JPH0348314A - Optical neural network circuit - Google Patents

Abstract

PURPOSE:To easily attain the multielement and compact constitution of a Hopfield type associative storage by forming a Hopfield circuit using two-dimen sional light emitting element and light receiving element arrays only based upon spatial position relation between incoherent light sources and light receiv ing elements and the positions of spatial filters. CONSTITUTION:The light emitting element array part 31 formed by arraying plural light emitting element subarray parts 41 and the light receiving element array part 33 formed by arraying plural light receiving element subarray parts 43 are used, and only the light receiving element parts 43-1 arrayed on the 1st parts of respective light receiving segments constituting the array part 33 are irradiated with light emitted from the light emitting element parts 41-1 arrayed on the 1st parts of respective light emitting segments constituting the array part 31. Similarly, only the light receiving element parts 43-2 arrayed on the 2nd parts of respective light receiving segments are irradiated with light emitted from the light emitting element parts 41-2 arrayed on the 2nd parts of respective light emitting segments. Consequently, the multielement constitution of the Hopfield type associative storage circuit can simply and compactly be obtained.

Description

[Detailed Description of the Invention] [Objective of the Invention] (Industrial Field of Application) The present invention relates to an optical neural network circuit, and more particularly, to multi-element Hopfield type content addressable memory circuit using optical spatial wiring technology. Regarding.

(Prior Art) In recent years, in fields such as pattern recognition and associative memory, there has been a growing movement to incorporate the structure of biological neural networks into data processing devices to realize data processing methods different from conventional ones. The structure of such an artificially reproduced neural network is generally called a neural network, and is usually
It consists of node parts called synapses that perform simple signal processing, and wiring parts called neurons that connect these synapses and transmit signals, and the overall number of synapses are complicatedly wired by neurons. It is shaped like this.

One form of neural network is called the Hopfield model. In this model, each synapse takes one of two states, O or 1, and its state changes over time depending on the connection state of neurons. The state change of each synapse is represented by a matrix T- (T L j ) indicating the connections of neurons. When the number of synapses is n, the state of the i-th synapse at a certain moment is Vt, and the state of the i-th synapse when one step has passed from this state is vi'. That is, each synapse receives the states of other synapses during the time of one step, adds weights defined by the matrix W, sums them, performs threshold processing, and uses the result to determine the white separation. Determine the condition of the white meat. Here, there are m pieces of n whose elements take two values, 0 or 1.
Consider the original vector U(1)U(21, ●... lJL+aゝ.At this time, if 1-(I i jl is an n-by-n unit matrix and the matrix T is defined by (2), then this The model assumes that when a certain initial state is given to a synabus, the state of the synabus is U (1
1, U(2ゝ, An associative memory function is obtained that outputs the vector closest to the input vector from several vectors rl1.

Also, even if the matrix T defined by the formula obtained by removing the jp-order matrix term from the right side of formula (2) is used instead of T,
Almost the same associative memory function can be obtained.

The associative memory method using the Hopfield model described above requires that all synapses are connected to all synapses other than one via neurons for n minutes each.Therefore, the number of synapses is n. In the case of n(n
-1)/Two neurons capable of bidirectional signal transmission are required. Therefore, if a Hopfield circuit is to be realized by an electric circuit, as the number of processors corresponding to synapses increases, the number of wires corresponding to neurons increases in proportion to the square of the number of processors. Therefore, a multi-element Hopfield circuit requires a very large number of intricately intersecting wiring lines, making it difficult to realize a high-speed circuit using electrical wiring.

In order to solve the above problem, a Hopfield circuit using optical spatial wiring technology has been proposed. Optical space wiring is a method that uses light propagating in space as a signal transmission medium, and can perform high-speed signal transmission without using physical signal transmission lines.

Figure 6 is from the document rD. Psalltls. et al. ．．
”OpticalInforsaLonProce
ssing Basod On An Assocla
tlye oneto parable ewory ＞dinodol −
z●, Opt. LoLt. ．． Vol. l
O. pp. 98-100 (published in 1985)" is a diagram illustrating the configuration of a Hopfield circuit using the optical spatial wiring technology disclosed in 1985. In the figure, 1 is a one-dimensional light emitting element array with n elements, 2 is a light attenuation filter array arranged in a matrix of n rows and n columns, and 3 is n elements.
It is a one-dimensional light-receiving element array, and simultaneously has a function of thresholding the received light intensity and driving the light-emitting element array 1. The light emitted from the j-th element of the one-dimensional light emitting element array 1 illuminates the entire j-th column of the light attenuation filter array 2, and only the light that has passed through the filter located at the l row and j column is one-dimensional. It is coupled to the i-th element of the light receiving element array 3. Here, when there is no filter array 2, the light intensity emitted from one element in the light emitting element array 1 and detected by one element in the light receiving element array 3 is p1.
The transmittance of the filter at the i-th row and j-column position in the filter array 2 is Sij, and V-iVil is a vector representing the light emission state of each element in the light-receiving element array 1 (when the i-th element emits light, Vi -1, and Vi-0 when no light is emitted.

At this time, m n-element vectors U (1) . whose elements take two values, 0 or 1. Ul2+ , , , *
, U′′″〉, the light intensity Pi detected by the l-th element in the light receiving element array 3 is Pl−lllp●Σ 1−1 S+IV1 Therefore, the light intensity received by the i-th element in the array 3 at the light intensity threshold p s (lh) is pH + lh
) or above, the i-th element in the light-emitting element array 1 is made to emit light, and when p 1 (+1 or less), it is processed so that it does not emit light, thereby resolving equation (1) and Oka. It is possible to realize threshold processing such as the above, and to perform the operation of the Hopfield circuit.

In the method described above, since both the light emitting element array 1 and the light receiving element array 3 are one-dimensional arrays, the number of elements can be reduced by 2 by arranging the light attenuation filter array 2 two-dimensionally.
The number of wires proportional to the power can be arranged spatially without difficulty. However, in this method, when the light-emitting element array and the light-receiving element array are made two-dimensional, it becomes difficult to reasonably arrange the filter array corresponding to the wiring portion in space. Therefore, with this method, it is difficult to increase the number of elements using a two-dimensional light emitting element array and a light receiving element array.

In order to solve the above-mentioned difficulties, a method has been disclosed in which light emitted from each element in a two-dimensional light-emitting element array is distributed across a hologram to each element in a two-dimensional light-receiving element array.

FIG. 7 is from the document rJu-Seog Jang. etal.
．． ”OpticalImplesentat1on
Or The Ilopl'eld Model
...OpL. LeLt. ．． Vol. l3. pp. 24
8-250 (published in 1988)" is a diagram showing the configuration of a Hopfield circuit using a hologram as described above. In the figure, 11 is an argon ion laser, 12 is a collimating lens, and 13 is 2
14 is a two-dimensional hologram array, and 15 is a two-dimensional light receiving element array. Laser 11
The light emitted from the switch array 13 is turned into a parallel light beam by the lens 12 and enters the switch array 13 . switch array 13
Each switch transmits or blocks light depending on its on or off state. Light transmitted through one switch in the switch array 13 is incident on one hologram in the hologram switch 14 corresponding to that switch. The first-order diffracted light generated by this hologram is distributed to each element in the light-receiving element array at a predetermined ratio.

Here, V-(Vil is a vector indicating the state of the switch array 13 (if the i-th switch is on, Vi
-1, Vi-0 in the case of off), p is the intensity of the laser light transmitted through the switch in the on state in the switch array 13, and Sij is the intensity of the light receiving element 15 from the i-th hologram in the hologram array 14. Let it be the ratio of the intensity of the diffracted light that reaches the j-th element in the hologram and the intensity of the incident light to this hologram.

At this time, if Sij is determined according to the above equation (4) and the output of each element in the light receiving element array 15 is subjected to the same threshold processing as in the case of FIG. It is possible to operate as a Hopfield circuit based on this principle. In this way, by using the method described above, a neural network model using two-dimensional arrayed elements can be realized.

(Problem to be Solved by the Invention) In the above-mentioned method, it is necessary to distribute the same number of lights as the number of elements in one hologram in the hologram array, so as the number of elements increases, it is necessary to downsize each hologram. This poses a problem in that it becomes difficult to reduce the size of the entire device, and also that the device becomes complicated because it requires a coherent optical system.

The present invention has been made in view of the above, and its purpose is to easily realize a multi-element Hopfield type content addressable memory circuit using optical wiring space technology and to realize a compact optical neural network. The purpose is to provide circuits.

[Structure of the invention]

(Means for Solving the Problems) In order to achieve the above object, the optical neural network circuit of the present invention includes a light emitting element array part in which a plurality of emitting element subarray parts each consisting of a light emitting element part that emits light, and an incident light emitting element subarray part. Detects the sum of the outputs of a light-receiving element array part in which a plurality of light-receiving element sub-array parts each consisting of a light-receiving element part that detects the intensity of the light to be detected, and a light-receiving element part belonging to one of the light-receiving element sub-array parts; a threshold processing circuit section that compares the detected output with a predetermined threshold value and generates a binary output for each of the light receiving element subarrays; a light-emitting element drive circuit unit that causes the light-emitting element parts to emit or extinguish light simultaneously; and a light-emitting segment having first and second parts that are provided in each of the light-emitting element sub-array parts and can have one light-emitting element part. , a light-receiving segment G that is provided in each of the light-receiving element sub-array parts and can have one light-receiving element part, and a light-receiving segment that belongs to one of the light-emitting element sub-array parts;
The light emitted from the light emitting element part disposed in the first part of each light emitting segment is coupled so that the light emitted from one light emitting element part is coupled only to one light receiving element part constituting the light receiving element subarray part. Only the light-receiving element part attached to the first part was irradiated, and the light from the light-emitting element part arranged in the second part of each light-emitting segment was arranged in the second part of each light-receiving segment. The main feature is to include an optical coupling circuit section that illuminates only the light emitting element section.

(Function) The optical neural network circuit of the present invention uses a light-emitting element array part in which a plurality of light-emitting element sub-array parts are arranged and a light-receiving element array part in which a plurality of light-receiving element sub-array parts are arranged. The light from the light emitting element section disposed in the first part of the segment illuminates only the light receiving element section disposed in the first part of each light receiving segment constituting the light receiving element array section. Light from the light emitting element portion disposed in the second portion of each light receiving segment irradiates only the light emitting element portion disposed in the second portion of each light receiving segment.

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

FIG. 1 is a diagram showing the configuration of a first embodiment of the present invention.

In the figure, 31 is a light emitting element array, 32 is a spatial filter corresponding to an optical coupling circuit section, and 33 is a light receiving element array. In the figure, the circles shown in the light emitting element array 31 and the light receiving element array 33 represent light emitting elements and light receiving elements, respectively. Furthermore, 34 is a threshold value processing circuit, which performs threshold value processing on each of the plurality of outputs from the light receiving element array 33. 35 is a light emitting element drive circuit, which generates an output proportional to the number of light emitting elements in a light emitting state. Both the light emitting element array 31 and the light receiving element array 33 are divided into a plurality of subarrays as shown by the thick solid lines. The number of subarrays in the light emitting element array 31 and the light receiving element array 33 are equal to each other, and the number of light emitting elements or light receiving elements included in each subarray is also mutually equal.

FIG. 2 shows a light emitting element array 31 and a light receiving element array 33.
FIG. 3 is a diagram showing an example of the structure of a subarray. In the figure, 41 is a light emitting element sub-array, and 42 is a spatial filter 3.
A part of 2, 43, indicates a light receiving element subarray.

Each subarray is divided into a plurality of segments each containing one light emitting element or one light receiving element.
,2. ... is numbered. Each segment has two positions H and L where the element should be placed, and the actual element is placed at either H or L. The spatial filter 42 is located at the same position within the segment with the same number of the light-emitting element sub-array 41 and the light-receiving element sub-array 43,
That is, it has a function of controlling the spatial optical path so that light is transmitted only between the elements arranged between H and H or between L and L. Such a function can be easily realized if each segment of the light emitting element subarray 41 and the light receiving element subarray 43 and the positions of the elements within the segment are arranged symmetrically with respect to the light transmission hole of the spatial filter 42. In addition, by arranging the centers of all the light emitting element sub-arrays and light receiving element sub-arrays and the light transmission holes of the spatial filter on mutually similar two-dimensional grid points, all functions of the spatial filter 42 are performed for the transmission and reception of light between the sub-arrays. It can be realized.

Here, the light emitting element array 31 and the light receiving element array 33
Assuming that the number of subarrays included in is N and the number of segments included in each subarray is M, this device can have an associative memory function that stores M N-element vectors. Let these storage vectors be [J ( l l
[J (2 1 UIM), the positions of the light emitting element and the light receiving element within each segment are determined as follows.

In other words, the i-th 11th element of the k-th storage vector is 0.
When , both the light-emitting element and the light-receiving element in the segment numbered k in the i-th subarray in each array are placed at position L. On the other hand, when the above element is 1, both the light emitting element and the light receiving element are placed at position H. When arranged in this way, the light emitted from the light emitting elements of the segment number k in the light emitting element subarray is located at the same position among the light receiving elements belonging to the segment number k in all the light receiving element subarrays, i.e. It reaches only the elements arranged between H and H or L and L, and does not reach the elements at mutually different positions, that is, the elements where one is H and the other is L.

In the operation of this device, first, when the i-th element of V is 1, all the light-emitting elements in the i-th light-emitting element sub-array emit light, and when it is 0, they do not emit light. Set to . At this time, if the optical power transmitted from one light-emitting element to one light-receiving element is p, the total optical power Pj received by the light-receiving elements included in the light-receiving element array of number j 1 is given by the following formula: It is expressed as

Pr −p” Σ V+ E Yn(U+
”' , IJ+ ”' ) (2U+ ”'
−1)+++/2(2U, 1) + l + . /2 (2U, -1)l/2+M/21 However, matrix 1-(Iijl represents a tIt-order matrix with N rows and N columns, and matrix Tij has the same definition as Equation (2).

Represents the number of arrays. Here, if n is large enough, the formula (
The term Vj in the last equation of 6) is negligibly small compared to the other terms in this equation. Therefore, the output of the light emitting element drive circuit 35 is set to p M / 2 for 1 m of light emitting element subarrays emitting light, and the output of each light receiving element subarray is compared with the above output in the threshold processing circuit 34. By controlling the light emitting state of the light emitting element, the circuit shown in FIG. 1 operates as a Hopfield circuit having an associative memory function.

By the way, in the above-mentioned first embodiment, p M is set to the number of emitting subarrays as a reference value for threshold processing.
/ 2 is used. Therefore, if there are variations in the light emitting power or its spatial distribution of individual light emitting elements, the value of p mentioned above will differ from element to element, and as a result, there will be a difference between the above reference value and the reference value that should truly be set. Since errors occur, performance as an associative memory device is impaired. This problem can be solved by adding a function to offset the optical power corresponding to the reference value.

FIG. 3 is a diagram showing a second embodiment of the present invention with the above-mentioned functions added. In the figure, 51 is a light emitting element array, 52 is a spatial filter corresponding to an optical coupling circuit section, and 53 is a light receiving element array. In the figure, the ● mark represents a light emitting element or a light receiving element, respectively. A threshold processing circuit 54 performs threshold processing on each of the plurality of outputs from the light receiving element array 53. 55 is a light emitting element drive circuit. Light emitting element array 51 and light receiving element array 53
Both are divided into subarrays as shown surrounded by thick solid lines. The number of subarrays in the light emitting element array 51 and the light receiving element array 53 is equal to each other, and is 1.
The number of light receiving elements included in one light emitting element subarray is twice the number of light emitting elements included in one light emitting element subarray.

FIG. 4 shows a light emitting element array 51 and a light receiving element array 53.
FIG. 3 is a diagram showing an example of the structure of a subarray. In the figure, 61 is a light emitting element sub-array, and 62 is a spatial filter 5.
A part of 2, 63, indicates a light receiving element sub-array.

The light emitting element sub-array 61 is divided into a plurality of segments each containing one light emitting element, and each segment is sequentially numbered 1.2, ●. Further, the light-receiving element sub-array 62 is divided into a plurality of segments each including two light-receiving elements, and these segments are also sequentially numbered. Each segment has two positions H and L in which the elements should be placed, the light emitting element being placed at either H or L, and the light receiving element being placed at both H or L. As in the first embodiment, the spatial filter 62 includes a light emitting element subarray 61.
The spatial optical path is controlled so that light is transmitted only between elements arranged at the same position within the segment with the same number of the light-receiving element sub-array 63, that is, H and H or L and L. ing.

Here, as in the case of the first embodiment, the number of subarrays in each array is N, the number of elements in the subarray is M, and the storage vector is U(1)UT2),●●◆
[J(Ml). At this time, the arrangement of the light emitting elements in each light emitting element subarray is determined by the same method as in the first embodiment. On the other hand, each light receiving element in the light receiving element subarray is set with an attribute of + or 1 according to the value of the element of the storage vector. That is, if the value of U(kl l is 1, the number k in the first photodetector subarray
Let the element at position H in the segment be 10, and let the element at position L
Let the element in . Further, when the value of U(k)1 is 0, the light receiving element located at position H in the segment is set as -, and the element located at position L is set as 10. After such settings are made, the sum of the outputs of ten elements and the sum of the outputs of one element are separately output from each light receiving element subarray. With this setting, light emitted from the light emitting element in the segment numbered k in the light emitting element subarray is transmitted to the corresponding memory vector element among all the light receiving elements belonging to the segment numbered k in the light receiving element subarray. are equal to each other, only ten elements are reached, and if the elements of the corresponding storage vectors are different from each other, only one element is reached.

In the operation of this device, the light emitting state of each light emitting element subarray is first set in accordance with the value of the element of the human power vector V, as in the case of the first embodiment. At this time, if the optical power transmitted from one light-emitting element to one light-receiving element is p, then the sum Pj of the optical power received by the ten light-receiving elements included in the j-th light-receiving element sub-array and - The total sum P-j of the optical powers received by the light-receiving elements is expressed by the following equations.

(Ul U+ ) (2U, =1) +11 /2 (7) (Ul U, ) (2U, -1)+l+/2 (8) Therefore, if we take the output difference between 10 light receiving elements and 1 light receiving element, we get , (2U, -1) N - p · Σ v1 (T,, + MI ■) p N · { (Σ ■ . Tt+) + MV, ) (9) is obtained as the output of the j-th light receiving element sub-array. Here, if N is sufficiently large, the Vj term in equation (9) becomes negligibly small compared to other terms. Therefore, if the threshold processing circuit determines whether the differential output obtained from each light-receiving element sub-array is positive or negative, and the light-emitting element drive circuit 55 is controlled using the result, the circuit of this embodiment is a hop with an associative memory function. Operates as a field circuit.

In the circuit of this example, the positive and negative light receiving elements are arranged in the light emitting element sub-array, and the difference in output between the two is used to cancel out the influence of variations in the light intensity of the light emitting elements. It is possible to operate more stably than in the first embodiment.

FIG. 5 shows a circuit for obtaining the output difference between + and 1 light-receiving elements from each light-receiving element sub-array.

In the figure, 71-1 to 71-n are ten light receiving elements;
2-1 to 72-n represent negative light receiving elements.

Generally, a light receiving element that converts optical output into an electrical signal outputs a photocurrent proportional to the optical intensity. Therefore, if the elements are connected as shown in Figure 7, the current flowing through point A in the figure is the sum of the photocurrents generated by ten light-receiving elements minus the total photocurrent generated by one light-receiving element. become. Therefore, if the resistance between points A and B in the same figure is R, then A
A voltage equal to the value obtained by multiplying the difference in photocurrent to be determined by R is observed between the point and the point B. Therefore, by observing this voltage, it is possible to know the difference in light intensity that should be determined.

In this embodiment, a spatial filter with light transmission holes provided on the lattice points is used as the optical coupling circuit section, but the present invention is not limited to this. For example, instead of the above spatial filter, the lattice points The optical coupling circuit section may be constructed by a lens array in which convex lenses are arranged so that the center thereof is located at the top.

Further, in this embodiment, an electric circuit is used as the threshold value processing circuit section and the light emitting element drive circuit section, but the present invention is not limited to this. An array of elements may be arranged in the light receiving element array portion, and light generated from this array may be fed back to a position facing the light receiving element array by an optical method.

〔Effect of the invention〕

As explained above, according to the present invention, it is possible to realize a Hopfield circuit that is superior to a two-dimensional light emitting element array and a light receiving element array by only using the spatial positional relationship of an incoherent light source and a light receiving element and the installation of a spatial filter. Therefore, an optical system can be configured only with a light emitting element array, a light receiving element array, and a spatial filter, simplifying the device, and easily realizing multi-element and miniaturization of the Hopfield type content addressable memory device.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the structure of an embodiment of the present invention, FIG. 2 is a partially enlarged configuration diagram of the embodiment of FIG. 1, and FIG. 3 is a diagram showing a second embodiment of the present invention.
Figure 5 is a circuit diagram of a circuit that outputs the difference between the photocurrents of the 10 and - light receiving elements. The figure is a block diagram of a Hopfield circuit using conventional optical space wiring technology.
The figure is a configuration diagram of a Hopfield circuit in which a two-dimensional array of elements is formed using the optical wiring function of a conventional hologram. 31...●Light emitting element array, 32... Spatial filter, 33... Light receiving element array, 34... Threshold processing circuit, 35... Light emitting element drive circuit, 41...●Light emitting element sub-array, 42.・● Spatial filter, 43... Light-receiving element sub-array.

Claims (3)

[Claims] (1) A light-receiving element array section in which a plurality of light-emitting element sub-array sections consisting of light-emitting element sections that emit light are arranged, and a plurality of light-receiving element sub-array sections consisting of a light-receiving element section that detects the intensity of incident light. Detecting the sum of the outputs of the element array section and the light receiving element section belonging to one of the light receiving element sub-array sections, and comparing the detected output with a predetermined threshold;
a threshold processing circuit section that generates a binary output for each of the light receiving element subarrays; and a light emitting element that simultaneously causes all light emitting element sections belonging to one of the light emitting element subarray sections to emit or extinguish light according to an output signal of the threshold processing circuit section. a light emitting segment provided in each of the light emitting element subarray parts and having first and second parts that can have one light emitting element part; and a light emitting segment provided in each of the light receiving element subarray parts. , a light-receiving segment having a first and a second portion capable of having one light-receiving element portion, and light emitted from one light-emitting element portion belonging to one of the light-emitting element sub-array portions, forming the light-receiving element sub-array portion. The light from the light emitting element part disposed in the first part of each light emitting segment is coupled to only one light receiving element part disposed in the first part of each light receiving segment. and an optical coupling circuit section for irradiating light from the light emitting element section arranged in the second part of each light emitting segment to only the light emitting element section arranged in the second part of each light receiving segment. An optical neural network circuit characterized by: (2) Each light emitting element part included in the light emitting element subarray part is provided only in either the first or second part of the light emitting segment, and each light receiving element part included in the light receiving element subarray part is provided in the first or second part of the light emitting segment. It is provided only in either the first or second portion of the light-receiving segment, and the predetermined threshold value is an amount proportional to the sum of the intensity of light from all light-emitting element portions in a light-emitting state. The optical neural network circuit according to claim (1). (3) Each light emitting element part included in the light emitting element subarray part is provided only in either the first or second part of the light emitting segment, and each light receiving element part included in the light receiving element subarray part is provided in the first or second part of the light emitting segment. The light receiving element sub-array section is provided in both the first and second parts of the light receiving segment, and the light receiving element sub-array part outputs the sum of the differences in the outputs of the two light receiving element parts included in each light receiving segment included in the light receiving element sub-array part. 2. The optical neural network circuit according to claim 1, wherein the threshold value processing circuit section executes the processing by setting the threshold value for the output of the light receiving element sub-array section to zero.