JPS6257266A - Neural network element - Google Patents

Abstract

PURPOSE:To obtain the neural network element of a simple structure capable of changing input-output operation pattern and of performing integration, by forming many junction parts to interconnect a conductive line pattern arranged in matrix with tunnel junction. CONSTITUTION:The junction part at crossing points of conductive line patterns 11 and 12 arranged in matrix is formed with a tunnel junction. For example, conductive line pattern 11 of aluminum is formed on insulative substrate 10, and its surface is oxidized by about 50Angstrom thickness to form thin film insulator 1 of alumina. On this film a conductive line pattern12 of gold, silver, etc. is formed so as to cross the line pattern 11. When this element is irradiated with a light 13, the tunnel current flows and a stable output signal pattern can be obtained. Thus, the input-output operation pattern of the element can be arbitrarily selected by changing freely the light irradiation to each junction position, and the element of simple structure suitable for integration can be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to neural network elements that perform input/output operations similar to those of the nervous system.

(Prior Art) It is generally believed that the operation of the nervous system refers to previously read data for several input patterns and provides a stable output pattern through some kind of feedback mechanism.

As a model of such a nervous system, for example, Proceedings of the National Academy of Sciences
(Proceedings of Natlonal
Academy Science ) USA Volume 79 2
Hopfield's model is known, as described on page 554 (1982).

Recently, based on such a model, an attempt to artificially form a neural network element using a wire matrix with a resistive pattern was presented in the Proceedings on 29th International Symposium.
On Electron, Ion and Photon Beams (Proceedingsof 29th In
ternational SyrrIposiumon
Electron.

Ion and Photon Bear's) A
-2 volume (1985). FIG. 3 shows the element, in which (a) is a plan view and (b) is a cross-sectional view.

As shown in FIG.
V, ,V2. ...■ゎ), the output line terminals P (Pl, P2 .
...Pt, l) and outputs a signal by passing a current through it. Here, a resistance pattern 21 consisting of a junction part 21id and amorphous silicon and a contact layer is shown.
A, and an insulating layer 22 is provided between the layers, thereby forming a capacitor. When a certain input signal pattern is input to the neural network element configured in this way, a certain output signal corresponding to the arrangement of the junction parts 21 on the matrix is obtained. It can be used as an element to find stable light for nonlinear problems.

(Problems to be Solved by the Invention) However, in the case of the neural network element shown in FIG. 3, the output operation pattern of the element is fixed by the arrangement of the junk 717 section 21, which is unsatisfactory as a neural network element. .

As a way to improve these points, random access memory IJ rather than read-only memory ('fLOM) is recommended.
It is conceivable to fabricate semiconductor elements such as - (RAM) on a substrate, but if we try to form a neural network, there is currently a risk that it will become a huge force system.

SUMMARY OF THE INVENTION An object of the present invention is to provide a neural network element that is structurally simple and can be integrated, and whose input/output operation patterns can be changed.

(Means for Solving the Problems) The neural network element of the present invention connects conductive line patterns arranged in a matrix at multiple junctions to connect a neural network that performs input/output operations similar to those of the nervous system. This is a network element in which the junction portion is formed by using a tunnel junction.

(Function) In a neural network element using a tunnel jumper, electrons in a conductive line pattern become a tunnel current due to the strong electric field generated at the jumper 52 and flow into the opposing conductive line pattern. At this time, Physical Review, Letters (Physical R
evlew Letters) Volume 37, Page 923 (1
976), when the junction portion is irradiated with light, the tunnel current changes. In fact, metal-oxide-metal junction elements are used as photodetecting elements by utilizing such current changes.

By irradiating some junctions of the neural network element with light to cause a tunneling current to flow from one side of the conductive line pattern to the other, the output signal relative to the input signal can be controlled. When there are N input lines, the number of input signal patterns is 2N even if the input signal is binary, which becomes a huge number when N is large, but it is possible to change the amount of light irradiation and the irradiation position. A small number of stable output signal patterns can be obtained for various input signals. That is, it can be applied as a memory device by classifying and simplifying a huge number of signals. It also operates as a memory circuit for the light irradiation pattern itself and as a pattern recognition circuit element.

(Example) Next, embodiments of the present invention will be described using the drawings.

FIGS. 1(a) and 1(b) are a plan view and a sectional view of a first embodiment of the present invention.

In FIGS. 1A and 1B, the junction portions at the intersections of the conductive line patterns 11 and 12 arranged in a matrix are formed by tunnel junctions.

This example was created by forming, for example, an aluminum conductive line pattern 11 on an insulating substrate 10, and covering the surface with approximately 50%
It can be easily formed by a process in which a thin film insulator 1 made of alumina is formed by oxidizing it by about λ, and then conductive line patterns 12 made of gold, silver, etc. are provided in an intersecting manner. A submicron pattern could be formed as a conductive line pattern. It was confirmed that when this element was irradiated with light, a tunnel current flowed and a stable output signal pattern was obtained.

6-FIGS. 2(a) and 2(b) are a plan view and a sectional view showing a second embodiment of the present invention. The feature of this embodiment is that a metal electrode pattern 2 in the shape of a cone or pyramid with a pointed tip is formed at the junction part, and a thin film insulator 1 is placed on top of the metal electrode pattern 2, and a conductive line pattern 12 made of gold, silver, etc. is formed on top of the metal electrode pattern 2. It was formed.

First, a conductive line pattern 11 is formed on an insulating substrate 10 using an ion implantation method or a lithography method. A metal electrode pattern 2 made of tungsten, aluminum, or the like is formed on this conductive monolithic line pattern 11 by a seven-off method. At this time, the metal electrode pattern 2 was close to a cone shape, but in order to further sharpen the tip, IJ was performed using argon ions while rotating the sample to obtain a cone or pyramid shape acute angle pattern. Here, tungsten was used as the metal, and many submicron tungsten cone patterns were formed. Then add argonium to 20-1
The thin film insulator 1 was deposited to a thickness of 0.00 mm and oxidized to form a thin film insulator 1. Furthermore, a conductive rough-in pattern of gold, silver, aluminum or the like having a thickness of 200 to 400 layers was formed to intersect with the conductive line pattern 11 in the lower layer. Thereafter, contact holes are made and wire bonding is performed to complete the element formation process.

An input voltage was applied between the electrodes of the neural network device thus fabricated, and it was found that the S/N ratio of the tunnel current caused by light irradiation to the dark current was better than that of the example shown in FIG. In addition, results were obtained that responded to shorter wavelength light.

(Effects of the Invention) As explained above, the tunnel junction neural network device of the present invention allows the input/output operation pattern of the device to be freely selected by freely changing the light irradiation to each junction position. Because it is structurally simple and suitable for integration, it is highly effective in developing elements with more advanced functions such as a type of associative memory that mimics the nervous system and pattern recognition.

[Brief explanation of the drawing]

FIGS. 1(a) and (b) are a plan view and a sectional view of a first embodiment of the present invention, and FIGS. 2(a) and (b) are a plan view and a sectional view of a second embodiment of the present invention. 3(a) and 3(b) are a plan view and a sectional view of an example of a conventional neural network element. In the figure, 1... Thin film insulator, 2...
Metal electrode pattern, 10...Substrate, 11.12
......Conductive line pattern, 13...
Light, 14...Current amplifier, 21...Junction section, 21A...Resistance pattern, 2
2...Insulating layer. $l concave

Claims (1)

[Claims] A neural network element that performs input/output operations similar to a nervous system by connecting conductive line patterns arranged in a matrix at a plurality of junction parts, characterized in that the junction parts are formed by tunnel junctions. neural network elements.