensitization performance of the perovskite nickelate thin film device of the present invention, it is used as an adaptive sensitized artificial neuron. Thus, a single perovskite nickelate thin film device undertakes the function of an artificial neuron and has an adaptive sensitization characteristic.

[0007] According to an embodiment of the present invention, the perovskite nickelate thin film device includes:

[0008] A substrate, a perovskite nickelate thin film on the surface of the substrate, a first electrode and a second electrode oppositely arranged on the perovskite nickelate thin film, and a perovskite nickelate channel is left between the first electrode and the second electrode;

[0009] The first electrode is a polar plate capable of catalyzing hydrogenation, and the second electrode is a polar plate incapable of catalyzing hydrogenation;

[0010] The perovskite nickelate thin film under the first electrode is doped with hydrogen ions. The hydrogen ions diffuse a distance L in the perovskite nickelate channel, and the concentration of hydrogen ions in the perovskite nickelate channel continuously decreases with the increase of the distance from the first electrode. Thus, a hydrogen ion concentration gradient structure is formed in the perovskite nickelate thin film device. Due to the high-sensitivity response of the hydrogen ion concentration gradient structure to electrical pulses, the distribution of hydrogen ions near the first electrode can be controlled by regulating the electrical pulses, so that the device has an adaptive sensitization characteristic to electrical pulse signals when used as an artificial neuron.

[0011] According to an embodiment of the present invention, the length of the perovskite nickelate channel between the first electrode and the second electrode is 0.1 μm to 100 μm.

[0012] According to an embodiment of the present invention, the distance L is 0.01 μm to 50 μm.

[0013] According to an embodiment of the present invention, in the perovskite nickelate thin film, the chemical formula of the perovskite nickelate is MNiO3, where M represents a metal element and is selected from at least one of Sm, Nd, Gd, Eu, Dy, Lu, Y, Pr, and La.

[0014] According to an embodiment of the present invention, the substrate is a LaAlO3 substrate, a SrTiO3 substrate, a Si substrate, or a SiO2 substrate.

[0015] According to an embodiment of the present invention, it has the function of randomly firing neurons, and the function of randomly firing neurons is manifested as:

[0016] When an electrical pulse signal is input, the adaptive sensitization artificial neuron randomly generates a resistance mutation, and the change in the resistance value exceeds one order of magnitude; while when the electrical pulse signal is input in the reverse direction, the adaptive sensitization artificial neuron is reset.

[0017] According to an embodiment of the present invention, within a preset electrical pulse intensity range, the firing probabilities of the artificial neurons under a series of electrical pulses with different intensities follow the S-shaped Sigmoid distribution in the biological nervous system.

[0018] According to an embodiment of the present invention, it has an adaptive sensitization characteristic;

[0019] The adaptive sensitization characteristic is manifested as: different adaptive sensitization states are shown for pulse signals with different intensities, and the corresponding firing probability functions change to different extents with the accumulation of pulse signals with different intensities. Specifically:

[0020] When a first pulse signal with an intensity less than or equal to E1 is input, the corresponding firing probability function remains unchanged;

[0021] When the input intensity is a second pulse signal greater than E1, the corresponding excitation probability function increases as the accumulation of the pulse signal increases.

[0022] Thus, this artificial neuron has an adaptive sensitization characteristic, can selectively filter out low-intensity background noise, while enhancing the sensitivity to weak signals, autonomously expanding the signal reception range to capture signals that could not be recognized originally.

[0023] According to an embodiment of the present invention, E1 is 0.05V / μm to 0.15V / μm. Thus, different usage scenarios can be satisfied.

[0024] In a second aspect of the present invention, a method for preparing an adaptive sensitization artificial neuron is provided, and a perovskite nickelate thin film device is provided.

[0025] According to an embodiment of the present invention, it includes the following steps:

[0026] (1) Grow a perovskite nickelate thin film on the surface of a substrate;

[0027] (2) Deposit a first electrode on the perovskite nickelate thin film through a first deposition, deposit a second electrode on the perovskite nickelate thin film through a second deposition, the first electrode and the second electrode are arranged oppositely, and a perovskite nickelate channel is left between the first electrode and the second electrode;

[0028] The first electrode is a pole piece capable of catalyzing hydrogenation, and the second electrode is a pole piece not capable of catalyzing hydrogenation;

[0029] (3) Place the device obtained in step (2) in a hydrogen atmosphere for annealing to perform hydrogen ion doping, so that the hydrogen ion concentration in the perovskite nickelate channel continuously decreases as the distance from the first electrode increases.

[0030] In the above embodiments, the first electrode is a pole piece capable of catalyzing hydrogenation, and the second electrode is a pole piece not capable of catalyzing hydrogenation, so that in the hydrogen atmosphere in step (3), hydrogen is doped into the perovskite nickelate thin film covered by the first electrode under the action of the first electrode, and diffuses a certain distance L around during the annealing process. On the diffusion path, as the distance from the first electrode increases, the hydrogen ion concentration continuously decreases, thereby obtaining a hydrogen ion gradient structure. At the same time, the second electrode is an electrode not capable of catalyzing hydrogenation to avoid the two electrodes simultaneously catalyzing hydrogen to form a symmetric hydrogen ion distribution in the perovskite nickelate channel, thereby eliminating the hydrogen ion gradient structure.

[0031] In a third aspect of the present invention, an application of the adaptive sensitization artificial neuron described in the second aspect in a neural network is provided.

[0032] According to an embodiment of the present invention, the application includes at least one of image edge recognition, image target detection, image target tracking, image segmentation, speech recognition, audio enhancement, and data classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0034] Figure 1 is a structural diagram of a perovskite neodymium nickelate thin film device;

[0035] Figure 2 is for the realization of the random neuron adaptive sensitization function, where Figure 2 a is the functional characteristic of a randomly excited neuron, Figure 2 b is the excitation probability distribution under different intensity electric pulses, Figure 2 c is the autonomous adjustment of the neuron state under different weak stimuli, Figure 2 d is the excitation probability function corresponding to different states;

[0036] Figure 3 is the mechanism characterization test of the adaptive sensitization characteristic, where Figure 3 a is the NRA (nuclear reaction analysis) test chart, Figure 3 b is the XAS (synchrotron X-ray absorption spectroscopy) test chart, Figure 3 c~ Figure 3 e are all cAFM (conducting atomic force microscopy) test charts;

[0037] Figure 4 is an application case of the inclusion and non-inclusion of artificial neurons in the edge recognition algorithm and the classification task of the neural network in different environments, where Figure 4 a is the comparison of the vehicle edge recognition effects of the traditional network and the adaptive sensitization network in different environments, Figure 4 b is the comparison of the building recognition effects of the traditional network and the adaptive sensitization network in a gradually changing environment, Figure 4 c is the comparison of the accuracy rates of the classification networks based on the edge recognition results using traditional neurons or adaptive sensitization neurons for different means of transportation. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present invention and should not be construed as limiting the present invention.

[0039] Traditional neuron devices have the following characteristics: 1. Traditional artificial neurons rely on complex circuits composed of multiple electronic components or artificial adjustment of external inputs to change the neuron state. They cannot achieve intelligent self-regulation from the device itself, resulting in a very complex implementation method for neuron state switching. 2. Traditional neuron devices cannot directly interact with the environment for information. They can only passively receive external signals. Under the same external conditions, the device can only exhibit one corresponding state, lacking different processing methods of the hardware itself for the same external signal, that is, the device cannot autonomously interact with environmental information. That is, traditional neurons can only passively receive signals and have no self-regulation ability. 3. Once a traditional neuron device is manufactured, it can only receive signals within an inherent limited range and cannot overcome the natural limitation of its own reception range. When facing the situation of signal intensity change in different environments, there are often many problems where signals cannot be received, resulting in information loss. 4. In the case of input changes in a neural network, a network using ordinary neurons does not have autonomous adaptation and sufficient generalization ability, or very complex algorithms are required to achieve self-adaptation. The realization of neural network self-adaptation depends on algorithms.

[0040] Accordingly, a first aspect of the present invention provides an adaptive sensitized artificial neuron, including a perovskite nickelate thin film device.

[0041] In the first aspect of the present invention, based on the adaptive sensitization performance of the perovskite nickelate thin film device, it is used as an adaptive sensitized artificial neuron. Thus, when a single perovskite nickelate thin film device is used as an artificial neuron, it has the characteristics of adaptive sensitization while greatly saving energy consumption. Specifically, when the perovskite nickelate thin film device is used as an artificial neuron, it can break through the limitation that traditional neurons can only passively receive signals and cannot autonomously process information. The adaptive sensitized artificial neuron of the present invention can achieve autonomous adjustment of the device state under the same signal.

[0042] First, the adaptive sensitization artificial neuron of the present invention breaks through the limitation of traditional artificial neurons relying on circuits and manual operation, and realizes biological neuron characteristics in a simple way through a single device. Traditional neurons often rely on complex circuits to realize some biological characteristics, or single-device neurons only have simple switching functions, while the neurons of the present invention only need a single device to realize neuron functions, which greatly saves energy consumption and realizes the important biological characteristic of adaptive sensitization, making up for the contradiction between the simplicity and versatility of traditional artificial neurons, and has broad practical application prospects in the fields of brain-like computing, new electronic chips, etc.

[0043] Second, break through the limitation that traditional neurons can only passively receive signals and cannot process information autonomously, and realize autonomous adjustment of the device state under the same signal. Traditional artificial neurons can only serve as signal receivers. When external environmental conditions such as light signals and electrical signals change, neurons produce corresponding changes to respond to environmental changes, but such neurons do not have the ability to process information autonomously and can only show a unique state corresponding to different environmental conditions. This new neuron can make self-adjustments according to different inputs, and can show different response states to the same input signal before and after autonomous processing. This adaptive ability and autonomous regulation ability from the hardware itself is one of the important foundations for realizing future biological-level intelligent devices.

[0044] Third, break through the limitation that traditional neurons can only work within their inherent range, autonomously enhance the ability to receive weak signals, broaden the signal reception range, and solve the problem of signal loss. Once traditional neurons are prepared, they will naturally form a very narrow signal reception range, which is very powerless when facing a changing environment, because environmental changes are often accompanied by large changes in the signal range, and the limited signal reception range of traditional neurons will lose a lot of dynamic information. The adaptive sensitized artificial neurons can make autonomous judgments on important information that recurs in a changing environment, enhance the ability to capture previously unrecognizable signals, autonomously broaden the signal reception range, and solve the problem of information loss of traditional neurons when facing a dynamic environment.

[0045] Fourth, breaking through the limitation that traditional neural networks need to rely on complex algorithms to achieve adaptability to inputs. The self - adaptability of this new neuron - inspired neural network does not depend on complex algorithms at all, enabling the neural network to have autonomous cognitive ability when facing changing inputs, directly solving the problem of data drift caused by input changes. For example, when a traditional classification network is trained with objects in a bright environment and tested with objects in a dark environment, obvious model failure will occur. The classification network using adaptive sensitized neurons can well adapt to various changing environments. This kind of adaptive sensitized neuron enables the neural network to greatly save computational costs and complete more complex tasks in a simple and efficient way.

[0046] In some embodiments, the perovskite nickelate thin - film device includes:

[0047] A substrate, a perovskite nickelate thin - film located on the surface of the substrate, a first electrode and a second electrode oppositely arranged on the perovskite nickelate thin - film, and a perovskite nickelate channel left between the first electrode and the second electrode;

[0048] The first electrode is a polar plate capable of catalyzing hydrogenation, and the second electrode is a polar plate incapable of catalyzing hydrogenation;

[0049] Hydrogen ions are doped in the perovskite nickelate thin - film under the first electrode. The hydrogen ions diffuse a distance L in the perovskite nickelate channel, and the concentration of hydrogen ions in the perovskite nickelate channel continuously decreases with the increase of the distance from the first electrode.

[0050] In this embodiment, the concentration of hydrogen ions in the perovskite nickelate channel continuously decreases with the increase of the distance from the first electrode. Specifically, it can linearly decrease or non - linearly decrease with the increase of the distance from the first electrode, forming a hydrogen - ion concentration gradient structure in the perovskite nickelate thin - film device. Due to the high - sensitivity response of the hydrogen - ion concentration gradient structure to electrical pulses, the distribution of hydrogen ions near the first electrode can be controlled by regulating electrical pulses. Thus, when the device is used as an artificial neuron, it has adaptive sensitization characteristics. Specifically, when an electrical pulse is input, the hydrogen ions migrate under the action of the electrical pulse. According to the different intensities of the input electrical pulses, the content of hydrogen ions at the interface of the first electrode is different due to the migration of hydrogen ions. More hydrogen content shows a non - excited state, and less hydrogen content shows an excited state; as the electrical pulses are cumulatively input, the corresponding excitation probability function changes to different degrees. Thus, it can be used as an artificial neuron.

[0051] Further, the first electrode is a Pd electrode or a Pt electrode, and the second electrode can be an Au electrode, an Ag electrode or a Pb electrode. Thus, catalytic hydrogenation occurs on one side of the perovskite nickelate thin film, and further, the hydrogen ion concentration in the perovskite nickelate channel continuously decreases with the increase of the distance from the first electrode.

[0052] Further, the perovskite nickelate channel between the first electrode and the second electrode may be fully diffused with hydrogen ions, or a section of hydrogen ions may be diffused from the first electrode into the perovskite nickelate channel. The hydrogen ion distribution with a hydrogen ion concentration gradient sequence structure can enable the device to have an adaptive sensitization characteristic. As a specific example, the width of the perovskite nickelate channel can be 0.1 μm to 100 μm, for example, 0.5 μm, 10 μm, 20 μm, 30 μm, 50 μm, 80 μm, 100 μm, etc., and the diffusion distance L of the hydrogen ions can be 0.01 μm to 50 μm, for example, 0.01 μm, 0.5 μm, 0.6 μm, 1 μm, 2 μm, 3 μm, 10 μm, 25 μm, 40 μm, 50 μm, etc.

[0053] In some embodiments, on a substrate, there may be a basic unit composed of a hydrogen-doped perovskite nickelate channel and two electrodes, or multiple such units may be distributed on the substrate.

[0054] The shape of the perovskite nickelate channel is not limited. In some embodiments, the distances at different positions of the perovskite nickelate channel may be equal or unequal.

[0055] In some embodiments, in the perovskite nickelate thin film, the chemical formula of the perovskite nickelate is MNiO3, where M represents a metal element and is selected from at least one of Sm, Nd, Gd, Eu, Dy, Lu, Y, Pr, and La. In a specific embodiment, when two or more metal elements are selected, it may be a composition containing the metal element or an alloy form containing the metal.

[0056] In some embodiments, the substrate is a LaAlO3 substrate, a SrTiO3 substrate, a Si substrate, or a SiO2 substrate. The above substrates all play a supporting role.

[0057] In some embodiments, the adaptive sensitization artificial neuron has the function of randomly firing neurons.

[0058] Specifically, the function of randomly firing neurons is manifested as:

[0059] When an input electrical pulse signal is applied, the adaptive sensitization artificial neuron randomly generates a resistance mutation, and the change in the resistance value exceeds one order of magnitude;

[0060] When a reverse electrical pulse signal is input, the adaptive sensitized artificial neuron is reset.

[0061] Furthermore, within a preset electrical pulse intensity range, the firing probabilities of artificial neurons under a series of electrical pulses with different intensities follow the S-shaped sigmoid distribution in the biological nervous system. Based on the different specifications of the perovskite nickelate channels and the different diffusion states of hydrogen ions therein, the firing probabilities of different adaptive sensitized artificial neurons show the S-shaped sigmoid distribution in the biological nervous system within a specific electrical pulse intensity range. For example, for a perovskite nickelate channel with a width of 20 μm and a device with a hydrogen ion diffusion length of 1 μm, in the initial state, it follows the S-shaped sigmoid distribution under electrical pulse stimulation between 3.2 V and 5.4 V.

[0062] In some embodiments, the adaptive sensitization is manifested as follows: different adaptive sensitization states are shown for pulse signals with different intensities, and the corresponding firing probability function changes to different degrees with the accumulation of pulse signals with different intensities. Specifically:

[0063] When a first pulse signal with an intensity less than or equal to E1 is input, the corresponding firing probability function remains unchanged;

[0064] When a second pulse signal with an intensity greater than E1 is input, the corresponding firing probability function increases with the increase in the accumulation of the pulse signal. Thus, it is possible to make an autonomous judgment on important information repeatedly appearing in a changing environment, enhance the capture ability of signals that cannot be recognized in the initial state, autonomously broaden the signal reception range, and solve the problem of information loss of traditional neurons when facing a dynamic environment.

[0065] According to the different selected materials and devices, the corresponding E1 is different, and the present invention does not make specific limitations. As an example, E1 is 0.05 V / μm to 0.15 V / μm, for example, 0.05 V / μm, 0.1 V / μm, 0.15 V / μm, etc. For electrical pulse signals less than or equal to E1, the excitation function does not shift with the cumulative input, thereby selectively filtering out low-intensity background noise; while for electrical pulse signals greater than E1, the excitation function shifts with the cumulative input to enhance the sensitivity to weak signals, autonomously expand the signal reception range to capture signals that could not be recognized originally, and can well solve the problem of information loss caused by the inability of traditional neurons to adapt when facing changes in input signals caused by environmental changes, and can greatly enhance the recognition ability for detailed information. A large number of experiments have confirmed that this adaptive sensitization phenomenon can be repeated on the same device and reproduced on devices with different specifications, thus ensuring good reproducibility of the function of the adaptive sensitized neuron on the device.

[0066] The second aspect of the present invention provides a method for preparing the adaptive sensitized artificial neuron described in the first aspect of the present invention, and provides a perovskite nickelate thin film device.

[0067] In some embodiments, the method includes the following steps:

[0068] (1) Grow a perovskite nickelate thin film on the surface of a substrate;

[0069] (2) Deposit a first electrode on the perovskite nickelate thin film by a first deposition, and deposit a second electrode on the perovskite nickelate thin film by a second deposition. The first electrode and the second electrode are arranged opposite to each other;

[0070] The first electrode is a pole piece capable of catalyzing hydrogenation, and the second electrode is a pole piece not capable of catalyzing hydrogenation;

[0071] (3) Place the device obtained in step (2) in a hydrogen atmosphere for annealing to perform doping of hydrogen ions, so that the hydrogen ion concentration in the perovskite nickelate channel continuously decreases with the increase of the distance from the first electrode.

[0072] In the above embodiments, the first electrode is a pole piece capable of catalyzing hydrogenation, and the second electrode is a pole piece not capable of catalyzing hydrogenation. So that in the hydrogen atmosphere in step (3), hydrogen is doped into the perovskite nickelate thin film covered by the first electrode under the action of the first electrode, and diffuses a certain distance L around during the annealing process. On the diffusion path, the hydrogen ion concentration continuously decreases with the increase of the distance from the first electrode, thereby obtaining a hydrogen ion gradient structure. At the same time, the second electrode is an electrode not capable of catalyzing hydrogenation to avoid the two electrodes simultaneously catalyzing hydrogen to form a symmetric hydrogen ion distribution in the perovskite nickelate channel, thereby eliminating the hydrogen ion gradient structure.

[0073] As an example, the preparation method further includes: performing ultraviolet lithography after the growth of the nickelate thin film in step (2) to form a plurality of rectangular nickelate channels, so as to prepare a basic unit composed of a plurality of hydrogen-doped perovskite nickelate channels and two electrodes on the substrate. Specifically, in combination with Figure 1 Look, spin-coat a photoresist thin film on the substrate on which the perovskite nickelate thin film is grown, and perform ultraviolet exposure using ultraviolet lithography technology. Then use ion beam etching to etch away the photoresist, leaving only the rectangular nickelate channels on the substrate, and the rest is the insulating LaAlO3 substrate.

[0074] As an example, the preparation method may be: grow a nickelate thin film on a LaAlO3 substrate by a metal-organic decomposition (MOD) method, deposit a gold electrode by thermal evaporation, and deposit a Pd electrode by electron beam evaporation.

[0075] The third aspect of the present invention provides an application of the adaptive sensitization artificial neuron according to the first aspect of the present invention in a neural network.

[0076] In some embodiments, all neurons in a traditional neural network can be replaced with adaptive sensitization artificial neurons. Based on the adaptive sensitization performance of the adaptive sensitization artificial neurons, the neural network has adaptive sensitization performance, and the trained neural network can autonomously adapt to training data and has a wider reception range. The above neural network breaks through the limitation that traditional neural networks need to rely on complex algorithms to achieve adaptation to inputs. The self-adaptability of the neural network inspired by the adaptive sensitization artificial neuron in the first aspect of the present invention enables the neural network to have autonomous cognitive ability when facing changing inputs without relying on complex algorithms at all, and solves the problem of data drift caused by input changes. For example, when a traditional classification neural network is trained using objects in a bright environment and tested using objects in a dark environment, obvious model failure will occur, while a classification network using adaptive sensitization neurons can well adapt to various changing environments. Such adaptive sensitization neurons enable the neural network to greatly save computational costs and be able to complete more complex tasks in a simple and efficient manner, opening up a new path for the future realization of autonomous disaster relief robots and autonomous exploration of unknown outer space environments.

[0077] In some embodiments, it is used for image edge recognition, image target detection, image target tracking, image segmentation, speech recognition, audio enhancement, and data classification.

[0078] Example 1

[0079] Example 1 provides a manufacturing method of a neodymium nickelate perovskite artificial neuron, and the specific operation is as follows:

[0080] (1) Grow a NdNiO3 thin film on a LaAlO3 substrate by the metal-organic decomposition (MOD) method

[0081] React Nd(AC)3, Ni(AC)2 with C7H 15 COOH and ammonia gas to generate Nd(C7H 15 COO)3 and Ni(C7H 15 COO)2. Dissolve Nd(C7H 15 COO)3 and Ni(C7H 15 COO)2 in xylene. In every 1 L of xylene, add 0.05 mol of Nd(C7H 15 COO)3 and 0.05 mol of Nd(C7H 15(COO)2 to obtain a precursor solution; spin-coat the precursor solution on a LaAlO3 substrate. Subsequently, anneal at 450 °C under an oxygen pressure of 15 MPa to decompose the precursor, and then anneal at 850 °C under an oxygen pressure of 5 MPa to crystallize the material.

[0082] (2) Fabricate the device through standard ultraviolet lithography technology

[0083] Spin-coat a photoresist film on the NdNiO3 film with a substrate, and perform ultraviolet exposure using ultraviolet lithography technology. Then use ion beam etching to remove the photoresist at specific positions to obtain a rectangular nickelate channel.

[0084] (3) Deposit the Au electrode by thermal evaporation and deposit the Pd electrode by electron beam evaporation.

[0085] (4) Finally, anneal the NdNiO3 device in a mixed hydrogen atmosphere at 120 °C to complete the doping of hydrogen ions.

[0086] The structure of the perovskite neodymium nickelate artificial neuron prepared in Example 1 is shown in Figure 1 , with a NdNiO3 film grown on a LaAlO3 substrate, the first electrode and the second electrode are oppositely arranged at both ends of the NdNiO3 film, and the electrode coverage areas are 100 μm 2 , respectively. Between the first electrode and the second electrode is a rectangular NdNiO3 film channel with a channel length of 20 μm.

[0087] The electrical stimulation test of the perovskite neodymium nickelate artificial neuron in Example 1 is as follows:

[0088] I. Use a Tektronix AFG31000 series arbitrary function generator to generate electrical pulses, use a Keithley 2636B to measure the resistance change of the artificial neuron, and minimize the interference of vibration during the measurement through an ultra-low noise triaxial cable and a vibration isolation table. The Tektronix AFG31000 series arbitrary function generator and the Keithley 2636B are controlled by a LabVIEW program for test parameters. All applied electric fields in the test are marked with reference to the Au electrode.

[0089] II. Stimulate the artificial neuron with an electrical pulse of 0.21 V / μm. Whether it is excited or not, that is, whether there is an obvious resistance change or not, it is followed by a reverse and equal-sized pulse for resetting. This is the test process for one test point. Figure 2 a shows the random excitation behavior of the neuron under 30 pulse stimulations.

[0090] After testing the random firing of neurons, the artificial neurons were stimulated 30 times with pulses of different intensities, and the number of times the neurons fired in the 30 stimulations was counted to obtain the proportion of the number of firings, which was used as the firing probability of the neurons at that signal intensity. The data obtained is as Figure 2 b.

[0091] The perovskite neodymium nickelate artificial neurons were cumulatively stimulated with different signals, and no reverse pulse was applied before the first firing. Only when firing occurred would a reverse pulse be used for resetting. The different self - adaptive sensitization states obtained are as Figure 2 shown in c, and the firing probability functions for each corresponding state are as Figure 2 shown in d.

[0092] As Figure 2 shown in a, the hydrogen - doped neodymium nickelate neurons in Example 1 randomly generated resistance mutations under the action of an electrical pulse, and the degree of resistance change (real - time resistance R - initial non - firing state high resistance R0) exceeded an order of magnitude, and could be reset by a reverse pulse. The device exhibited the function of a randomly firing neuron.

[0093] As Figure 2 shown in b, the firing probability under different - intensity pulses followed the S - shaped Sigmoid distribution common in biological nervous systems. Here, the firing probability at each test point was the percentage of the number of times the neuron fired in 30 independent signal inputs.

[0094] As Figure 2 shown in c, the artificial neuron was also able to show self - adjustment of its state for weak signals that originally had a very low firing probability and repeated. That is, when facing weak pulses, the neuron initially showed a very low firing probability, and then when multiple weak pulses of the same intensity were cumulatively applied without any reverse reset pulse, it could be observed that the neuron autonomously adapted to these signals, changing from a low - firing - probability state to a higher - firing - probability state.

[0095] The autonomous switching of the artificial neuron in Example 1 under different states can also be represented by the change in the firing probability function. By accumulating signals of different intensities, different degrees of change in the firing probability function can be achieved, that is, it shows different self - adaptive sensitization states, as Figure 2 shown in d. Here, the firing probability at each test point was the percentage of the number of times the neuron fired in 30 independent signal inputs after the neuron state transition. It should be noted that the artificial neuron in Example 1 does not respond to all weak signals from the external environment. When a very weak signal (0.10 V / μm) was applied to stimulate the neuron, even if it was repeated more than 30 times, there was no shift in the self - adaptive firing probability function, which means that the neuron can selectively filter out low - intensity background noise, as shown in Figure 2c The last figure. This bio - adaptive sensitization characteristic achieved by a single device indicates that the neurons in Example 1 can autonomously adjust their firing probability function without human intervention, enhance the sensitivity to weak signals, autonomously expand the signal reception range to capture information that could not be recognized originally, well solve the problem of information loss caused by the inability of traditional neurons to adapt when facing input signal changes caused by environmental changes, and can greatly enhance the ability to identify detailed information. This phenomenon can be repeated on the same perovskite nickelate thin - film device and can also be reproduced on perovskite nickelate thin - film devices of different specifications, thus ensuring good reproducibility of the adaptive - sensitization neuron function on the device.

[0096] III. Test characterization of the adaptive - sensitization characteristic of the artificial neuron in Example 1

[0097] Figure 3 a shows the absolute content distribution of hydrogen ions in the depth direction after hydrogen doping using an NRA (nuclear reaction analysis) test device. Due to the limitations of the test method, the diffusion of hydrogen ions in the depth of the artificial neuron in Example 1 was selected for testing to observe the natural diffusion situation. Figure 3 a It can be seen that the hydrogen ions show a gradient distribution in the thickness direction, which most intuitively proves the diffusion of hydrogen ions in the NdNiO3 lattice.

[0098] Figure 3 b shows the peak change of the K - edge energy spectrum of Ni element in the NdNiO3 device measured by XAS (synchrotron X - ray absorption spectroscopy). The higher the hydrogen - ion concentration, the lower the peak energy of the XAS spectrum. Thus, the hydrogen - ion concentration can be characterized by the valence state of Ni. Tests were carried out at different positions near the Pd electrode. Specifically, the abscissa represents the distance from the interface between the Pd electrode and the device channel as the origin to both sides. Specifically, - 2μm to 0μm is the distance to the NdNiO3 thin film covered by the Pd electrode, 0μm is the junction of the Pd electrode and the device channel, and 0μm to 20μm is the distance of the device channel. It was observed that the peak energy of the K - edge of Ni element under the Pd electrode is at a low level, which is due to a slight decrease in the valence state of Ni caused by the injection of extra electrons through hydrogen doping, while the peak energy of the K - edge of Ni element gradually increases near the junction of the Pd electrode and the device channel, which is due to diffusion. There are obvious differences in the observed movement of the peak energy of the Ni element K - edge between the excited state and the non - excited state. The larger decrease in the energy peak in the non - excited state indicates a higher hydrogen - ion concentration in the device channel, resulting in a higher resistance state, which is consistent with the cAFM test results.

[0099] Figure 3 c -Figure 3 Figure e shows the cAFM (conducting atomic force microscopy) test diagrams of samples in different states. Figure 3 c- Figure 3 The abscissa of Figure e is the distance from the junction of the Pd electrode and the device channel as the origin into the perovskite nickelate channel. Among them, there is no test data at 0 - 0.15 μm due to the sensitivity limitation of the test equipment. The local resistance of the device at a corresponding position can be directly reflected by the local current distribution at different positions, which is directly related to the hydrogen ion concentration. The smaller the current and the larger the resistance at a position, the higher the hydrogen ion concentration at that position. Figure 3 The difference between hydrogenated and non-hydrogenated samples can be seen from Figure c. The non-hydrogenated device does not have a current gradient, indicating no resistance change, while the hydrogenated device has a current gradient, indicating that the hydrogen ion concentration varies in a gradient at different positions. Figure 3 Figure d shows the difference between the excited state and the non-excited state. The current at the position close to the Pd electrode in the non-excited state is significantly smaller, indicating a higher hydrogen ion concentration and corresponding to the high resistance of the device, that is, the non-excited state. While the current at the position close to the Pd electrode in the excited state becomes larger, indicating a relatively lower hydrogen ion concentration and corresponding to the low resistance of the device, that is, the excited state. Figure 3 Figure e shows the difference between the initial state and the sensitized state. By comparing the normalized currents of the initial state and the sensitized state in different states before and after the completion of adaptive sensitization, it shows that the morphology of the hydrogen ion distribution gradient has changed significantly, and it is speculated that this is an important mechanism for the device to have adaptive sensitization. Figure 3 c, Figure 3 d and Figure 3 Figure e together show that the device of Example 1 is doped with hydrogen ions at the contact interface between the perovskite nickelate thin film and the Pd electrode therein, and the diffusion distance L of the hydrogen ions from the contact interface to the perovskite nickelate thin film is 0.6 μm - 0.8 μm.

[0100] Test Example

[0101] The adaptive sensitization neuron of Example 1 was used for edge recognition in a neural network, resulting in a significant performance improvement.

[0102] Traditional edge recognition neural networks mainly consist of two parts, a convolutional module and a fully connected module. The convolution is based on the Alex Net architecture and contains 5 convolutional layers. Subsequently, there is a fully connected layer that uses traditional neurons with only a single probability activation function (the function cannot be switched). The adaptive sensitization network of Example 2 replaces all ordinary neurons in the traditional edge recognition neural network with artificial neurons with adaptive sensitization characteristics. Further, when applied, Example 2 assigns the data of the randomly firing neuron function and adaptive characteristics tested by the artificial neurons of Example 1 to the neurons in the traditional network through modeling to simulate the adaptive sensitization network.

[0103] When performing edge recognition algorithms and neural network classification, after the picture is read, the picture data of the red, green, and blue channels are saved separately. Subsequently, the Sobel operator is used to calculate the gradient. By calculating the magnitude and direction of these gradient images, the gradient information of each pixel in the image is obtained. Then, this processed data is input into the adaptive sensitization neuron layer for calculation. Finally, the pulse rate value is converted into a grayscale pixel value to visualize the edge detection result. The obtained detection result is shown in Figure 4 , Figure 4 It is an application case of edge recognition algorithms and neural networks with and without adaptive sensitization neurons in classification tasks under different environments.

[0104] Traditional edge recognition and neural networks often rely on complex algorithms to achieve adaptability to environmental changes. However, the neurons of the present invention can directly endow neural networks with the ability to autonomously recognize input changes, enabling them to autonomously adjust the signal reception range in different environments, thus solving the problems of information loss and data drift caused by input changes that plague traditional neural networks.

[0105] As Figure 4 a, when traditional edge recognition sets relevant parameters to adapt to objects in a bright environment, it cannot recognize objects in a dark environment well and loses a lot of key edge information. However, edge recognition using adaptive sensitization neurons can fully autonomously adapt to this change, not only improving the recognition of buildings in a detailed environment in a bright environment, but also significantly improving performance in a dark environment.

[0106] In Figure 4 b, further, when performing building edge recognition tasks in a changing environment, the significant effect brought by the autonomous adjustment of the activation function of the adaptive sensitization neurons can be more intuitively seen. This adaptive ability can dynamically and autonomously switch states to capture important edge information in a changing environment while maintaining excellent recognition ability in a bright environment.

[0107] In Figure 4In b, the classification neural network built based on the edge recognition results quantifies the difference in the edge information extraction capabilities between the adaptive sensitized neurons and the traditional neurons. The edge recognition results of vehicles in a bright environment are used as the training set to train the classification network, and the edge recognition results in a dark environment are used as the test set to test the network. In the classification tasks of different means of transportation in such a different environment, it can be seen that the adaptive sensitized neurons can provide more complete structural information for the classification network through autonomous adjustment of their states, enabling the neural network to make more accurate classifications such as Figure 4 c.

[0108] In this document, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of this application, "a plurality" means two or more, unless otherwise specifically and clearly defined.

[0109] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0110] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. An adaptive sensitized artificial neuron, characterized in that, Including a perovskite nickelate thin film device.

2. The adaptive sensitized artificial neuron according to claim 1, wherein The perovskite nickelate thin film device includes: A substrate, a perovskite nickelate thin film located on the surface of the substrate, a first electrode and a second electrode oppositely arranged on the perovskite nickelate thin film, and a perovskite nickelate channel left between the first electrode and the second electrode; The first electrode is a pole piece capable of catalyzing hydrogenation, and the second electrode is a pole piece incapable of catalyzing hydrogenation; Hydrogen ions are doped in the perovskite nickelate thin film below the first electrode, and the hydrogen ions diffuse a distance L in the perovskite nickelate channel, and the concentration of hydrogen ions in the perovskite nickelate channel continuously decreases with the increase of the distance from the first electrode.

3. The adaptive sensitized artificial neuron according to claim 2, wherein The length of the perovskite nickelate channel between the first electrode and the second electrode is 0.1 μm to 100 μm; Optionally, the distance L is 0.01 μm to 50 μm.

4. The adaptive sensitized artificial neuron according to claim 2, wherein In the perovskite nickelate thin film, the structural formula of perovskite nickelate is MNiO3, where M represents a metal element and is selected from at least one of Sm, Nd, Gd, Eu, Dy, Lu, Y, Pr, and La; Optionally, the substrate is a LaAlO3 substrate, a SrTiO3 substrate, a Si substrate, or a SiO2 substrate.

5. The adaptive sensitized artificial neuron according to any one of claims 1-4, characterized in that Having the function of randomly exciting neurons; The function of randomly exciting neurons is manifested as: When an input electrical pulse signal is applied, the adaptive sensitized artificial neuron randomly generates a resistance mutation, and the change in the resistance value exceeds one order of magnitude; while when the electrical pulse signal is input in the reverse direction, the adaptive sensitized artificial neuron is reset; Optionally, within a preset electrical pulse intensity range, the excitation probability of the artificial neuron under a series of electrical pulses with different intensities follows the S-shaped Sigmoid distribution in the biological nervous system.

6. The adaptive sensitized artificial neuron according to any one of claims 1-4, characterized in that, Having an adaptive sensitization characteristic; The adaptive sensitization characteristic is manifested as: different adaptive sensitization states are shown for pulse signals with different intensities, and the corresponding excitation probability function changes to different degrees with the accumulation of pulse signals with different intensities. Specifically: When a first pulse signal with an intensity less than or equal to E1 is input, the corresponding excitation probability function remains unchanged; When a second pulse signal with an intensity greater than E1 is input, the corresponding excitation probability function increases with the increase of the accumulation of the pulse signal; Optionally, E1 is 0.05 to 0.15 V / μm.

7. A method for preparing an adaptive sensitized artificial neuron, characterized in that, Providing a perovskite nickelate thin film device.

8. The preparation method of the adaptive sensitized artificial neuron according to claim 7, wherein Including the following steps: (1) Growing a perovskite nickelate thin film on the surface of a substrate; (2) Depositing a first electrode on the perovskite nickelate thin film by a first deposition, and depositing a second electrode on the perovskite nickelate thin film by a second deposition. The first electrode and the second electrode are oppositely arranged, and a perovskite nickelate channel is left between the first electrode and the second electrode; The first electrode is a pole piece capable of catalyzing hydrogenation, and the second electrode is a pole piece incapable of catalyzing hydrogenation; (3) Annealing the device obtained in step (2) in a hydrogen atmosphere to perform doping of hydrogen ions, so that the concentration of hydrogen ions in the perovskite nickelate channel continuously decreases with the increase of the distance from the first electrode.

9. Application of the adaptive sensitized artificial neuron according to any one of claims 1-6 or the adaptive sensitized artificial neuron obtained by the preparation method according to claim 7 or 8 in a neural network.

10. The application according to claim 9, wherein The application includes at least one of image edge recognition, image target detection, image target tracking, image segmentation, speech recognition, audio enhancement, and data classification.

Images