JPH06131484A - Optical information processing circuit - Google Patents

Abstract

PURPOSE:To eliminate the need for many connections for a local coupling type, to reduce the size of a processing system, and to transmit a minus-number signal. CONSTITUTION:Constituent units of an optical neural network as an optical information processing circuit consisting of LCLVs 21 as spatial optical modulating elements and weighting means 25 which weight information light are coupled. Input information light I0 is inputted from the side W of an LCLV21, which is irradiated with read/write light H1 from the side R to read the information light I0 out as information light I1; and the light is weighted by the weighting mask W1 of a weighting means 25 and the information light I1' weighted by the LCLV 22 is inputted through a lens L1. Similarly, this process is repeated to obtain a finally processed output O1.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information processing circuit, and more particularly to an optical information processing circuit using light as a medium for carrying information.

[0002]

2. Description of the Related Art Since the invention of the laser in the 1960s, optical information processing which realizes a large amount of information processing by utilizing parallel propagating property and high speed of light, for example, image emphasis and image matching has been actively researched. It was

On the other hand, in recent years, research on neurocomputers simulating excellent information processing functions of living things has been actively conducted. This information processing circuit performs massively parallel information processing by paying attention to the connection state between a huge number of elements called neurons. Neurocomputer is a dedicated LSI
It can be realized as hardware by using the or optical element.

The neurocomputer utilizing the optical information processing technology (1) has essentially good compatibility with the neurocomputer because the light waves have spatial parallelism.

(2) Mutual wiring between a large number of neurons is easy. Moreover, if a real-time hologram or a spatial modulation element is used, it is possible to make a variable wiring between a huge number of elements.

(3) Moreover, the light waves propagate without being affected by crosstalk, and the transmission capacity is large.

It has the following features. The goal of optical neurocomputers is to realize intelligent computers that handle problems that current computers are not good at, such as associative processing and pattern recognition, at extremely high speed and in parallel.

As a basic model of an optical neurocomputer using such optical information processing technology,
As shown in FIG. 19, a light emitting element array 90 such as an LED (light emitting diode) or an LD (laser diode) and an optical mask
A model including 91 and a light receiving element array 92 has been proposed. The vector information V input to the light emitting element array 90 is
Light is emitted from the light emitting element array 90 as light intensity. The output from each light emitting element (for example, Vj (j = 1,2, ..., N)) is wavefront converted into a fan-shaped beam by a lens system (not shown), and only the j-th column component of the optical mask 91 corresponding to the matrix T is converted. Irradiate evenly. If the magnitude of the (i, j) component Tij of T is given as the light transmittance, its output intensity is Tij × Vj
Will be proportional to. Next, in the output light from the optical mask 91, all the i-row components are focused on one of the light receiving element arrays 92 by the lens system. Therefore, the output Ui of the i-th light receiving element array is

[0009]

[Equation 1]

Then, the matrix-vector product is obtained at the output of the light-receiving element.

By using the optical method in this model, operations can be executed in parallel at extremely high speed.
For example, assuming that a matrix having 100 × 100 = 10 ⁴ components is used and the light emitting element is driven at 100 MHz, this corresponds to 10 ² calculations per second.

A model using a hologram has been proposed as another model. A hologram is one in which a refractive index is periodically changed by exposing and then fixing a photosensitive agent by utilizing an interference effect between two coherent light waves. Therefore, by forming a plurality of diffraction gratings in one hologram, a model as shown in FIG. 20 can be formed. In this model, the input surface
A light beam emitted from any one of 93 points (corresponding to a neuron) can be routed to any of 95 points on the output surface. Since the spatial resolution of the hologram 94 is about 2000 lines / mm,
In principle, a 2 cm ² hologram element will have 10 ⁹ or more independent diffraction gratings. This model utilizing hologram, which means that it has the possibility of freely connect 10 ⁴ or more inputs and 10 ^four or more output points.

On the other hand, in addition to the optical wiring element, many optical threshold elements have been proposed. For example, an element called MQW (Multiple Quantum Well) layer, in which a very thin GaAs and AlGaAs layers of about 100 angstroms are alternately laminated, which utilizes a strong nonlinear optical effect of a medium, has been proposed. Here, the non-linear optical effect refers to a phenomenon in which the refractive index changes according to the intensity of input light. An element can be formed by sandwiching this MQW layer with two partial mirrors.

Various attempts have been made with respect to an optical information processing circuit using the optical information processing technology using the above-mentioned model and the optical threshold device (Ota, Kuma: Optical Neurocomputer, Journal of Television Society Vol.42). ， No9 (1988) ， pp
931 ~ 936, Takeda: Optical Neurocomputing, Information Processing Vol.29, No.9 (1988), pp984 ~ 992, etc.).

[0015]

In the information processing circuit in the above-described model, the connections between neurons are all connections in which one neuron and all neurons in the next layer are connected. The number of lines becomes enormous. For example, for a 1000 × 1000 pixel image, as many as 10 ¹² connecting lines are needed. This is not a value that can be easily realized even if a large capacity coupling method such as a volume hologram is adopted in the model using the light emitting element or the model using the hologram described above.

In the information processing circuit, the weights of the connections between the neurons are all different, and in order to realize this optically, it is necessary to prepare separate connection weights for each neuron. For this reason, there is a problem that the size of the optical mask, hologram, etc. in the above model is limited due to the diffraction phenomenon, and the processing system using this information processing circuit becomes large.

Furthermore, in a general optical information processing circuit, not only so-called positive number signals but also negative number signals are transmitted. The model described above was able to transmit only a positive number of signals, and the information that could be processed was very limited.

In view of the above circumstances, the present invention does not require a large number of coupling lines, downsizes the processing system, and is capable of transmitting a negative number of signals.
An object of the present invention is to provide an optical information processing circuit and an information transmission method of the optical information processing circuit.

[0019]

A structural unit of a first optical information processing circuit according to the present invention is a spatial light that modulates and outputs information light (light having information) in accordance with input information light. It is characterized by comprising a modulator and a weighting means for weighting the information light modulated and output by the spatial light modulator.

Here, the spatial light modulator (Spatial Light M
odulator (SLM) means that when one-dimensional or two-dimensional information (for example, an image) is input by an optical or electrical method, optical characteristics such as complex light transmittance or reflectance change in real time, and this element It is an element in which another light uniformly illuminated on the surface is modulated according to the input information (Matata: Optronics (1988) No.7, pp124-130). As the spatial light modulator, there are various types such as an electro-optical spatial light modulator and a liquid crystal spatial modulator. The electro-optic spatial light modulator modulates light with an electro-optic crystal, and the generated image is a polarization plane rotation image or a scattering image. The electro-optic spatial light modulator includes DKDP, PLZT, BSO, microchannel plate, LCLV (Liquid Crystal Light V).
alve) and the like.

As the liquid crystal spatial modulation element, there are various types such as an optical writing type, a thermal writing type, an electric writing type and a CCD writing type. By using such a spatial light modulator, two-dimensional information processing can be performed in real time.

The second optical information processing circuit according to the present invention is the same as the first optical information processing circuit according to the present invention, in which the spatial light modulator is the information input from one surface. It is a spatial light modulator that modulates read-out light emitted from another surface in accordance with light.

The third structural unit of the optical information processing circuit according to the present invention is the same as the first structural unit of the optical information processing circuit according to the present invention, in which the spatial light modulation element is responsive to input information light. , A spatial light modulator that outputs modulated emission light.

Further, the constituent unit of the fourth optical information processing circuit according to the present invention is the first, second or third constituent unit according to the present invention.
In the configuration unit of the optical information processing circuit, the weighting means has a switching means for switching the weight.

Further, the fifth optical information processing circuit according to the present invention is the first, second, third or fourth optical information processing circuit according to the present invention, wherein the weighting output means is It is characterized in that it is a weighting means for weighting different plural lights respectively.

Furthermore, the first optical information processing circuit according to the present invention comprises the above-mentioned constituent units of the optical information processing circuit,
The plurality of constituent units are combined so that the information light output from one constituent unit among the plurality of constituent units is sequentially input to one or more other constituent units. is there.

The sixth structural unit of the optical information processing circuit according to the present invention includes a first spatial light modulator which modulates and outputs the information light in accordance with the input information light. Weighted output means in the structural unit of the information processing circuit, a second spatial light modulator that modulates and outputs the information light according to the input information light, and the second spatial light modulator outputs the information light. And an information light re-input means for re-inputting the information light into the first spatial light modulator.

Further, a second optical information processing circuit according to the present invention comprises at least one constituent unit of the first, second or third optical information processing circuit according to the present invention and at least one sixth optical information processing circuit according to the present invention. And a constituent unit of an optical information processing circuit, and these constituent units are combined so that information light output from one constituent unit of these constituent units is sequentially input to one or more other constituent units. It is characterized by being done.

Further, the third optical information processing circuit according to the present invention comprises a plurality of constituent units of the fourth optical information processing circuit according to the present invention, the plurality of constituent units being among the plurality of constituent units. It is characterized in that the information lights output from one structural unit are sequentially coupled to be input to another structural unit.

Furthermore, the information transmission method of the first optical information processing circuit according to the present invention uses the third optical information processing circuit according to the present invention, and a predetermined information light is supplied to the third optical information processing circuit according to the present invention. The information light is input to the spatial light modulator in the processing circuit, the input information light is output as the first information light from the spatial light modulator, and the output first information light is weighted by the weighting means. And outputs the first weighted information light as the first weighted information light, inputs the output first weighted information light to the other spatial light modulator, and thereafter switches the weight of the weighting means, and outputs the other weighted information light. The input characteristic of the spatial light modulator is changed as necessary, the outputted first information light is weighted by the weighting means whose weight is switched, and is outputted as the second weighted information light. Output And it is characterized in that the second weighted information beam the input characteristics is inputted to the other spatial light modulator that has changed as needed.

Further, in the fourth optical information processing circuit according to the present invention, the information light output from two units of the constituent units of the first, second or third optical information processing circuit according to the present invention is another It is characterized in that the constituent units are combined so as to be input to the constituent unit of one unit.

Further, a second optical information processing circuit information transmission method according to the present invention uses the fourth optical information processing circuit according to the present invention. Of the constituent units of the optical information processing circuit of one unit, the predetermined information light is input to the first constituent unit, the information light is weighted and output as the first weighted information light, The weighted information light of the other 1
Input to the constituent unit of one unit and input to the spatial light modulator in the constituent unit, and after the input, change the input characteristics of the spatial light modulator in the constituent unit of the other unit as necessary. Of the constituent units of the optical information processing circuit of the two units, the predetermined information light is input to a second constituent unit, the information light is weighted and output as a second weighted information light, It is characterized in that the second weighted information light is inputted to the spatial light modulator whose input characteristic is changed as necessary.

Further, a fifth optical information processing circuit according to the present invention comprises the constituent units of the fifth optical information processing circuit according to the present invention, wherein the plurality of constituent units are one of the plurality of constituent units. It is characterized in that the information lights outputted from the constituent units are sequentially coupled so as to be inputted to another constituent unit.

Further, a third optical information processing circuit information transmitting method according to the present invention uses the fifth optical information processing circuit according to the present invention, and a predetermined information light is supplied to the fifth optical information processing circuit according to the present invention. A first weighting is performed by inputting to the spatial light modulator in the processing circuit, outputting the information light modulated by the spatial light modulator by a predetermined light, weighting the output information light by the weighting means. The information is output as information light, the output first weighted information light is input to another spatial light modulation element, and then the input characteristic of the other spatial light modulation element is changed as necessary, and the predetermined value is set. Second predetermined information light by weighting the output information light by the weighting means, by outputting the predetermined information light by a light different from the predetermined light from the spatial light modulator to which the information light of Output as , It is characterized in that for receiving the second weighted information light the output to the other spatial light modulator the input characteristics is changed as needed.

Further, a sixth optical information processing circuit information transmitting method according to the present invention uses the fifth optical information processing circuit according to the present invention, wherein predetermined information light is supplied to the fifth optical information processing circuit according to the present invention. A first weighting is performed by inputting to the spatial light modulator in the processing circuit, outputting the information light modulated by the spatial light modulator by a predetermined light, weighting the output information light by the weighting means. The information light is output as information light, and at the same time, the spatial light modulator to which the predetermined information light is input causes the predetermined information light to be output as light different from the predetermined light, and the output information light is the weighting means. And outputs the first weighted information light and the second weighted information light, which are output as the second weighted information light. Different from In the is characterized in that simultaneously input to the other spatial light modulator having different input characteristics.

[0036]

The structural unit of the first optical information processing circuit according to the present invention comprises a spatial light modulator and weighting means for weighting the information light modulated by the spatial light modulator.

As a result, the constituent unit of the optical information processing circuit according to the present invention transmits the information at one point of the spatial light modulator to the corresponding one point of the other spatial light modulator and a plurality of points in the vicinity thereof. Since information is transmitted by local coupling, that is, even when a large number of these constituent units are coupled,
Information processing can be performed in real time without requiring a large number of wires for connecting the respective constituent units, and the processing system can be downsized.

Further, this constitutional unit is combined so that the information light outputted from the two constitutional units is inputted to the constitutional unit of the other one unit, and is combined into the first constitutional unit of the two constitutional units. After inputting a predetermined information light (light having information), weighting this information light and outputting it as a first weighted information light, and after inputting this first weighted information light to another constituent unit , The input characteristic of the spatial light modulator in the other constituent unit is changed as necessary, and a predetermined information light is input to the second constituent unit of the above-mentioned two constituent units, The information light having a weight different from that of the first information light is output as the second weighted information light, and the second weighted information light is input to the spatial light modulator whose input characteristic is changed as necessary. By doing so, it is possible to transmit a plurality of types of signals.

Further, the spatial light modulator in the other structural unit is a spatial light modulator having different input characteristics from the first weighted information light and the second weighted information light, and
The second weighted information light may be simultaneously transmitted to the spatial light modulator in another structural unit.

Further, the second optical information processing circuit according to the present invention is configured so that the spatial light modulator in the first optical information processing circuit according to the present invention is applied to the information light input from one surface. Accordingly, the spatial light modulator for modulating the read light emitted from the other surface is used. Therefore, even when a large number of these structural units are combined, the read light can be further amplified and transmitted without attenuating the information light by irradiating each spatial light modulator.

Further, the constituent unit of the third optical neural network according to the present invention is modulated by the spatial light modulator in the constituent unit of the first optical information processing circuit according to the present invention in accordance with the input information light. This is a spatial light modulator that outputs emitted light. Therefore, it is not necessary to provide a light source that emits the reading light for reading the information light, and the information light can be transmitted by the light emission of the spatial light modulator itself.

Further, a fourth structural unit of the optical information processing circuit according to the present invention is the above-mentioned weighting means provided with a switching means for switching the weight. Therefore, by combining a plurality of these structural units, one signal is weighted by one weight, and another signal is weighted by another weight, as in the information transmission method of the third optical information processing circuit according to the present invention described above. By transmitting the information light by performing the above-mentioned procedure, it is possible to transmit a plurality of signals in one structural unit.

In the fifth structural unit of the optical information processing circuit according to the present invention, a plurality of different weighting means in the first, second, third or fourth optical information processing circuit according to the present invention are used. This is a weighting means for weighting light differently. Therefore, a plurality of constituent units of the optical information processing circuit are combined to form an information processing circuit, and the information light is transmitted in the same manner as the information transmission method of the fifth optical information processing circuit according to the present invention. For example, a plurality of signals can be transmitted by one structural unit.

Furthermore, the sixth structural unit of the optical information processing circuit according to the present invention is the first spatial light which modulates the read light emitted from the other surface in response to the information light input from the one surface. Modulation element and first, second or third according to the invention
Of the optical information processing circuit, a second spatial light modulator of the same type as the first spatial light modulator described above, and an information light modulated by the second spatial light modulator. And the information light re-inputting means for re-inputting into the spatial light modulator 1. If the constituent unit of this optical information processing circuit and the constituent unit of the first, second, or third optical information processing circuit according to the present invention are combined and combined to form the information processing circuit, the information processing circuit outputs one constituent unit. Since it is possible to input the information light that has been input to this structural unit again and repeat the weighting performed in this structural unit as many times as necessary, for example, perform processing such as cooperation and competition for the input information light. You can

[0045]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a schematic view showing a spatial light modulator used in an embodiment of the present invention.

As shown in FIG. 1, the embodiment of the present invention uses an LCLV (Liquid Crystal Light Valve) as a spatial light modulator.
1 is used. The information light (light having information) is input from one surface of the LCLV 1, and the information light is input to the LCLV 1. The read light is emitted from the other surface, and the read light is modulated in accordance with the input information. The liquid crystal layer 2, the photoconductor 3 such as a photorefractive crystal, the reflection film 4 that reflects the input light, and the transparent electrodes 5 and 6 are layered. The transparent electrodes 5 and 6 are applied with a DC voltage. The LCLV1 is divided into a W (write) side to which the information light is input and an R (read) side to which the information light is output. Since such a type of LCLV can supply light from the outside, a large number of each structural unit can be connected even if the LCLV itself does not have an amplification function. When light is irradiated from the transparent electrode 5 side of the LCLV 1, a voltage is applied to the liquid crystal layer 2 (accurately, there is a difference in applied voltage between a part irradiated with light and a part not irradiated with light), and the liquid crystal is aligned. The state changes. In this state, the liquid crystal layer 2 is changed to black in a portion irradiated with light. That is, the light pattern is input to the liquid crystal layer 2 of the LCLV 1 as a polarization pattern. Therefore, when information light such as a two-dimensional image is input to the LCLV 1 from the W side, this information can be directly input, and by irradiating the read light from the R side, the information light is output as a light-dark pattern. Is something that can be done. Once information light is input to this LCLV1, it will not disappear for a while, so if you want to delete this information, you can
The input characteristics can be inverted by inverting the potential of the voltage applied to 1, and the image once input can be erased by irradiating uniform light.

Next, an optical neural network in the optical information processing circuit using this LCLV will be described.

FIG. 2 is a diagram showing structural units of the optical neural network according to the first embodiment of the present invention.

As shown in FIG. 2, the constituent unit of the optical neural network according to the first embodiment of the present invention is the weighting output means 12 including the LCLV1 and the mask W for weighting and reflecting the information light and the lens L. It consists of The information light I0 input from the W side of the LCLV1 is once input to the LCLV1 and is irradiated with the read light H from the R side, and the read light H is modulated according to the input information light I0 and output. It passes through 12 and is weighted and reflected by the mask W, and is output as the weighted information light I ′.

FIG. 3 shows a state in which a plurality of such constituent units of the optical neural network are combined to form an optical neural network.

As shown in FIG. 3, in the optical neural network 20, three constituent units are connected so as to face each other, and the information light is reflected and folded back by the weighting output means of each constituent unit, and this information light is passed to the next stage. It is intended to be transmitted to the constituent units of. By arranging the constituent units so as to face each other in this manner, the neural network can be made compact and the accuracy in arranging the constituent units can be improved.

First, the information light I0 is inputted and inputted to the W side of the LCLV21 in the constituent unit of the first optical neural network. Next, the R side of the LCLV21 is irradiated with the read / write light H1 and the read / write light H1 is modulated according to the information light input to the LCLV21, and the information light is read. This read information light I1 is reflected by the lens L1.
And a weighting mask W1 and weighted and reflected by the weighting output means 25. The weighted information light I1 'is input and input to the W side of the LCLV22 in the constituent unit of the second optical neural network. At this time, transmission of the information light from one point with LCLV21 to LCLV22 corresponds to one point with LCLV21
The lens L1 and the LCLV are arranged so that the information light is transmitted only to the vicinity of a certain point on the LV22.
21,22 are located. Next, the read / write light H2 is irradiated from the R side of the LCLV 22, the read / write light H2 is modulated according to the information light input to the LCLV 22, the information light is read, and the read information light I2 is read. Is reflected by a weighting output means 26 including a lens L2 and a weighting mask W2, which is weighted differently from the weighting mask W1. This weighted information light I
2'is input to the LCLV23 in the third structural unit of the optical neural network, read by the read / write light H3, weighted by the weighting output means 27 including the lens L3 and the weighting mask W3, and input to the output LCLV24. Is input. The information light I3 'input to the LCLV 24 is read by being irradiated with the read light R1 from the R side of the LCLV 24, and finally the processed information light O1 is output.

After the information light processing is completed in this way, the input characteristics are inverted by inverting the potential of each LCLV of the optical neural network 20, and each LCL is inverted.
V is irradiated with uniform light, and the information light input to each LCLV is erased.

In this neural network 20, each LCLV is one neuron layer, and the connection weight is common at any point of each layer, so that space invariant weight connection is realized between the layers. Can be done.

Next, the information light modulated and output by the first spatial light modulator, the weighted output means, the second spatial light modulator and the second spatial light modulator is output as the first spatial light modulator. The optical neural network having the constituent unit of the optical neural network, which is composed of the information light re-inputting means for re-inputting the optical neural network, will be described.

FIG. 4 is a diagram showing a state in which a plurality of constituent units of the optical neural network having the above-described information light re-inputting unit and the constituent units of the optical neural network shown in FIG. 2 are combined to form the optical neural network. is there. Since the optical neural network shown in FIG. 4 has substantially the same configuration as the optical neural network shown in FIG. 3, the same parts as those of the optical neural network shown in FIG. The detailed description is omitted.

The optical neural network shown in FIG. 4 has half mirrors M1 and M2 and lenses L4 and L5 as information light re-inputting means.

The information light I0 'input to the optical neural network 20' is weighted by each structural unit in the same manner as the optical neural network described above, and LCL is added.
It is transmitted to V23 '. The information light I2 ″ transmitted to the LCLV 23 ′ is inputted and inputted to the W side of the LCLV 23 ′. Then, the read / recursive writing light H4 is irradiated on the LCLV 23 ′ via the half mirror M2 and the lens L5, and L
The read / recursive writing light H4 is modulated according to the information light I2 ″ input to the CLV 23 ′, and the information light I2 ″ is read. The read information light H4 'is reflected by the half mirror M1 via the lens L4, and LCLV23'.
It is input again to the W side of the LCLV 22 'in the previous stage and input. At this time, the information light I1 ″ transmitted from the LCLV 21 ′ to the LCLV 22 ′ and input is erased by irradiating uniform light with the potential of the LCLV 22 ′ inverted and the input characteristics inverted.

The information light H4 'input again to the LCLV 22' is irradiated with the read / write light H2 'from the R side of the LCLV 22' and read / written according to the information light H4 'input to the LCLV 22'. The writing light H2 'is modulated and the information light H4' is read. Information light H read
Reference numeral 4'is weighted again by a weighting output means composed of a lens L2 'and a weighting mask W2' and is reflected by L.
Input to the W side of CLV23 '. Also at this time, since the information light I2 ″ input to the LCLV 23 ′ is erased, new information light is input to the LCLV 23 ′. By repeating the above processing, the information light I2 ″ is input to the LCLV 22 ′. Weighting by the weight mask W2 'can be repeatedly performed on the information light I1 "thus obtained.

After performing the above-described processing a required number of times, R of the LCLV23 'is passed through the half mirror M1 and the lens L4.
The reading / writing light H5 is irradiated from the side, and LCLV23 '
The read / write light H5 is modulated in accordance with the information light I2 'input to the optical disk I2' and the information light I3 'is read. The read information light I3 'is reflected by the half mirror M2 via the lens L5, input to the output LCLV 24', and input. Information light I input to the LCLV 24 '
3 ″ is read out in the same manner as the optical neural network shown in FIG. 3 and finally processed information light O1 ′.
Is output.

Further, as a configuration of the optical neural network having the above-mentioned information light re-inputting means, the configuration shown in FIG. 5 may be used in addition to the configuration shown in FIG.

Optical neural network 30 shown in FIG.
Arranging the LCLVs 31 and 33 on the same plane 41, the weighting output means 32 on the cylindrical surface 41 surrounding the plane 41, and the weighting output means 34 and the mirror 35 which is the information light re-inputting means. It is arranged at a position substantially opposite to the weighting output means 32 on 41.

The information light input to the LCLV31 is
The CLV 31 is read by irradiating the read / write light H6, weighted and reflected by the weighting output means 32, and input to the LCLV 33. Next, the read / recursive write light H7 is input to the optical neural network 30, is irradiated on the LCLV 33, is modulated according to the information input to the LCLV 33, and outputs this information light. The output information light is reflected by the mirror 35 and is again LCLV.
It is input to 31 and input. By repeating the above processing in the same manner as the optical neural network shown in FIG. 4, weighting by the weighting output means 32 can be repeated.

After performing the above-described processing a required number of times, the read / write light H8 is irradiated on the LCLV33, and the LCLV33 is irradiated.
The information light is read out after being modulated in accordance with the information input to. The read information light is weighted by the output means 34.
Are weighted and output as the final processed information light O2. At this time, the read / recursive write light H7 and the read / write light H8 are set to cross L
Since the CLV 33 is irradiated and transmitted to the mirror 35 and the weighting output means 34, respectively, the read / recursive write light H7 and the read / write light H8 do not overlap with each other.

If the constituent unit of the optical neural network having the information light re-inputting means is constructed as the optical neural network 30 as shown in FIG. 5, it is repeatedly weighted as compared with the optical neural network 20 'shown in FIG. Since the weighted information light can be immediately weighted, the LCLV 24 ′, which is provided in the optical neural network 20 ′, for inputting the repeatedly weighted information light once becomes unnecessary, and the configuration can be simplified.

In the above-mentioned embodiments, the optical neural network in which the LCLV and the weighted output means are arranged opposite to each other has been described, but the constituent unit of the optical neural network according to the present invention is not limited to such a structure. . For example, as shown in FIG. 6, the weighting output means may be a transmission type weighting output means, and the LCLV and the weighting output means may be arranged in series to form a neural network.

Next, the unit of information transmission between the constituent units of the optical neural network described above will be explained. That is, the optical neural network according to the present invention can transmit not only a positive number signal but also a negative number signal. The method will be described below.

FIG. 7 is a diagram showing an optical neural network for explaining the information light transmitting method according to the present invention. As shown in FIG. 7, the optical neural network 50
Is a combination of the structural unit of the optical neural network according to the first embodiment of the present invention and the LCLV 52. The weight mask W6 in the weighting output means 53 of the optical neural network is made of liquid crystal, and the weight mask can be rewritten. Here, the optical neural network 50 is a neural network which performs absolute value type differential processing, and the weight mask is as shown in FIG. The image input to the optical neural network 50 is a bright and dark image as shown in FIG.
56.

First, the weight mask W6 displays the mask 55a shown in FIG. 8 (a). First, the image 56 is taken as the information light I.
Input as 5 on the W side of the LCLV51 and input. Next, the read / write light H9 is irradiated from the R side of the LCLV51, the read / write light H9 is modulated according to the information light I5 input to the LCLV51, and the information light I5 'is read. The read information light I5 'is weighted as shown in FIG. 10 (a) by the weighting mask W6 in which the mask 55a is displayed via the lens L6 of the weighting output means 53, and the positive component of the image 61 LCLV as information light I6
Input to the W side of 52.

After the input of the information light I6, the display of the weight mask W6 is rewritten from the mask 55a shown in FIG. 8A to the mask 55b shown in FIG. 8B, and at the same time, the potential of the LCLV 52 is inverted and the input state is inverted. . Since the polarization state of the image input to the LCLV is maintained even if the potential of the LCLV is inverted, the information light I6 is not erased unless light is irradiated, and the information light I6 remains in the input state. Has become.

Then, the read / write light H9 emits LCLV51.
The information light I5 'is read out again from the R side of. The read information light I5 'is weighted as shown in FIG. 10 (b) by the weight mask W6 in which the mask 55b is displayed via the lens L6 of the weight output means 53. The information light I5 ′ thus weighted is the information light I5.
LCLV in which the potential is inverted and the input state is inverted as 7.
It is input to the W side of 55, and is input to be superimposed on the information light 6. At this time, since the potential of the LCLV 52 is inverted, the portion irradiated with light is polarized in the direction opposite to the normal direction. Therefore, when the information light I7 is written so as to overlap the portion where the information light I6 was input, the information light I7 erases the input information light I6. As a result, CLV
As shown in FIG. 10 (c), the information light in which only the light and dark boundary portions are extracted, that is, the information light subjected to the absolute value type differential processing is input to 52.

Then, the read / write light H10 is LCLV55.
The light is emitted from the R side, modulated, and output as the information light O3 that has been subjected to the differential processing.

In the above-described embodiment, the weight mask in the weighting output means is switchable so that the positive component and the negative component of the information light are respectively weighted. However, it is particularly necessary to switch the weight mask. Absent. For example
As shown in 11, the optical neural network is constructed by combining the constituent units so that the information light output from the constituent units of the two optical neural networks is input to one constituent unit. 8 shows the weight mask W7 of the weight output means 64 in one of the constituent units of FIG.
The mask 55a shown in (a) and the weight mask W8 of the weighting output means 65 in the other one unit are the mask 55b shown in FIG. 8 (b). The information light input to the LCLVs 61 and 62 is assumed to be bright and dark information light 56 as shown in FIG. L
The information light I8 output from the CLV 61 is weighted by the weighting mask W7, and the information light I8 is given as a lens L7 as the information light I8 '.
And it is inputted to the LCLV 63 from the W side of the LCLV 63 through the beam splitter 70 and inputted. Then LC
Inverts the potential of LV63 and inverts the input state, then LC
The information light I9 is output from the LV62. The output information light I9 is weighted by the weight mask W8
9'is reflected by the lens L8 and the beam splitter 70, and is input to the LCLV 63 from the W side of the LCLV 63 and input. If the information light is transmitted to the optical neural network 60 as described above, the LCLV 63 shows the information light in which only the light-dark boundary portion is extracted, that is, the absolute value type differential as shown in FIG. 10 (c). The processed information light is input.

As another example, the weighting mask W6 of the weighting output means 53 in the optical neural network 50 shown in FIG. 7 may be a mask that gives different weights to lights of different wavelengths. For example, in the weighting mask W6, the mask 55a shown in FIG. 8A is displayed with a red filter, and the mask 55b is displayed with a blue filter.
In addition, the read / write light H9 is used as a red light and the LCLV51
The information light I5 input to is read out, weighted by the weighting output means 53, and input to the LCLV 52. After that, the potential of the LCLV 52 is inverted and the input state is inverted, the information light I5 input to the LCLV 51 is read using the read / write light H9 as blue light, weighted by the weighting output means 53, and input to the LCLV 52. The I5 read by the red light is weighted by the red filter portion of the weight mask W6, so that the information light as shown in FIG. 9 is converted into the information light I6 as shown in FIG. 10 (a). Also, since I5 read out by the blue light is weighted by the blue filter portion of the weight mask W6, FIG.
It is converted into information light I7 as shown in (b). Therefore, the LC in which the information light I6 and the information light I7 are overlapped and input
To the LV 52, as shown in FIG. 10C, the information light in which only the clear boundary portion is extracted, that is, the information light subjected to the absolute value type differential processing is input.

At this time, the spatial light modulator may be a spatial light modulator having different input characteristics between blue light and red light, and the blue light and the red light may be transmitted at the same time. Thus, by transmitting different information lights at the same time, the calculation time is shortened, which is more preferable.

Further, as a mask for weighting different lights differently, not only the above-mentioned red filter and blue filter but also lights having different polarizations are separated by using a polarizing plate, and positive signals are respectively separated. , A negative signal may be transmitted.

As the spatial light modulator, in addition to the LCLV described above, for example, MSLM manufactured by Hamamatsu Photonics
Any spatial light modulator such as (Micro channel Spatial Light Modulator) may be used.

In the above-mentioned embodiment, the weight mask is used as a mask for performing the differential processing of the absolute value type.
For example, if the weight mask is a mask as shown in FIG. 12, and the mask 70a weights the positive component of the information light and the mask 70b weights the negative positive component of the information light,
It is possible to extract the edge of the object included in the information light. Further, not limited to these masks, any mask can be used according to the processing performed by the neural network.

Further, in the above-mentioned embodiment, the spatial light modulation element is a spatial light modulation element which once inputs the input information light and modulates and outputs the separately input light according to the input information. However, the spatial light modulator is not limited to this type. For example, a plurality of optical neuron elements as shown in FIG. 13 may be two-dimensionally arranged to form a spatial light modulation element, and the input information may be immediately read by the reading light. Examples using the optical neuron element shown in FIG. 13 will be described below.

As shown in FIG. 13, the optical neuron element 71
Is a photoelectric operation element 78 including a positive / negative signal separation filter 72, a light receiving element pair 73 and an operation element, and transparent electrodes 79a and 79b.
It is composed of a reflection-type light modulation element 81 composed of a reflection-type liquid crystal 80 sandwiched between and a light input to the light modulation element 81, and a polarizer 82 that polarizes the reflected light.

When the optical neuron element 71 is simultaneously irradiated with the optical signals 83 representing positive and negative information light as blue light and red light, respectively, the blue filter 72a of the positive / negative signal separation filter 72 of the light receiving / separating element 74 and the red filter 72a The filter 72b separates the positive and negative information lights. The separated information lights are respectively received by the light receiving elements 73a and 73b and output as an electric signal 75a representing positive information light and an electric signal 75b representing negative information light. The output electric signals 75a and 75b are the calculation elements 76
Addition / subtraction and non-linear processing are performed by this, and the electric signal 77 corresponding to the calculation result is output by this.
The light modulation element 81 is modulated in accordance with 77. When the read light 84 such as white light is irradiated to the light modulation element 81 thus modulated, the read light 84 is polarized by the polarizer 82, and the polarized read light 84 is reflection-modulated by the light modulation element 81. It is again polarized by the polarizer 82 and output as information light 85.

A plurality of optical neuron elements 71 having such a function are arranged one-dimensionally or two-dimensionally, and a spatial light modulation element for outputting information light composed of light output from each optical neuron element is provided. The spatial light modulation element and the information light modulated by the spatial light modulation element are weighted, and a constituent unit of an optical neural network including weighting means for outputting the weighted information light is configured, and these are configured. An optical neural network can be configured by combining a plurality of optical neural networks.

FIG. 14 is a diagram showing an embodiment of a structural unit of an optical neural network configured by using the above-mentioned optical neuron element 71.

As shown in FIG. 14, the constituent units of the optical neural network configured by using the above-mentioned optical neuron element 71 are the optical neuron layer 86 and the optical neuron layer 86 formed by connecting the above-mentioned optical neuron elements 71 two-dimensionally. The weighting means 87 includes a weighting mask W and a lens L. The optical neuron layer 86 is arranged so that the light receiving / separating element faces the right side in the drawing. The information light I, in which the positive information light is blue light and the negative information light is red light, is applied to the optical neuron layer 86 from the W side of the optical neuron layer 86. The irradiated information light I is subjected to the above-mentioned arithmetic processing by each optical neuron element that constitutes the optical neuron layer 86, and the light modulation element in each optical neuron element is modulated according to the information light input to each optical neuron element. It Thereafter, writing / reading light H such as white light is emitted from the R side of the optical neuron layer 86. The irradiated writing / reading light H is reflection-modulated by the optical modulation element of each optical neuron element forming the optical neuron layer 86, and is output as information light I ′. The output information light I ′ passes through the lens L, is weighted by the weight mask W, and is output as the weighted information light I ″. The weight mask W can be weighted as desired. It is composed of a blue filter and a red filter. In this embodiment, it is a reflection type weight mask.

If a plurality of structural units of such an optical neural network are combined, a neural network can be constructed in the same manner as the above-mentioned embodiment for inputting information light. For example, as shown in FIG. 15, the weighting output means may be a transmissive weighting output means, and the optical neuron layer and the weighting output means may be arranged in series to form a neural network.

Further, in the above-described embodiment, FIG.
A plurality of light emitting optical neuron elements as shown in FIG. 2 are arranged two-dimensionally to form a spatial light modulator, and the information light modulated according to the input information light is transmitted by the light emission. Good. Hereinafter, an example using the optical neuron element shown in FIG. 16 will be described.

As shown in FIG. 16, the optical neuron element 10
Positive / negative signal separation filter 7 that constitutes photoelectric operation element 78 of 1
2, the light receiving element pair 73 and the arithmetic element 76 are shown in FIG.
Since the configuration is the same as that of the optical neuron element in the embodiment of the present invention shown in FIG. The optical neuron element 101 according to the present embodiment is the optoelectronic operation element 78.
And a light emitting element 102 that emits light with an intensity according to an electric signal corresponding to the calculation result output from the photoelectric calculation element 78.

When the optical neuron element 101 is simultaneously irradiated with the optical signals 83 representing positive and negative information in blue light and red light, respectively, the blue filter 72a and the red filter 72a of the positive / negative signal separating filter 72 of the light receiving / separating element 74 are irradiated. 72b separates the positive and negative information lights. The separated information lights are respectively received by the light receiving elements 73a and 73b and output as an electric signal 75a representing positive information light and an electric signal 75b representing negative information light. The output electric signals 75a and 75b are the calculation elements 76
Addition / subtraction and non-linear processing are performed by this, and the electric signal 77 corresponding to the calculation result is output by this.
The light emitting element 102 emits light with an intensity corresponding to 77. The light emitted from the light emitting element 102 is output as the information light 104.

A plurality of optical neuron elements 101 having such a function are arranged one-dimensionally or two-dimensionally to form an optical neuron layer which outputs information light composed of light output from each optical neuron element. Then, a constituent unit of the optical neural network including the optical neuron layer and the weighting means for outputting the weighted information light by weighting the information light output by being modulated by the optical neuron layer is constructed. An optical neural network can be configured by combining a plurality of optical neural networks.

FIG. 17 is a diagram showing an embodiment of a constituent unit of an optical neural network constituted by using the above-mentioned optical neuron element 101.

As shown in FIG. 17, the constituent unit of the optical neural network constructed by using the above-mentioned optical neuron element 101 is the above-mentioned optical neuron element 101.
A weighting output means 10 including an optical neuron layer 105 formed by connecting a plurality of dimensions and a weighting mask W and a lens L.
It consists of 6 and 6. The optical neuron layer 105
Is arranged so that the light receiving and separating element faces the right side in the drawing. The information light I, in which the positive information is represented by blue light and the negative information is represented by red light, is applied to the optical neuron layer 105 from the W side of the optical neuron layer 105. The irradiated information light I is subjected to the above-mentioned arithmetic processing by each optical neuron element 101 constituting the optical neuron layer 105, and the light emitting element in each optical neuron element emits light in accordance with the information light input to each optical neuron element. . This emitted light is output as information light I ′, and the output information light I ′ passes through the lens L, is weighted by the weighting mask W, and is output as weighting information light I ″. Is composed of a blue filter and a red filter so that desired weighting can be performed.

FIG. 18 shows a state in which a plurality of constituent units of such an optical neural network are combined to form an optical neural network.

As shown in FIG. 18, in the optical neural network 110, two constituent units are connected in series, and the information light is weighted by the weighting output means of each constituent unit. It is designed to be transmitted to the constituent units.

By constructing the optical neural network as described above, the information light can be transmitted without using the reading light, and the neural network can be constructed more compactly.

In the above embodiment, the optical neural network is constructed by combining a plurality of constituent units of the optical neural network.
When only the differential processing described above is performed, the processing can be performed with only one structural unit, and therefore it is not necessary to combine a plurality of units. In this case, although the optical neural network has only one unit, the one unit constitutes the neural network.

Further, in the above embodiment, two types of positive and negative information light are transmitted by the neural network according to the present invention, but the present invention is not limited to positive and negative information, and a plurality of information are different from each other. You may make it transmit by the light of a wavelength. In this case, the spatial light modulator preferably changes the input characteristic according to the wavelength of light.

The optical information processing circuit according to the present invention can be applied not only to the neural network but also to a circuit for performing various information processing.

[0099]

As described in detail above, according to the constituent unit of the optical information processing circuit, the optical information processing circuit and the information transmission method of the present invention, it is possible to reduce the number of connections between the spatial light modulators. Further, since it is composed of the spatial light modulator and the weighted output means, high-speed parallel processing is possible and the optical information processing circuit according to the present invention can be miniaturized. Also, by using the spatial light modulator and changing the input characteristics of the spatial light modulator as necessary, not only the positive component of the information light input to the information processing circuit but also the negative component are transmitted and input. Therefore, various weighting processes can be performed.

[Brief description of drawings]

FIG. 1 is a schematic diagram showing a spatial light modulator used in an embodiment of the present invention.

FIG. 2 is a diagram showing structural units of an optical neural network according to the first embodiment of the present invention.

FIG. 3 is a diagram showing an embodiment in which a plurality of constituent units of an optical neural network are combined to form an optical neural network.

FIG. 4 is a diagram showing an embodiment in which a plurality of constituent units of an optical neural network having information light re-inputting means are combined to form an optical neural network.

FIG. 5 is a diagram showing another embodiment in which a plurality of structural units of an optical neural network having information light re-inputting means are combined to form an optical neural network.

FIG. 6 is a diagram illustrating an example in which a plurality of constituent units of an optical neural network are connected in series to configure an optical neural network.

FIG. 7 is a diagram showing a structural unit of an optical neural network which is a weighting output means capable of switching the weighting output means.

FIG. 8 is a diagram showing an example of a weight mask.

FIG. 9 is a diagram showing bright and dark information light.

FIG. 10 is a diagram showing a state in which bright and dark information lights are weighted by different masks.

FIG. 11 is a diagram showing an example in which optical neural networks are configured by combining constituent units so that information light output from the constituent units of two optical neural networks is input to one constituent unit.

FIG. 12 is a diagram showing another example of the weight mask.

FIG. 13 is a diagram showing an optical neuron element that constitutes a spatial light modulator according to another embodiment of the present invention.

FIG. 14 is a diagram showing a structural unit of an optical neural network configured by using a spatial light modulator according to another embodiment of the present invention.

FIG. 15 is a diagram showing an optical neural network using constituent units of an optical neural network configured by using a spatial light modulator according to another embodiment of the present invention.

FIG. 16 is a diagram showing an optical neuron element forming a spatial light modulator according to still another embodiment of the present invention.

FIG. 17 is a diagram showing a structural unit of an optical neural network configured by using a spatial light modulator according to still another embodiment of the present invention.

FIG. 18 is a diagram showing an optical neural network using a constituent unit of an optical neural network configured by using a spatial light modulator according to still another embodiment of the present invention.

FIG. 19 is a diagram showing an example of a model of an optical neurocomputer.

FIG. 20 is a diagram showing another example of the model of the optical neurocomputer.

[Explanation of symbols]

1, 21, 21 ', 22, 22', 23, 23 ', 24, 24', 31, 3
3, 51, 52, 61, 62, 63 Spatial light modulator (LCL
V) W, W1, W1 ', W2, W2', W3, W6, W7,
W8 Weight mask L, L1, L1 ', L2, L2', L3, L4, L5
L6, L7, L8 lens 12,25,25 ', 26,26', 27,32,34,53,64,65
Weighting output means I0, I0 ', I5 Input information light H, H1, H1', H2, H2 ', H3, H5, H6
H8, H9, H10 Read / write light H4, H7 Read / recursive write light 20, 20 ', 30, 50, 60, 110 Optical neural network

Claims (15)

[Claims] 1. A spatial light modulator for modulating and outputting the information light according to input information light, and a weighting means for weighting the information light modulated and output by the spatial light modulator. A constituent unit of an optical information processing circuit, which is composed of 2. The spatial light modulation element is a spatial light modulation element that modulates read-out light emitted from another surface in accordance with information light input from one surface. 1. A constituent unit of the optical information processing circuit described in 1. 3. The optical information processing circuit according to claim 1, wherein the spatial light modulation element is a spatial light modulation element that outputs a modulated light emission according to the input information light. Building unit. 4. The structural unit of an optical information processing circuit according to claim 1, 2 or 3, wherein the weighting means has a switching means for switching the weight. 5. The structural unit of the optical information processing circuit according to claim 1, wherein the weighting output means is a weighting means for weighting different plural lights respectively. . 6. A plurality of the constituent units of the optical information processing circuit according to claim 1, 2 or 3, wherein the plurality of constituent units are output from one constituent unit of the plurality of constituent units. An optical information processing circuit, characterized in that it is coupled so as to sequentially input information light to one or more other structural units. 7. A first spatial light modulator that modulates and outputs the information light in accordance with the input information light, and the optical information processing circuit according to claim 1, 2, 3, 4 or 5. Weighting output means in the constitutional unit, a second spatial light modulator for modulating and outputting the information light according to the input information light, and the information light output by the second spatial light modulator And an information light re-input means for re-inputting into the first spatial light modulator. 8. At least one of claims 1, 2, 3, 4
Or 5 and at least one of the constituent units of the optical information processing circuit according to claim 7, wherein these constituent units are output from one constituent unit of these constituent units. The information light that has been generated is sequentially transferred to the other 1
An optical information processing circuit, characterized in that the optical information processing circuit is connected so as to be an input of one or more constituent units. 9. A plurality of constituent units of the optical information processing circuit according to claim 4, wherein the plurality of constituent units sequentially output the information light output from one constituent unit among the plurality of constituent units. An optical information processing circuit, wherein the optical information processing circuit is connected so as to be an input of one of the structural units. 10. A predetermined information light is input to the spatial light modulator in the optical information processing circuit according to claim 9, and the input information light is output from the spatial light modulator as the first information light. The outputted first information light is weighted by the weighting means and outputted as first weighted information light, and the outputted first weighted information light is outputted to the other spatial light modulator. After that, the weight of the weighting means is switched, and the input characteristic of the other spatial light modulator is changed as necessary, and the output first information light is switched to the weighting means. To output as the second weighted information light, and input the output second weighted information light to the other spatial light modulator whose input characteristics are changed as necessary. A contract characterized by Information transmission method for an optical information processing circuit in claim 9. 11. The information light output from two units of the constituent units of the optical information processing circuit according to claim 1, 2 or 3 is input to the constituent unit of another unit. An optical information processing circuit characterized by combining constituent units. 12. A predetermined information light is input to a first constituent unit of the constituent units of the optical information processing circuit of the two units in the optical information processing circuit according to claim 11, and the information light is converted into the information light. Weighted and output as the first weighted information light, the first weighted information light is input to the constituent unit of the other one unit, and is input to the spatial light modulator in the constituent unit, After the input, the input characteristics of the spatial light modulator in the other one of the constituent units are changed as necessary, and among the constituent units of the optical information processing circuit of the two units,
The predetermined information light is input to the second structural unit, the information light is weighted and output as second weighted information light, and the second weighted information light is input as required by the input characteristic. 12. The information transmission method for an optical information processing circuit according to claim 11, wherein the information is input to the spatial light modulator that has been changed. 13. A plurality of constituent units of the optical information processing circuit according to claim 5, wherein the plurality of constituent units sequentially output the information light output from one constituent unit among the plurality of constituent units. An optical information processing circuit, wherein the optical information processing circuit is connected so as to be an input of one of the structural units. 14. A predetermined information light is input to the spatial light modulation element in the optical information processing circuit according to claim 13, the information light modulated by the spatial light modulation element is output by the predetermined light, and the output The outputted information light is weighted by the weighting means and outputted as the first weighted information light, the outputted first weighted information light is inputted to another spatial light modulation element, and then the other information light is outputted. The input characteristic of the spatial light modulation element is changed as necessary, and the predetermined information light is output from the spatial light modulation element to which the predetermined information light is input, as light different from the predetermined light, and the output is performed. The information light thus obtained is weighted by the weighting means and outputted as second weighted information light, and the outputted second weighted information light is the other spatial light modulation in which the input characteristic is changed as necessary. Input to the element 14. The information transmission method for an optical information processing circuit according to claim 13. 15. A predetermined information light is input to the spatial light modulator in the optical information processing circuit according to claim 13, the information light modulated by the spatial light modulator is output by the predetermined light, and the output The obtained information light is weighted by the weighting means and output as the first weighted information light, and at the same time, the predetermined information light is changed to the predetermined light from the spatial light modulator to which the predetermined information light is input. Are output by different lights, the output information light is weighted by the weighting means and output as second weighted information light, and the output first weighted information light and second weighted information light are output. 14. The optical information according to claim 13, wherein the light and the predetermined light and the light different from the predetermined light are simultaneously input to another spatial light modulator having different input characteristics, if necessary. processing Information transmission method of the road.