JPH0261709A - Optical arithmetic unit - Google Patents

Abstract

PURPOSE:To quickly execute the algorithm of a neurocomputer by using interconnection of light to connect respective interlayer neurons. CONSTITUTION:The optical switch of an optical arithmetic unit is provided with optical function element arrays 1 to 4 of light emitting elements, light receiving elements, etc., spatial optical modulators 5 to 7 like liquid crystal televisions, condenser lenses 8 to 11, a beam splitter 12, a driving device, modulator driving devices 17 to 19, and arithmetic units 20 to 23 like personal computers. The light from this light emitting part is modulated by the output of neurons of the input layer and is condensed on light receiving parts of facing arrays 1 to 4 and is outputted. Patterns of modulators 5 to 7 are changed by the weight value of this network, and the pattern of the modulator 3 is changed in accordance with a learning rule by an arithmetic unit, thereby performing learning.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] This invention relates to a device that processes data at high speed using a neural network using light.

[Prior art and its challenges] Research is progressing on computers that can perform calculations at high speed in order to process large-scale information, but methods that use sequential processing using electric circuits are already approaching their performance limits. . Therefore, supercomputers, array processors, etc.
Research is progressing on parallel processing architectures that perform multiple operations simultaneously. On the other hand, light has a spatial expanse and its physical properties do not interfere with each other, so operations using light have excellent parallelism. As means for modulating light, amplitude, phase, frequency, polarization, etc. can be considered, and spatial light modulators are being developed.

So far, as a neurocomputer using light,
A method of optically realizing the Hopfield model's ink connection by connecting a one-dimensional light emitting element and a one-dimensional light receiving element with an anamorphic optical system, and a method of optically realizing the ink connection of the Hopfield model, and a method that modulates input data with a spatial light modulator to create a bidirectional A method is known that enables the calculation of . For information on how to optically realize the Hopfield model, see, for example, the magazine OPTIC5LETTERS, Vol. 10, 1.
985, pages 98 to +00, the paper ``Optical Information Processing Using an Associative Memory Model Using a Bi-ral Net with Threshold Processing and Feedback Functions (0ptical
Information process ba
sed on an associative-mem
ory model of neura,l
nets vlth threshotdlng
andfeedback) is described in detail in J. Further, regarding the method of modulating input data with a spatial light modulator, see, for example, the magazine Applied Optics, Vol. 26, 19.
1987, pages 5055-5060 of the paper "Design and Device of Optical Bidirectional Associative Memory (Deslgns
and devices for optlcat
bldlrectlonal assoclatlv
e mellorles) J. However, these methods cannot perform learning in real time and cannot create more general neuromodels.

The purpose of this invention is to use optical ink connections to
An object of the present invention is to provide an optical arithmetic device that can execute a learnable multilayer neurocomputer algorithm at high speed.

[Means for Solving the Problem] This optical processing device is a two-way computer using an optical ink connection, in which an optical functional element in which a light emitting part and a light receiving part are arranged adjacent to each other is arranged in an IJ box shape. a first surface having the same structure as the first surface and connecting each of the optical functional elements in the row direction or the column direction;
The surface is arranged to face the first surface so that the light emitted from each light emitting section of the surface enters each light receiving section, and the light emitted from each light emitting section enters each light receiving section of the first surface, A second surface has optical functional elements connected in a direction perpendicular to the connection direction of the optical functional elements on the first surface, and has the same structure as the first surface, and emits light from each light emitting part on the second surface. The light receiving section is arranged to face the second surface so that the light emitted from each light emitting section enters each light receiving section, and the light emitted from each light emitting section enters each light receiving section on the second surface. A third surface has the same structure as the first surface and a third surface in which each optical functional element is connected in a direction perpendicular to the connection direction of the elements, and the light emitted from each light emitting section on the third surface is directed to each light receiving section. It is arranged to face the third surface so that the light incident thereon and emitted from each light emitting section enters each light receiving section on the third surface, and is orthogonal to the connection direction of each optical functional element on the third surface. a fourth surface on which each optical functional element is connected in the direction;
Light emitted from the light emitting section on the first surface enters the light receiving section on the second surface, light emitted from the light emitting section on the second surface enters the light receiving section on the third surface, and the light emitted from the light emitting section on the second surface enters the light receiving section on the third surface. a beam splitter arranged so that the light emitted from the light emitting part of the surface enters the light receiving part of the fourth surface, and modulating the intensity of each light emitted from the first surface, the second surface, and the third surface. a plurality of optical modulators that calculate the difference between the output from each light receiving section of the optical functional element and a desired output, and calculate a factor of the optical modulator that reduces the difference; It is characterized by comprising: an arithmetic device; and a plurality of modulator driving means for changing the pattern of the light modulating means based on the output from the arithmetic device.

[Operation] The principle of this invention will be explained with reference to FIGS. 2, 3, and 4. FIG. 2 is an example of a three-layer neural network. Input layer 101, intermediate layer 102 and output layer 10
Each of 3 has a plurality of neurons, and each neuron of the input layer is connected to each neuron of the intermediate layer, and each neuron of the intermediate layer is connected to each neuron of the output layer. The strength of the connections between each neuron can be changed independently, and by changing this strength,
The network can be trained.

FIG. 3 shows the configuration of a device that optically realizes a three-layer neural network.

The input surface 201 corresponds to an input layer, the intermediate surface 202 corresponds to an intermediate layer, and the output surface 203 corresponds to an output layer. Signals are transmitted between neurons through light. The strength of the network connections is varied by spatial light modulation elements 204, 205 between each plane. FIG. 4 shows (a) the relationship between the output A of each neuron in the input layer and the input surface 201, (b) the relationship between the state of each neuron in the intermediate layer and the intermediate surface 202, and (C) the relationship between the output layer and the input surface 201. FIG. 3 is a diagram showing the relationship between the input B and the output surface 203 of each neuron. The outputs of the four neurons in the input layer A, , A2 . A, , A4 are connected to the light emitting elements in the row direction, and depending on the size of input data,
The intensities of the light emitting elements in the row direction are simultaneously modulated. for example,
When focusing on input A2, A, which is an element in the row direction of A2
211 A 221 A 231 A 2J all emit light. After the emitted light is collimated, it is focused on light receiving sections B21, B2°, B23, and B24 on the output surface. The other elements operate in the same way, and the elements in the column direction of the output surface are connected, so for example, when focusing on output B3, the elements in the column direction of B3, B13, B
23, B 33, and B 43 is output. This relationship can generally be expressed as follows.

Therefore, in order to realize a delta learning agent, it is sufficient to change the weight W according to the following equation.

Δ W ,=n ('r, I-o#I) ■
(3) At this time, if a spatial light modulator is inserted between the input surface and the output surface and the amplitude transmittance of the portion corresponding to each element is changed independently, (1)
The formula is as follows.

Here, Wl represents the amplitude transmittance of the spatial light modulator.

Equation (2) shows that the input of the neuron in the output layer is the output A of the neuron in the input layer, and the weight W of the neural network.
This means that it is the product of . This is equivalent to a general neural model, and various neural networks can be constructed using this optical calculation device. For example, using this arithmetic device, n is the number of times of learning NT11 is the desired result, 0° is the output of the optical arithmetic device, and IP is the weighting of learning. For details on the delta learning agent, see, for example, the magazine BYTE, 1987 IO issue, Nos. 155-162.
The paper written on page ``Bac Propagage Gin (Bac
k-Propagatlon) J. Furthermore, in this device, the light emitting element and the light receiving element are arranged adjacent to each other as shown in FIG. 3, and the data received can be processed to immediately modulate the adjacent light emitting element. Therefore, data can be processed sequentially from the first side to the second side, from the second side to the third side, and from the third side to the fourth side, and a neural network with a maximum of four layers can be realized.

In this way, if the intensity of the light emitting element is changed according to the output of the neuron in the input layer, and the amplitude transmittance of the spatial light modulator is changed according to the weight value, it is possible to optically create a multilayer structure such as a packed propagation type. Neural network interconnection can be realized. For details on Pack Properage Day, please see the magazine Nature + (Nature).
The article “Learning representations by backward transfer of errors (Lea
rnlng Representations by 1
3ack-Propagating Errors)
It is detailed in J.

[Example code] The present invention will be explained in detail below.

FIG. 1 is a perspective view showing an example of an embodiment for realizing the optical arithmetic device of the present invention. This optical switch is composed of an optical functional element array 1.2.3.4 such as an 0EIC in which a light-emitting element such as a semiconductor laser and a light-receiving element such as an Sl photodetector are arranged adjacent to each other, and a TN liquid crystal, for example. Spatial light modulator 5 capable of expressing gradations for LCD TVs, etc.
．． 6.7, and a condensing lens 8. such as a microlens array, which collimates the light emitted from the light emitting part of the optical functional element and focuses the light transmitted through the spatial light modulator onto the optical functional element array. 9. Apply voltage to the spatial light modulator, a device consisting of a beam splitter 12 that connects the optical functional element arrays 1 and 2. 2 and 3 and 3 and 4, and a circuit that applies voltage to the optical functional elements. modulator drive devices 17, 18, and 19 composed of circuits that perform 20, 2
1, 22, and 23.

In the optical arithmetic device having the above configuration, the light emitting section of the optical functional element is modulated by the output of the neuron in the input layer. All the elements in the row direction of the optical functional element array l are electrically connected, and each element in the column direction of the optical functional element array 2, each element in the row direction of the optical functional element array 3, and the optical functional elements All elements in the column direction of the array 4 are also electrically connected.

The light emitted from the light emitting section of the optical functional element is modulated by the spatial light modulator, focused on the light receiving section of the opposing optical functional element array, and becomes an output. Furthermore, the pattern of the spatial light modulator is changed depending on the weight value of the neural network. Furthermore, learning is possible by changing the pattern of the spatial light modulator according to the learning rule determined by the neural model such as equation (3) using the arithmetic unit with respect to the obtained output and desired result. . In this way, neural network interconnection can be realized optically. At this time, since the light receiving section and the light emitting section are arranged adjacent to each other, the received light data can be processed by the arithmetic unit and the adjacent light emitting section can be immediately modulated.

[Effects of the Invention] As detailed above, by using the optical arithmetic device of the present invention, it is possible to execute a learnable multilayer neurocomputer algorithm at high speed.

[Brief explanation of the drawing]

FIG. 1 is a perspective view showing an embodiment of the optical calculation method of the present invention;
Figure 2 shows an example of a three-layer neural network, Figure 3 shows the configuration of a device that optically implements the neural network, and Figure 4 shows the relationship between input/output data and the number of people. It is a diagram. In the figure, 1, 2, 3, 4 ●● Optical functional element array 5, 6, 7
●●Spatial light modulation element 8, 9, 10, 11●Collimating lens 12
●●●● Beam splitter 17, 18, 19 ●●
●●Modulator drive device 20, 21, 22, 23●● Arithmetic unit 101 I●●●●Input layer 102 ●●Life @ conflict middle layer 103 ●●●●●Output layer 201 ●●0●●Input surface 03 ．． Half figure of Nofuno vomit walking (a) Ha 2014 person n-side

Claims (1)

[Claims] In a neurocomputer using optical interconnection, optical functional elements each having a light emitting part and a light receiving part arranged adjacent to each other are arranged in a matrix, and each of the optical functional elements is connected in a row direction or a column direction. one side, and the first side
The surface has the same structure as the first surface, and the light emitted from each light emitting section on the first surface is incident on each light receiving section, and the light emitted from each light emitting section is incident on each light receiving section on the first surface. a second surface that is disposed facing the first surface and has optical functional elements connected in a direction orthogonal to the connecting direction of the optical functional elements on the first surface; and a second surface that has the same structure as the first surface. , facing the second surface so that the light emitted from each light emitting section on the second surface enters each light receiving section, and the light emitted from each light emitting section enters each light receiving section on the second surface. a third surface having the same structure as the first surface and having the optical functional elements connected in a direction perpendicular to the connection direction of the optical functional elements on the second surface; The third surface is disposed facing the third surface so that the light emitted from the light emitting section enters each light receiving section, and the light emitted from each light emitting section enters each light receiving section on the third surface.
a fourth surface in which each optical functional element is connected in a direction perpendicular to the connection direction of each optical functional element on the surface, and light emitted from a light emitting section on the first surface enters a light receiving section on the second surface; The light emitting unit is arranged so that the light emitted from the light emitting part on the second surface enters the light receiving part on the third surface, and the light emitted from the light emitting part on the third surface enters the light receiving part on the fourth surface. a beam splitter, a plurality of light modulating means for modulating the intensity of each light emitted from the first surface, the second surface, and the third surface, and an output from each light receiving section of the optical functional element and a desired output. a plurality of arithmetic devices that calculate a difference between the two and a pattern of the light modulator that reduces the difference; and a plurality of modulation devices that change the pattern of the light modulation means according to the output from the arithmetic device. 1. An optical arithmetic device comprising a device driving means.