JPH04328726A - Optical neural network device - Google Patents

Abstract

PURPOSE:To provide the optical neural network device which eliminates the problem of heat radiation from constituent elements of a light emission part, a spatial modulation part, etc., the problem of the layout of electric network and the problem of signal deterioration due to dust, etc. CONSTITUTION:In the optical neural network device equipped with a light emission part 22 which has a one- or two-dimensional array, a lens part 23 which collimates the light emitted by the light emission part 22, a spatial modulation part 24 which imposes two-dimensional modulation on the collimated light, and a one- or two-dimensional light reception part 26 which receives the modulated light and obtains the result of vector matrix arithmetic between the light emission and spatial modulation, one or plural groups of light emission parts 22, lens parts 23, spatial modulation parts 24, and light reception parts 26 are arranged on the surface of one transparent substrate 21.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an optical neural network device that constitutes the hardware of a neural network applied to image recognition, character recognition, and the like.

0002

[Prior Art] Conventionally, LSI or optical technology has been used as hardware for neural networks. Of these, optical technology utilizes the spatial parallelism of light to realize large-scale neural networks. there's a possibility that. As an example of this, an optical neurochip in which a light emitting part, a spatial modulation part, and a light receiving part are stacked has been reported (for example, IEICE Technical Report (CP)
SY) Vol. 88 No. 107 (1988) PP
．． 53-60, JP-A No. 2-226127, etc.). Here, FIG. 6 shows a conventional example of an optical neurochip, and 11
is a light emitting layer (light emitting part) consisting of a line-shaped light emitting diode array, and 12 is an Electro-Abso with an MQW structure.
A spatial modulation layer (spatial modulation section) utilizing the rption effect, and a light receiving layer (light receiving section) 13 made of a linear photodiode array. Further, 14 is an electrical insulating layer provided between the light emitting layer 11 and the spatial modulation layer 12 and between the spatial modulation layer 12 and the light receiving layer 13, and 15 is a substrate. Next, the operation of the optical neurochip will be explained. When a signal is input to the light emitting diode array of the light emitting unit 11,
A light emitting diode array emits light with an intensity proportional to the signal. The spatial modulation unit 12 performs intensity modulation proportional to the value expressed by a two-dimensional matrix of the connection strength in the neural network. The light emitted by the light emitting diode array of the light emitting section 11 is modulated by the spatial modulation section 12 and then enters the light receiving section 13 . The light emitting section 11 and the light receiving section 13 form a one-dimensional strip-shaped array, and are arranged so as to be orthogonal to each other as shown in the figure. As a result, the signal obtained from the light receiving section 13 is a vector matrix calculation of light emission (vector) and modulation (matrix). After the signal is converted into an electrical signal by the light receiving section 13, it is subjected to electrical threshold processing to complete the calculation of the neural network.

0003

[Problems to be Solved by the Invention] Conventional optical neurochips are characterized by high speed and high integration (large number of neurons); Lamination becomes difficult due to the electrical wiring of the light emitting section, spatial modulation section, and light receiving section. Also,
Even in neural networks, when networks are connected in a cascade like a hierarchical one, it becomes difficult to stack devices like chips. Although it is possible to construct an optical neural network using bulk elements, it has the drawbacks of being difficult to miniaturize, difficult to align, and susceptible to dust. The present invention has been made in view of the above circumstances, and the present invention solves the above-mentioned problems by arranging each component on a transparent substrate with a certain thickness and allowing light to travel through the transparent substrate. The purpose is to provide neural net devices.

0004

[Means for Solving the Problems] In order to achieve the above object, the invention according to claim 1 includes a light emitting section arranged one-dimensionally or two-dimensionally, a lens section collimating light emitted from the light emitting section, In an optical neural network device that includes a spatial modulation section that two-dimensionally modulates collimated light, and a one-dimensional or two-dimensional light receiving section that receives the modulated light and obtains a vector matrix calculation result between the emission and the spatial modulation, It is characterized in that one or more sets of the light emitting section, lens section, spatial modulation section, and light receiving section are arranged on the surface of one transparent substrate. The invention as set forth in claim 2 is characterized in that, in the optical neural network device as set forth in claim 1, the light emitting section includes one light source and a spatial modulation section that modulates light from the light source. Furthermore, the invention as set forth in claim 3 is characterized in that, in the optical neural network device as set forth in claim 1, a device that optically performs threshold processing is used in the light receiving section, and a processing result is extracted as optical information. Further, the invention according to claim 4 provides an optical neural network device in which optical neural network devices are connected in cascade, in which both sides of a transparent substrate are provided in the output light portion of the first optical neural network device according to claim 3. A second optical neural network device having a spatial modulation section and a light receiving section arranged thereon is superimposed on the optical neural network device, and the output optical signal of the first optical neural network device is input to the second optical neural network device.

[0005]

[Function] In the optical neural network device according to claim 1, since the constituent devices of the light emitting section, the lens section, the spatial modulation section, and the light receiving section are arranged on the surface of one transparent substrate,
The light from the light emitting section travels through the transparent substrate and is collimated by the lens section.The collimated light travels through the transparent substrate and enters the spatial modulation section.After being modulated by the spatial modulation section, it travels through the transparent substrate and is collimated at the light receiving section. Light is received. In this manner, in the optical neural network device according to the first aspect, since light travels within the transparent substrate, it is not affected by dust or the like. Furthermore, since one surface of each component device is in contact with the air, the heat generated from the light emitting section and the spatial modulation section can be radiated into the air, and the electrical wiring from each component device can be freely laid out.

In the optical neural network device according to the second aspect, the light emitting section includes one light source and a spatial modulating section that modulates the light from this light source, and in order to express the input vector by the spatial modulating section, the light source is can be made into Further, in the optical neural network device according to the third aspect, since the threshold value processing is performed using light, the result can be obtained as an optical signal. Further, in the optical neural network device according to the fourth aspect, by connecting the optical neural network devices according to the third aspect in a cascade, it is possible to easily realize a hierarchical neural network among the neural networks.

[0007]

Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a schematic perspective view of an optical neural network device according to an embodiment of the invention. In FIG. 1, 21 is a transparent substrate, but unlike a waveguide, this transparent substrate is a medium with a constant refractive index, and has a thickness and size that allows the use of normal light propagation phenomena. For example, glass or crystal. Further, 22 is a one-dimensional or two-dimensional vector light emitting unit such as a semiconductor laser, 23 and 25 are flat plate lenses, and 24 is a two-dimensional vector light emitting unit such as a semiconductor laser.
The dimensional matrix spatial modulation section 26 is a one-dimensional or two-dimensional vector light receiving section, and each of these constituent elements is arranged on both sides of the transparent substrate 21. Here, the light emitting section 22
The light emitted from the substrate travels diagonally within the substrate as shown by the arrow in the figure. Therefore, it is necessary for the light emitting section 22 to input light obliquely to the surface of the substrate 21. For this reason, Figure 1
In this configuration, a grating 27 is provided on the output surface of the light emitting section, and the diffracted light by the grating 27 is directed toward the flat lens 23. Furthermore, this grating 2
7 is created, for example, by electron beam drawing on a resist thin film. Further, the flat lenses 23 and 25 are made by, for example, forming a Fresnel zone plate on a resist film by electron beam lithography and depositing a reflective film thereon. Furthermore, the spatial modulation section 24 uses a reflecting type whose input and output are in the same direction. In the case of liquid crystal, a reflective film using Al may be provided on the surface not in contact with the substrate 21. Furthermore, the pitch of each element array is the same, and the positional relationship is precisely matched for each array.

Next, the operation of the optical neural network device having the configuration shown in FIG. 1 will be explained. In FIG. 1, the light emitting section 22 emits light with an intensity proportional to an input signal. The emitted light then enters the grating 27 and is diffracted. At this time, the undiffracted light travels straight, but the diffracted light travels 2
It is diffracted in the direction of the flat plate lens No. 3. Here, the light is collimated and reflected towards the spatial modulator 24 . The light reflected by the spatial modulation section 24 is then intensity-modulated. Further, the reflected light from the spatial modulation section 24 is focused by a flat lens 25, enters the light receiving section 26, is photoelectrically converted, and an electric signal proportional to the light intensity is output. Finally, this electrical signal is subjected to threshold processing by a signal processing circuit (not shown) provided outside the board and is used as the output of the neural network.

Next, FIG. 2 shows another embodiment in which the method of incidence on the light emitting section is changed, and is an example in which a prism 28 is used in place of the grating 27 described above. In this case, since the surface of the prism 28 in contact with the light emitting section 22 is inclined with respect to the substrate 21, the light from the light emitting section 22 can easily be obliquely incident. Now, in the above embodiment, an example was shown in which one set of each component element was provided on the substrate, but many sets of the above configurations may be formed on one substrate. Furthermore, if a flat lens with a short focal length can be used, it is also possible to integrate the light emitting section and the light receiving section, and in that case, the optical neural network device can be made smaller.

Next, FIG. 3 is a schematic perspective view of an optical neural network device showing an embodiment of the invention as claimed in claim 2. In this device, a light emitting section includes one light source and light emitted from this light source. It is characterized by comprising a spatial modulation section that modulates. In the figure, parts similar to those described above are indicated by the same numbers. Further, in the drawing, reference numeral 41 indicates collimated light having a uniform intensity distribution, and reference numeral 42 indicates a vector spatial modulation section that modulates the transmitted light intensity of input light, and this spatial modulation section is composed of, for example, a liquid crystal. In FIG. 3, the collimated light 41 enters the spatial modulation section 42, and the collimated light 41 enters the spatial modulation section 42.
In No. 2, the intensity is modulated in advance by an amount proportional to the input vector, and the incident light is intensity-modulated and emitted toward the substrate. Thereafter, as described above, the light travels within the substrate 21, is modulated and reflected by the spatial modulation section 24, enters the light receiving section 26, and is photoelectrically converted. However, since the incident light is collimated in advance, no flat lens is used. Note that when uncollimated light is used as incident light, a flat lens may be used as in the embodiments shown in FIGS. 1 and 2.

Next, the invention according to claim 3 will be explained. FIG. 4 is a schematic perspective view of an optical neural network device according to an embodiment of the invention. still,
In the figure, the above-mentioned components are indicated by the same numbers. Also, 51
is a one-dimensional array of devices that perform threshold processing on input signals and output the results as optical signals. An example of this device is an optical bistable element (using differential properties).
is typical. The operation of the optical neural network device having the configuration shown in FIG. 4 will be described below. In FIG. 4, the light emitting section 22 emits light with an intensity proportional to the input signal. The emitted light then enters the grating 27 and is diffracted. At this time, the undiffracted light travels straight, but the diffracted light is diffracted in the direction of the flat lens 23. Here, the light is collimated and reflected towards the spatial modulator 24 . The light reflected by the spatial modulation section 24 is then intensity-modulated. Furthermore, the reflected light from the spatial modulation section 24 is 25
The light is condensed by a flat plate lens, and enters a threshold device 51. Then, it changes nonlinearly depending on the intensity of the incident light. In other words, the output light intensity suddenly increases after reaching a certain light intensity. In this way, the calculations of the neural network are obtained as a light intensity signal. As described above, the optical neural network device according to the third aspect is characterized in that a device that optically performs threshold processing is used in the light receiving section 51, and the processing result is extracted as optical information.

Next, the invention according to claim 4 will be explained. FIG. 5 is a schematic perspective view of an optical neural network device according to an embodiment of the invention. still,
In the figure, the above-mentioned components are indicated by the same numbers. Also, 61
62 is the optical neural network device according to claim 3, with the light emitting portion removed. Reference numeral 63 is the optical neural network device according to claim 1, with the light emitting portion removed. Further, the optical neural net devices 61 and 62 are in contact with each other based on the position of the light receiving section 51 of 61 and the position of the light emitting section of 62, and the optical neural network devices 62 and 63 are in contact with each other based on the position of the light receiving section 51 of 62. and 63 are in contact with each other based on the position of the light emitting section. Therefore, the light emitting unit 22 of the 61 optical neural network device
The emitted light propagates through each device 61, 62, and 63.

The operation of the optical neural network device shown in FIG. 5 will be explained below. Light emitted from the light emitting section 22 of the device 61 with an intensity proportional to the input signal enters the grating 27 and is diffracted. The diffracted light travels inside the substrate, is collimated by 23 reflective lenses, and is collimated by 23 reflective lenses.
Proceed toward the spatial modulation section No. 4. Here, the light is intensity-modulated and reflected. Then, the reflected light is collected by 25 reflective lenses, enters a threshold processing device 51, undergoes threshold processing, and enters an optical neural network device 62. The light incident on the optical neural network device 62 travels through the substrate, is reflected by the reflective lens 23, becomes collimated light, and enters the spatial modulation section 24. Here the light is modulated and reflected again,
The light is condensed by a reflective lens 25, subjected to threshold processing by a threshold processing device 51, and then input to an optical neural network device 63. The light incident on the optical neural network device 63 travels through the substrate, is reflected by the reflective lens 23, and becomes collimated light.
24 into the spatial modulation section. Then, the light is further modulated and reflected by the spatial modulation section, becomes condensed light by the reflective lens 25, and enters the light receiving section 26. Here, the light intensity signal is converted into an electrical signal and electrically thresholded to complete the neural network calculation. Note that if it is desired to obtain an optical signal as the final result, a threshold processing device may be used instead of a light receiving element as the final stage device 63. In that case, the same device as 62 may be used for 63.

[0014]

As described above, in the optical neural network device according to claim 1, the constituent devices of the light emitting section, the lens section, the spatial modulation section, and the light receiving section are arranged on the surface of one transparent substrate. The light from the light emitting section travels through the transparent substrate, is collimated by the lens section, the collimated light travels through the transparent substrate, enters the spatial modulation section, is modulated by the spatial modulation section, and then travels through the transparent substrate. Since the light travels through the transparent substrate as it is received by the light receiving section, it is completely unaffected by dust and the like, and results with good reproducibility can always be obtained. In addition, since one surface of each component device is in contact with the air, the heat generated from the light emitting section and spatial modulation section can be radiated into the air, and the electrical wiring from each component device can be freely laid out. Can be done. Further, in the optical neural network device according to claim 2, the light emitting section includes one light source and a spatial modulation section that modulates the light from this light source,
Since the input vector is expressed by the spatial modulator, only one light source can be used. Therefore, there is no need to consider variations in characteristics that occur when using an array light source. Further, in the optical neural network device according to claim 3,
Since threshold processing is performed using light, the results can be obtained as optical signals. As a result, cascading properties can be achieved. Further, in the optical neural network device according to the fourth aspect, by connecting the optical neural network devices according to the third aspect in a cascade, it is possible to easily realize a hierarchical neural network among the neural networks.

[Brief explanation of drawings]

FIG. 1 is a schematic perspective view of an optical neural network device according to an embodiment of the invention.

FIG. 2 is a schematic perspective view of an optical neural network device according to another embodiment of the invention.

FIG. 3 is a schematic perspective view of an optical neural network device according to an embodiment of the invention;

FIG. 4 is a schematic perspective view of an optical neural network device according to an embodiment of the invention;

FIG. 5 is a schematic perspective view of an optical neural network device according to an embodiment of the invention.

FIG. 6 is a schematic perspective configuration diagram showing an example of an optical neural network device according to the prior art.

[Explanation of symbols]

21 Transparent substrate 22　　　　　　　　　　Light-emitting part 23, 25 Lens section 24 Spatial modulation section 26 Light receiving part

Claims (4)

[Claims] Claim 1: A light emitting unit arranged in one or two dimensions, a lens unit that collimates the light emitted from the light emitting unit, a spatial modulation unit that two-dimensionally modulates the collimated light, and a light receiving unit that receives the modulated light. In an optical neural network device equipped with a one-dimensional or two-dimensional light receiving section that obtains vector matrix calculation results between light emission and spatial modulation, the light emitting section, lens section, and spatial modulation section are arranged on the surface of one transparent substrate. An optical neural network device characterized in that one or more sets of light receiving sections are arranged. 2. The optical neural network device according to claim 1, wherein the light emitting section comprises one light source and a spatial modulation section that modulates light from the light source. 3. The optical neural network device according to claim 1, wherein a device for optically performing threshold processing is used in the light receiving section, and the processing result is extracted as optical information. Claim 4: An optical neural network device formed by connecting optical neural network devices in a cascade,
A second optical neural network device having a spatial modulation section and a light receiving section arranged on both sides of a transparent substrate is superimposed on the output light portion of the first optical neural network device according to claim 3, and a first optical neural network device is formed. An optical neural network device, characterized in that an output optical signal of the device is input to a second optical neural network device.