Machine learning reservoir computing element, machine learning method

Abstract

This invention provides a safe and stable reservoir computing element for machine learning that can be manufactured simply and inexpensively, and whose elements can be made small and lightweight, as well as a machine learning method using this element. [Solution] The device comprises a substrate, a gate electrode layer and a gate insulating layer formed sequentially on one side of the substrate, a source electrode layer and a drain electrode layer formed on one side of the gate insulating layer and spaced apart from each other, and a channel layer formed in contact with the gate insulating layer, the source electrode layer and the drain electrode layer. The gate insulating layer contains a solid electrolyte and ions, and the channel layer contains InO. An electrical double layer is formed in the region including the junction between the gate insulating layer and the channel layer. An electrical signal is input to the gate electrode layer, and machine learning is performed using at least the gate current signal output from the gate electrode layer.

Description

[Technical Field]

[0001] The present invention relates to a reservoir computing element for machine learning and a machine learning method using the same. [Background technology]

[0002] In recent years, neural networks and artificial intelligence that mimic brain neural networks have been actively researched, and for example, artificial intelligence technologies that utilize multi-layered neural networks, including deep learning, are well known. These artificial intelligence technologies generally have problems such as requiring a large amount of computing resources, consuming a lot of power, and requiring large-scale equipment. Therefore, there are significant obstacles to their use, especially in small, portable devices where power and size are limited.

[0003] On the other hand, another approach to neural networks is reservoir computing. This reservoir computing consists of a three-layer structure: an input layer, a recurrent layer, and an output layer. The weights and thresholds between the input layer and the recurrent layer, and between the recurrent layers themselves, are kept constant, and machine learning is performed to calculate weights and thresholds only between the recurrent layer and the output layer, allowing for high-speed computation.

[0004] Furthermore, physical reservoir elements, which replace the recurrent layer with physical phenomena, are also being considered. These reservoir elements utilize physical phenomena to perform machine learning necessary for artificial intelligence (AI), and because they can perform calculations at extremely high speeds compared to conventional large-scale computing elements, they are expected to have applications in edge AI computing.

[0005] The properties required for such reservoir elements include nonlinear response, short-term memory, and high dimensionality. Known physical reservoir elements include electrical and electronic elements that utilize neural mimic circuits and memristors, and electrochemical elements that utilize the delayed response of ionic liquids and electrolytes (see, for example, Patent Documents 1 and 2). [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2022-182198 [Patent Document 2] Japanese Patent Publication No. 2023-013949 [Overview of the Initiative] [Problems that the invention aims to solve]

[0007] However, the reservoir elements for machine learning that have been considered so far have had several drawbacks, including (a) the devices are large and difficult to miniaturize, making them unsuitable for edge AI where portability is essential (dynamical systems, optical systems, mechanical systems), (b) the circuits are complex and expensive (electrical and electronic systems), and (c) there are problems with safety and stability (biological systems, electrochemical systems).

[0008] This invention was made in view of the above background, and aims to provide a safe and stable reservoir computing element for machine learning that can be manufactured simply and inexpensively, and whose elements can be made smaller and lighter, as well as a machine learning method using the same. [Means for solving the problem]

[0009] To solve the above problems, the reservoir computing element for machine learning of one embodiment of the present invention proposes the following means. (1) The reservoir computing element for machine learning according to embodiment 1 of the present invention comprises a substrate, a gate electrode layer and a gate insulating layer formed in order on one side of the substrate, a source electrode layer and a drain electrode layer formed on one side of the gate insulating layer and spaced apart from each other, and a channel layer formed in contact with the gate insulating layer, the source electrode layer and the drain electrode layer, wherein the gate insulating layer contains a solid electrolyte and ions, and the channel layer contains InO, and an electric double layer is formed in the region including the junction between the gate insulating layer and the channel layer, and an electrical signal is input to the gate electrode layer to perform machine learning using at least the gate current signal output from the gate electrode layer.

[0010] (2) A second aspect of the present invention is a reservoir computing element for machine learning according to the first aspect, in which the current-voltage characteristic showing the change in the drain current value between the source electrode layer and the drain electrode layer in response to a change in the voltage value of the gate electrode layer that crosses 0V, the drain current value when the voltage value of the gate electrode layer is 0V is greater when the voltage decreases than when the voltage increases.

[0011] (3) A third aspect of the present invention is a reservoir computing element for machine learning according to aspect 1 or 2, wherein the solid electrolyte is LaZrO.

[0012] (4) Aspect 4 of the present invention is a reservoir computing element for machine learning according to any one of aspects 1 to 3, wherein the ion is a hydrogen ion.

[0013] (5) The machine learning method of aspect 5 of the present invention is a machine learning method using any one of the machine learning reservoir computing elements of aspects 1 to 4, wherein an electrical signal is input to the gate electrode layer, and machine learning is performed using both the gate current signal output from the gate electrode layer and the drain current signal output from the drain electrode layer. [Effects of the Invention]

[0014] According to the present invention, there can be provided a safe and stable reservoir computing element for machine learning, which can be easily and inexpensively manufactured, and the elements can be reduced in size and weight, and a machine learning method using the same.

Brief Description of Drawings

[0015] [Figure 1] FIG. 8 is a schematic cross-sectional view showing the layer structure of the reservoir computing element according to an embodiment of the present invention. [Figure 2] FIG. 11 is a schematic diagram schematically showing the action of an electric double layer. [Figure 3] FIG. 14 is a graph showing the results of an example.

Modes for Carrying Out the Invention

[0016] Hereinafter, a reservoir computing element for machine learning, which is an embodiment to which the present invention is applied, and a machine learning method using the same will be described. The following embodiments are specifically described in order to better understand the gist of the invention, and do not limit the present invention unless otherwise specified. In addition, the drawings used in the following description may show, for convenience, the main parts enlarged in order to make the features of the present invention easy to understand, and the dimensional ratios of the respective components are not necessarily the same as the actual ones.

[0017] (Reservoir Computing Element) FIG. 1 is a schematic cross-sectional view showing the layer structure of the reservoir computing element for machine learning according to an embodiment of the present invention. As shown in FIG. 1, a reservoir computing element for machine learning (hereinafter simply referred to as a reservoir computing element) 10 composed of a thin film transistor according to the present embodiment has a gate electrode layer 12 and a gate insulating layer 13 formed in order on one surface 11a side of a substrate 11, a source electrode layer 14 and a drain electrode layer 15 formed on one surface 13a side of the gate insulating layer 13 and arranged apart from each other, and a channel layer 16 arranged between the source electrode layer 14 and the drain electrode layer 15 and formed so as to contact the gate insulating layer 13, the source electrode layer 14, and the drain electrode layer 15.

[0018] In this embodiment, the source electrode layer 14, drain electrode layer 15, and channel layer 16 are formed side by side on one side 13a of the gate insulating layer 13. However, other configurations are also possible, for example, in which the channel layer is formed to cover one side of the gate insulating layer, and the source electrode layer and drain electrode layer are formed on top of this channel layer, spaced apart from each other.

[0019] The substrate 11 can be, for example, an insulating substrate or a semiconductor substrate. Examples of insulating substrates include high heat-resistant glass, silicon substrates (SiO2 / Si substrates) having a thermal oxide film (silicon oxide film), alumina (Al2O3) substrates, STO (SrTiO) substrates, and Si substrates on which an STO (SrTiO) layer is formed via an SiO2 layer and a Ti layer. Examples of semiconductor substrates include Si substrates, SiC substrates, and Ge substrates. The thickness of the substrate 1 can be, for example, within the range of 10 μm to 1 mm.

[0020] The gate electrode layer 12 may be a single-layer film consisting of one conductive film, or a multilayer film consisting of two or more conductive films. For example, the conductive film constituting the gate electrode layer 12 can be made of a metal, an alloy containing that metal, or a metal oxide. Examples of metals used in the conductive film include platinum, gold, silver, copper, aluminum, molybdenum, palladium, ruthenium, iridium, tungsten, and titanium.

[0021] Examples of metal oxides for conductive films include indium tin oxide (ITO) and ruthenium oxide (RuO2). The thickness of the gate electrode layer 12 can be, for example, within the range of 10 nm to 500 nm. Furthermore, to improve the adhesion between the gate electrode layer 12 and the substrate 11, an adhesion layer made of an oxide film such as titanium oxide (TiO2) may be formed between the gate electrode layer 12 and the substrate 11.

[0022] The gate insulating layer 13 is composed of a solid electrolyte and ions. Examples of solid electrolytes include oxides containing lanthanum and zirconium (LaZrO), and oxides containing lithium (LiPO, LiMnO, LiCoO). In this embodiment, a LaZrO-containing film was used as the solid electrolyte. Such a LaZrO-containing film may be a sputtered film formed by a sputtering method, or it may be formed by a coating and firing method.

[0023] The lanthanum content of the LaZrO-containing film constituting the gate insulating layer 13 is preferably the same as or less than the zirconium content. The lanthanum and zirconium content of the LaZrO-containing film is, for example, in the range of 1:9 to 5:5 (atomic ratio). The LaZrO-containing film may also contain hydrogen. The hydrogen content of the LaZrO-containing film is, for example, in the range of 2 atom% to 20 atom%.

[0024] The thickness of the gate insulating layer 13 is, for example, within the range of 10 nm to 500 nm. The hydrogen (H) content can be measured, for example, using Rutherford backscattering spectrometry (RBS analysis), hydrogen forward scattering spectrometry (HFS analysis), and nuclear reaction analysis (NRA analysis) using a Pelletron 3SDH manufactured by National Electrostatics Corporation.

[0025] The source electrode layer 14 and the drain electrode layer 15 may each be a single-layer film consisting of one conductive film, or a multilayer film consisting of two or more conductive films. Examples of materials for the conductive films include metals, alloys containing the metal, and metal oxides. The examples of metals and metal oxides may be the same as those for the gate electrode layer 12.

[0026] The thickness of the source electrode layer 14 and the drain electrode layer 15 may be, for example, within the range of 10 nm to 500 nm. The distance between the source electrode layer 14 and the drain electrode layer 15, which is the distance in the first direction (x direction in Figure 1) where they face each other (electrode distance LT), may be, for example, within the range of 1 μm to 100 μm. The lengths of the source electrode layer 14 and the drain electrode layer 15 in the first direction (electrode length LS of the source electrode layer 14, electrode length LD of the drain electrode layer 15) may be, for example, within the range of 1 μm to 1000 μm, and preferably within the range of 10 μm to 200 μm.

[0027] The lengths of the source electrode layer 14 and the drain electrode layer 15 in the second direction (y-direction in Figure 1), which are perpendicular to the length in the first direction (electrode width WS of the source electrode layer 14 and electrode width WD of the drain electrode layer 15), may be, for example, within the range of 1 μm to 1000 μm, and preferably within the range of 10 μm to 100 μm.

[0028] The channel layer 16 consists of an InO-containing film containing indium oxide (InO). The InO-containing film is crystalline and may be a sputtered film formed by sputtering or a coated and fired film formed by coating and firing. The shape of the crystal grains of the InO-containing sputtered film may be, for example, columnar, spherical, or ellipsoidal.

[0029] The average diameter of the InO crystal grains constituting the channel layer 16 is, for example, 10 nm or more, and preferably within the range of 10 nm to 10 μm. The average diameter of the crystal grains is the average of the longest diameters of the crystal grains obtained by observing the cross-section of the InO-containing sputtered film with a TEM (transmission electron microscope). The thickness of the channel layer 16 is, for example, within the range of 10 nm to 200 nm. The channel layer 16 may be patterned to form a rectangle with sides of several tens to several hundreds of μm.

[0030] The operation of the reservoir computing element of this embodiment will be explained above. Typical characteristics required for a reservoir computing element are (a) nonlinearity, (b) short-term memory characteristics, and (c) high dimensionality. The reservoir computing element 10 of this embodiment utilizes the electric double-layer (EDL) generated in the region including the junction between the gate insulating layer 13 and the channel layer 16, and uses its drain current for reservoir computing.

[0031] Figure 2 is a schematic diagram illustrating the function of the electric double layer. When a gate voltage is applied to the gate electrode layer 12, hydrogen ions in the gate insulating layer 13, which is made of LaZrO, move to the vicinity of the interface (junction) with the channel layer 16, which is made of InO, and an electric double layer ED is formed. This induces a huge capacitance, and a large number of carrier electrons are generated in the channel layer 16.

[0032] When the gate voltage is reduced, hydrogen ions move away from the interface between the gate insulating layer 13 and the channel layer 16, and the capacitance of the electric double layer ED decreases. However, because the movement of hydrogen ions is slower than that of electrons, the state of the hydrogen ions does not immediately return to normal even when the gate voltage becomes 0V, and hysteresis occurs. Due to this delay in hydrogen ion movement, a relaxation phenomenon occurs in which the response of the drain current to the application of a pulsed voltage is also slightly delayed.

[0033] Similarly, the gate current, which flows in proportion to the change in gate voltage and capacitance, also experiences a delayed response and relaxation phenomenon due to the delay in hydrogen ion movement. These relaxation phenomena of drain current and gate current provide the nonlinearity, short-term memory characteristics, and high dimensionality required for reservoir computing elements.

[0034] Because this type of electric double-layer ED utilizes phenomena occurring in very thin regions, it is also suitable for high-level integration of the reservoir computing element 10.

[0035] (Manufacturing method for reservoir computing elements for machine learning) When manufacturing the reservoir arithmetic element 10 of the above-described embodiment, for example, a gate electrode layer 12 and a gate insulating layer 13 are stacked in this order on one surface 11a of the substrate 11. Then, a channel layer 16 is formed on one surface 13a of the gate insulating layer 13, and a source electrode layer 14 and a drain electrode layer 15 are formed in contact with this channel layer 16, thereby enabling manufacturing.

[0036] The conductive film constituting the gate electrode layer 12 can be formed using methods that are commonly used in conventional thin-film transistor gate electrode formation methods, such as sputtering, vapor deposition, or coating and firing. The coating and firing method involves applying a conductive film-forming coating solution to create a conductive film by heating, forming a coating film, and then heating the coating film to deposit the conductive film.

[0037] The LaZrO film, which is the solid electrolyte constituting the gate insulating layer 13, can be formed by, for example, sputtering or coating and firing. In the case of sputtering, a sintered target made of a sintered body of La2O3 and ZrO can be used as the LaZrO target. Such a LaZrO target may further contain carbon. The carbon content is preferably in the range of 0.1% by mass or more and 5.0% by mass or less.

[0038] In the sputtering method, a mixed gas containing argon and oxygen can be used as the plasma gas supplied to the sputtering apparatus. The content ratio of argon to oxygen is preferably in the range of 95:5 to 80:20 (partial pressure ratio). The plasma gas may further contain hydrogen gas or water vapor. The content ratio of the argon-oxygen mixed gas to the hydrogen gas or water vapor is preferably in the range of 99:1 to 95:5 (partial pressure ratio).

[0039] The coating and firing method involves applying a coating solution for forming a LaZrO-containing film, which generates a LaZrO-containing film upon heating, to form a coating film, and then heating the coating film to deposit the LaZrO-containing film. As the coating solution for forming the LaZrO-containing film, for example, a solution containing a lanthanum source, a zirconium source, and propionic acid can be used.

[0040] Lanthanum acetate can be used as the lanthanum source. Zirconium butoxide can be used as the zirconium source. The total content of lanthanum and zirconium in the coating solution for forming the LaZrO film is preferably in the range of 0.05 mol / kg to 2.0 mol / kg in terms of molar concentration on a LaZrO basis.

[0041] As a method for applying the coating solution for forming the LaZrO-containing film, for example, a spin coating method can be used. Alternatively, as a method for adjusting the thickness of the LaZrO-containing film, a method may be used in which the coating solution for forming the LaZrO-containing film is applied to form a coated film, and then the coated film is calcined to form a calcined film, and this process is repeated multiple times until the thickness of the calcined film laminate reaches the desired thickness, at which point the final calcination is performed. In this case, it is preferable that the calcination temperature is in the range of 100°C to 300°C, and the final calcination temperature is in the range of 300°C to 500°C.

[0042] The InO-containing film constituting the channel layer 16 can be formed, for example, by sputtering or coating and firing. In the case of sputtering, the InO target used in the film formation process can be, for example, a sintered body target made of a sintered body of In2O3 or a sintered body target made of a sintered body of In2O3 and carbon. The carbon content of the carbon-containing InO target is preferably in the range of, for example, 0.1% by mass or more and 5.0% by mass or less. The film thickness of the InO-containing sputtered film formed in the film formation process is preferably, for example, 10 nm or more.

[0043] In the sputtering method, a mixed gas containing argon and oxygen can be used as the plasma gas supplied to the sputtering apparatus. The ratio of argon to oxygen is preferably in the range of 95:5 to 80:20 (partial pressure ratio). The plasma gas may also contain hydrogen gas or water vapor. The ratio of the argon-oxygen mixed gas to the hydrogen gas or water vapor is preferably in the range of 99:1 to 99:5 (partial pressure ratio).

[0044] It is preferable to heat-treat an InO-containing sputtered film formed by the sputtering method. The heat treatment of the InO-containing sputtered film may be carried out in air. The heat treatment temperature is preferably in the range of 300°C to 400°C. The heat treatment time can be, for example, in the range of 0.5 hours to 1 hour. This heat treatment process may be carried out after the formation of the channel layer 16, or after the formation of the source electrode layer 14 and drain electrode layer 15 adjacent to the channel layer 16.

[0045] In the coating and firing method, the InO-containing film is formed by applying an InO-containing film-forming coating solution to form a coating film, and then heating the coating film. As the InO-containing film-forming coating solution, for example, a solution containing indium nitrate can be used.

[0046] As a method for applying the coating solution for forming the InO-containing film, for example, a spin coating method can be used. Alternatively, as a method for adjusting the film thickness of the InO-containing film, a method may be used in which the coating solution for forming the InO-containing film is applied to form a coated film, and then the coated film is calcined to form a calcined film, and this process is repeated multiple times until the film thickness of the calcined film laminate reaches the desired thickness, at which point the final calcination is performed. In this case, it is preferable that the calcination temperature is in the range of 80°C to 300°C, and the final calcination temperature is in the range of 180°C to 500°C.

[0047] Photolithography can be used as a method for forming the channel layer 16 into a desired shape.

[0048] The conductive films forming the source electrode layer 14 and the drain electrode layer 15 can be formed, for example, by sputtering, vapor deposition, or coating and firing. The coating and firing method involves applying a conductive film-forming coating solution that generates a conductive film by heating to form a coating film, and then heating the coating film to form the conductive film. As methods for forming the source electrode layer 14 and the drain electrode layer 15 into desired shapes, wet etching using photolithography, dry etching, lift-off method, and mask deposition using a contact mask can be used.

[0049] (Machine learning methods) The machine learning method of one embodiment of the present invention is performed using the machine learning reservoir arithmetic element of the embodiment described above. That is, an electrical signal is input to the gate electrode layer 12 of the reservoir arithmetic element 10, and machine learning is performed using the gate current signal output from the gate electrode layer 12. Alternatively, an electrical signal is input to the gate electrode layer 12 of the reservoir arithmetic element 10, and machine learning is performed using the drain current signal output from the drain electrode layer 15. Alternatively, machine learning is performed using both these gate current signals and drain current signals.

[0050] Specifically, for example, multiple time-series data are converted into pulse voltage signals and applied to the gate electrode layer 12. Then, the gate current signal output from the gate electrode layer 12 or the drain current signal output from the drain electrode layer 15 is obtained as a response value.

[0051] The obtained response values ​​are used as virtual nodes in the final recurrent layer of the reservoir operation. Weights are multiplied by these values ​​to perform a sum-of-products operation to calculate the learned values, and the weights are optimized (machine learning) to match the correct values ​​as closely as possible. This optimization uses methods such as backpropagation (BP method), gradient descent, and ridge regression. Finally, the optimized weight values ​​are multiplied by the response values ​​used for prediction, and a sum-of-products operation is performed to obtain the predicted values ​​for the time series data. [Examples]

[0052] We verified the effectiveness of a reservoir computing element for machine learning according to one embodiment of the present invention. The reservoir computing element of the embodiment was fabricated according to the manufacturing method of the reservoir computing element for machine learning described above. The configuration of the fabricated reservoir computing element (thin-film transistor) is as follows. • Substrate: Silicon wafer with thermal oxide coating (4-inch diameter) • Gate electrode layer: Pt sputtered film (100 nm) • Gate insulating layer: LaZrO coated and fired film (100 nm) • Source and drain electrode layers: ITO / Pt sputtered film (10nm / 100nm) • Channel layer: InO sputtered film (20nm) Channel length (LT) / Channel width (WS, WD): 30 μm / 40 μm

[0053] Using the obtained reservoir computing elements, we conducted a nonlinear time series data prediction test using a NARMA model. The NARMA model is a Nonlinear Autoregressive Moving Average Model, and its prediction is based on the following equation (1), which predicts y(k+1) from u(k). y(k+1)=0.4y(k)+0.4y(k)y(k-1)+0.6u(k) 3 +0.1 ···(1)

[0054] In this example, first, 400 time series data points (u(k), k=1 to 400) were randomly generated between 0 and 0.5, and then 400 time series data points (y(k), k=1 to 400) were created using equation (1).

[0055] This u(k) is multiplied by 10 to convert it to a voltage value of 0-5V, and for each k, the converted voltage value is used as the amplitude to assign a square wave with a pulse length of 1ms (data is recorded every 50μs, for a total of 20 points) and four different duty cycles (25%, 50%, 75%, 100%), and this is input as the input pulse Vg(k) to the gate electrode of the reservoir arithmetic element, and the output gate current I gand drain current I d For each k, 400 data points (total 400×20 = 8000 data) were measured. A pulse voltage and current measuring device from National Instruments was used for the measurement.

[0056] For each duty ratio corresponding to each obtained k, 20 points of each I g (k), I d (k), or 80 points of each I obtained by combining data of four different duty ratios g (k), I d (k), or each of those 80 points of I g (k), I d (k) were used as virtual nodes in the final recurrent layer of machine learning, multiplied by weights (weights, 20, 80, and 160 respectively), and the sum of products was calculated to obtain the value of y(k + 1)_L (learning value). Using the y(k) data from k = 1 to 200, optimization (machine learning) of the weights was performed to make it as close as possible to the correct value of y(k + 1). The ridge regression method was used for optimization.

[0057] Using the obtained weight values, for the subsequent 200 points of y(k)_P (predicted value) from k = 200 to 400, for each I g (k), I d [[ID=2 (k), 4I g (k), 4I d (k), 4I g (k)+4I d (k) values, the sum of products was calculated, the error from y(k) was obtained, and the normalized mean square error (NMSE) was calculated. Figure 3 shows representative graphs of y(k), y(k)_L, and y(k)_P. In the graph, for clarity, 100 points of learning and prediction are shown respectively.

[0058] As shown in Figure 3, the values of y(k), y(k)_L (learning), and y(k)_P (prediction) are almost the same. Table 1 shows the NMSE values for each data used in the calculation. As shown in Table 1, for I with a duty ratio of 50% dExcept for the prediction using the 20 virtual nodes in (k), all NMSE values ​​are 0.1 or less, with the minimum value being 0.018. From these findings, I g (k), four I with different duty cycles g (k) or I d (k), and four of each I g (k) and I d When both (k) are used, it was confirmed that highly accurate machine learning and prediction are possible.

[0059] [Table 1]

[0060] Although one embodiment of the present invention has been described above, this embodiment is presented as an example and is not intended to limit the scope of the invention. This embodiment can be carried out in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. This embodiment and its variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of Symbols]

[0061] 10… Machine learning reservoir computing element 11… Circuit board 12…Gateplate layer 13…Gate insulating layer 14…Source electrode layer 15…Drain electrode layer 16…Channel Layer

Claims

1. The device comprises a substrate, a gate electrode layer and a gate insulating layer formed sequentially on one side of the substrate, a source electrode layer and a drain electrode layer formed on one side of the gate insulating layer and spaced apart from each other, and a channel layer formed in contact with the gate insulating layer, the source electrode layer and the drain electrode layer. The gate insulating layer comprises a solid electrolyte and ions. The channel layer includes InO, An electric double layer is formed in the region including the junction between the gate insulating layer and the channel layer. A reservoir computing element for machine learning, which inputs an electrical signal to the gate electrode layer and performs machine learning using at least the gate current signal output from the gate electrode layer.

2. The machine learning reservoir computing element according to claim 1, wherein, in the current-voltage characteristics showing the change in the drain current value between the source electrode layer and the drain electrode layer with respect to a change in the voltage value of the gate electrode layer across 0V, the drain current value when the voltage value of the gate electrode layer is 0V is greater when the voltage decreases than when the voltage increases.

3. The machine learning reservoir computing element according to claim 1 or 2, wherein the solid electrolyte is LaZrO.

4. The reservoir computing element for machine learning according to claim 1 or 2, wherein the ion is a hydrogen ion.

5. A machine learning method using a machine learning reservoir computing element as described in claim 1, A machine learning method that inputs an electrical signal to the gate electrode layer and performs machine learning using both the gate current signal output from the gate electrode layer and the drain current signal output from the drain electrode layer.

Images