JP2827378B2 - Optical neuro element - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural element for performing pattern recognition and associative processing of information at a high speed, and more particularly, to an optical neural element using parallel operation by an optical waveguide. Things.

[Prior art] A neurocomputer that simulates an excellent information processing function of a biological system is relatively easy to process associative, learning, pattern recognition, combination optimization problems, etc., which are difficult with a conventional von Neumann-type computer. It is said that it can be realized. The basis of a neurocomputer excellent in intelligent information processing functions such as association, inference, learning and the like is parallel processing, and information is accumulated in connection strength between elements called many neurons. This is because the matching with the light having the spatial parallelism is good.
10, the optical mask 81, the light receiving element 82, the comparator
83, one neuron is optically formed by the light emitting element 84,
Optical wiring between neurons is realized using a matrix-shaped optical mask 81. The excitation state of each neuron corresponds to the blinking state of the LEDs 84 arranged in a line. The output V _j (j = 1, 2,... N; N is the number of neurons) from each light emitting element 84 is wavefront transformed into a fan-shaped beam using a lens system (not shown), and is converted into a coupling intensity matrix T _ij . Only the j-th column component of the corresponding optical mask 81 is uniformly irradiated. T
_ij is, for example, using accumulated information vectors V ^(A) , V ^(J) , V ^(E) representing A, J, E shown in FIG. When the magnitude of T _ij is given as the light transmittance of the optical mask 81, the output light intensity is proportional to T _ij V _j . Next, all the i-row components of the output light from the optical mask 81 are focused on one of the light receiving element arrays 82 by a lens system (not shown). Therefore, the i-th light receiving element output u
_i is And a matrix-vector product is obtained at the output of the light receiving element.
The signal photoelectrically converted by the light receiving element array 82 is subjected to threshold processing by the comparator 83, and is fed back to the LED array 84.
By this repetitive operation, for example, for the incomplete input 85, the most similar A among the accumulated complete information A, J, E is selected, and a complete output 86 is obtained. In practice, the matrix T _ij is bipolar having positive and negative components corresponding to excitatory and inhibitory synaptic connections, and therefore only T _ij ^(+), which is a collection of only positive components of T _ij , and only negative components Using a two-channel optical system corresponding to the collected T _ij ^(-) , the difference between them is taken before the comparator 83. Also, as described in Miyazaki, 1989 IEICE Spring National Convention SD-1-6 and Miyazaki, 1989 IEICE Autumn National Convention D-217, an optical neuron device using a thin film waveguide was used. Has also been proposed.

[Problem to be Solved by the Invention] However, since this optical neurocomputer is composed of individual optical components such as a light emitting element, a light receiving element, and an optical mask, the dimensions are large, the optical axis adjustment is troublesome, and mass production is difficult. And had problems with reliability. Further, the pattern of the optical mask is fixed, and its application range is limited in advance, and there is a problem that it cannot be applied to other applications.

The present invention has been made in order to solve the above-described problems, and applies a magnetostatic wave to a plurality of intensity-modulated light fluxes, and further, the refractive index of the waveguide surface is perpendicular to the light propagation direction. The parallel operation is performed by focusing and detecting the light beam by the light focusing waveguide having distribution in the direction, and the purpose is to integrate the optical components on one substrate, An object of the present invention is to provide an optical neuro element which is small in size, does not require optical axis adjustment, has excellent mass productivity and reliability, and can cope with various processes by controlling a magnetostatic wave.

[Means for Solving the Problems] To achieve this object, an optical neuron device according to the present invention comprises a light source for emitting a plurality of light beams, a light modulating means for modulating the intensity of the light beam, and a static light for diffracting the light beam. Magnetostatic wave exciting means for applying a magnetic wave, a light focusing waveguide for condensing a light beam diffracted by the magnetostatic wave, a plurality of light receiving elements for detecting the condensed light beam, and an output signal of the light receiving element. And means for performing threshold processing.

[Operation] In the optical neuron device of the present invention having the above-described configuration, a plurality of light beams emitted from the light source are modulated by the incomplete input signal by the light modulation unit. When a magnetostatic wave acts on the plurality of modulated light fluxes, the light flux is diffracted in a direction of an angle corresponding to the frequency of the magnetostatic wave with a diffraction intensity corresponding to the intensity of the magnetostatic wave. Of the diffracted components diffracted in different directions from each light beam, diffracted light components diffracted in the same direction are condensed at one point by a light-converging waveguide, and detected by a light receiving element.
Thus, a vector given by the intensity of the light beam and a matrix given by the intensity of the magnetostatic wave are calculated, and the result is given by the output of the light receiving element. The output of the light receiving element is subjected to threshold processing, and is fed back to the light modulating means to perform feedback. By this repetition processing, the most similar to the incomplete input among the stored complete information is obtained as a complete output.

[Embodiment] An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows the configuration of an optical neuron element of the present invention. An optical neuron element 11 comprises a slab type waveguide 12,
Light focusing waveguide 13, semiconductor lasers 14a, 14b, 14c, collimator lenses 15a, 15b, 15c, strip electrode 16 for exciting a magnetostatic surface wave, diffracted by a magnetostatic surface wave, and collected by the light focusing waveguide 13. Light receiving elements 17a, 17b, 17 for detecting emitted light
c and comparators 18a, 18b, 18c.

The light focusing waveguide 13 is made of LiNb as shown in FIG.
_On a crystal substrate 26 of O ₃ or the like, a metal 28 such as Ti is formed by a sputtering method so that the film thickness distribution becomes larger toward the center. Thereafter, the waveguide 13 having a refractive index distribution as shown in FIG. Incidentally, the film thickness distribution of Ti28 shown in FIG.
And a shadow mask method of performing sputtering. Light waves emitted from a point light source propagating through such a gradient index waveguide are shown in Yamauchi, Miyazaki,
As described in IEICE MW84-128, it can be used as a light-converging waveguide because it is condensed every certain length and propagates as a periodic beam as shown in FIG.

In FIG. 1, semiconductor lasers 14a, 14b, 14
The light emitted from c becomes parallel light beams 31a, 31b, and 31c by collimator lenses 15a, 15b, and 15c such as Selfoc lenses, and propagates through the slab waveguide 12. Slab type optical waveguide 12
Is facilitated by growing the Gd ₃ Ga ₅ O liquid phase growth method on a substrate 35 on such _12, Y by RF sputtering ₃ Fe ₅ O _12, Bi substituted Y ₃ Fe ₅ magnetic garnet films 36 O _12, etc. Can be manufactured. Semiconductor lasers 14a, 14b, incomplete input 32a of current supplied to 14c, 32 b, by controlling the magnitude corresponding to 32c, incomplete input 32a, 32b, the intensity V ₁ corresponding to 32c, V _2, V
Light beams 31a, 31b and 31c having ₃ are obtained.

The light beams 31a, 31b, 31c are diffracted by the magnetostatic surface waves excited by the strip electrode 16. That is, as shown in FIG. 4, a DC magnetic field H is applied to the magnetic garnet thin film 36 in a direction parallel to the surface, and a microwave is applied to the strip electrode 16, so that the periodic increase of the magnetization M of the magnetic garnet thin film 36 is achieved. The differential motion propagates in a direction perpendicular to the direction of the magnetic field, and a so-called magnetostatic surface wave is excited. This periodic change in magnetization causes mode conversion between TE and TM modes, and also causes light diffraction. As shown in FIG. 5, by modulating the intensity of the magnetostatic surface wave 41 so that the coupling strength matrix becomes T _i1 , T _i2 , and T _i3 , the diffracted light 33 diffracted at an angle of θ _i is obtained.
a, 33b, the strength of 33c becomes _{T i1 V 1, T i2 V} 2, T i3 V 3 , respectively.
Here, the diffraction angle θ _i can be changed by changing the frequency of the microwave applied to the strip electrode 16. Incidentally, the magnetostatic surface wave can be converted into a microwave by the receiving electrode 30 and monitored.

The diffracted lights 33a, 33b, 33c are applied to the light focusing waveguide 1 as shown in FIG.
It is collected while propagating through 3. The intensity of the collected light is the sum of the intensities of the diffracted lights 33a, 33b, and 33c: u _i = T _i1 V ₁ + T _i2 V ₂ + T _i3 V ₃ (3) By changing the frequency of the microwave applied to the strip electrode 16 to f ₁ , f ₂ , f ₃ , the diffraction angle changes, and the condensing positions become the positions of the photodetectors 17a, 17b, 17c, respectively, and the three diffractions light 33a, 33b, the sum u _i of the intensity of 33c is obtained as a photodetector output. Thereby, the vector-matrix operation of the expression (2) can be performed.

The outputs of the photodetectors 17a, 17b, 17c are subjected to threshold processing by the comparators 18a, 18b, 18c, and fed back to the modulation currents of the semiconductor lasers 14a, 14b, 14c. Thereby, the luminous fluxes 31a, 31b, 31c
Is changed, the same calculation is repeatedly performed, and the most similar information is selected from the complete information accumulated as the magnetostatic surface wave intensity T _{ij with} respect to the incomplete inputs 32a, 32b, and 32c, Complete outputs 44a, 44b, 44c are obtained. Here, in the present embodiment, the semiconductor lasers 14a, 14b, 14c as light sources
From the collimator lenses 15a, 15b,
After being converted into parallel lights at 15c, the light is diffracted by the magnetostatic surface waves excited by driving the single strip electrode 16, respectively. At this time, the coupling intensity matrix that is affected by each light beam is as shown in FIG.
It is T _i1 for a, T _i2 for the luminous flux 31b,
The light flux 31c is _Ti3 . In order to function as a neuro element, the combination of the coupling intensity matrices corresponding to one light flux must be unchanged, so in the present embodiment, it is understood that feedback is performed when the coupling intensity matrix for each light flux is the same. Should.

In the above-described configuration, it is possible to perform an operation when the coupling strength matrix T _ij is positive. When T _ij takes positive and negative values, two optical neuro elements 51 and 52 having the above-described configuration are used as shown in FIG. 6, while T _ij is positive at 51 and T _ij is at the other 52.
_ij performs operations when negative, whereas photodetector 17a Neuro elements 51, 17b, the output of 17c, i.e., the photodetector 54a of the other neuro device 52 from the calculation result when T _ij is positive, 54b, 54c
, The operation result when T _ij is negative,
a, 55b, by pulling by 55c, operation results when the T _ij takes a positive or negative value is obtained. This result is compared with the comparators 56a, 56b,
The threshold processing may be performed at 56c, and the light may be fed back to the modulation currents of the semiconductor lasers 14a, 14b, 14c, 58a, 58b, 58c of the neuro elements of the respective lights.

The present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the gist of the present invention.

For example, as shown in FIG. 7, light from a semiconductor laser 61 is made parallel by a collimator lens 62, collected by a cylindrical lens 63, made incident on the slab type waveguide 12, and intensity-modulated by a light modulator 64. It may be converted into a plurality of light beams. Optical modulator 64
Is a metal clad 71 using Al or the like as shown in FIG.
It comprises an electrode 72 and a buffer layer 73. Buffer layer 73
Functions to prevent the light wave from being attenuated by the electrode 72. Electrode 72
, A magnetic field is generated parallel to the light propagation direction and proportional to the magnitude of the current. The magnetic garnet 36 of the slab waveguide 12 is magnetized by this magnetic field. Here, if a DC magnetic field H is applied in advance in a direction perpendicular to the film surface, when no current flows through the electrode 72, no mode conversion occurs because the magnetization of the magnetic garnet 36 is perpendicular to the light propagation direction. When a modulation current is applied to the electrode 72, the generated magnetic field causes a component in the direction of light propagation in the magnetization of the magnetic garnet 36, and the Faraday effect converts the TE mode to the TM mode, which is attenuated by the metal cladding 74. Thus, a light beam having an intensity corresponding to the modulation current is obtained. Here, since the metal cladding 75 below the portion without the metal 72 is long, both the TE and TM modes are attenuated, and no light wave is output from the portion without the electrode 72. Further, in a portion of the electrode 72, the TE mode is hardly attenuated because the metal clad 74 is short, and only the TM mode is attenuated. As described above, by using the optical modulator, the number of light beams can be made very large, and a large amount of information can be processed. The following operation is the same as in the above-described embodiment. However, when a DC magnetic field is applied perpendicular to the film surface, a magnetostatic volume wave is excited by applying a microwave to the strip electrode 16 as shown in FIG. The magnetostatic wave can also diffract the light, and the diffracted light is collected by the light-converging waveguide 13 in FIG. Further, after threshold processing is performed by the comparator 77, the result is fed back to the optical modulator 64. This allows incomplete input from stored information
The one most similar to 78 is obtained as output 79. Although the Faraday effect, which is a magneto-optical effect, is used for the optical modulator, another magneto-optical effect, an electro-optical effect, an acousto-optical effect, or the like may be used.

Further, the number of light beams and the material of the waveguide are not particularly limited.

In the signal processing section, a sample-and-hold circuit, a delay circuit, and the like may be appropriately added.

The type of the magnetostatic wave used is not particularly limited.

Further, the light focusing waveguide may be manufactured by changing the concentration of Bi or the like in the magnetic garnet.

[Effects of the Invention] As is apparent from the above description, according to the present invention, a magnetostatic wave is applied to a light wave propagating through an optical waveguide, and diffracted light is condensed by a light focusing waveguide. Perform a parallel operation to configure an optical neuron element. That is, since the optical neuron device of the present invention can perform parallel operation using an optical waveguide, it is compact and does not require optical axis adjustment,
Excellent mass productivity and reliability. Furthermore, since the coupling strength matrix can be changed by controlling the magnetostatic wave, it is possible to cope with various processes.

[Brief description of the drawings]

FIGS. 1 to 9 show an embodiment embodying the present invention. FIG. 1 is a structural view of an optical neuron element according to an embodiment of the present invention, and FIG. 2 is a light focusing waveguide. FIG. 3 is an explanatory diagram showing propagation of a light wave in a light focusing waveguide, FIG. 4 is an explanatory diagram showing propagation of a magnetostatic surface wave,
FIG. 5 is an explanatory diagram showing diffraction of a light wave by a magnetostatic wave, FIG. 6 is a configuration diagram showing a configuration when two neuro elements corresponding to positive and negative coupling strength matrices are used, and FIG. FIG. 8 is a configuration diagram showing another embodiment, FIG. 8 is a configuration diagram of an optical modulator, FIG. 9 is an explanatory diagram showing propagation of a magnetostatic volume wave, and FIG.
FIG. 10 is a block diagram showing another embodiment of the conventional optical neuro element, and FIG. 11 is an explanatory diagram showing accumulated information. In the figure, 11 is an optical neuro element, 12 is a slab type waveguide, 13 is a light focusing waveguide, 14a, 14b, 14c are semiconductor lasers, 15a, 15
b and 15c are selfoc lenses, 16 is a strip electrode, 17
a, 17b, 17c are light receiving elements, 18a, 18b, 18c are comparators, 31a, 31b,
31c is a light beam, and 33a, 33b, and 33c are diffracted lights.

Claims (1)

(57) [Claims] A light source for emitting a plurality of light beams; a light modulating means for modulating the intensity of the light beam; a magnetostatic wave exciting means for exciting a magnetostatic wave for diffracting the light beam; A light-converging waveguide for condensing the light beam, a plurality of light-receiving elements for detecting the condensed light beam, and means for performing threshold processing on an output signal from the light-receiving element. Characteristic optical neuro element.