JP2889275B2 - Optical associative identification device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a learnable optical associative identification device used in the field of optical information processing.

[Problems to be Solved by Conventional Technique and Invention] Conventionally, as a method for optically associating, identifying, and classifying information, a method shown in FIG. 2 has been proposed. This method
It was conceived to realize the Hopfield model optically. Since the Hopfield model is configured by matrix vector calculation, threshold processing, and a non-linear feedback system, optically, according to the state of neurons, light emitted from the LED array 21 is transmitted by the microlens array 22 to each of them. LED array element 21
After forming an image on each pixel of the mask 24 corresponding to the above, one-dimensionally expanding by the cylindrical lens 23 and performing a product operation with the mask 24 having a possible transmittance, the cylindrical lens 25 is again connected to the cylindrical lens 23. The light flux is collected by arranging the light beam in a direction perpendicular to the direction, and further converged on the PD array 5 by the cylindrical lens 26 to perform a sum operation. The output of each element of the PD array 5 represents the output of the next-stage neuron group. Each state of the neuron is
Although it takes a state of +1 or -1, it is difficult to realize an optically negative state, so that the output of the group of neurons having the state of +1 and the output of the group of neurons having the state of -1 are: After two systems of the devices were prepared and separately calculated, an electrical subtraction was performed, and then a threshold process was performed, and the results were fed back to the LED array 21.

However, in this method, +
Since it is necessary to express the state of 1 or -1, two systems of the device have to be prepared, so that the device has to be complicated.

Also, in order to arrange the micro lens array and the cylindrical lens, it is necessary to make the optical adjustment accurate,
A correct operation result could not be derived. Further, the device corresponds to a single-layer perceptron, but it is theoretically impossible to perform an EXOR operation as is generally known. Required at least a three-layer structure.
In this case, the above-mentioned problem is further increased.

The present invention has been made in order to solve the above-described problems, and does not use a cylindrical lens or separate the excitatory neuron and the inhibitory neuron into two, and without using two optical systems. It is an object of the present invention to provide an optical device that can be used as an associative identification device that can be learned with a simple configuration.

Means for Solving the Problems In order to solve the above technical problem, the present invention provides at least a light source for generating coherent light and one or more light sources for detecting an optical cross-correlation coefficient. A detection unit (for example, 1 to 7); and a modulation unit (for example, 7) for changing the optical complex amplitude distribution output from the image display area included in the detection unit. A pattern presenting part for converting an input pattern into an optical complex amplitude distribution, wherein the pattern presenting part comprises at least spatially divided image display areas (for example, 1a, 1b, 1c);
The first image display area of the divided image display area (for example,
In 1a), the output light complex amplitude is continuously changed in accordance with the input first input pattern. In the second and third image display areas (for example, 1b and 1c), the input second light input amplitude is changed. A first spatial light modulator group (for example, 1) capable of changing the output light complex amplitude into two values according to the third input pattern and
Means (for example, 2) for obtaining a first Fourier transform image of a two-dimensional distribution pattern of output light complex amplitude output through the first spatial light modulator group (for example, 1); and the first Fourier transform A first light receiving means (for example, 3) for detecting an intensity pattern of an image, and an intensity pattern obtained by the first light receiving means are written to a second spatial light modulator (for example, 4);
Means for reading it out with coherent light to obtain a two-dimensional distribution pattern of the second output light complex amplitude (for example, four output lights), and a Fourier transform image of the two-dimensional distribution pattern of the second output light complex amplitude From the intensity pattern (e.g., according to 5) (e.g., output of 6), a second pattern corresponding to the pattern drawn in the first image display area of the three divided image display areas of the first spatial light modulator group is obtained. A cross-correlation coefficient of the pattern drawn in the image display area and a cross-correlation coefficient of the pattern drawn in the third image display area with respect to the pattern drawn in the first image display area are obtained. Take the difference, and set the pixels corresponding to the difference between the cross-correlation coefficients having a value equal to or greater than a certain threshold to a bright state, and use the pixels as the second input pattern, and set the pixels having a value equal to or less than the certain threshold to a bright state. And the third Means for the input pattern (e.g., 7
Wherein the modulating means changes the pattern of each first image display area included in the one or more detecting means so that the cross-correlation coefficient has a desired output. (E.g., output from 7 to 1), and means (e.g., 7) for setting the second input pattern and the third input pattern as modulation inputs of the second and third image display areas.
To an output from 1 to 1).

[Operation] With the configuration of the optical associative identification device that can be learned according to the present invention as described above, the second and third input patterns are made to be patterns in which light and dark are inverted from each other, and each of the first and second patterns is cross-correlated. By taking the coefficients and taking the difference between the cross-correlation coefficients for each pixel corresponding to the first pattern, a substantial sum of products operation can be performed simultaneously and easily, and the output after threshold processing is desirable. Since the first pattern is determined so as to satisfy the following condition, when new information to be associated, identified, or classified is newly input to the second input pattern, a desired output can be extracted in accordance with the newly input information. This makes the information processing device highly flexible.

According to the invention, a learnable optical associative identification device comprises:
It has the following configuration.

That is, it has at least a light source for generating coherent light, irradiates the coherent light, and has one or more detecting means for detecting an optical cross-correlation coefficient, from an image display area included in the detecting means. The main configuration is a modulating means for changing the output optical complex amplitude distribution.

The detecting means has a pattern presenting part for converting the input pattern into an optical complex amplitude distribution, and the pattern presenting part comprises at least a spatially divided image display area, and the three divided image In the first image display area of the display area, the output light complex amplitude is continuously changed according to the input first input pattern, and in the second and third image display areas, the input light complex amplitude is changed. A first spatial light modulator group capable of changing the output light complex amplitude into two values in accordance with the second and third input patterns, and an output light complex amplitude output through the first spatial light modulator group. Means for obtaining a first Fourier transform image of a two-dimensional distribution pattern (for example, a Fourier transform lens), first light receiving means for detecting an intensity pattern of the first Fourier transform image, and the first light receiving means Strength obtained Means for writing a turn to a second spatial light modulator, reading it out with coherent light, and obtaining a two-dimensional distribution pattern of a second output light complex amplitude, and a two-dimensional distribution of the second output light complex amplitude From the intensity pattern of the Fourier transform image of the pattern, the first spatial light modulator group is drawn in a second image display area corresponding to the pattern drawn in the first image display area of the three divided image display areas. A cross-correlation coefficient of the pattern drawn in the third image display area with respect to the pattern drawn in the first image display area, and a difference between the cross-correlation coefficients is obtained. A pixel corresponding to the difference between the cross-correlation coefficients having a value equal to or larger than a threshold value is set to a bright state, and a second input pattern is set. Toss It is made by the means.

Further, the modulating means of the main component includes means for changing a pattern of each first image display area included in the one or more detecting means so that the cross-correlation coefficient becomes a desired output. , And the second input pattern and the third input pattern are used as modulation inputs of the second and third image display areas.

Next, the optical associative identification device of the present invention will be specifically described with reference to examples, but the present invention is not limited thereto.

Embodiment FIG. 1 is a schematic configuration diagram showing a main part of an example of an optical associative identification device according to the present invention.

In the optical arrangement diagram of FIG. 1, coherent light a from a semiconductor laser or the like is applied to a liquid crystal light valve (hereinafter abbreviated as LCLV) or, particularly, a liquid crystal panel for a liquid crystal television (hereinafter referred to as an LVTV) which can be used most easily. , Etc.), and output via a spatial light modulator 1 capable of electrical input. As shown in FIG. 3, this spatial light modulator 1 has an image display area 1a in which the transmittance can be continuously changed and a light-dark 2
Consisting of image display areas 1b and 1c from which value output is obtained,
A certain pattern is drawn on 1a, 1b, and 1c according to a rule described later. The light transmitted through the spatial light modulator 1 is two-dimensionally Fourier-transformed by a Fourier transform lens 2 into a complex amplitude distribution on the spatial light modulator 1 and is input to a light receiving element 3 such as a CCD camera. .

The light receiving element 3 converts the light intensity on the Fourier transform plane of the pattern on the spatial light modulator 1 into an electric signal. By changing the transmittance of a spatial modulator 4 such as an LCTV based on the electric signal, a pattern in which a Fourier transform pattern incident on the light receiving element 3 is written onto the spatial modulator 4 is irradiated with coherent light b, Fourier transform is again performed by the Fourier transform lens 5, and a light intensity peak corresponding to the cross-correlation coefficient of the patterns drawn in 1a, 1b, 1c can be obtained by the light receiving element 6 such as a CCD.

Now, briefly consider a Hopfield model of a neuron network. As shown in FIG. 4, the output state b _i of neuron groups and output states a _j groups of neurons of the first layer the second layer are connected by the following relationship.

_{b i = θ (Σw ij a} j -h i) ...... (1) where, W _ij is a bond linking the output state a _j of the neurons of the first layer and the output state b _i of neurons of the second layer Is the intensity, _hi is a self-specific threshold, and θ represents θ (x) = + 1 (x> 0) or −1 (x ≦ 0).

As is clear from equation (1), each neuron is either in the +1 excited state or in the -1 suppressed state.

Now, in order to examine equation (1), the absolute value of the output state of the neuron group whose output state of the neurons in the first layer is +1 is represented as a _j ^+, and similarly, the output state of the neuron group whose value is -1. the absolute value a _j of ^- expressed as, (1) _{formula, b i = θ (Σw ij} (a j + + a j -) -h i) can be expressed as ‥‥ (2).

As is clear from the equation (2), the i-th row of the coupling strength w _ij is a
_In the case of a pattern close to _j ⁺ , the neurons in the second layer
b _i is excited, whereas, the i-th row of coupling strength W _ij is a _j ^- in the case of almost close pattern neuron b _i of the second layer, is suppressed. That is, the pattern of bond strength w _ij is, a _j ⁺ and a _j ^- depending closer either pattern, is determined the state of the neuron b _i of the second layer.

Therefore, as shown in FIG. 5, the states of neurons in the first layer are divided into excitatory neurons and inhibitory neurons, and light and dark patterns ε and η are written in 1b and 1c in correspondence with both neuron states. And one pixel state can be represented by the corresponding pixels 1b and 1c. ○ in the figure represents a bright state. On the other hand, in the region 1a, the coupling strength w _ij
And the coupling strength pattern is α, and the cross-correlation f (y ′) = ∫α (y) ε (y−y ′) d _y is calculated, and y ′ = 0, i.e., at the peak position of the cross-correlation, f (0) = become ∫α (y) ε (y) d y ≒ Σw ij a j + that was calculated, which represents the degree of neuronal excitation b1 Will be. Similarly, when the bond strength pattern leading to suppression state of neurons b1 and α, g (0) = | α (y) η (y) d y ≒ Σw ij a j - will be a calculated, neuronal b1 This indicates the degree of suppression. The same calculation can be performed for the neurons b2, b3, and b4, assuming that the connection strength patterns associated with the neurons are β, γ, and δ, respectively.

Then, each pattern shown in FIG. 5 is irradiated with coherent light and Fourier-transformed by the Fourier transform lens 2, and the pattern I (fx) on the light receiving element 3 becomes each pattern β, γ, δ with respect to the pattern α. , Ε, and η, in the X direction, β _x , γ _x , δ _x , ε _x , η _x, and I (f _x ) = | A + Bexp (j2πf _x β _x ) + Γexp (j2πf _{x γ x) + Δexp (j2πf x} δ x) + Εexp (j2πf x ε x) + Ηexp (j2πf x η x) | 2 become.

Here, A, B, Γ, Δ, Ε, Η are α, β,
It is a Fourier transform pattern of γ, δ, ε, η.

Now, take out only claim ^{_{notable, I (f x) + =}} A * Eexp (j2πf x ε x) + B * Eexp (j2πf x (ε x -β x)) + Γ * Eexp (j2πf x (ε x ^{_{. -γ x)) + Δ *}} Eexp (j2πf x (ε x -δ x)) + C.C I (f x) - = A * Еexp (j2πf x η x) + B * Hexp (j2πf x (η x -β ^{_{x)) + Γ * Hexp (}} j2πf x (η x -γ x)) + Δ * Hexp (j2πf x (η x -δ x)) + C.C. , and the spatial light modulator these Fourier transform intensity pattern 4
And irradiates the output pattern with coherent light, and then Fourier-transforms the output pattern again.
Intensity distribution corresponding to each neuron b _i layer is obtained, the signal related to the excitation state of each neuron of the second layer of its ^{output, I (x) + = α} * ε (δ- (x-ε _x ) + δ (x + ε _x )) + β * ε (δ (x−ε _x + β _x ) + δ (x + ε _x −β _x )) + γ * ε (δ (x−ε _x + γ _x ) + δ (x + ε _x −γ _x )) + Δ * ε (δ (x−ε _x + δ _x ) + δ (x + ε _x −δ _x )). Similarly, the signal related to the suppression state of the neurons in the second layer is I (x) ⁻ = α ^* ε (δ (x−ε _x ) + δ (x + ε _x )) + β * ε (δ (x−ε _x + β _x ) + δ (x + ε _x −β _x )) + γ * ε (δ (x−ε _x + γ _x ) + δ (x + ε _x −γ _x )) + δ * ε (δ (x−ε _x + δ _x ) + δ (x + ε _x −δ _x )).

 Here, an asterisk indicates a correlation operation.

As described above, the peak value of α * ε is
This indicates the degree of excitement. The same applies to other neurons of the second layer.

Similarly, the peak value of α * η indicates the degree of suppression of neuron b1. The same applies to other neurons of the second layer.

In this way, the degree of excitation and the degree of inhibition of each neuron of the second layer are represented. For example, for the neuron b1 of the second layer, the outputs of α * ε and α * η feeding the processing device 7, wherein the calculating the difference between the outputs of the alpha * epsilon and alpha * eta, (1) formula, or (2) as expressed in expression, with the threshold h ₁ specific to neuronal b1 By taking the difference, it is determined whether the neuron b1 is in an excited state or a suppressed state. The same operation is performed for the other neurons in the second layer. Thus, the output state of each neuron of the second layer is determined.

Therefore, by learning the above-mentioned w _ij pattern by a known back propagation or orthogonal learning method and writing it to the spatial light modulator 1 so that the neuron state of the second layer becomes a desired output, the learning is performed. A possible associative identification device can be obtained.

FIG. 6 is a configuration diagram showing a main part of the configuration of another example of the optical associative identification device of the present invention. In this example, the configuration up to sending the output from the light receiving element 6 to the image processing device 7 is the same as the example in FIG.

In the image processing device 7, the neuron state of the second layer is determined, and this is displayed on a new spatial light modulator 8 by dividing it into excitatory neurons and inhibitory neurons as shown in FIG. . Also, the coupling constant w with the next neuron group
_{ij is} also newly input, and Fourier-transformed by the lens 9,
The light receiving element 10 obtains a Fourier transform intensity pattern. By preparing three sets for performing the following series of operations, a so-called three-layer perceptron can be optically constructed,
It will also be possible to perform operations such as EXOR that were not possible with further perceptrons. When constructing a multi-layered perceptron in this way, even if the optical system is divided into two groups, an excitatory neuron group and an inhibitory neuron group, the operations are performed substantially simultaneously, which is simple. It goes without saying that the same operation as that of a single-layer perceptron can be performed. At this time, the neurons in the second layer can be written into the first spatial light modulator 1 as new inputs.

Next, FIG. 7 is a schematic diagram showing the configuration of another optical associative identification device of the present invention. This will further explain the present invention.

Coherent light a such as semiconductor laser or gas laser
Is incident on the beam splitter 11, and the reflected light passes through a reflective spatial light modulator 12 capable of electrical input such as LCTV,
The light is reflected and further incident on the Fourier transform lens 2.
The following description of the configuration is the same as that of the optical associative identification device of FIG. In this case, unlike the example shown in FIG. 1, it goes without saying that not the transmittance but the reflectance is changed. As described above, the spatial light modulator may be a reflection type or a transmission type.

In the present invention, although there is a difference in the specification of the part functioning as a spatial light modulator, in principle, the same electrical address type can be used. As an example of the electric address type, in addition to the liquid crystal panel described above,
A ceramic or crystal having an electro-optical effect, such as PLZT, KDP, or BSO (Bi ₁₂ SiO ₂₀ ), or a material obtained by adding a matrix electrode to a crystal is often used.

As can be seen from the above description, these spatial light modulators can be used in any combination, and thus the optical associative identification device of the present invention has a combination of:
Many embodiments will be possible.

[Effects of the Invention] The above-described effects are obtained by the optical associative identification device according to the present invention. In summary, the following remarkable technical effects are obtained.

That is, an associative identification device that can be learned with a simple configuration without using a cylindrical lens or dividing it into excitatory neurons and inhibitory neurons and using two optical systems is not required. Equipment could be provided.

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram showing a configuration of an example of an optical associative identification device of the present invention. FIG. 2 is a schematic diagram showing a conventional optical associative identification device. FIG. 3 is a schematic conceptual diagram showing an area drawn on a spatial light modulator used in the optical associative identification device of the present invention. FIG. 4 is an explanatory diagram showing a connection state of neurons in the optical associative identification device of the present invention. FIG. 5 is an explanatory diagram showing an example of a pattern drawn on a spatial light modulator in the optical associative identification device of the present invention. FIG. 6 is a schematic configuration diagram showing the configuration of another example of the optical associative identification device of the present invention. FIG. 7 is a schematic configuration diagram showing a configuration of still another example of the optical associative identification device of the present invention. [Description of Signs of Main Parts] 1, 4, 8, 12 ... Spatial light modulator 2, 5, 9 ... Fourier transform lens 3, 6, 10 ... Light receiving element 7 ... Image processing device 11 ... Beam splitter

Claims (1)

(57) [Claims] At least a light source for generating coherent light, at least one detecting means for detecting an optical cross-correlation coefficient, and a light complex amplitude distribution output from an image display area included in the detecting means The detecting means has a pattern presenting part for converting an input pattern into an optical complex amplitude distribution, the pattern presenting part comprises at least a spatially divided image display area, In the first image display area of the three divided image display areas, the output light complex amplitude is continuously changed according to the input first input pattern, and the second and third image display areas are changed. Then, according to the input second and third input patterns, a first spatial light modulator group capable of changing the output light complex amplitude into two values, and an output signal via the first spatial light modulator group Means for obtaining a first Fourier transform image of a two-dimensional distribution pattern of output light complex amplitude, first light receiving means for detecting an intensity pattern of the first Fourier transform image, and first light receiving means. Means for writing the obtained intensity pattern to the second spatial light modulator, reading it out with coherent light, and obtaining a two-dimensional distribution pattern of the second complex amplitude of the output light; A second image display area corresponding to a pattern drawn in a first image display area of the three divided image display areas of the first spatial light modulator group based on an intensity pattern of a Fourier transform image of a dimensional distribution pattern The difference between the cross-correlation coefficient of the pattern drawn on the third image display area and the cross-correlation coefficient of the pattern drawn on the third image display area with respect to the pattern drawn on the first image display area is calculated as a value equal to or greater than a certain threshold value. A pixel corresponding to the difference between the cross-correlation coefficients having a bright state and a second input pattern, and a pixel having a value equal to or less than the certain threshold being a bright state and a third input pattern. A modulating means for changing a pattern of each first image display area included in the one or more detecting means so that the cross-correlation coefficient has a desired output; and a second input means. An optical associative identification device, comprising: means for using a pattern and a third input pattern as modulation inputs of respective second and third image display areas.