JPH0667241A - Light vector matrix arithmetic unit - Google Patents

Abstract

PURPOSE:To easily obtain an integrated structure and to make light beams of respective elements of a matrix equal in intensity. CONSTITUTION:Four input parallel light beams 24 which are arrayed in an (x)-axial direction are each split into two light beams through a half-mirror 25 and a mirror 26 in an optical path converting element array 21a and further split into two light beams similarly by an optical path converting element array 21b to generate 4X4 light beam arrays. Those light beams are optically modulated by the matrix mask 22 consisting of a polarizing plates 22a and 22b and a liquid crystal element array 22c. Those 4X4 optically modulated light beams are multiplexed by the light signal multiplexing element 23 consisting of beam splitters 271, 272, and 273 which have a 1/2, a 1/3, and a 1/4 reflection factor respectively.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an optical vector matrix computing device for performing matrix computation on an optical signal sequence arranged in a matrix and propagating in space and obtaining the result as an optical signal sequence.

[0002]

2. Description of the Related Art With respect to the complexity of wiring in electronic circuits such as LSI, attention has been paid to the spatial parallelism of light, and studies have been made to perform calculations using light. For example, in a neural network, the number of wirings between neurons dramatically increases as the number of neurons increases. Therefore, when attempting to realize a large-scale neural network with an LSI, the complexity of wiring becomes a major obstacle to high integration, and at the same time, crosstalk due to interference between electrical signals transmitted through the wiring becomes a problem.

On the other hand, if an optical signal propagating in space is used as wiring, crosstalk between the wirings can be eliminated, and at the same time, three-dimensional wiring or cross wiring can be performed, facilitating a large-scale and complicated wiring system. realizable. Further, since changing the transmittance of the optical signal and collecting a plurality of optical signals correspond to the product and the sum, respectively, the product-sum operation can be easily performed by using light. For the above reasons, optical vector matrix arithmetic devices using spatially propagating light are being studied for application to neural networks and crossbar switches.

FIG. 5 shows the principle of light vector matrix calculation using spatially propagating light. This device uses an input signal vector, an output signal vector, and a matrix mask as V
An optical arithmetic circuit that realizes V _o = T V _i , where _i , V _o , and T. LED as shown
Consider a light emitting element array 11 in which N light emitting elements such as (light emitting diodes) are arranged in a line in the x-axis direction. The optical input signal from this is elongated and elongated only in the y-axis direction by a lens system (not shown), and is made incident on each column of the matrix mask 12 of M rows and N columns. After each incident light passes through the mask of each element having the transmittance corresponding to each integration, it is condensed for each row of the matrix of the matrix mask 12 by the lens system not shown in the figure, and is detected by the detection element array 13. An output signal as a calculation result is obtained.

Conventionally, such vector matrix calculation by light has been realized by using a lens system. FIG. 6 shows a conventional example of a light vector matrix computing device constructed using a lens system. The configuration shown in FIG.
odman, A .; R. Dias, and L.D. M. W
woody, “Fully Parallel, high
-Speed incoherent optical
method for forming disc
rete Fourier transforms ”,
Opt. Lett. , 2, 1 (1978)). As the light emitting element array 11, an LED array or the like arranged in the x-axis direction (row direction) is used, and each L
Each optical input signal from the ED is converted into parallel rays by the spherical lens L1. The cylindrical lens L2 is placed at the focal point of the spherical lens L1 and x is taken on the xy section of each ray.
The light is focused only in the axial direction, the shape of the cross section is made into an ellipse elongated in the y-axis direction, and each is made incident on each column of the matrix mask 12. At this time, the cylindrical lens L3 is used as a field lens to adjust the way of incidence. Each of the light rays transmitted through the matrix mask 12 and arranged in a lattice pattern is extended in the y-axis direction, and is returned to a parallel light ray in the x-axis direction by the cylindrical lens array L4 arranged in the x-axis direction. Further, a set of a cylindrical lens L5 and a spherical lens L6 collects each line of the lattice-shaped light flux and makes it enter the detection element array 13. With such a configuration, the vector matrix calculation shown in FIG. 5 can be realized.

[0006]

In the above-mentioned prior art, the vector matrix operation is realized by the lens system. Therefore, there is a problem that the optical system becomes complicated, high precision is required for position setting of optical elements such as lenses, and it is difficult to configure as an integrated device. Further, what is transmitted through the matrix mask is that the cross section of the light beam spread over one column of the matrix mask and the mask element of the matrix mask overlap with each other, and the light intensity of the light beam cross section is not uniform. There is also a problem that the intensity of incident light per area differs depending on the position of the mask element when light is incident.

[0007]

According to the present invention, a parallel light beam propagating in a space is used as a signal input, the light beams are combined by an optical signal branching element and are incident on a matrix mask, and the light emitted from the matrix mask is converted into a light beam. Combine with a signal combining element. The optical signal branching element and the optical signal multiplexing element are each composed of an optical path conversion element array, and the optical path conversion element array is composed of a combination of a plurality of reflection elements / partial reflection elements, and the partial reflection element converts a signal into transmitted light and reflected light. Split into. In the optical signal branching element, the input signal is sequentially demultiplexed by arranging the optical path conversion element array in multiple stages. Further, in the optical signal multiplexing element, the partial reflection element combines the transmitted light of a certain signal and the reflected light of another signal to perform multiplexing.

According to the first aspect of the invention, the signal division in the partial reflection element is performed by intensity division of the light beam by using a beam splitter as the partial reflection element. According to the invention of claim 2, a polarization beam splitter is used as the partial reflection element,
When used in combination with a circular polarization conversion element (for example, a wave plate), polarization division is performed. Further, in order to provide compatibility with an optical circuit, it is inputted as a light beam train and outputted as a light beam train.

[0009]

Embodiment 1 As Embodiment 1 of the present invention, FIG. 1 shows an example in which four inputs and four outputs are demultiplexed and combined by light intensity division. The device of the present invention comprises an optical signal branching element 21, a matrix mask 22,
It is composed of the optical signal multiplexer 23. In the first embodiment, the optical signal branching device 21 is a two-stage optical path conversion device array 21.
a and 21b, and the matrix mask 22 may be the same as the conventional one. For example, the matrix mask 22 includes polarizing plates 22a and 22b and a liquid crystal element array 22c, and an optical signal multiplexing element 23 is used.
Consists of a single-stage optical path conversion element array 23a.

In this embodiment, the four optical input signals 24 are demultiplexed into four by the optical signal branching element 21 into 4 × 4, and then the 4 × 4 optical signals are respectively divided by the matrix mask 22. Intensity modulation is applied to the four signals, and the signals are combined by the optical signal combining element 23 to extract four optical outputs. 2A is a cross-sectional view of the optical signal branching device 21 and the matrix mask 22 in FIG. 1 taken along the yz plane, and FIG.
A cross-sectional view on the y-plane is shown in FIG. 2B.

In practice, four parallel light beam arrays arranged in the x-axis direction using a laser diode array or the like are input as an input signal (vector) 24, and each light beam is respectively converted into an optical path conversion element array 21a. , 21b, it is demultiplexed into four and expanded into a 4 × 4 light beam bundle. That is, the light intensity 2 as an optical path changing element in the optical path changing element array 21a
With 1/25 beam splitter (half mirror) 25,
Each one light beam is divided into two, the reflected light is bent at a right angle by the mirror 26 and arranged in parallel with the transmitted light in the z-axis direction, and the transmitted light and the reflected light are further arranged in the optical path conversion element array 2 of the next stage.
The half beam splitter 25 and the mirror 26 in 1b similarly similarly divide each into two light beams and arrange them in the z-axis direction.

These 4 × 4 light beams are transmitted through the polarizing plate 22.
a, 4 × 4 array liquid crystal element array 22c, polarizing plate 22b
After passing through the matrix mask 22 consisting of, a 4 × 4 light beam bundle corresponding to the state where the elements are multiplied by the product of the matrix and the vector is obtained. This is realized by controlling the polarization direction of the optical signal by each liquid crystal element of the liquid crystal element array 22c and changing the transmittance of each part of the matrix mask 22 for each light beam. The calculation is completed by combining each of these four light beams in the x-axis direction by the optical path conversion element array 23a, combining them into four, and outputting them. This corresponds to the summation of the product of the matrix and the vector. Here, one light beam is transmitted through the optical path conversion element array 23a and the beam splitters 27 ₁ , 27 ₂ , and 27 ₃ , and the other one light beam is transmitted by the beam splitter 2
7 is reflected by ₁ by transmitting the beam splitter 27 _2, 27 _3, by reflecting the other light beam by the beam splitter 27 ₂ through the beam splitter 27 _3, and further reflects the other light beam by the beam splitter 27 ₃ When the four light beams are combined with each other by appropriately selecting the reflectance of each beam splitter, the mixing ratios of the respective light beams can be made equal. In the case of this embodiment, the beam splitters 27 ₁ and 2
The reflectances of 7 ₂ and 27 ₃ are 1/2, 1/3, and 1 respectively.
When set to / 4, 1/4 of each input component to the optical path conversion element array 23a is multiplexed and output. This is obtained by setting (transmittance) + (reflectance) = 1 to the optical path conversion element array 2 for the output signal 28.
The simultaneous equations may be set so that the contribution ratios of the respective input components to 3a become equal.

In general, if there are K input components to the optical path conversion element array 23a, the (K-1) beam splitters in the optical path conversion element have reflectances of 1/2, 1/3, and
It may be set to 1/4, ..., 1 / K. As described above, the present embodiment is characterized in that the device configuration is simple and that the optical beam train is used for input / output, so that it can be directly connected to another optical circuit. Embodiment 2 As Embodiment 2 of the present invention, an example of performing demultiplexing / combining using polarization with four inputs and four outputs is shown in FIG. Also, FIG.
4A is a sectional view of the optical signal branching device 31 and the matrix mask 32 in the yz plane, and FIG.
A cross-sectional view on the y-plane is shown in FIG. 4B, respectively. The configuration of the device includes an optical signal branching element 31, a matrix mask 32,
It is composed of three parts of the optical signal multiplexing element 33, and in each part, the optical signal branching element 31 has the optical path changing element arrays 31a, 3
The matrix mask 32 includes a polarizing plate 32a and a liquid crystal element array 32b.
And the optical signal multiplexing element 33 is the optical path conversion element array 3
3a, 33b and quarter wave plates 33c, 33d.

The function of this embodiment is the same as that of the first embodiment, but in the first embodiment, the demultiplexing / combining is performed by the intensity division of light, whereas in this embodiment, the demultiplexing is performed by using the polarized light.・
The difference is that they combine waves. That is, a polarization beam splitter is used as the partial reflection element used for the demultiplexing / multiplexing portion in the first embodiment, instead of the beam splitter half mirror or the like. As a result, it becomes unnecessary to control the reflectance at the combining portion.

In practice, the optical signal branching element 31 has four linearly polarized incident light beam trains 24 each.
Circularly polarized light is polarized by the half-wave plate 31c, and the respective circularly polarized beams are demultiplexed into S-polarized light and P-polarized light with equal light intensity by the polarization beam splitter 34 inside the optical path conversion element array 31a, and the reflected light Of the polarized beam of
It is bent parallel to the transmitted polarized beam and arranged in the z-axis direction, and these four S-polarized and P-polarized light beams are each converted into circularly polarized light by the quarter-wave plate 31d. ,
These eight circularly polarized light beams are converted into an optical path conversion element array 3
Similarly in 1b, the polarized beam splitter 34 and the mirror 35 split the light into S-polarized light and P-polarized light. The 4 × 4 light beam array thus obtained is incident on the matrix mask 32 and undergoes individual light modulation.

Each of the light beams of the light-modulated 4 × 4 light beam train is incident on the optical path conversion element array 33a as circularly polarized light by the quarter-wave plate 33c. Optical path conversion element array 33a is a front polarizing beam splitter 34a arranged in the light beam every other, and a subsequent polarizing beam splitter 34b which are arranged for each light beam, P-polarized light P ₁ of the light beam B ₁ is The S-polarized light S ₁ transmitted through the front and rear polarization beam splitters 34a and 34b is reflected twice by the front polarization beam splitter 34a and twice by the rear polarization beam splitter 34b, and the adjacent light beam B ₂
P-polarized light P ₂ is transmitted through the rear-stage polarization beam splitter 34b and S-polarized light S ₂ is transmitted through the rear-stage polarization beam splitter 34b.
Is reflected twice. Similarly, S polarization and P polarization of adjacent light beams
The polarized light is combined with each other. 4 of each of these combined lights
Circularly polarized light is generated by the half-wave plate 33d, and similarly, the polarized beam splitter inside the optical path conversion element array 33b mixes the reflected S-polarized light and the P-polarized light that has been transmitted and goes straight to combine different beams. To do.

When the output destination is terminated by a detector array or the like, a lens system may be used only for the output destination. In FIG. 1, the optical signal branching element 21
Alternatively, the optical signal branching device 31 shown in FIG. 3 may be used. In this case, the matrix mask 22 is also constructed as shown in FIG. Further, only the optical signal multiplexing element 23 in FIG. 1 may be replaced with the optical signal multiplexing element 33 in FIG. In the optical signal branching device 21 of FIG. 2A, for example, the beam splitter for ⅓ reflection is used to divide the beam into two, and the two-thirds of the transmitted light is further divided into ½ and one light beam is divided into three. It is generally composed of a beam splitter, such as a book.

[0018]

As described above, according to the present invention, the lens system is composed of a reflective element and a partial reflective element,
The structure of the device is simplified and integration is facilitated.
Since the light beam train is input and the calculation result for the light beam train is output as a light beam train, it is possible to connect elements in multiple stages as one element of an optical circuit.

Since the light beam is demultiplexed and made incident on the matrix mask, one whole light beam is made incident on one element of the matrix, and the intensity of each beam can be made constant in all the elements. I can.

[Brief description of drawings]

FIG. 1 is a perspective view showing a first embodiment of the present invention.

2A is a sectional view of the optical signal branching element 21 and the matrix mask 22 in FIG. 1 taken along the yz plane, and B is a sectional view of the optical signal multiplexing element 23 shown in FIG. 1 taken along the xy plane. .

FIG. 3 is a perspective view showing a second embodiment of the present invention.

4 is a sectional view of the optical signal branching element 31 and the matrix mask 32 in FIG. 3 taken along the yz plane, and B is a sectional view of the optical signal multiplexing element 33 shown in FIG. 3 taken along the xy plane.

FIG. 5 is a diagram showing the principle of an optical vector matrix computing device.

FIG. 6 is a perspective view showing an optical vector matrix arithmetic device using a conventional lens system.

Claims (5)

[Claims] 1. N number of incident parallel light beams (N
Is an integer greater than or equal to 2), and each of the optical input signals is divided into M (M is an integer greater than or equal to 2) optical signals by using an optical path conversion element array composed of a combination of reflective elements and partially reflective elements.
An optical signal branching element that emits a parallel light beam of M rows and N columns, a matrix mask that individually modulates each of the M rows and N columns of optical signals that are emitted from the optical signal branching element, and a matrix mask that emits the light signal. An optical signal multiplexing element that combines the optical signals of M rows and N columns for each row by using an optical path conversion element array including a combination of a reflective element and a partially reflective element, and outputs the combined signal as M optical signals. An optical vector matrix computing device comprising: 2. The optical vector matrix computing device according to claim 1, wherein the partially reflecting element used in the optical signal branching element is a beam splitter for intensity-dividing a light beam. 3. The partial reflection element of the optical signal multiplexing element is N beam splitters having reflectances of 1/2, 1/3, ... 1 / N, respectively. Optical vector matrix computing device. 4. The partial reflection element used for the optical signal branching element is a polarization beam splitter, and a circular polarization conversion element is provided in the preceding stage of the optical path conversion element array. Optical vector matrix computing device. 5. The partial reflection element used in the optical signal multiplexing element is a polarization beam splitter, and a circular polarization conversion element is provided in the preceding stage of the optical path conversion element array. The optical vector matrix computing device described.