JPH02157832A - Optical neural processor - Google Patents

Abstract

PURPOSE:To execute pattern processings in parallel with the use of non-interference and a broad band which are the characteristics of light, to recognize a complicated voice and a pattern, and to attain the high-speed processing of a picture by using an optical element for a neuron and wiring the section of each element by means of the light. CONSTITUTION:One-dimensional and multidimensional input signals 111 are first inserted to a positive input port 11, then an optical analog operation is executed by an optical analog arithmetic part 13, and the signals are outputted. Signals 201 obtained by processing the signals 111 by a threshold processing part 20 are outputted. A part of them are divided as signals 301 by a branching part 30, for the respective signals, a positive feedback signals 311 and negative feedback signals 321 are generated by optical connectors 31 and 32, and respectively inputted to the positive input port 11 and a negative input port 12. In such a way, the neural processor can be realized with the use of the optical element, and the parallel characteristic, non-interference and broad band, etc., of the light can be sufficiently applied.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an optical neural processor, and more particularly to an optical neural processor in which a neural computer is constructed of optical and electrical elements and interacts with the outside world during processing.

[Conventional technology]

Recently, processing devices using neural networks have been developed. For example, Nikkei Electronics Magazine N (1427
(19&7.8.lO), PP, 115-124,
A device using neural networks for pattern recognition, signal processing, and knowledge processing is described.

A neural processor is a processor that uses a neural network, and is equipped with parallel distributed processing, learning functions, and analog processing characteristics, imitating the brain's excellent information processing system. Therefore, it is good at non-verbal processing, such as speech processing and pattern recognition, which conventional artificial intelligence (AI) based on Von Neumann computers are weak at.

[Problem to be solved by the invention]

Conventionally, when constructing a neural processor with the characteristics described above using hardware and having a somewhat complex recognition function, processors corresponding to neurons in several directions or tens of directions are mutually connected. In addition, using a processor with an electronic device such as a transistor requires an enormous amount of wiring, and there are problems such as delay time due to wiring, mutual electrical interference, and insufficient bandwidth. As a result, it was not possible to take full advantage of the parallelism that is the original feature of neural processors.

Various proposals have been made to introduce optical technology into this neural processor and take advantage of the inherent characteristics of light, such as parallelism, non-coherence, and broadband properties. There has been no progress in hardware due to reasons such as the inability to find suitable elements and the immaturity of optical wiring.

The present invention has been made in view of the above circumstances, and its purpose is to solve the above-mentioned problems in the conventional technology, and to solve complex speech or pattern recognition that is difficult to realize with the conventional von Neumann type. An object of the present invention is to provide an optical neural processor capable of processing two images at high speed.

[Means to solve the problem]

The above-mentioned object of the present invention is to provide a plurality of processing elements configured to add weights to signals input from a plurality of input ports, perform nonlinear processing on the sum signal, and output the sum signal to an output port. , a neural network having a plurality of connection elements that connect the plurality of elements to transmit signals, characterized in that the elements are constituted by an analog operator, a threshold processing element, and an optical distributor. Achieved by an optical neural processor.

[Effect]

In the optical neural processor according to the present invention, the neural processor is realized using an optical element, so that the inherent properties of light such as parallelism, non-coherence, and broadband properties can be fully utilized. That is, unlike conventional neural processors that use electronic devices as processors corresponding to neurons, the optical neural processor according to the present invention uses optical elements for neurons and uses light to conduct wiring between each element. , it is possible to eliminate various drawbacks of metal wiring.

〔Example〕

Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.

In this embodiment, one-dimensional and multidimensional information of the external world is limited to positive input. Note that here, a positive signal is a signal that acts cooperatively so that neurons are excited together, and a negative signal is a signal that acts competitively to suppress the excitement of other neurons.

In FIG. 1, 111 is a one-dimensional and multidimensional input optical signal from the outside world, 11 is a positive input port, 12 is a negative input port, and 13 is an optical system that simultaneously performs analog operations on one-dimensional and multidimensional optical signals. An analog calculation section 131 indicates the output of the optical analog calculation section 13. Further, 20 is a threshold processing unit;
201 is the output of the threshold processing unit 20, 30 is the branching unit, 31 is the optical connection that is responsible for the feedback of the positive part, 311 is the output of the optical connection 31, 32 is the optical connection that is responsible for the feedback of the negative part, 321 is the optical connection 32 outputs are shown.

The operation will be explained below. One-dimensional and multidimensional input optical signals 111 from the outside world are first input to the positive input port 11, then subjected to optical analog calculation by the optical analog calculation section 13, and output to 131. This signal is processed by the threshold processing section 20 and a signal 201 is output, a part of which is input to 301 and branched into two by the branching section 30 . The respective signals are connected by optical connections 31, 32 to a positive feedback signal 31] and a negative feedback signal 3.
21 are generated and input into the positive input port 11 and the negative input port 12, respectively.

Positive input port 11. Negative input port 12. Optical analog calculation section 1
Part 3 is realized by an optical system as shown in an enlarged view in FIG. What is shown in FIG. 2 is an example using a reflective spatial modulator. In Figure 2, 4001 to 4
003 is a half mirror; 15001 and 5002 are spatial modulators that input light intensity information and convert it into polarized light;
60 indicates a polarizer. Further, 501 is light for reading polarization information of the spatial modulator 5002, 502 is light carrying polarization information corresponding to the intensity information of the output signal from the positive input port II, and 503 is the negative input port. 131 is a light carrying polarization information corresponding to the intensity information of the output signal from 12, 504 is the output obtained by combining them, and 131 is the polarization information carried by the output 504 converted into intensity information by a polarizer 60. The optical calculation output is shown.

Input optical signal 111 and positive feedback signal 112 (3
11) are synthesized by a half mirror 4003, converted into optical polarization information by a spatial modulator 5002, and converted into polarization information 502 by a readout light 501. This light passes through the half mirror 4001 and becomes the readout light for the spatial modulator 5001. The information read by this is passed through the half mirror 4002.

It is output in the form of polarization information 504 and converted into intensity information 131 by a polarization grating 60. By this, (strength of 131)
An analog calculation corresponding to = (intensity of input signal 111) + (intensity of positive feedback signal 112) - (intensity of negative feedback signal 121) is performed.

Above, positive input port 11. Negative input port 12. The optical analog calculation section 13 is a reflective spatial modulator 5001.

5002, a transmissive spatial modulator can also be used. For example, as shown in FIG. 3, it can be realized by a hybrid optical/electrical calculation mechanism. In FIG. 3, 51 is a calculation means using an electric circuit, which corresponds to (intensity of 131) = (intensity of input signal 111) + (intensity of positive feedback signal 112) - (intensity of negative feedback signal 121). Perform analog calculations. 5211 and 5212 are TV cameras, and 53 is an electric signal controlled spatial modulator. The arithmetic circuit 51 can also be realized by a computer or the like.

Branch portion 30 in FIG. Optical connection 31. , 32 can be realized by an optical system as shown in FIG.

In FIG. 4, 4010 is a half mirror, 170 is a total reflection mirror, 71.72 is an optical component for realizing positive and negative feedback connections, respectively, and 5011.501
2 is a spatial modulator, 501 is light for reading information from the spatial modulator 5011, and 311 and 321 are positive and negative feedback outputs, respectively. In this embodiment, the output 131 of the optical analog calculation section 13 is a half mirror.
A predetermined optical connection is obtained by passing through parts 71 and 72.

As a specific example of operation, consider an operation in which edges are extracted from the optical input signal 111 and output to the optical input signal 201 using the system shown in FIG. This requires a feedback connection having the characteristics shown in FIG.
．． As the pinhole 72, an array of pinholes with different diameters (diameter=D1.D) as shown in FIG. 6 can be considered. In FIG. 6, 80 is a light shielding plate, 81 is parallel input light, 82 is a pinhole with a diameter of D or D2, and 83 is a far field pattern. In addition, a pinhole 82 with the same diameter as shown in FIG.
Even if the light shielding plate 80 with
A similar effect can be obtained by changing the distance Lz between the optical component 72 and the spatial modulator 5011 (see FIG. 4), and the distance between the negative optical component 72 and the spatial modulator 5012 (see FIG. 4). can get. Further, as the optical components 71 and 72, it is also possible to use lenses arranged in an array, a plurality of fibers bundled together, or the like. Furthermore, it is also possible to use a polarizing beam splitter instead of the half mirrors 4001, 4002, 4003, and 4010. In this case, an optical system that can effectively utilize optical power can be constructed.

〔Effect of the invention〕

As described in detail above, according to the present invention, signals input from a plurality of input ports are weighted and then added, and the sum signal is nonlinearly processed and output to an output port. In a neural network having a plurality of processing elements and a plurality of connection elements that connect the plurality of elements and transmit signals, the elements are configured by an analog operator, a threshold processing element, and an optical distributor. Therefore, unlike conventional computers, pattern processing can be performed in parallel by taking advantage of the non-coherence and broadband properties of light, making it possible to process complex audio or pattern recognition two images at high speed. This has the remarkable effect of realizing an optical neural processor that can perform

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a first embodiment of the present invention, and FIG.
The figure shows a specific configuration example of the optical calculation section, which is the main part,
FIG. 3 is a diagram showing another specific example of the configuration of the optical calculation section, and FIG. 4 is a branching section in FIG. 1. FIG. 5 is a diagram showing a specific example of the configuration of the optical connection, FIG. 5 is a diagram showing characteristics featuring the feedback connection, and FIG. 6 is a diagram showing a specific example of the configuration of the optical component shown in FIG. 4. 11: Positive input port, 12: Negative input port, 13: Optical analog calculation unit, 20: Threshold processing unit, 30: Branch unit, 31: Optical connection responsible for feedback of positive part, 32: Negative part Optical connections responsible for feedback, 4001, 4002
゜4003.4010-: Half mirror, 5001,5
002: Spatial modulator, 60: Polarizer, 70: Total reflection mirror, 71: Optical component for realizing positive feedback connection, 72: Optical component for realizing negative feedback connection, 111: Input optical signal , 131: Output of optical analog calculation section, 201: Output of threshold processing section, 311
, 321: Output of optical connection. Diagram: Synaptic connection distribution

Claims (1)

[Claims] (1) A plurality of processing elements configured to weight and add signals input from a plurality of input ports, nonlinearly process the sum signal, and output it to an output port, and the plurality of elements. 1. An optical neural processor, characterized in that the neural network has a plurality of connection elements that are interconnected and transmit signals, wherein the elements are constituted by an analog operator, a threshold processing element, and an optical distributor.