JPH0511303A - Optical neuronetwork - Google Patents

Abstract

PURPOSE:To decrease wiring treatments and to provide a smallsized device. CONSTITUTION:Light is generated from a light emitting element array 2 and a light emitting element array 4 at the time of learning. The respective units of a matrix-type variable optical mask 1 is irradiated with this light. The respective units of the matrix-type variable optical mask 1 contain photodiodes and capacitors to accumulate the output voltages thereof and the optical properties corresponding to the voltages accumulated in the capacitors are set.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention uses a computer (neurocomputer) that imitates the neural network of the brain of an organism and has an associative function, a pattern recognition function, etc., as an optical neurocomputer that is achieved by using optical technology. It relates to a suitable optical neuro network.

[0002]

2. Description of the Related Art Regarding an optical neurocomputer, for example, JP-A-2-45815 and JP-A-2-20142.
5 and JP-A-2-2972045. These optical neurocomputers basically emit a light-emitting element array, a matrix-shaped variable optical mask that transmits the light emitted from the light-emitting element array, and receive the light emitted from the matrix-shaped variable optical mask. And a light receiving element array. The transmittance of the matrix variable optical mask is adjusted so as to correspond to the synapse weight by controlling the voltage applied from the outside. The light emitted from the light emitting element array is
After being adjusted to a predetermined light amount for each unit by the matrix-shaped variable optical mask, the light is guided to the light receiving element array.
Therefore, a predetermined calculation can be performed by controlling the transmittance of each unit of the matrix variable optical mask to a predetermined value.

[0003]

However, in the conventional device, a voltage is applied from the outside to control the light transmittance of each unit in the matrix variable optical mask.
Since the voltage is controlled from the outside, the wiring amount becomes enormous, which is one of the major causes of hindering the miniaturization of the device.

The present invention has been made in consideration of such a situation, and provides an optical neuro network which can be further miniaturized.

[0005]

In the optical neuro-network according to claim 1, each unit of the variable optical mask comprises:
It is characterized by having a storage unit for storing the irradiation amount of light at the time of learning. As this storage unit, a magneto-optical recording material or a capacitor is used in the embodiment.

In the optical neuronetwork according to a second aspect of the present invention, the storage section detects a light, a photodetector, a switching element that switches in response to the output of the photodetector, and charging / discharging is controlled by the switching element. It is characterized by including a charge storage unit.

In the embodiment, the photodetector is composed of the photodiodes 11 and 12, the switching element is composed of the FET 22, and the charge storage portion is the capacitor 2.
It is composed of four.

The optical neuro-network according to a third aspect is characterized by comprising a learning light irradiation array for irradiating the variable optical mask with light during learning.

In the embodiment, the learning light irradiation array is composed of the light emitting element array 4.

An optical neuro-network according to a fourth aspect is characterized in that the learning light irradiation array and the light receiving array are integrally formed in a shape corresponding to the variable optical mask.

In the embodiment, the light receiving array is the light receiving element array 3 and the learning light irradiation array is the light emitting element array 4.
It is composed by.

In the optical neuronetwork according to a fifth aspect of the present invention, a lens matrix for injecting the light emitted from the recognition light irradiation array so as to correspond to each unit of the variable optical mask, and the variable optical mask is irradiated with light during learning. And a learning light irradiation array formed in a shape corresponding to the variable optical mask.

In the embodiment, the lens matrix is composed of the lens matrix 63, the learning light irradiation array is composed of the light emitting element array 4, and the light receiving array is composed of the light receiving element array 3.

[0014]

In the optical neuro-network according to the first aspect, the variable optical mask is irradiated with light during learning, and the amount of the irradiated light is stored in the storage unit of each unit of the variable optical mask. Therefore, an enormous amount of wiring is not required to set the optical properties of each unit of the variable optical mask, and the device can be downsized.

In the optical neuro-network according to the second aspect of the invention, light is irradiated to cause the switching element to switch so that a predetermined charge is stored in the charge storage section. Therefore, only a power supply line is required for each storage unit, and the wiring can be reduced.

In the optical neuro-network according to the third aspect, at the time of learning, the variable optical mask is irradiated with the light emitted from the learning light irradiation array, and the optical properties of each unit are changed. Therefore, it is possible to reduce the wiring for setting the optical property of each unit, and it is possible to downsize the device.

In the optical neuronetwork according to the fourth aspect, the light receiving array that receives the light that has passed through the variable optical mask and the learning light irradiation array that irradiates the variable optical mask with light during learning correspond to the variable optical mask. It is formed integrally with the shape. Therefore, it is advantageous to make the device IC, and the size can be reduced.

In the optical neuro-network according to the fifth aspect, the light emitted from the recognition light irradiation array is made incident by the lens matrix so as to correspond to each unit of the variable optical mask. This lens matrix is formed in a shape corresponding to the variable optical mask. Further, a light receiving array that receives light that has passed through the variable optical mask at the time of recognition and a learning light irradiation array that irradiates the variable optical mask at the time of learning are integrally formed in a shape corresponding to the variable optical mask. Therefore, it is advantageous for the IC, and the size can be reduced.

[0019]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows the basic configuration of an optical neuro network of the present invention. This optical neuronetwork includes a light emitting element array 2, a matrix-shaped variable optical mask 1 that transmits the light emitted from the light emitting element array 2, and a light receiving element array 3 that receives the light emitted from the matrix variable optical mask 1. , And a light emitting element array 4 for irradiating the variable matrix optical mask 1 with light during learning.

The light emitting element array 2 is vertically divided into _five units 2a _{1 to} 2a ₅ . In the matrix-shaped variable optical mask 1, the units 1w _{11 to} 1w ₅₅ are 5 × 5.
They are arranged in a matrix. The light receiving element array 3 is horizontally divided into light receiving elements 3u _{1 to} 3u ₅ . Further, the light emitting element array 4 also includes the light emitting elements 4a _{1 to} 4a.
a ₅ is arranged in the horizontal direction. The light emitting element array 2, the light receiving element array 3, and the light emitting element array 4 are arranged on opposite sides of the matrix-shaped variable optical mask 1.

The light emitted from the light emitting element 2a ₁ is the uppermost five units 1w of the matrix variable optical mask 1.
It is designed to be incident on _{11 to} 1w ₁₅ . Similarly, the light emitted from the light emitting element 2a ₂ of the second stage is incident on the units 1w _{21 to} 1w ₂₅ of the second stage, and the light emitted from the light emitting element 2a ₃ of the third stage is the unit 1w ₃₁ of the third stage. Through 1w
The light that is incident on ₃₅ and emitted from the light emitting element 2a ₄ on the fourth stage is incident on the units 1w _{4 1 to} 1w ₄₅ on the fourth stage, and the light emitted by the light emitting element 2a ₅ on the fifth stage is the fifth stage. Unit 1w ₅
It is adapted to be respectively incident on the _first to 1 w _55.

On the other hand, in the light receiving element array 3, the leftmost light receiving element 3u ₁ receives light from the leftmost units 1w ₁₁ , 1w ₂₁ , 1w ₃₁ , 1w ₄₁ , 1w _{51 of the} variable optical mask 1. Has been done. Similarly, the second light receiving element 3u ₂ from the left receives light from the second units 1w _{12 to} 1w ₅₂ from the left, and the third light receiving element 3u ₃ from the units 1w _{13 to} 1w _{53 in the} third row. Receive light,
The fourth light receiving element 3u ₄ receives the light from the units 1w _{14 to} 1w ₅₄ in the fourth row, and the fifth light receiving element 3u ₅ receives 5
Light from each of the units 1w _{15 to} 1w _{55 in the} row is received.

In the light emitting element array 4, the leftmost light emitting element 4a ₁ irradiates the units 1w _{11 to} 1w ₅₁ in the leftmost column with light.
Similarly, the second light emitting element 4a ₂ is the unit 1w in the second row.
_{12 to} 1w ₅₂ , the third light emitting element 4a ₃ irradiates the units 1w _{13 to} 1w ₅₃ in the third row, and the fourth light emitting element 4a ₄ irradiates the units 1w _{14 to} 1w ₅₄ in the fourth row. However, the fifth light emitting element 4a ₅ is the unit 1w ₁₅ of the fifth row.
To 1 w ₅₅ are irradiated respectively.

Each unit of the matrix-shaped variable optical mask 1 is constructed as shown in FIG. 2, for example. In this embodiment, each unit corresponds to the output of the photodiode 11 which receives the light incident from the light emitting element array 2, the photodiode 12 which receives the light incident from the light emitting element array 4, and the photodiodes 11 and 12. And a voltage control variable optical mask 14 controlled by the control circuit 13. The voltage-controlled variable optical mask 14 can be made of, for example, liquid crystal.

The control circuit 13 is constructed, for example, as shown in FIG. The outputs of the photodiodes 11 and 12 are supplied to the gate of the FET 22 via the NAND gate 21. One end of the FET 22 has a predetermined voltage source VDD
And the other end is connected to the capacitor 24 via the resistor 23. The other end of the capacitor 24 is grounded. One end of the capacitor 24 is connected to the other end of the voltage controlled variable optical mask 14 whose one end is grounded.
The connection point between the resistor 23 and the FET 22 is connected via the resistor 25 to the other end of the FET 26 whose one end is grounded. The output of the photodiode 11 is the AND gate 28.
Is supplied to one input of the AND gate 27 and, after its logic is inverted, is also supplied to one input of the AND gate 27. The output of the photodiode 12 is supplied to the other input of the AND gate 27 and, after its logic is inverted, is supplied to the other input of the AND gate 28. The outputs of the AND gates 27 and 28 are supplied to the gate of the FET 26 via the OR gate 29.
The AND gates 27 and 28 and the OR gate 29 form an exclusive OR gate.

Next, the operation will be described. In the learning process of setting a predetermined coefficient (weighting) for each unit of the matrix-shaped variable optical mask 1, the light emitting element array 2 and the light emitting element array 4 are driven. For example, in the light emitting element array 2, the light emitting elements 2a ₂ and 2a ₄ emit light, and the light emitting elements 2a ₁ , 2a ₃ and 2a ₅ are turned off. Similarly, in the light emitting element array 4, the light emitting elements 4a ₂ and 4a ₄ emit light, and the other light emitting elements 4a ₁ and 4a 4
a _3, 4a ₅ and turned off, respectively.

By doing so, the units 1w _{21 to} 1w _{25 in the} second row from the top and the units 1w _{41 to} 1w in the fourth row from the top
₄₅ receives light from the left side in FIG. Further, the units 1w _{12 to} 1w _{52 in the} second column from the left and the units 1w _{14 to} 1w _{54 in the} fourth column receive the incident light from the right side in FIG. As a result, the units 1w ₂₂ , 1w ₄₂ , 1w
₂₄ , 1w ₄₄ receives light from both sides, and unit 1
w ₂₁ , 1w ₂₃ , 1w ₂₅ , 1w ₄₁ , 1w ₄₃ , 1w ₄₅ , 1w
₁₂ , 1w ₃₂ , 1w ₅₂ , 1w ₁₄ , 1w ₃₄ , and 1w ₅₄ receive the incident light from only one surface. The other units will not receive light from any direction.

In the unit where light is incident from both surfaces, both photodiodes 11 and 12 are ON.
Then, the output of the NAND gate 21 becomes logic 0 and FET2
2 turns on. As a result, the predetermined voltage charges the capacitor 24 via the FET 22 and the resistor 23. At this time A
Since the exclusive OR gate 30 including the ND gates 27 and 28 and the OR gate 29 outputs a logic 0, the FET 26 is off.

On the other hand, in the unit in which light is incident only from one surface, one of the photodiodes 11 and 12 outputs logic 1. Therefore, at this time,
The output of the NAND gate 21 becomes logic 1 and the FET 22
Is off. Then, since the exclusive OR gate 30 outputs a logic 1, the FET 26 is turned on.
As a result, the charging voltage of the capacitor 24 is discharged through the route of the resistor 23, the resistor 25, and the FET 26.

In the unit in which light is not incident from any surface, the outputs of the photodiodes 11 and 12 are logic 0. Therefore NAND gate 2
The output of 1 becomes logic 1, and the FET 22 turns off. Further, since the output of the exclusive OR gate 30 also becomes a logic 0, the FET 26 is also turned off. As a result, the capacitor 24 is in a state in which neither charging nor discharging is performed, that is, a state in which the already charged voltage is held.

Therefore, by controlling the number of times light is incident on the photodiodes 11 and 12 per unit time, it is possible to make the capacitor 24 hold a predetermined voltage. The voltage control variable optical mask 14 is set to have a light transmittance corresponding to the charging voltage of the capacitor 24.

In this way, the coefficient w _ij in each unit can be set to a predetermined value. Positive reinforcement learning can be performed by irradiating each unit with light from both, and negative reinforcement learning can be performed by irradiating light from only one surface. Here, positive reinforcement learning means weighting the data w _ij (= α), and negative reinforcement learning means weighting the data w _ij (= −α) (here α is a positive coefficient). This Δw _ij can be appropriately adjusted by setting the values of the resistors 23 and 25 to a predetermined value.

In the above embodiments, the incident light amount of light at the time of learning is stored in the capacitor 24. However, the incident light amount is stored by using a fiber metal recording film made of, for example, amorphous as a mask. It is possible. In the case of this embodiment, the molecules inside the metal film undergo spin phase transition, and as a result, the optical characteristics such as the light transmittance change and the characteristics are maintained. Therefore, the light transmittance of the metal film can be set to a predetermined value by appropriately adjusting the relationship between the intensity of light applied to the metal film and the critical temperature of the spin phase transition of the metal. .

After the learning in each unit of the matrix variable optical mask 1 is completed as described above, FIG.
As shown in FIG. 3, light corresponding to input information is generated from the light emitting element array 2, and the light is generated in the matrix variable optical mask 1
The light is received by the light receiving element array 3 through each unit. Thereby, the sum of the values obtained by multiplying the outputs of the light emitting elements 2a _{1 to} 2a ₅ from the respective light receiving elements of the light receiving element array 3 by the coefficient of each column in the matrix variable optical mask 1 can be obtained. That is, the calculation of the following equation is performed.

U _i = Σw _ij a _j

FIG. 5 shows the configuration of the second embodiment of the optical neuro network of the present invention. In the embodiment of FIG. 1, the light emitting element array 2 and the light emitting element array 4 are arranged before and after the matrix-shaped variable optical mask 1, but in the embodiment of FIG. 5, the light emitting element array 4 is the light emitting element. Like the array 2, it is arranged on the left side of the matrix-shaped variable optical mask 1. Then, the light emitting element array 2
The emitted light is made incident on the matrix variable optical mask 1 via the half mirror 41, and the light emitted from the light emitting element array 4 is reflected by the half mirror 41 and made incident on the matrix variable optical mask 1. Has been done. In this embodiment, the photodiode 11 of the photodiodes 11 and 12 shown in FIGS. 2 and 3 can be omitted. Then, the output level of the photodiode 12 may be compared with a predetermined reference level to determine whether light is incident from only one of the light emitting element arrays 2 and 4 or light is incident from both.

FIG. 6 shows the configuration of the third embodiment. In this embodiment, the light emitting element array 2 is formed in a shape corresponding to the matrix variable optical mask 1. That is, the height A of the light emitting element array 2 is the same as the height A of the matrix variable optical mask 1, and the width B of the light emitting element array 2 is the same as the width B of the matrix variable optical mask 1. In addition, the light receiving element array 3 and the light emitting element array 4 are integrated, and these are formed in a shape corresponding to the matrix variable optical mask 1. That is, each of the light receiving elements 3a to 3e and the light emitting element 4a
4 to 4e have the same height A as the matrix variable optical mask, the light receiving elements 3a to 3e and the light emitting elements 4a to 4e are alternately arranged, and the entire width thereof is the matrix variable optical mask 1. The width B is set to be the same as the width B. By doing this, it will be advantageous for IC conversion,
More miniaturization is possible.

FIG. 7 shows still another embodiment.
In this embodiment, the light emitting element array 2 is formed as an N × M (3 × 3 in the embodiment) light emitting matrix 61, and the light emitted from the light emitting matrix is also P × Q (3 in the embodiment). It corresponds to each unit of the lens matrix 63 formed in the matrix of (3). This lens matrix 6
Each unit of 3 corresponds to each unit of the matrix-shaped variable optical mask 1, and light emitted from each unit of the lens matrix 63 is incident on the corresponding unit of the matrix-shaped variable optical mask 1. It is designed to Further, the light receiving element array 3 and the light emitting element array 4 are also integrally formed as a light receiving and emitting matrix 62 so as to correspond to each unit of the matrix variable optical mask 1. In this example,
The upper part of each unit of the light receiving and emitting matrix 62 is a light emitting element, and the lower part is a light receiving element. Even in the case of such a configuration, it is advantageous for making into an IC, and the device can be downsized.

[0039]

As described above, according to the optical neuro-network of the first aspect, each unit of the variable optical mask is provided with the storage unit for storing the irradiation amount of light at the time of learning. It is possible to reduce the size of the device because it requires less. Further, local and parallel learning can be executed, and the learning time can be shortened.

According to the optical neuro-network of the second aspect, the electric charge corresponding to the amount of incident light at the time of learning is accumulated in the electric charge accumulating portion, so that the learning can be executed quickly and surely. .

According to the optical neuro-network of the third aspect, at the time of learning, the variable optical mask is irradiated with light from the learning light irradiation array to change the optical property for each unit. It is possible to execute learning quickly and surely without increasing the size.

According to the optical neuro-network of the fourth aspect, the light receiving array and the learning light irradiation array are formed in a shape corresponding to the variable optical mask.
It is advantageous for downsizing, downsizing and mass production are possible.

According to the optical neuro-network of the fifth aspect, the lens matrix is formed in a shape corresponding to the variable optical mask, and the learning light irradiation array and the light receiving array are integrally corresponded to the variable optical mask. Since it is formed in a shape, it is advantageous for IC production and mass production.

[Brief description of drawings]

FIG. 1 is a perspective view showing a configuration of an embodiment of an optical neuro network of the present invention.

FIG. 2 is a side view showing the configuration of the unit in the embodiment of FIG.

3 is a block diagram showing a configuration example of a control circuit 11 in the embodiment of FIG.

FIG. 4 is a diagram illustrating an operation at the time of recognition in the embodiment of FIG.

FIG. 5 is a perspective view showing a configuration of a second embodiment of the optical neuro network of the present invention.

FIG. 6 is a perspective view showing a configuration of a third embodiment of the optical neuro network of the present invention.

FIG. 7 is a perspective view showing a configuration of a fourth embodiment of the optical neuro network of the present invention.

[Explanation of symbols]

1 Matrix variable optical mask 2 Light emitting element array 3 Light receiving element array 4 Light emitting element array 11,12 Photodiode 13 Control circuit 14 Voltage control variable optical mask 22 FET 24 capacitors 26 FET 30 Exclusive OR Gate

Claims (5)

[Claims] 1. A variable optical mask for applying a predetermined optical change to incident light and emitting the light, a light irradiation array for irradiating the variable optical mask with light, and light passing through the variable optical mask. An optical neuro-network comprising a light-receiving array for receiving light, wherein each unit of the variable optical mask has a storage unit for storing an irradiation amount of light at the time of learning. 2. The storage unit includes a photodetector that detects light, a swinging element that switches in response to the output of the photodetector, and a charge storage unit whose charge / discharge is controlled by the switching element. The optical neuro network according to claim 1, further comprising: 3. Upon learning, the optical properties of each unit are changed corresponding to the incident light, and at the time of recognition, the incident light is given a predetermined optical change and emitted. A variable optical mask, a learning light irradiation array that irradiates the variable optical mask with light during learning, a recognition light irradiation array that irradiates the variable optical mask with light during recognition, and a recognition light irradiation array that passes through the variable optical mask during recognition An optical neuro network comprising a light receiving array for receiving light. 4. When learning, the optical properties of each unit are changed corresponding to the incident light, and when recognized, the incident light is given a predetermined optical change for each unit and emitted. A variable optical mask; a recognition light irradiation array formed in a shape corresponding to the variable optical mask for irradiating the variable optical mask with light during recognition; and a light receiving device for receiving light passing through the variable optical mask during recognition. An optical neuronetwork, comprising: an array; and a learning light irradiation array that irradiates the variable optical mask with light during learning and is formed in a shape corresponding to the variable optical mask together with the light receiving array. 5. When learning, the optical properties of each unit are changed in response to the incident light, and at the time of recognition, the incident light is given a predetermined optical change and emitted. A variable optical mask, a recognition light irradiation array for irradiating the variable optical mask with light at the time of recognition, and light emitted from the recognition light irradiation array is incident so as to correspond to each unit of the variable optical mask. A lens matrix formed in a shape corresponding to the variable optical mask; a light receiving array that receives light that has passed through the variable optical mask during recognition; and a light receiving array that irradiates the variable optical mask with light during learning And a learning light irradiation array formed in a shape corresponding to the variable optical mask.