JPH02293987A - Optical neural net circuit - Google Patents

Abstract

PURPOSE:To facilitate the multi-element conversion by placing a light emitting element and a light receiving element on a two-dimensional array, and attaining the setting of a coupling coefficient between both of them by controlling a plane of polarization of a linearly polarized light radiated by the light emitting element and a plane of polarization of a linearly polarized light component detected by the light receiving element. CONSTITUTION:Light intensity of a linearly polarized light radiated from an element array part 21 is detected by a light receiving element array part 24, a photocurrent generated from the light receiving element array part 24 is integrated, this integral value is compared with a threshold, and in accordance with a result of this comparison, a control is executed so that each light emitting element part of the light emitting element array part 21 is brought independently to light emission or extinction. Accord ingly, by only covering each element of the light emitting element array 21 and the light receiving element array 24 with space filter arrays 23a, 23b having a function for rotating a plane of polarization, a hop field type associative storage device can be realized. In such a way, comparing with a conventional technology using a holo gram, a necessary space per one element on the array can be reduced remarkably, and the multi-element can be realized easily.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Field of Application) The present invention relates to an optical neural network circuit that can realize multi-element Hopfield type content addressable memory circuits using optical spatial wiring technology.

(Prior art) In recent years, in fields such as pattern recognition and associative memory, efforts have been made to incorporate the structure of biological neural networks into data processing devices to realize data processing methods different from conventional methods. . The structure of such an artificially reproduced neural network is generally called a neural network, and it usually has a node part called a synapse that performs simple signal processing, and a node part called synapses that performs simple signal processing and connects these synapses to process the signal. It consists of wiring parts called neurons that carry out transmission, and as a whole, many synapses are intricately wired by neurons.

One form of neural network is called the Hopefield model. In this model, each synapse takes one of two states, 0 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=(Tijl) indicating the connections of neurons.

When the number of synapses is n, the state of the i-th synapse at a certain moment is Vi, and the state of the i-th synapse when one step has passed from this state is V1°, then V1=L
h(.Σ TIjVj) ...(1)j-
1 However, th(x)=1 (X≧0)0 (x<O
) becomes. 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 love processing on this, and uses the result to determine its own state. determine the state of Here, m n-element vectors U(11.[J(21,...
... Consider U (m). At this time! = (Ilj) as n
As an identity matrix with rows and n columns, TB=Σ (2UI ”'
-1)(2Ul (1 -1)k-1 -mIij...(2)
When the matrix T is defined by
．． [J(2),... IJ fm+ has the effect of converging to the one closest to the given initial state. Therefore, when a certain input vector is given, an associative memory function is obtained that outputs the one closest to the input vector from among a number of predefined vectors. In addition, the following formula 'r'' +j=Σ (2U1 lkゝ -1) (2U
Even if the matrix T8 defined by l'kゝ -l)k-1...(3) 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 via all neurons other than their own. Therefore, if the number of synapses is n, the total is n (n-1)/2
We need neurons that can transmit signals in both directions. Therefore, if we try to implement a Hopteeld circuit using an electric circuit, as the number of processors corresponding to synapses increases, the number of wires corresponding to neurons will increase in proportion to the square of the number of processors. . Therefore, a multi-element hoptailed circuit requires a very large number of wirings that intersect in a complicated manner, 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 a physical signal transmitter.

Figure 3 is based on the literature (D. Psalts, et al.
“opttca+lnl'ormatlon proc
essing based on an assocI
atelve-memory model-, Opt
．． Lett. ．． l/of. 10. pp. 98-1
00. 1 is a diagram illustrating the configuration of a Hopfield circuit using the optical spatial wiring technology disclosed in 1985 (published 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 a one-dimensional light-receiving element array with n elements. It simultaneously has the function of thresholding the light intensity and driving the light emitting element array 1. The light emitted from the j-th element of the one-dimensional optical element array 1 illuminates the entire J-th element of the optical attenuation filter array 2, and only the light that has passed through the filter at the j row and j column position 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, one-dimensional light emitting element array 1
The light intensity emitted from one element in the one-dimensional light-receiving element array 3 and detected by one element in the one-dimensional light receiving element array 3 is p, and the transmission through the filter at the i-th row and j-th column in the optical attenuation filter array 2 is Let the rate be SIj and V = fVi) be a vector representing the light emitting state of each element of the one-dimensional light emitting element array 1 (Vi = 1 when the i-th element is emitting light, Vi = 0 when it is not emitting light)
). At this time, m n-element vectors U"u". whose elements take two values, 0 or 1.・・・・・・ U (m
) for S ij=1/2+(2Ul ” −1)(
2Uj (kゝ-1)/2m-Iij/2
...If Sij is determined as shown in (4), the light intensity Pi detected by the i-th element in the one-dimensional light-receiving element array 3 will be as shown in the following equation.

PI = p ・fisIjVj j−1 (5) Therefore, the threshold value pH (lh) of light intensity is p14lhl
The light intensity received by the i-th element in array 3 is pH +
Ih) or above, the i-th element in the one-dimensional light emitting element array 1 is made to emit light, and if Pl(lh) or less, the above equation (1) is performed.
It is possible to realize threshold processing equivalent to 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, this method has a drawback in that it is difficult to increase the number of elements using a two-dimensional light emitting element array and a light receiving element array.

As a method for solving the above-mentioned drawbacks, a method has been disclosed in which light emitted from each element in an example of a two-dimensional light emitting element is distributed to each element in a two-dimensional light receiving element array using a hologram.

Figure 4 shows the literature (Ju-Seog Jang. et a
t. , “Opalcal Implementat
ion of the Hopf'Ield mode
I...゜Opt. Lett. , Vol. l3. p
p. 248-250. Published in 1988) is a diagram showing the configuration of a Hopfield circuit. In the figure, 11 is an argon ion laser, 12 is a collimating lens, 13 is a two-dimensional spatial optical switch array, 14 is a two-dimensional hologram array, 15 is a two-dimensional light receiving element array, 16 is a threshold circuit, and 17 is a switch array. It is a drive circuit.

In FIG. 4, the light emitted from the laser 11 is transmitted through the lens 12.
The light becomes a parallel light beam and enters the optical switch array 13. Each switch in the optical switch array 13 transmits or blocks light depending on its on or off state.

Light transmitted through one switch in the optical switch array 13 is incident on one hologram in the hologram array 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 15 at a predetermined ratio.

Here, V = {Vi} is a vector indicating the state of the optical switch array 3 (If the i-th switch is on, Vi = 1
, if off, Vi=O),! ) optical switch array 13
The intensity of the laser light transmitted through the on-state switch in ,
Let Slj be the ratio of the intensity of the diffracted light that reaches the j-th element in the light-receiving element array 15 from the i-th hologram in the hologram array 14 and the intensity of the light incident on 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 threshold processing similar to that in the case of FIG. 1 described above, the result shown in FIG. It can operate as a Hopfield circuit using the same principle as in the example.

By using the method described above, a 221 ral net 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 an object thereof is to provide an optical neural network circuit that is simple and easy to miniaturize and can be applied to a Hopfield type content addressable memory circuit.

[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 emits linearly polarized light into the surrounding space, and at least one component that arbitrarily controls the polarization direction. a light-emitting element array part having one or more light-emitting element parts; a light-receiving element array part having one or more light-receiving elements for detecting the light intensity of a specific linearly polarized component in the incident light and controlling the polarization direction; For each of the light receiving elements belonging to the light receiving element array section, an integrating circuit section generates an output proportional to an integral value over time of a photocurrent generated from the light receiving element, and an output is supplied separately from the integrating circuit. and means for performing threshold processing and driving of the light emitting element sections to control each light emitting element section constituting the light emitting element array section to independently emit or extinguish light according to a comparison result with a threshold value. The gist is:

(Function) In the optical neural network circuit of the present invention, the light intensity of linearly polarized light emitted from the light emitting element array is detected by the light receiving element array, the photocurrent generated from the light receiving element array is integrated, and the integrated value is is compared with a threshold value, and each light emitting element section of the light emitting element array section is controlled to emit or extinguish light independently according to the comparison result.

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

FIG. 1 is a diagram showing the configuration of an embodiment of the present invention. In the figure, 21 is a light emitting element array in which light emitting diodes are arranged in a two-dimensional array, 22a and 22b are polarizing plates, and 23
23a and 23b are spatial filter arrays having a function of arbitrarily rotating the plane of polarization of spatially propagating light, and 24 is a light receiving element array in which photodiodes are arranged in a two-dimensional array.

Here, it is assumed that the light emitted from each light emitting element in the light emitting element array 21 uniformly illuminates each element in the light receiving element array 24. Further, 25 is a control device for controlling the spatial filter arrays 23a and 23b, 26 is an integrating circuit that outputs the integral value of the photocurrent of each element in the light receiving element array 24 with respect to time, and 27 is the output of the integrating circuit 26. A threshold processing/light emitting element drive circuit 28 performs threshold processing on the light emitting element array 21 and controls the light emitting state of each light emitting element in the light emitting element array 21 based on the result. This is a threshold generation circuit that detects the number of elements in a light emitting state and generates an output proportional to this number.

The polarizing plates 22a and 22b are arranged so that the polarization planes of transmitted polarized light coincide with each other. In addition, the light emitting element array 21
The light-emitting state and non-light-emitting state of each light-emitting element in are defined as states 1 and 0, respectively. Each filter in the spatial filter array 23a123b covers each element in the corresponding light emitting element array 21 and light receiving element array 24, and is in a state in which the polarization plane of transmitted light is not rotated by the control device 25 (state 0). or a state in which the plane of polarization is rotated by 90° (state 1). That is, in this embodiment, the light emitting element array 21,
The polarizing plate 22a and the spatial filter array 23a constitute a light emitting element array section having the function of emitting linearly polarized light and controlling the polarization direction arbitrarily, and the linearly polarized light is incident on the light receiving element array 24, the polarized light 1i1E22b and the spatial filter array 23b. It constitutes a light-receiving element array section that has the function of detecting the light intensity of a specific linearly polarized component in light and arbitrarily controlling its polarization direction.

Next, the effect will be explained. Let n be the number of light emitting element array 21, spatial filter arrays 23a, 23b, and light receiving element array 24, and m n-element vectors U(11, U(21,...) in which each element takes a value of 0 or 1.・・U (
Consider m l and one n-element vector V.

At this time, the i-th (1≦i≦n
) to match the i-th element Vi of the vector V, and the spatial filter arrays 23a, 23
The vector 7J of the state of the i-th filter (1≦i≦n) in b (matches the i-th element ux (k) of If the photocurrent generated in the photodetector by light reaching the photodetector is IIj(kl), then jjlj(kl = Io, XOR (ll (k)
．． lJj (k) )""Io ((2U1"'
-1)(2Uj (k'-1)+1)/2...(6) holds true. However, Io represents the photocurrent when the i-th and j-th filters are in the same state, and XOR represents the exclusive logical sum. From the above equation (6), I for all combinations of light emitting elements and light receiving elements. If the values of are equal, then in this state j
The total sum of photocurrents flowing through the th light-receiving element ■j'k'' is expressed by the following equation.

Ij ”' = ΣI IjLkli=1 (2Uj (kゝ -1) +l+/2...(7
) Here, the photocurrent generated from the j-th light receiving element is supplied to the integrating circuit 26, and the spatial filter array 23a, 23b is
Uf sequentially changes the state of each filter in
Ll. Switch to match U(2),...1) +m). At this time, if the amount of charge accumulated in the integrating circuit 26 is Qj, then Qj is expressed by the following equation.

=t oio a Σ Vi Σ ((2 Ul
"-1) (2Uj (k'-1)+11
/2) n Σ (( 2Uj ”' -1)・(2Uj (
"-1)}k-1 = (t ・ ■o/2)・(mN.,,, where, N
0. is the number of elements emitting light in the light emitting element array 21 (the number of “1” elements of the vector V), and T” is the formula (3)
is the coupling coefficient between neurons defined by .

The threshold processing/light emitting element drive circuit 27 receives N0 from the threshold generation circuit 28.

Using an output proportional to The state of the light emitting element is set to "1" (light emission), and otherwise it is set to "0" (quenched). The value subjected to threshold processing from equation (8) matches the output result of the Hopf mold model defined by equation (1). Therefore, by repeating the above operation again for the light emitting state of the light emitting element array 21 newly obtained by the above operation, the device of this embodiment operates as a normal Hopfield type content addressable memory device.

As described above, according to this embodiment, the light emitting element array 21
A Hopfield type content addressable memory device can be realized simply by covering each element of the light receiving element array 24 with the spatial filter arrays 23a and 23b having the function of rotating the plane of polarization. Therefore, the space required per element on the array can be significantly reduced compared to the conventional technology using holograms. Therefore, in this embodiment, two-dimensional high-density mounting of elements on the light emitting and light receiving element array is possible, and multi-element implementation, which has been a major problem in realizing a practical neural network, can be easily realized. can.

FIG. 2 is a diagram showing the configuration of another embodiment of the present invention. In the figure, 31a, 3lb are light emitting element arrays in which light emitting diodes are arranged in a two-dimensional array; 32a, 32b,
32c and 32d are polarizing plates, 33a to 33d are spatial filter arrays having a function of arbitrarily rotating the plane of polarization at the destination of spatial propagation, and 34a and 34b are light receiving element arrays in which photodiodes are arranged in a two-dimensional array. Here, all the light emitted from each light emitting element in the light emitting element array 31a uniformly illuminates each element in the light receiving element array 34a,
It is assumed that all light emitted from each light emitting element in the light emitting element array 31b uniformly illuminates each element in the light receiving element array 34b. Further, 35 is a spatial filter array 33
a, 33b, 33c, and 33d, and 36a136b is the light receiving element array 34a, respectively.
An integrating circuit 34b outputs the integral value of the photocurrent of each element with respect to time; 37 compares the outputs of the integrating circuits 36a and 36b, and based on the result, the light emitting element array 31
This is a threshold processing/light emitting element drive circuit that controls the light emitting state of each light emitting element in the 3 lb.

In this embodiment, the light emitting element array 31a, the polarizing plate 3
2a, 32b and spatial filter arrays 33a, 33b
Optical section (hereinafter referred to as +optical section) consisting of
is exactly the same as the optical part shown in FIG. On the other hand, the optical section (hereinafter referred to as one optical section) constituted by the light emitting element array 31b, the polarizing plates 32c, 32d, and the spatial filter arrays 33c, 33d is different from the optical section shown in FIG. The only difference is that the polarized light components transmitted through the polarizing plates 32c and 32d are arranged so as to be perpendicular to each other. When both optical sections are set as described above and both optical sections are operated simultaneously in the same manner as in the first embodiment, for ten optical sections, the photocurrent of the light receiving element and the accumulation in the integrating circuit are Equations (6) to (8) hold true regarding the amount of charge.

On the other hand, in one optical section, when the states of the spatial filters covering the light emitting element and the light receiving element are different from each other, the light emitted from the light emitting element reaches the light receiving element. Therefore, in the optical department,
The photocurrent generated in the light receiving element by the light emitted from the i-th light emitting element and reaching the j-th light receiving element is l ljl
-1, the following equation holds true.

11j'k-'=Io-XOR (UI Lk'
, Uj (k》)Io (-(2Ui Lk'
−1)−(2Uj ” −1)+ll/2 (9) Therefore, if the charge pot accumulated in the integrating circuit 36b for one optical section is Qj, then Qj is expressed as the following equation. Become.

-1》・(2Ul ”' -1)+L+/2””
t' IO /2)" Σ Vl(m -i-
1 ±(2Ui Lkl-1)・(2Uj Lk'-1)
k-1 = (t- Io/2)・(ffiN.,.1Σ TI
jVI>...(10) Also, let Q be the photocurrent integral value of the j-th photodetector of the ten optical parts.
If j4, n...(11) Therefore, the threshold processing/light emitting element drive circuit 37
hides Qj1 and Qj-, and if positive, sets the state of the j-th light emitting element in the light emitting element arrays 31a and 3lb to "1".
(light emission), otherwise it is set to "0" (quenching), and if the above operation is repeated again for this light emission state, ζ
The device of this embodiment operates as a normal Hopfield type content addressable memory device without requiring the threshold generation circuit 28 that was necessary in the embodiment of FIG.

[Effects of the Invention] As explained above, according to the present invention, a light emitting element and a light receiving element are arranged on a two-dimensional array, and the coupling coefficient between the two is set by adjusting the polarization plane of linearly polarized light emitted by the light emitting element and the light receiving element. This is achieved by controlling the plane of polarization of the linearly polarized light component detected by the light-receiving element, allowing for miniaturization of the light-emitting element and the light-receiving element.
High-density arrays can be realized, and multi-element implementation, which was a big problem in realizing conventional neural networks, can be easily realized. In particular, it is clearly different from the prior art in that it realizes a Hopfield circuit using a two-dimensional light emitting element array and a light receiving element array by controlling and identifying the polarization state of incoherent light.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the configuration of one embodiment of the present invention, FIG. 2 is a diagram showing the configuration of another embodiment of the present invention, and FIG. 3 is a diagram showing the configuration of a conventional Hopfield circuit using optical spatial wiring technology. FIG. 4 is a diagram showing the structure of a conventional Hopfield circuit in which elements are formed into a two-dimensional array using the optical wiring function of a hologram. 21...Light emitting element array 22a, 22b...Polarizing plate 23a, 23b...Spatial filter array 24...Light receiving element array 25...Control device 26...Integrator circuit 27... Threshold processing/light emitting element drive circuit 28/
...Threshold generation circuits 31a, 3lb...Light emitting element arrays 32a, 32b, 32c, 32d...Polarizing plate 33a,
33b, 33c, 33d - Spatial filter array 34... Light receiving element array 35... Control device 36a, 36b... Integrating circuit 37... Threshold processing/light emitting element drive circuit Agent Patent attorney 3 Yoshihide Kazu 1 z ≧ 2 figures

Claims (3)

[Claims] (1) A light emitting element array section having at least one light emitting element section that radiates linearly polarized light into the surrounding space and arbitrarily controls the polarization direction, and the light intensity of a specific linearly polarized light component in the incident light. 1 light receiving element that detects the polarization direction and controls the polarization direction.
an integrating circuit section that generates an output proportional to an integral value over time of a photocurrent generated from the light receiving element for each of the light receiving elements belonging to the light receiving element array section; Threshold processing and light emitting element section for controlling each light emitting element section constituting the light emitting element array section to independently emit or extinguish light according to a comparison result between the output of the circuit and a separately supplied threshold value. An optical neural network circuit characterized in that it has a means for driving. (2) A threshold generation circuit section that detects the number of light emitting element sections in a light emitting state in the light emitting element array section and generates a threshold value proportional to this number; 2. The optical neural network circuit according to claim 1, wherein the output of the circuit section is supplied to the threshold value processing and light emitting element section driving means. (3) Two sets each of the light emitting element array section, the light receiving element array section and the integrating circuit section are provided, and the output of one of the integrating circuit sections is supplied as a threshold value to the threshold processing and light emitting element section driving means. A claim characterized in (
1) Optical neural network circuit described.