Computing element and machine learning system including the same

Abstract

To provide an easy-to-manufacture arithmetic element provided with a signal transmission unit having nonlinear electrical characteristics formed using organic materials, and a machine learning system comprising the same.SOLUTION: In a machine learning system 10 comprising a signal transmission unit 13 electrically connected to an input electrode 11 and an output electrode 12, the signal transmission unit 13 comprises: one or both of a large number of conductive self-doped first organic materials 14 and a large number of conductive second organic materials doped with a dopant; and water. Here, it is preferable that the water is gas and that the signal transmission unit 13 is housed within a sealed space 21.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an arithmetic element formed using a material and a machine learning system including the arithmetic element.

Background Art

[0002] In recent years, research on machine learning systems such as neural network systems has been actively conducted. In an arithmetic element of a machine learning system (for example, a neuron element in a neural network system or a classifier in a classification system), a signal transmission unit that converts an electrical signal input to an input electrode and transmits it to an output electrode can be designed by an electronic circuit or configured using a material (such as an organic material or an inorganic material) (see Patent Documents 1 and 2).

[0003] To mainly design a signal transmission unit using a material, the material is required to have non-linear electrical characteristics in addition to conductivity. This is because a material having non-linear electrical characteristics can serve as a neural element that converts a given value by an activation function and outputs it as a different value in a neural network system, and can perform non-linear regression and classification to obtain a target variable. Further, in a machine learning system having a classifier such as Naive Bayes or a support vector machine, the classifier can perform feature extraction by complex input conversion due to non-linear electrical characteristics and can exhibit high arithmetic performance.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] Regarding this, the signal transmission unit (network structure) described in Patent Document 2 has been confirmed to have non-linear electrical characteristics. However, in Patent Document 2, since the signal transmission unit is formed from two types of specific substances, procurement of the specific substances and the like are necessary and the manufacturing is not easy. The present invention has been made in view of such circumstances, and an object thereof is to provide an easily manufacturable arithmetic element provided with a signal transmission unit having non-linear electrical characteristics formed using an organic material, and a machine learning system including the same.

Means for Solving the Problems

[0006] The arithmetic element according to the first invention meeting the above object is an arithmetic element including an input electrode, an output electrode, and a signal transmission unit that converts an electrical signal input to the input electrode and transmits it to the output electrode, wherein the signal transmission unit includes either one or both of a large number of self-doped type first organic materials having conductivity and a large number of second organic materials having conductivity to which a dopant is added, and water.

[0007] The machine learning system according to the second invention meeting the above object is a machine learning system including an arithmetic element including an input electrode, an output electrode, and a signal transmission unit that converts an electrical signal input to the input electrode and transmits it to the output electrode, and information processing means that performs a process of deriving a solution from the electrical signal output from the output electrode, wherein the signal transmission unit includes either one or both of a large number of self-doped type first organic materials having conductivity and a large number of second organic materials having conductivity to which a dopant is added, and water.

Effects of the Invention

[0008] The arithmetic element according to the first invention and the machine learning system according to the second invention are such that the signal transmission unit includes either one or both of a large number of self-doped type first organic materials having conductivity and a large number of second organic materials having conductivity to which a dopant is added, and water. Therefore, the signal transmission unit can be configured using easily available water, and easy manufacturing is possible. Further, in the signal transmission unit, movement of ions occurs through a transmission path formed by water, and the signal transmission unit comes to have non-linear electrical characteristics.

Brief Description of Drawings

[0009]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Embodiments for Carrying Out the Invention

[0010] Next, with reference to the attached drawings, embodiments of the present invention will be described to facilitate understanding of the present invention. As shown in FIG. 1, a machine learning system 10 according to an embodiment of the present invention includes an arithmetic element 15 having an input electrode 11, an output electrode 12, and a signal transmission unit 13 that converts an electrical signal input to the input electrode 11 and transmits it to the output electrode 12, and information processing means 17 that performs a process of deriving a solution from the electrical signal output from the output electrode 12. The signal transmission unit 13 includes self-doped polyaniline (SPAN) 14, which is an example of a large number of self-doped type first organic materials having conductivity, and water in contact with the self-doped polyaniline 14.

[0011] In the present embodiment, as shown in FIG. 1, the arithmetic element 15 has one input electrode 11, a plurality of output electrodes 12, and a signal transmission unit 13 electrically connected to the input electrode 11 and each output electrode 12 via conductive wires 16. One input electrode 11, a plurality of output electrodes 12, and the signal transmission unit 13 are provided on a plate-shaped Si-based substrate 18. The input electrode 11 is an electrode to which an input signal is given as an electrical signal from the outside.

[0012] The signal transmission unit 13 has an aggregate of self-doped polyaniline 14 (that is, a large number of self-doped polyaniline 14) and water, converts an input signal (electrical signal) given to the input electrode 11 into an electrical signal corresponding to the input signal, and transmits it to each of the plurality of output electrodes 12 via the conductive wire 16. Each output electrode 12 outputs the transmitted electrical signal. The information processing means 17 performs necessary processing using the electrical signals output from each output electrode 12. For example, when the information processing means 17 synthesizes the electrical signals output from each output electrode 12 to derive a waveform signal, in the learning phase, the information processing means 17 determines an optimal weighting for the electrical signals output from each output electrode 12, and in the prediction phase, based on the determined weighting, synthesizes the electrical signals output from each output electrode 12 to derive a waveform signal.

[0013] In principle, different electrical signals are applied to each output electrode 12, but it is also possible that the same electrical signal is applied to different output electrodes 12. In FIG. 1, the enlarged view of the signal transmission unit 13 is drawn with a solid line instead of a dashed line. Further, when one arithmetic element 15 has a plurality (N) of output electrodes 12 as in the present embodiment, compared with the case of forming a machine learning system including N output electrodes by using a large number of arithmetic elements each having one output electrode in parallel, it is possible to make the entire machine learning system more compact and power-saving.

[0014] Self-doped polyaniline 14 is a kind of conductive polymer and has a chain-like one-dimensional structure as shown in FIG. 2. In self-doped polyaniline 14, protons (hydrogen ions) of sulfonic acid groups are desorbed, and oxidation occurs on nitrogen atoms to form a self-doped structure. Therefore, different from general conductive polymers, self-doped polyaniline 14 does not require the addition of a dopant for the expression of conductivity, and electric conduction occurs by the movement of the desorbed protons along the bonds of the six-membered ring.

[0015] The electric conduction in self-doped polyaniline 14 is not only due to the movement of protons along the bonds of the six-membered ring. When water is supplied to self-doped polyaniline 14, a water conduction path 19 is formed in self-doped polyaniline 14, and electric conduction also occurs by the movement of protons from one self-doped polyaniline 14 to another through the conduction path 19. The water supplied to self-doped polyaniline 14 for the formation of the conduction path 19 may be water as a gas (i.e., water vapor) or water as a liquid.

[0016] Here, it has been experimentally verified that the signal transmission unit 13 including a large number of self-doped polyaniline 14 and water has non-linear electrical characteristics, and that the non-linear electrical characteristics become more prominent as the amount of water with respect to self-doped polyaniline 14 increases. Therefore, it is considered that the non-linear electrical characteristics of the signal transmission unit 13 are brought about by the movement of protons through the conduction path 19.

[0017] In this embodiment, the self-doped polyaniline 14 is brought into contact with air containing water vapor so that the conduction path 19 is formed. Specifically, as shown in FIG. 1, a dome-shaped cover body 20 is provided on the substrate 18, and the aggregate of the self-doped polyaniline 14 and water vapor (signal transmission unit 13) are housed in a sealed space 21 covered by the substrate 18 and the cover body 20, so that the amount of moisture applied to the aggregate of the self-doped polyaniline 14 (the humidity of the air in contact with the aggregate of the self-doped polyaniline 14) is not affected by the external environment. Thereby, the reproducibility of the electrical characteristics (computing characteristics) of the signal transmission unit 13 is ensured.

[0018] In manufacturing the computing element 15, as shown in FIGS. 3(A) and (B), first, an etching process is performed on the substrate 18 whose surface of the plate-like object 22 made of Si is oxidized to form the SiO2 layer 23. As shown in FIG. 3(C), a photoresist layer 24 is formed on the entire surface of the substrate 18 (SiO2 layer 23). As shown in FIG. 3(D), the regions such as the signal transmission unit 13 where no metal film is finally provided are covered with the photomask 25 and exposed, and the photoresist layer 24 other than the regions covered with the photomask 25 is removed.

[0019] Next, as shown in FIG. 3(E), after a metal film (Au / Cu film in this embodiment) 26 is formed on the entire surface of the substrate 18 by vapor deposition, the metal film 26 formed on the photoresist layer 24 together with the photoresist layer 24 is removed by lift-off, and as shown in FIG. 3(F), the input electrode 11, the output electrode 12, and the conductive line 16 are formed on the surface of the substrate 18 by the remaining metal film 26. In FIG. 3(F) (the same applies to FIGS. 3(G) and (H)), the description of the conductive line 16 is omitted.

[0020] Then, as shown in FIGS. 3(G) and 3(H), a self-doped polyaniline aqueous solution is dropped onto the planned formation area of the signal transmission part 13 on the surface of the substrate 18 and dried to form the signal transmission part 13 having a large number of self-doped polyaniline 14 on the surface of the substrate 18. Thereafter, the cover body 20 is fixed to the substrate 18 so as to cover the signal transmission part 13 in a predetermined humidity environment, and water vapor is confined in the cover body 20.

[0021] In the present embodiment, the concentration of the self-doped polyaniline aqueous solution dropped onto the surface of the substrate 18 is 1×10 -5 mass % or more and 1×10 -3 mass % or less, and the relative humidity in the sealed space 21 covered by the cover body 20 and the substrate 18 is in the range of 30% or more and 95% or less in a 20°C atmosphere. The higher the concentration of the self-doped polyaniline aqueous solution, the greater the density of the self-doped polyaniline 14 in the signal transmission part 13. Therefore, the density of the self-doped polyaniline 14 in the signal transmission part 13 can be adjusted according to the concentration of the self-doped polyaniline aqueous solution.

[0022] When the humidity in the sealed space 21 increases, the number of conduction paths 19 increases, and the non-linear electrical characteristics of the signal transmission part 13 become prominent. Therefore, the non-linearity of the electrical characteristics of the signal transmission part 13 can be adjusted by the humidity in the sealed space 21. In view of the non-linearity of the electrical characteristics suitable for the machine learning system 10, the relative humidity in the sealed space 21 is preferably 50% or more, more preferably 60% or more, and still more preferably 65% or more in a 20°C atmosphere.

[0023] Here, since the self-doped polyaniline 14 has water solubility, if the humidity in the sealed space 21 becomes too high, the self-doped polyaniline 14 in the sealed space 21 may liquefy, and it is considered that the reproducibility of the electrical characteristics of the signal transmission unit 13 cannot be maintained. Therefore, in the present embodiment, the relative humidity in the sealed space 21 is set to 95% or less in a 20°C atmosphere. Also, from the same viewpoint, the water included in the arithmetic element is preferably in a gaseous state (gaseous state) rather than in a liquid state (liquid state) (however, when the signal transmission unit is composed of a material having no water solubility, liquid water can be adopted). Note that the relative humidity in the sealed space 21 is preferably 90% or less, more preferably 85% or less, in a 20°C atmosphere.

[0024] Also, although the signal transmission unit 13 is composed only of a large number of self-doped polyaniline 14 and water, it is not limited thereto, and it can be composed of water, self-doped polyaniline, and other conductive materials. Examples of the first organic material other than the self-doped polyaniline include self-doped organic materials of polythiophene-based, polypyrrole-based, polyaniline-based, and polyphenylene-based.

[0025] Furthermore, the signal transmission unit may be constituted by a large number of second organic materials having conductivity to which water and a dopant are added. Examples of the second organic material include materials obtained by adding a dopant (for example, halogen, Lewis acid, protonic acid, electrolyte anion) to any one or a combination of polyacetylene, polythiophene, and polypyrrole.

[0026] All of the self-doped polyaniline, polythiophene, polyacetylene, polythiophene, and polypyrrole mentioned here have electron conductivity in addition to ion conductivity, but the signal transmission unit can also be constituted by an ion conductive organic material having no electron conductivity (for example, sodium polystyrene sulfonate, polymer-salt complex) and water. Also, a signal transmission unit having both a large number of self-doped first organic materials and a large number of second organic materials in addition to water may be adopted.

[0027] Next, an experiment conducted to confirm the effects of the present invention will be described. In the first experiment, a 5×10 -5 mass% self-doped polyaniline aqueous solution dropped on a substrate was dried to provide an aggregate of self-doped polyaniline. The aggregate of self-doped polyaniline was placed under vacuum, and then the aggregate of self-doped polyaniline was allowed to contact the indoor air by releasing the vacuum. In this process, the I-V characteristics of the aggregate of self-doped polyaniline or the signal transmission part were measured at each timing. The measurement results are shown in FIGS. 4(A) and (B).

[0028] In FIGS. 4(A) and (B), (1) represents the measured value of the signal transmission part during drying, (2) represents the measured value of the aggregate of self-doped polyaniline placed in a vacuum environment, and (3) represents the measured value of the signal transmission part after about 5 minutes have elapsed since the vacuum was released. Note that in FIG. 4(B), the measurement results of FIG. 4(A) are shown with the magnification of the vertical axis changed. From the measurement results, it was confirmed that non-linear electrical characteristics occur in the signal transmission part when the aggregate of self-doped polyaniline comes into contact with liquid water or water vapor, and non-linear electrical characteristics do not appear in a vacuum state where the aggregate of self-doped polyaniline is substantially not in contact with water.

[0029] In the second experiment, an aggregate of self-doped polyaniline obtained by drying a 5×10 -5 mass% self-doped polyaniline aqueous solution dropped on a substrate was placed in different humidity environments at room temperature (20 °C), and the I-V characteristics of the signal transmission part were measured. The measurement results are shown in FIGS. 5(A), (B), and 6. Note that the values with % in FIGS. 5(A), (B), and 6 indicate the relative humidity. Also, when the measurement result at a relative humidity of 50% is described in FIG. 5(A), since the measurement results at a relative humidity of 50% and a relative humidity of 55% overlap, the measurement result at a relative humidity of 50% is described in FIG. 5(B) instead of FIG. 5(A). In FIG. 6, the measurement results of FIG. 5(B) are shown with the magnification of the vertical axis changed. From the measurement results, non-linear electrical characteristics were confirmed at a relative humidity of 50% or more and 85% or less.

[0030] In the third experiment, the impedance of the signal transmission unit used in the second experiment was measured under different humidity environments at room temperature (20°C). The DC bias voltage was set to 3.0 V, the AC bias voltage was set to 0.3 V, and the impedance was measured while varying the frequency in the range of 0.005 Hz or more and 1 MHz or less. The measurement results are shown in Fig. 7. In Fig. 7, the vertical axis represents the component related to capacitance, the horizontal axis represents the component related to resistance, and the values with % indicate the relative humidity respectively.

[0031] From the measurement results, at a relative humidity of 96%, a straight-line portion representing the term of ion diffusion appeared following a protruding (mountain-shaped) curve portion on the upper side. At the measured value of a relative humidity of 96%, the inflection point where the curve portion changes to the straight-line portion was 0.08 Hz. Also, at the measured values of relative humidities of 37% and 63% respectively, a straight-line portion also seems to appear following the curve portion.

[0032] In the fourth, fifth, and sixth experiments, as shown in Fig. 8, an experimental system was used in which an aggregate of self-doped polyaniline formed by drying a 5×10 -4 mass% self-doped polyaniline aqueous solution was connected to 16 square electrodes by conductive wires on a substrate. One of the 16 electrodes was used as the input electrode, and the remaining 15 were used as output electrodes. Note that the thread-like material seen in the enlarged photograph on the right side in Fig. 8 is self-doped polyaniline.

[0033] In the fourth experiment, at room temperature (20°C), a sine-wave electrical signal was input to the input electrode of the experimental system, and it was measured how well the synthesized wave obtained by weighting and synthesizing the waveforms of the electrical signals output from the 15 output electrodes respectively matched the target waveform of the set synthesized wave. The target waveforms of the synthesized wave were five types: cosine wave, triangular wave, rectangular wave, sawtooth wave, and second harmonic. For each target waveform, a learning process was provided, and the weights of each output electrode were determined so that the actually obtained synthesized wave would approach the target waveform. The measurement was performed when the experimental system was placed under vacuum and when it was placed in an environment with a relative humidity of 65%, and the two were compared.

[0034] Figures 9(A) and (B) show the synthesized wave and the target waveform actually obtained at a relative humidity of 65%. Figures 9(A) and (B) are the measurement results of the target waveforms being a cosine wave and a rectangular wave accuracy respectively, and "accuracy" in the figures means the consistency of the obtained synthesized wave with respect to the target waveform. Figure 10 shows the consistency in the cases of relative humidity 65% and vacuum respectively for various target waveforms. From the experimental results, it was confirmed that for any target waveform, the constancy at a relative humidity of 65% exceeded the consistency in vacuum.

[0035] In the fifth experiment, first, six speakers were made to pronounce each of the ten numbers from 0 to 9 fifty times each, and the voice signals generated from each pronunciation were converted into electrical signals, obtaining a total of 3000 electrical signals. These 3000 electrical signals were divided into 2400 training data (training electrical signals) and 600 prediction data (prediction electrical signals). After training to classify the numbers corresponding to the electrical signals using the training data, the classification of the numbers corresponding to the prediction data was performed and the correct rate was measured.

[0036] The measurement of the correct rate was performed in the following two patterns. Pattern 1 (Example): In an environment of room temperature (20 °C) and relative humidity 65%, the prediction data was given to the input electrode of the experimental system, and the electrical signals output from the output electrode were input into a DAQ (data acquisition) system. Each electrical signal stored in the DAQ was classified into the corresponding number by a classifier. Pattern 2 (Comparative Example): The prediction data was directly input into the DAQ system without passing through the experimental system, and each electrical signal stored in the DAQ was classified into the corresponding number by a classifier. The results of comparing Pattern 1 and Pattern 2 are as shown in Figure 11(A). For all six speakers, the correct rate of Pattern 1 exceeded the correct rate of Pattern 2.

[0037] In the sixth experiment, for 500 electrical signals of one speaker selected from the electrical signals used in the fifth experiment, the classification of the numbers using the experimental system, the DAQ system, and the classifier was performed in the following two patterns. Pattern 3: Classification of numbers was performed by changing the number of output electrodes to be used among the 15 output electrodes of the experimental system under the environment of room temperature (20 °C) and relative humidity of 65%. Pattern 4: Classification of numbers was performed using all 15 output electrodes of the experimental system under the environment of room temperature (20 °C) and vacuum.

[0038] In addition, each measurement of Pattern 3 and Pattern 4 was performed by changing the ratio of training data and prediction data. The experimental results are shown in Fig. 11(B). In Fig. 11(B), (i) shows the result when the number of output electrodes of Pattern 3 is 3, (ii) shows the result when the number of output electrodes of Pattern 3 is 6, (iii) shows the result when the number of output electrodes of Pattern 3 is 9, (iv) shows the result when the number of output electrodes of Pattern 3 is 12, (v) shows the result when the number of output electrodes of Pattern 3 is 15, and (vi) shows the result of Pattern 4, respectively. From the experimental results, in all cases of Pattern 3, the correct answer rate exceeded that of Pattern 4. Also, overall for Pattern 3 and Pattern 4, the correct answer rate was the highest when the ratio of training data was 80% and the ratio of prediction data was 20%.

[0039] As described above, the embodiments of the present invention have been explained. However, the present invention is not limited to the above-described forms, and all changes and the like without departing from the gist are within the scope of application of the present invention. For example, there may be a plurality of input electrodes, and there may be one output electrode. Also, a plurality of arithmetic elements can be provided.

Explanation of Signs

[0040] 10: Machine learning system, 11: Input electrode, 12: Output electrode, 13: Signal transmission unit, 14: Self-doped polyaniline, 15: Arithmetic element, 16: Conductive wire, 17: Information processing means, 18: Substrate, 19: Conductive path, 20: Cover body, 21: Sealed space, 22: Plate-like object, 23: SiO2 layer, 24: Photoresist layer, 25: Photomask, 26: Metal film

Claims

1. In an arithmetic element comprising an input electrode, an output electrode, and a signal transmission unit that converts an electrical signal input to the input electrode and transmits it to the output electrode, the signal transmission unit is characterized in that it includes either one or both of a number of self-doped type first organic materials having conductivity and a number of second organic materials having conductivity to which a dopant is added, and water.

2. The arithmetic element according to Claim 1, wherein the output electrode is plural.

3. The arithmetic element according to Claim 1 or 2, wherein the water is a gas and the signal transmission unit is housed in a sealed space.

4. The arithmetic element according to Claim 3, wherein the relative humidity in the sealed space is 30% or more and 95% or less in a 20°C atmosphere.

5. In a machine learning system having an arithmetic element comprising an input electrode, an output electrode, and a signal transmission unit that converts an electrical signal input to the input electrode and transmits it to the output electrode, and information processing means that performs a process of deriving a solution from the electrical signal output from the output electrode, the signal transmission unit is characterized in that it includes either one or both of a number of self-doped type first organic materials having conductivity and a number of second organic materials having conductivity to which a dopant is added, and water.

6. The machine learning system according to Claim 5, wherein the output electrode is plural.

7. The machine learning system according to Claim 5 or 6, wherein the water is a gas and the signal transmission unit is housed in a sealed space.

8. The machine learning system according to Claim 7, wherein the relative humidity in the sealed space is 30% or more and 95% or less in a 20°C atmosphere.

Images