JPH0253122A - Signal processor - Google Patents

Abstract

PURPOSE:To secure a more desirable weight coefficient through the learning by forming a unit with the light transmissivity variable elements which vary their transmissivity according to the levels of the optical input signal received from a front stage and the learning control signal sent in the adverse direction from a rear stage. CONSTITUTION:The optical buses containing the transparent glasses D are used as the transmission lines set among the units 1 of neural network. The optical input received from an input layer is transmitted to an output layer of a rear stage from a front stage as shown in the broken lines. At the same time, the optical input of the rear stage received from an output layer is adversely sent to the input layer of the front stage as the learning signals as shown in the solid lines. In this case, the weight coefficient value can be changed more properly at the connecting parts of the unit 1 since a light transmissivity variable element 2 uses a variable light transmission cell whose transmissivity is changed by the optical input signal and the learning control signal.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a signal processing device, and particularly to a signal processing device for causing each unit of a neuron network to perform learning.

[Conventional technology]

Functional circuits using neural networks (neural networks) have been developed. For example, “Nikkei Electronics, April 18, 1988 (No. 445)
” pp. 106-107 describes a pattern identification device using a neural network. This network has a three-layer structure, including 4 outgoing units, 1 to 8 intermediate units, and 64 incoming units.
It consists of ~256 pieces. A handwritten kanji of 64 dots x 64 dots is divided into 8 x 8 meshes, and the density of each of these areas is calculated and input to the input unit of the neural network. As a learning algorithm, pack propagation is used and the number of intermediate units is increased to improve the learning effect of exceptional patterns.

FIG. 2 is a block diagram of a general three-layer neural network, and FIG. 3 is a schematic internal block diagram of each unit configuring the network in FIG.

As is clear from FIG. 2, for example, one intermediate layer unit 1b receives input signals from a plurality of input layer unit ICs, and transfers the outputs thereof to a plurality of output layer units 1a.

Now, if the expected output value is 1.0 for the input cover pattern p of the network, and if the output Opj is 0.4, then in each output layer, middle layer, and input layer,
It is hoped that education will be provided so that the remaining 0.6 is added. In other words, in each unit, it is necessary to calculate the output from multiple units in the lower layer and send it to multiple units in the upper layer, so that the sum of the units in the output layer becomes 1.0. It requires similar nerves and movements.

As shown in FIG. 3, each unit is comprised of a section 21 that receives output from other units, a section 22 that converts the input according to certain rules, and a section 23 that outputs the result. Note that a variable weighting coefficient (W
i J ) is provided.

When a certain unit i receives input from a plurality of units, the sum of the inputs and weights becomes the input value. That is, the input value IN is determined by a linear equation.

lN1=ΣW engineering, ・01 ・・・・・・・・・・ (1
) Here, OJ is the output of the other unit j, and WIJ is the weighting coefficient due to the combination of the other unit j and the unit i.

Further, the output Oi of this unit i is expressed by applying the input value IN to the function f, for example, by the following equation.

0 engineering = f (I Ni) = f (ΣWiJ・OJ
)... (2) This function f includes, for example, a step function, a broken line function, a nonlinear saturated quantity, etc.

When input data (input cover pattern) is given to each unit of the input layer, this data is converted in each unit, transmitted to the intermediate layer, and finally output from the output layer. Compare the actual output turn with the desired output turn (expected value),
In order to reduce the difference, the weighting coefficient WiJ of the coupling section 24 of each unit is changed.

[Problem to be solved by the invention]

In this way, in a neural network, the operation of changing the weighting coefficient value of the connection part of these units is
It is necessary to make the neural network work in a manner similar to that of living organisms, but there is a problem in that the neural network does not work well because there are no suitable objects in the natural world.

An object of the present invention is to solve such conventional problems and provide a unit configuration that can give desirable weighting coefficient characteristics through learning, and a wiring structure that includes a variable resistance element that can be used as a weight coupling section. There is a particular thing.

[Means to solve the problem]

In order to achieve the above object, the signal processing device of the present invention is a neural network configured by interconnecting units in a plurality of stages, in which each unit receives an optical input signal sent from the previous stage and an optical input signal sent from the subsequent stage. A variable light transmittance element whose transmittance changes depending on the magnitude of a learning control signal sent in the opposite direction, and a photoelectric conversion element that converts the input signal into an electrical signal.

It is composed of an arithmetic integrated circuit that performs arithmetic operations on the electrical signal, and a light emitting element that converts the electrical signal into an output optical signal,
The feature is that optical signals are used for interconnection between each unit. Each of the 5-nit variable light transmittance elements has a photodiode that is sensitive to the optical input signal from the front unit, a photodiode that is sensitive to the learning control signal from the rear unit, a metal, and a transparent metal of the metal. It consists of a salt and a transparent container, and is also characterized in that the larger the optical input signal and learning control signal, the more the metal is dissolved in the metal salt solution, and the easier it is for light to pass through. be.

[For production]

In the present invention, the transmission path between each unit of the neural network is an optical bus through transparent glass, and at the same time, optical input from the input layer is transmitted from the previous stage to the subsequent output layer, and at the same time, it is transmitted as a learning control signal to the output layer. The optical input from the subsequent stage is conversely transmitted to the input layer at the previous stage. In addition, each unit is equipped with a circuit configuration that changes the control output value to the previous stage as a function of all the control input light from the subsequent stage that is input to this unit and the output light output from this unit, and each circuit is In the coupling section that combines the light and provides a weighting coefficient, a transmission cell structure is provided in which the transmitted light increases in proportion to the intensity of the light from the previous input layer and the control output light from this unit. With the units having such a configuration and the transmission cells provided between these units, it is possible to provide the neural network with an output value close to the expected value.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 4 is a diagram of a wiring structure between each unit of a neural network showing an embodiment of the present invention.

In FIG. 4, 1c is an input layer unit, lb
are intermediate layer units, and all signal transmission between them uses an optical bus. Here, the optical bus refers to light itself, and does not use fibers. Each unit 1 consists of an optical LSI that forms a three-dimensional multilayer structure, and an optical mask (L
light mask), that is, the variable optical transmission cell (V
riable light transmis
sion Ce1l (WlJ) (This is the joint part 2
4) 2 and 1 weight accumulation circuit (Weightea
S u+mC1rcut), the A layer (which is
(corresponds to the input section 21 of the unit), the B layer (which corresponds to the conversion section 22 of the unit) which is a threshold, and the 0 layer (which corresponds to the light emitting element/driver) (LED/DRIVER). this is,
(corresponding to the output section 23 of the unit) and a glass substrate. Although not shown in FIG. 4, a photodiode that receives light (control light) arriving from the opposite direction is also provided in the 0 layer, and an LED that sends out output light in the opposite direction is also provided in the A layer. (described later in Figure 1)
.

Between each unit 1 and variable optical transmission cell 2 and unit 1
There is transparent glass between them, through which light is transmitted. That is, the light from the LED in the input layer 1c passes through the hole, or aperture, of the variable light transmission cell 2 through the transparent glass, and passes through the glass substrate (
G 1ass Substrata) After D,
The light is received by a photodiode in the input section and converted into an electrical signal. In unit 1b, the received signal undergoes AD conversion processing etc. in the A layer, input and output calculation according to the present invention is performed in the B layer, and output signal creation processing and DA conversion processing are performed in the 0 layer. We are the next stage output layer unit 1a from the LED.
The light is emitted towards. The wiring structure between the intermediate layer unit 1b and the output layer unit 1a is also the same as that in FIG. 4.

FIG. 1 is a block diagram of a neuron LSI showing an embodiment of the present invention.

In Figure 1, the input light in the forward direction (
(broken line) and input control light in the opposite direction (solid line) sent from the rear stage to the front stage.

In FIG. 1, 2 is a variable optical transmission cell, 1 is a unit,
D is a glass substrate, 11 is a photodiode, 12 is an amplifier, 13 is a threshold holder, 14 is a driver, 15
is an LED, and 16 is a multiplier.

In this embodiment, the variable optical transmission cell 2 trains the resistance of the optical path to be lowered, and the unit 1 trains the control light to depend on the forward direction output light and the reverse direction control input light.

The structure and function of the variable optical transmission cell 2 will be described later with reference to FIGS. 5 and 6.

In the neuron LSI (unit) 1, a control output according to the following equation is sent from the LED 15 to the previous unit.

AJ=XJ* (1-XJ) Umbrella (ΣWJk・δk)...
...(3) Here, AJ is the reverse direction control light output value of unit j, XJ is the forward direction output light from this unit j, Σ
WJk·δ is the sum of the values obtained by multiplying the backward light output values from all subsequent units by the weighting coefficient WJk.

That is, in FIG. 1, the hole (
The input light from the previous unit that has passed through the apertures (1 to 6) is multiplied by a weighting coefficient Wij in cell 2 and input to this unit j. When the light is received by the photodiode 11, it passes through the amplifier 12, and the thresholder 13 inputs (I
D15 converts it into light and sends it to the next stage unit.

On the other hand, in the same way, from the latter unit to the variable optical transmission cell 2
The light ΣWjk·δ, which is multiplied by a weighting coefficient and input through the glass substrate, is received by the photodiode 11, passes through the amplifier 12, and is input to the multiplier 16, where xJ and (I is multiplied. The multiplication result passes through the amplifier 14, is converted into an optical signal by the LED 15, and is sent to the previous unit.

Therefore, the control output from the neurons LS, I is an optical signal having the value of the above equation (3).

FIG. 5 is a cross-sectional structural diagram of a variable optical transmission cell showing one embodiment of the present invention.

The light with the downward arrow indicates the forward direction from the previous stage (ForwardLig).
ht), and the light indicated by the upward arrow is the control light input in the reverse direction (Back'vard Light) from the rear stage. A photodiode that is sensitive only to light input in the forward direction and a photodiode that is sensitive to light input in the reverse direction are connected in series, and as shown in the equivalent circuit in Figure 6, a resistor, a copper electrode, a copper sulfate solution, and a tin oxide and a variable transmittance electrode constituted by a wire. The photodiode becomes conductive in response to forward light input and reverse learning control light, and the copper of the variable transmittance electrode dissolves in the copper sulfate solution, making it easier for light to pass through.

FIG. 6 is an equivalent circuit diagram of the variable optical transmission cell of FIG. 5.

〔Effect of the invention〕

As explained above, according to the present invention, each unit of the neuron network sends out a control output as a function of the control light output from the subsequent stage and the forward direction light input from the previous stage, and the control output and the input light Since the light transmittance of the variable optical transmission cell is changed according to the change in the optical transmittance, each unit and the coupling section can be provided with a learning function, and it is possible to provide an output close to the expected value as the output of the network.

[Brief explanation of the drawing]

Figure 1 is a diagram of the internal structure of two units in a neural network showing an embodiment of the present invention, Figure 2 is a diagram of the configuration of a general neural network, and Figure 3 is the internal configuration of each unit in Figure 2. 4 is an overall configuration diagram showing the wiring configuration between units of the present invention, FIG. 5 is a cross-sectional structural diagram of the variable optical transmission cell in FIG. 4, and FIG. 6 is a diagram of the variable optical transmission cell in FIG. 5. FIG. 1: Unit, 2: Variable optical temporary feeding cell, 11: Photodiode, 12: Amplifier, 13: Threshold holder, 1
4: Driver, 15: LED, 16: Multiplier (
multiplier), A, B, C: first, second, and third layers, D glass substrate, 21: input section within the unit, 22: conversion section within the unit, 23: output section within the unit, 24: coupling section. Patent applicant Rico Co., Ltd.

Claims (2)

[Claims] (1) In a neuron network configured by interconnecting units in multiple stages, each unit has a transmittance that is determined by the magnitude of the optical input signal sent from the previous stage and the learning control signal sent in the opposite direction from the subsequent stage. A variable light transmittance element that changes, a photoelectric conversion element that converts the input signal into an electrical signal, an arithmetic integrated circuit that performs an operation on the electrical signal, and a light emitting element that converts the electrical signal into an output optical signal. What is claimed is: 1. A signal processing device comprising: and using optical signals for interconnection between each unit. (2) The variable light transmittance element of each unit includes a photodiode that is sensitive to the optical input signal from the previous unit;
It is composed of a photodiode that is sensitive to a learning control signal from a subsequent stage, a metal, a transparent metal salt of the metal, and a transparent container. The signal processing device according to claim 1, wherein the signal processing device is configured to be dissolved in a solution and easily transmit light.