JPH0519322A - Neural network circuit - Google Patents

Abstract

PURPOSE:To perform highly precise optical arithmetic processing, e.g. pattern recognition, associative storage, and parallel arithmetic processing by providing a display device such as a lens system for receiving an input image and a screen to the neural network circuit consisting of a two-dimensional lens array, a spatial optical modulating element, and an optical threshold element, and then forming the input image of various size to constant size and imposing spatial optical modulation. CONSTITUTION:The neural network circuit consists of the lens system 20 for receiving the input image, the screen 21 where the image formed by the lens system 20 is projected, the two-dimensional lens array 10 which forms images as many as neurone on the screen 21, the spatial optical modulating element 12 which is equivalent to the synapse coupling of nerve cells, and the optical threshold element 13 which calculates the sum of inputs and performs threshold processing. The lens system 20 moves in the direction of the optical axis to make the size of the input image constant.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an input / output operation similar to an optical arithmetic unit and a nervous system, such as pattern recognition, associative memory,
The present invention relates to a neural network circuit that performs parallel arithmetic processing and the like, and particularly to a neural network circuit that includes an optical system capable of adjusting an image forming magnification of an input image.

[0002]

2. Description of the Related Art In recent years, information processing devices using neural network circuits have been actively researched, and in particular, the realization of optical neural network circuits for processing two-dimensional image information at high speed is expected. Has been done. This uses optical signals as a medium of information,
It focuses on the spatial parallelism and high speed of light.

As shown in FIG. 6, a conventional optical nerve network circuit has a two-dimensional lens array for creating an input image by the number of neurons, a spatial light modulator corresponding to the synaptic connection of nerve cells, and a threshold for obtaining a sum of inputs. It consists of an optical threshold element for processing.

[0004]

However, in the conventional optical nerve network circuit, since the number and arrangement of the pixels of the spatial light modulator and the optical threshold device are fixed, the image formed by the two-dimensional lens array is The shape and the repeat interval should always be constant. Therefore, in order to perform high-precision optical arithmetic processing, the shape of the image input to the optical nerve network circuit must always be constant, and there is a problem that it cannot handle input images having various shapes. .

Here, the image formation relationship of the input image will be specifically described. The size of the image formed and the repetition interval are determined by the distance from the input image to the two-dimensional lens array and the lens focal length of the two-dimensional lens array, and are approximately represented by (Equation 1).

[0006]

[Equation 1] However, b is the size of the image formation, p is the repetition interval of the image formation, and h
Is the size of the input image, L ₁Is a two-dimensional lens array from the input image.
Distance to a, L₂Is the imaging position from the two-dimensional lens array
Distance, f is the lens focal length of the two-dimensional lens array
Distance, a is the repetition interval of the lenses of the two-dimensional lens array
Represent.

From the formula (1), the size h of the input image changes
Also, the distance L from the input image to the lens ₁Or lens focal length
By adjusting the separation f, the image formation size b can be kept constant.
I know what I can do. However, the adjusted L₁And in f,
Formula (1) f / (L on the right side₁-F) term and L on the right side of equation (2)
₁/ (L₁-Repeat the image formation due to the difference from the item
Since the interval p is changed, the size b of the image formation and the image formation
There is no simultaneous solution that keeps the repetition interval p of
Be understood.

Therefore, when the size of the input image is different, the size b of the image formation and the repetition interval p of the image formation cannot be matched with the pixel arrangement of the spatial light modulator and the optical threshold device, and the optical nerves are not able to be arranged. There is a problem that the network circuit does not operate normally.

[0009]

In order to solve the above problems, the neural network circuit of the present invention captures an input image in a neural network circuit composed of a two-dimensional lens array, a spatial light modulator and an optical threshold element. And a screen.

Further, the neural network circuit of the present invention is
A neural network circuit including a two-dimensional lens array, a spatial light modulator and an optical threshold element is characterized by including a lens system for capturing an input image and a display device including at least a light-receiving layer and a light-emitting layer.

In the above structure, it is preferable that the display device comprises a light receiving layer and a light emitting layer laminated on a transparent substrate. Further, in the above configuration, the display device is a light receiving layer,
It is preferable that the pattern electrode and the light emitting layer are laminated in this order.

Further, the neural network circuit of the present invention is
In a neural network circuit composed of a two-dimensional lens array, a spatial light modulator and an optical threshold element, a lens system for capturing an input image and a light receiving element array in which a plurality of light receiving elements are two-dimensionally arranged, and the same number as the light receiving element array. Is provided with a light emitting element array in which the light emitting elements are two-dimensionally arranged, and the output of the light receiving element and the input of the light emitting element are connected via a gate circuit.

The neural network circuit of the present invention is
A neural network circuit including a two-dimensional lens array, a spatial light modulator and a light threshold element is characterized by including a lens system for capturing an input image and an optical fiber bundle.

In the above structure, it is preferable that the lens system is movable in the optical axis direction.

[0015]

According to the above construction, the light from the input image can be condensed by the lens system to form an image on the screen, and this image formation can be taken into the neural network circuit. Therefore, when input images with various shapes are to be captured, by adjusting parameters such as the focal length of the lens system, the distance from the input image to the lens system, and the distance from the lens system to the image formation position An image can be formed on the screen. By taking this image formation into the neural network circuit as an input image of the two-dimensional lens array,
The size of the image formation and the repetition interval of the image formation coincide with the pixel arrangement of the spatial light modulator and the optical threshold element, and stable and highly accurate optical calculation processing can be performed.

Further, in the above-described structure, as a substitute for a screen, a display device including at least a light-receiving layer and a light-emitting layer is provided, so that the light input / output characteristic of the light-receiving layer and the light output characteristic of the light-emitting layer are combined. Optical calculation processing having characteristics can be performed. That is, the light input / output characteristic of the screen is generally linearly proportional, whereas the light input / output characteristic obtained by combining the light input characteristic of the light receiving layer and the light output characteristic of the light emitting layer is the composition of the light receiving layer or the light emitting layer. A display device having desired light input / output characteristics can be obtained by selecting the above composition. For example, when the light input characteristics of the light-receiving layer are linearly proportional and the light output characteristics of the light-emitting layer have a non-linear relationship with a threshold value, when light above the threshold value is input to a certain area of the display device, The area emits light, and does not emit light even when light below the threshold is input. Therefore, even if noise light other than the original image information is input, light below the threshold is removed, and the S / N ratio of the input image can be improved.

Further, according to a preferable configuration in which the display device is composed of the light-receiving layer and the light-emitting layer laminated on the transparent substrate, the transmissive display device outputs light in the same direction as the light input direction. Therefore, the neural network circuit of the present invention can be easily obtained only by inserting the lens and the display device into the conventional neural network circuit shown in FIG. 5, for example.

Further, according to a preferable structure in which the display device has the light-receiving layer, the pattern electrode and the light-emitting layer laminated in this order, the light incident on the region where the pattern electrode exists can be integrated over the region. The light output characteristics of the light emitting layer function with respect to this integrated value. Therefore, by forming the pattern electrode in a desired shape or size, desired spatial light modulation can be performed.

Further, in the neural network circuit configuration including the screen described above, as an alternative to the screen, there are a light receiving element array in which a plurality of light receiving elements are two-dimensionally arranged, and the same number of light emitting elements as the light receiving element array. A light emitting element array arranged two-dimensionally is provided, and the output of the light receiving element and the input of the light emitting element are connected through a gate circuit, so that the same arithmetic processing as the spatial light modulation is performed using the gate circuit. It can be done electrically. For example,
As the gate circuit, a comparator for performing binarization processing for the number of pixels is used, and each pixel of the light receiving element array and each pixel of the light emitting element array are connected in a one-to-one relationship to maintain the shape of the input image. However, threshold processing can be performed on all pixels, and noise light below the threshold can be removed to improve the S / N ratio of the input image.

Further, by providing an optical fiber bundle as an alternative to the screen, it is not necessary to arrange each element constituting the neural network circuit in a linear arrangement, so that the degree of freedom in the arrangement of the circuit is increased and the space utilization efficiency is improved. Can be improved.

Further, according to the preferable construction in which the lens system is movable in the optical axis direction, the distance from the input image to the lens system and the connection from the lens system are fixed while the distance from the input image to the neural network circuit is fixed. Since the distance to the image position can be adjusted at the same time, the shape of the image can be adjusted smoothly.

[0022]

Embodiments of the present invention will be described below with reference to the drawings. In each of the embodiments, the case of character recognition will be described as an example, but the same applies to general image pattern recognition, associative memory, parallel arithmetic processing, and the like.

(Embodiment 1) FIG. 1 is a perspective view schematically showing an embodiment of a neural network circuit of the present invention. The input image 4 is, for example, the character “A”, the character portion is black, and the non-character portion is white. Although FIG. 1 shows a neural network circuit composed of two layers of an input layer and an output layer, a neural network composed of three layers is formed by adding a lens array, a spatial light modulator and an optical threshold element and connecting them in series. A circuit can also be configured.

As shown in FIG. 1, a light source 1 and an irradiation lens system 2 for illuminating a document, a lens system 20 for capturing an input image, a screen 21 for projecting an image formed by the lens system 20, and a screen 21. A neural network circuit is composed of a two-dimensional lens array 10 for creating an image corresponding to the number of neurons, a spatial light modulator 12 such as an optical mask corresponding to synapse coupling of nerve cells, and an optical threshold device 13 for thresholding the sum of inputs. To be done.

Light emitted from the light source 1 is condensed by the irradiation lens system 2 and uniformly illuminates the periphery of the input image 4 such as characters on the document surface 3. Light from the document surface passes through the aperture 5 for removing crosstalk and stray light. The transmitted light is
An image is formed on the screen 21 by the lens system 20. The lens system 20 may have a single lens structure, but it is also preferable to use a lens system composed of two or more lenses in order to reduce the aberration of the image forming system. Particularly, by using a zoom lens system. The imaging magnification can be changed in a state where the imaging focus is matched.

The lens system 20 is movable in the optical axis direction, and is controlled so that the size of the image formed on the screen 21 is always the same even if characters of different sizes are input. . The screen 21 is made of a phosphor so that only the place where it shines when exposed to light is illuminated. In particular, it is preferable to use a phosphor that matches the sensitivity wavelength of the light receiving system after the screen 21. As the screen 21, it is also preferable to use a diffusive screen that diffuses incident light from one side and emits it from the other side.

Light from the screen 21 enters each lens 11 of the two-dimensional lens array 10. In the two-dimensional lens array 10, since the optical axes of the lenses 11 are all arranged in the same direction in a matrix, the images formed by the lenses are formed in a matrix on the spatial light modulator 12, The document surface 3, on which the character images corresponding to the number of neurons are formed, the screen 21, the two-dimensional lens array 10 and the spatial light modulation element 12 are arranged in parallel with each other, and the central axis of the two-dimensional lens array and the characters to be read are displayed. The central axes of the input images (object planes) 4 are matched. The two-dimensional lens array 10 includes lenses 11 of the two-dimensional lens array 10 in order to obtain an image of the same size on the spatial light modulator 12.
Are made equal in focal length, and the main planes of the respective lenses 11 are all on the same plane.

The distance between the images formed by the two-dimensional lens array 10 is (L ₄ / L ₃ +1) times the distance between the lenses 11. Here, L ₃ represents the distance from the screen 21 to the two-dimensional lens array 10, and L ₄ represents the distance from the two-dimensional lens array 10 to the spatial light modulator 12. For this reason,
The space between the spatial light modulators corresponding to one neuron must be (L ₄ / L ₃ +1) times the space between the two-dimensional lens arrays.

Further, in order to prevent the images formed on the spatial light modulator 12 from overlapping each other, it is necessary to prepare the neural network circuit under the condition shown in (Equation 2).

[0030]

[Equation 2] Here, a is the repeating interval of the lenses of the two-dimensional lens array, k is the size of the image formed on the screen 21, f is the lens focal length of the two-dimensional lens array, and L ₃ is the screen 21.
To the two-dimensional lens array 10 and m represents the magnification.

The output light from the spatial light modulator 12 is input to the optical threshold element 13. That is, the light that has passed through one neuron of the spatial light modulator 12 is input to one pixel of the optical threshold element 13. The optical threshold element 13 is composed of, for example, a photoconductive layer and a light emitting layer, and is the sum of the intensities of light input to the optical threshold element (1
This element is turned on and emits light when the number of neurons is equal to or greater than a predetermined reference value, and does not emit when the value is equal to or less than the reference value.

(Embodiment 2) FIG. 2 is a perspective view schematically showing another embodiment of the neural network circuit of the present invention.
In this embodiment, a light source 1 for illuminating a document, an illuminating lens system 2, and a lens system 2 for capturing an input image.
0, a display device 30 that receives an image formed by the lens system 20,
From a two-dimensional lens array 10 that creates an output image from the display device 30 by the number of neurons, a spatial light modulator 12 such as an optical mask corresponding to synapse coupling of nerve cells, and an optical threshold element 13 that performs threshold processing by summing the inputs. , A neural network circuit is constructed. The display device 30 is the same as that of the first embodiment except for the display device 30.

FIG. 3 is a sectional view of the display device 30, which is composed of at least a light receiving layer 31 and a light emitting layer 32. First, on a transparent insulating substrate 33 such as a glass substrate, I
A transparent conductive electrode 34 such as a TO thin film is formed by an electron beam evaporation method so as to have a film thickness of about 100 nm.

Next, the light emitting layer 32 is formed on the transparent conductive electrode 34. As the light emitting layer 32, a thin film EL (electroluminescence) element made of ZnS: Mn or the like is preferable. On the transparent conductive electrode 34, a dielectric layer 35 such as a Y ₂ O ₃ film is vapor-deposited by an electron beam vapor deposition method so as to have a film thickness of about 40 nm. Furthermore, ZnS: Mn
The phosphor layer 36, etc., was formed by a sputtering method so that the film thickness was about 200 nm. Then, heat treatment was performed in vacuum at a temperature of 620 ° C. for about 1 hour. And
A dielectric layer 37 such as a Y ₂ O ₃ film is vapor-deposited by an electron beam vapor deposition method so as to have a film thickness of about 40 nm.

Next, the light from the phosphor layer 36 is converted into the light receiving layer 31.
The metal film 38 having both a light-shielding property for preventing the light from entering the light source and a reflectivity for effectively extracting the light from the phosphor layer 36.
(For example, Al, Cr, or an Al / Cr laminated layer) is formed to have a film thickness of about 100 nm. The shape of the pattern electrode of the metal film 38 is determined based on the integration region of incident light, and for example, the size of the pattern electrode is about 10 μm to 1 μm.
Patterning is performed so as to form an n × m matrix arrangement (n and m are natural numbers) having a size of about mm. In this way, the light emitting layer 32 including the dielectric layer 35, the phosphor layer 36, the dielectric layer 37, and the metal film 38 is formed.

Next, a procedure for forming the light receiving layer 31 will be described. On the metal film 38, a plasma CVD method is used at a substrate temperature of 250 ° C., a film thickness of 0.5 μm to 2 μm, and p / i /
A photoconductive layer 39 is formed by laminating an a-Si: H film having an n-diode structure. As the photoconductive layer 39, amorphous Ge or Si _1-x G containing hydrogen or a halogen element is also used.
It is also preferable to use e _x , Si _1-x C _x , and Ge _1-x C _x . Finally, the transparent conductive electrode 40 is formed on the entire surface to obtain the display device 30 constituting the neural network circuit of the present invention. Since the light receiving layer 31 and the light emitting layer 32 are stacked on the transparent substrate, a so-called "transmissive" display device 30 can be configured.

Next, the operation of the display device 30 will be described. An AC power source 43 is provided between the transparent conductive electrodes 34 and 40.
The AC voltage is applied by. When the light receiving layer 31 is not irradiated with light, the photoconductive layer 39 has a high resistance, and most of the AC voltage is applied between the light receiving layers 31 rather than the light emitting layer 32. Therefore, the voltage applied to the light emitting layer 32 becomes less than the threshold value and no light is emitted.

On the other hand, when the light receiving layer 31 is irradiated with light, the resistance of the photoconductive layer becomes low, and when the light having the total light intensity across the region of the metal film 38 is higher than a predetermined reference value, Most of the AC voltage is applied to the portion of the light emitting layer 32 corresponding to the region, and the region emits light. The emitted light is received by the light receiving layer 3
Since it is reflected by the first metal film 38, it can be efficiently extracted to the light output side. Since the metal films 38 are arranged in a two-dimensional matrix, characters formed by the lens 20 are displayed by the display element 30 as matrix pixels corresponding thereto.

As for the wavelength of emitted light, a desired spectrum can be obtained by selecting the type of phosphor or the type of impurities in the phosphor. Therefore, the emission spectrum with good sensitivity of the light receiving layer of the optical threshold element 13 is obtained. It is preferable to select Further, when the sensitivity of the light receiving layer 31 of the display element 30 is low with respect to the emission wavelength, or when a dielectric multilayer film is formed between the light receiving layer and the light emitting layer to form a reflective film structure, the metal film 38 is omitted. It is possible to use the above display device, and in this case, the operation is such that only the region where the incident light amount distribution exceeds the threshold value emits light.

The composition of the light receiving layer 31 or the light emitting layer 32
The display device 30 having desired light input / output characteristics can be obtained by selecting the above composition. For example, in the case of a light input / output characteristic having a threshold value, noise light below the threshold value can be removed to improve the S / N ratio of the input image.
Further, when the output light intensity becomes larger than the input light intensity, the light input / output characteristic becomes a proportional straight line graph having a slope of 1 or more, so that the light information can be amplified and the S / N ratio of the input image is further improved. be able to.

The lens system 20 is movable in the direction of the optical axis, and is controlled so that the image on the display device has the same size even if the size of the input character is different. Therefore, the size of the image formation and the repetition interval of the image formation coincide with the pixel arrangement of the spatial light modulator and the optical threshold device, and stable and highly accurate optical arithmetic processing can be performed.

(Embodiment 3) FIG. 4 is a perspective view schematically showing another embodiment of the neural network circuit of the present invention.
In this embodiment, a light source 1 for illuminating a document, an illuminating lens system 2, and a lens system 2 for capturing an input image.
0, a light receiving element array 50 including a plurality of light receiving elements 51 two-dimensionally arranged on the image plane of the lens system 20, a light emitting element array 60 in which the same number of light emitting elements 61 as the light receiving element array 50 are two-dimensionally arranged, and a light receiving element Spatial light such as a gate circuit 52 that connects the output of 51 and the input of the light emitting element 61, a two-dimensional lens array 10 that creates an output image from the light emitting element array 60 by the number of neurons, and an optical mask that corresponds to synapse coupling of nerve cells. A neural network circuit is composed of the modulation element 12 and the optical threshold value element 13 for thresholding the sum of the inputs. It should be noted that, except for the light receiving element array 50, the light emitting element array 60, and the gate circuit 52, it is the same as the first embodiment.

The light receiving element array 50 is formed by the lens system 20.
A character image is formed on the image. Each light receiving element 51 of the light receiving element array 50 includes a comparator, an AND circuit, an OR circuit,
It is electrically connected to each light emitting element 61 of the light emitting element array 60 via a gate circuit 52 composed of a knot circuit or the like. The circuit configuration of the gate circuit 52 is the same as that of each light receiving element 5.
The output of 1 and the input of each light-emitting element 61 may be connected by a one-to-one mapping at equal magnification, but they should be connected by a mapping relationship such as an anti-transfer image, a mirror image, a transfer image, a magnifying map, and a reducing map. Is also possible. Further, the output of each light receiving element 51 is subjected to image processing such as pattern matching, window processing, Fourier transform, etc. by combining AND circuits, OR circuits, knot circuits and the like to output to each light emitting element 61. It is also preferable to

When the amount of light incident on the light receiving element 51 is greater than or equal to a predetermined reference value, the gate circuit 52 is turned on and the light emitting element 6 is turned on.
1 emits light. Depending on the content of learning of the neural network circuit, it is possible to use the gate circuit by turning off the light when the amount of light incident on the light receiving element 51 is equal to or more than a predetermined reference value. Further, various spatial light modulations can be performed by changing and setting the reference value serving as a threshold for each gate circuit 52 corresponding to each light receiving element 51.

In the configurations of the first and second embodiments, the lens 2
While 0 to the optical threshold element 13 are linearly arranged on one optical axis, in the present embodiment, the light receiving element array 50 is used.
Since the light emitting element array 51 is electrically wired, the respective elements do not necessarily have to be arranged linearly, the degree of freedom in the arrangement of the neural network circuit is increased, and the space utilization efficiency can be improved.

(Fourth Embodiment) FIG. 5 is a perspective view schematically showing another embodiment of the neural network circuit of the present invention.
In this embodiment, a light source 1 for illuminating a document, an illuminating lens system 2, and a lens system 2 for capturing an input image.
0, an optical fiber bundle 70 such as an image fiber, and an output image of the optical fiber bundle 70 is created by the number of neurons 2
A neural network circuit is composed of the three-dimensional lens array 10, a spatial light modulator 12 such as an optical mask corresponding to synapse coupling of nerve cells, and an optical threshold element 13 for thresholding the sum of inputs. The optical fiber bundle 70 is the same as in the first embodiment except for the optical fiber bundle 70.

The optical fiber bundle 70 is formed by the lens system 20.
A character image is formed on the incident surface of, and the image information is transmitted by a one-to-one mapping of the same size, and is output from the exit surface of the optical fiber bundle 70. Therefore, similarly to the configuration of the third embodiment, it is not always necessary to arrange the elements in a linear arrangement, the degree of freedom in the arrangement of the neural network circuit is increased, and the space utilization efficiency can be improved.

[0048]

As described in detail above, the neural network circuit of the present invention can form and capture an input image of a fixed shape on a screen even if the size of the image to be input is different. Since stable and highly accurate optical calculation processing is realized, the character recognition rate and the accuracy of image processing can be dramatically improved.

By using a display device including at least a light-receiving layer and a light-emitting layer as an alternative to the screen, it becomes possible to provide desired light input / output characteristics, especially in the case of input / output characteristics having a threshold value. The S / N ratio of the input image can be improved, and image deterioration can be prevented when the neural network circuit is hierarchized. Also, by stacking the light-receiving layer and the light-emitting layer on a transparent substrate,
The display device is a transmissive display device, and the neural network circuit can have a simpler structure than that of a reflective display device. Further, by selecting the shape of the pattern electrode existing between the light-receiving layer and the light-emitting layer, desired spatial light modulation becomes possible, and the degree of freedom of spatial light modulation contents increases.

Since a light receiving element two-dimensional array, a light emitting element two-dimensional array and a gate circuit connecting them are used as an alternative to the screen, various light input / output characteristics and spatial calculation processing can be performed. The S / N ratio can be improved and image processing can be performed, and the degree of freedom in the arrangement of the neural network circuit can be increased.

Further, by using an optical fiber bundle as an alternative to the screen, the degree of freedom in the arrangement of the neural network circuit can be increased. In addition, since the lens system is movable in the optical axis direction, the size of the input image captured by the neural network circuit can always be set to a constant size, so compared to electrical processing such as image size conversion processing by a computer, It can be operated at high speed.

[Brief description of drawings]

FIG. 1 is a perspective view schematically showing an embodiment of a neural network circuit of the present invention.

FIG. 2 is a perspective view schematically showing another embodiment of the neural network circuit of the present invention.

3 is a cross-sectional view of a display device 30 used in the neural network circuit shown in FIG.

FIG. 4 is a perspective view schematically showing another embodiment of the neural network circuit of the present invention.

FIG. 5 is a perspective view schematically showing another embodiment of the neural network circuit of the present invention.

FIG. 6 is a perspective view schematically showing a conventional optical nerve network circuit.

[Explanation of symbols]

1 light source 2 irradiation lens system 3 Original side 4 Input image (object plane) 5 aperture 10 Two-dimensional lens array 11 lenses 12 Spatial light modulator 13 Optical threshold element 20 lens system 21 screen 30 display 31 light receiving layer 32 light emitting layer 33 Transparent glass substrate 34 Transparent conductive electrode 35 Dielectric layer 36 phosphor layer 37 Dielectric layer 38 Metal film 39 Photoconductive layer 40 Transparent conductive electrode 41 light emission 42 incident light 43 AC power supply 50 Light receiving element array 51 Light receiving element 52 gate circuit 60 light emitting element array 61 Light emitting element 70 Optical fiber bundle

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Akio Takimoto             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. (72) Inventor Koji Nomura             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. (72) Inventor Kuni Ogawa             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd.

Claims (7)

[Claims] 1. A neural network circuit comprising a two-dimensional lens array, a spatial light modulator and an optical threshold device,
A neural network circuit comprising a lens system and a screen for capturing an input image. 2. A neural network circuit comprising a two-dimensional lens array, a spatial light modulator and an optical threshold device,
A neural network circuit comprising a lens system for capturing an input image and a display device including at least a light-receiving layer and a light-emitting layer. 3. The neural network circuit according to claim 2, wherein the display device comprises a light-receiving layer and a light-emitting layer laminated on a transparent substrate. 4. The neural network circuit according to claim 3, wherein the display device has a light-receiving layer, a pattern electrode, and a light-emitting layer stacked in this order. 5. A neural network circuit comprising a two-dimensional lens array, a spatial light modulator and an optical threshold element,
A lens system for capturing an input image and a light receiving element array in which a plurality of light receiving elements are two-dimensionally arranged, and a light emitting element array in which the same number of light emitting elements as the light receiving element array are two-dimensionally arranged, and the output of the light receiving element A neural network circuit, wherein the input of the light emitting device and the input of the light emitting device are connected via a gate circuit. 6. A neural network circuit comprising a two-dimensional lens array, a spatial light modulator and an optical threshold device,
A neural network circuit comprising a lens system for capturing an input image and an optical fiber bundle. 7. The neural network circuit according to claim 1, 2, 5 or 6, wherein the lens system is movable in the optical axis direction.