JPH0251713A - Optical neural computer - Google Patents

Abstract

PURPOSE:To perform an intelligent process at a high speed for the two dimensional information on recognition of the complicated voices or patterns, the picture processing, etc., by preparing an optical arithmetic unit, a signal input part, an optical distributor, and an optical connection line which connects the signal input part and the optical distributor. CONSTITUTION:The excitation and suppression signals received from a signal input part 13 are separated from each other by an optical distributor 16 to perform a process equivalent to that carried out in case the difference is secured between both signals. An optical arithmetic unit 12 performs an operation to output 1 only when one of both signals is equal to 1. In other words, the excitation signal which secures the positive synapse connection is produced separately from the suppression signal which performs the negative synapse connection. The difference between both signals is optically calculated for each optical neuron. Then only the output signal obtained when the exitation signal and the suppression signal are kept ON and OFF respectively is extracted out of the calculated output optical signal. These separated output signals are fed back to other optical neurons as the input signals via an optical connection line. Thus it is possible to perform an intelligent process at a high speed for the two dimensional information on recognition of the complicated voices or patterns, the picture processing, etc.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to an optical neural computer, and in particular to an optical neural computer that can purposefully increase its processing capacity through interaction with the outside world during the processing process. This invention relates to an optical neural computer in which the neural computer is composed of optical elements.

[Conventional technology]

Conventionally, processing devices using neural networks have been developed. For example, “Nikkei Electronics J No.
, 427 (1987, 8, 10) pp, 115-12
4 uses neural networks for pattern recognition, signal processing,
The equipment used for knowledge processing is described.

A neural computer is a computer that uses a neural network, and it emulates the brain's excellent information processing mechanism and has the characteristics of parallel distributed processing, learning function 2, and analog processing6.Therefore, a neural computer AI (Artificial Intelligence) based on von Neumann computer
enca (artificial intelligence brain) is weakest at non-verbal processing 1. For example, it is best at speech processing, for which algorithms are difficult to define, and pattern recognition processing.

Additionally, neural computers have a learning function to increase their processing power through learning.

(7) The learning function is performed using synapses that connect output or feedback connections between neurons that make up the neural network, and synapses that connect to the input layer. That is, the learning function is realized by utilizing the plasticity of synaptic connections to change the connections between neurons or the weights of the connections in response to external stimuli.

Note that there are two types of learning methods: a first method with a teacher signal and a second method without a teacher signal. Here, we will discuss learning without a teacher signal, that is, a self-organizing neuron computer.

[One problem that the invention seeks to solve] Conventionally, when constructing a neuron computer with hardware to create a practically meaningful machine with a certain degree of complex recognition function, the number of neuron computers was tens of thousands to hundreds of thousands. Since it is necessary to connect processors corresponding to neurons to each other, if a processor of an electronic device such as a transistor is used.

In addition to the huge number of wires, problems such as delay time in the wires, interference with each other's power levels, and insufficient bandwidth have arisen. As a result, the advantages of parallel processing, which is an original feature of neural computers, could not be fully utilized.

Therefore, various optical neural computers have been proposed for the purpose of introducing optical technology into this neural computer and taking advantage of the inherent characteristics of light such as parallelism, non-coherence of optical paths, and high bandwidth. It was done. however,
In either case, there is currently no progress in hardware due to reasons such as the lack of suitable optical elements that play the role of neurons and the immaturity of optical wiring technology.

Currently, there is a desire to develop an optical neural computer that has a learning function and is capable of performing complex processing.

The purpose of the present invention is to solve such conventional problems and to perform complex voice or pattern recognition, image processing, etc. that could not be realized with conventional Von-Neumann type computers.
The purpose of this invention is to provide an optical neural computer that can perform intelligent processing of dimensional information at high speed.

[Means to solve the problem]

In order to achieve the above object, the optical neural computer of the present invention has a plurality of neurons constituting the processing device of the computer, each of which is a processor with nonlinear input/output characteristics of multiple inputs and one output, and has a feedback loop to itself. , the output line to Iya's neuron and multiple input lines from the other two neurons are interconnected, and a cooperative action in which the neurons are excited together through positive synaptic connections in response to sensory input received from the outside world. In a neural computer that performs parallel distributed processing by performing a competitive action that suppresses the excitation of other neurons through negative synaptic connections and outputting a response to the outside world.

Each of the neurons described above is an optical neuron using an optical element, and the optical neuron is composed of a plurality of optical logical operation units arranged on a two-dimensional plane, and the synaptic connections between the optical neurons are connected to optical signals. It is characterized by being configured with an optical branching element that branches and an optical connection that transmits the optical signal from the optical branching element. Further, in the optical logic operation unit, each optical neuron receives optical input signals from a plurality of other optical neurons inputted via optical connections, and generates positive synaptic connections based on the received optical input signals. An excitation signal for synaptic activation and an inhibitory signal for negative synaptic connection are generated separately, and the difference between the two signals is optically calculated for each optical neuron. Another feature is that only the output signal when the signal is ON and the suppression signal is OFF is separated, and this output optical signal is extracted and fed back as an input signal to a plurality of other optical neurons via optical connections.

[For production]

In the present invention, a neural computer is realized using optical elements, and the inherent properties of light such as parallelism, optical path incoherence, and high bandwidth can be fully utilized. Unlike conventional neural computers that use electronic devices as processors corresponding to neurons, the optical neural computer of the present invention uses optical elements for neurons and uses optical space wiring between each element, so there is no need for metal wiring. It is possible to eliminate various disadvantages of. In other words, since it is an all-optical neural computer that performs everything optically, it eliminates the various drawbacks that occur with metal wiring, requires only a small number of actual wires, has little delay time due to wires, and eliminates interference from mutual optical signals. This has the advantage that there is no need to use the network, and there is no shortage of bandwidth.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

First, we will explain the dynamics of excitation and inhibition of the neural network.

FIG. 3 is a model diagram showing mutual connections between neurons.

In Figure 3, 1 and 2 are neurons that constitute a neural network, 3 is a synaptic connection that performs feedback,
4 is a synaptic connection from another neuron or a synaptic connection with the input layer, 5 is a synaptic connection of output to another two-neelon, and 6 and 7 are two-neelon #1 to #2, respectively.
Alternatively, it is the synaptic connection strength of neurons #2 to #1.

Synaptic connections include excitability with connection strength w > O and w <
There are two types of 00 suppressive properties. In that case, if the connection Wij from one neuron to the other neuron and the connection W, ku from the other neuron to one neuron are both positive, both neurons will be easily excited, and a cooperative effect will occur. .

Furthermore, if the connections Wjj and Wj are both negative, the firing of one neuron suppresses the excitation of the other neuron, so it is difficult for the two neurons to be excited at the same time<<, and only one neuron is excited. This causes a competitive effect. These cooperative effects and competitive effects are the basic dynamics of parallel information processing in neural networks.

FIG. 4 is a diagram showing a neural network with a one-layer structure.

A neural network generally has a layered structure as shown in FIG. In FIG. 4, 8 is a neuron, 9 is a synaptic connection, 10 is an input signal, and 11 is an output signal. As is clear from this, synaptic connections are
Found in neurons within the same layer and between different layers.

Fig. 5 is a diagram showing a one-dimensional neural network, Fig. 6 is a spatial distribution diagram of each synaptic connection strength in Fig. 5, and Fig. 7 is a characteristic diagram showing a sigmoid type threshold function. be.

For simplicity, we focus on the dynamics of neurons within the same layer. Considering a one-dimensional neural network as shown in Fig. 5, the spatial distribution of the strength of each synaptic connection is a lateral inhibitory connection as shown in Fig. 6.

In other words, it is assumed that inputs near the input position have positive synaptic connections, and inputs on both sides have negative synaptic connections.

In FIG. 5, among the inputs 10, white circles indicate excitability and black circles indicate inhibition, and each output 11 shows a synaptic connection distribution as shown in FIG. 6. That is, excitatory input is given to both sides including the self, and inhibitory input is given to both sides. In Figure 5, the state is 1
Although only the output of one neuron is shown, the outputs of other neurons are also the same.

In neurophysiology, this model explains the coordination of edges and contrast in vision. in general,
A multilayer neural network.

Although it is known to have more complex processing functions, the fifth
Since it is easy to expand from the single layer shown in the figure to the multilayer shown in FIG. 4, the model shown in FIG. 5 will be considered here, focusing on the side-inhibiting connections.

The operation of individual neurons is as follows.

The internal state 5i(t
;) is expressed by the following equation using the input T (1) and the input Oi-m(t i) from another input at time (t-1).

S □ (t) = Ii (t) 10Σγ, ○i, k (t-1)
To ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ (1)
Here, γ3 is the spatial distribution of each synaptic connection strength in FIG. The output O of the neuron is obtained by thresholding this internal state using a sigmoid-type threshold value σ(·) as shown in FIG.

Oyo (t)=σ (Si (t)) ・ ・ ・
・ ・ (2) FIG. 2 is a diagram showing an example of calculation of temporal changes in output signals of neurons at each position.

The output level of the neuron group centered on the neuron that receives input from the outside world gradually becomes more excitatory over the course of an hour, while the excitation of the surrounding neuron group is gradually suppressed. So-called “clus”
tering' occurs, and learning occurs such that only a specific group of neurons responds to a specific stimulus from the outside world. That is, in FIG. 2, when 1=0, a specific neuron group is given a slightly stronger excitation than the surrounding area, but as time passes at t=1.2.

Gives stronger excitation to specific neurons compared to surrounding areas. The intensity of excitement for t=1 to 5 is shown.

In this way, in the process of processing stimuli (inputs) from the outside world, increasing processing capacity through interaction with the outside world is called self-organization of neural networks.

FIG. 1 is a block diagram of an optical neural computer showing one embodiment of the present invention.

The present invention realizes the dynamics of the neural network shown in FIG. 5 with side-inhibiting connections using an optical neural network.

In FIG. 1, 12 is an optical calculation unit, 13 is a signal input section, 16 is an optical distributor for the input signal from the signal input section 13 when the excitation signal is ON and the inhibition signal is OFF, and 18
is a completion line connecting the optical splitter 16 and the signal input section 13.

The optical neural computer of the present invention separates the input of the excitation signal and the input of the inhibition signal by the optical distributor 16, and
Performs processing equivalent to taking the difference between the numbers. The optical operation unit 12 operates as 1 only when one of the two operators is 1.
Performs an operation that outputs , that is, an exclusive OR operation.

FIG. 8 is a specific configuration diagram of the optical operation unit in FIG. 1.

In FIG. 8, 19 is an optical input/reflection type spatial light modulator;
20 is a birefringent plate, 21 is a half mirror, 22 is a calculation kernel, 23 is an input cover turn A, 24 is an input cover turn B
, 25 is output light, and 26 is a reading laser beam.

This optical operation unit performs arbitrary logical operations (16
It is possible to perform various types of pool logical operations). Regarding the principle of calculation, see 1 for example.

Since it is explained in detail in the specification of Japanese Patent Application No. 63-26887 and the drawings, the detailed description will be omitted. As shown in FIG. 8, the logic of 0 and 1 of each pixel of the input pattern is expressed by polarized waves, and the four combinations of the logic of each pixel of two people are encoded by spatially separating them. The operations are performed according to arithmetic instructions that perform spatially separated polarization components by an arithmetic kernel.

This can be done by selective extraction.

The arithmetic processing to be performed is to take the difference between the 52 inputs, but since negative expression cannot be expressed with light intensity, exclusive OR,
In other words, an operation is performed that outputs 1 only when either one of the two signals is 1, and as described later, these two types of output signals are separated by the optical splitter 16 shown in FIG. Performs processing equivalent to taking .

In FIG. 8, the laser light incident on the optical input/reflection type spatial light modulator 19 is transmitted to the optical input/reflection type spatial light modulator 19 in which binary information (0 or 1) of the input coverage pattern A is written.
The polarized light is modulated into either ↑→ by the optical input/reflection type spatial light modulator 19 in which binary information of the input pattern B is written, and the birefringent plate 20 spatially separates the light into upper and lower directions. be done. Therefore, one pixel is (00)
The result is a combination of logical values (01), (11), and (10), and the calculation result is obtained by opening any of the four windows by the calculation kernel 22 according to the logical calculation. For example, when A AND B, only the upper right of the four windows opens and the rest close. Also, in the case of A ORB, the lower two of the four windows are open and the upper two are closed.

FIG. 9 is a diagram showing the output of one pixel of exclusive OR obtained by the calculation method of the present invention.

27 in Fig. 9 is the excitation signal ON and the inhibition signal OFF.
The output light in the case of 28 is the excitation signal OFF and the inhibition signal O
This is the output light in the case of N. In this way, it can be seen that the two outputs are spatially separable.

FIG. 10 is a configuration diagram of the calculation kernel in FIG. 1.

The calculation kernel 22 in FIG. 8 is a mask having slits as shown in FIG. In the figure.

29 is an opaque plate that blocks light, 30 is an area occupied by a pixel, and 31 is a slit that allows light to pass through. That is, as shown in diagram M9, the excitation signal is ON and the inhibition signal is O.
This is a slit that allows output light to pass in the case of an FF. The positions and dimensions of these slits and calculation kernels must be matched so that each pixel has a one-to-one correspondence.

If the two dimensions do not match, adjustment can be made by inserting a lens system to enlarge or reduce the size.

FIG. 11 is a block diagram of the optical distributor in FIG. 1.

In FIG. 11(a), 33 is input light, 34 is a mirror, 35 and 36 are slits, 37 is output light, and 21 is a half mirror. - Also, as shown in (b) and (c), the shape of the slits 35 and 36 is circular or annular. Note that 29 is an opaque iFi that blocks light.

The light distributor 16 is for distributing light to a plurality of pixels corresponding to neurons on the incident surface of the reflective spatial light modulator 19 shown in FIG.

FIGS. 12 and 13 are diagrams showing the slit and the spread of the light beam.

FIGS. 12(a) and 12(b) show the state of diffraction of the light that has passed through the slit shown in FIG. 11 and the spatial light intensity distribution in the near field.

In the figure, 37 is the diffracted light, and 38 is the spatial light intensity distribution of the diffracted light. In addition, 33 is an input light, 35 is a slit, 2
9 is a light shielding plate.

The spread angle of the diffracted light is as shown in FIG. 13, assuming that the light beam has a Gaussian distribution.

It is inversely proportional to the opening diameter of the slit 35, and a linear relationship holds true.

Δθ=1.22(λ/D) ・ ・ ・ ・ ・ ・
(3) However, λ is the wavelength.

Therefore, the aperture diameter, the distance between the slit and the input surface of the spatial light modulator, and the intensity ratio of the light input to both slits can be set as desired by appropriately determining the transmittance of the half mirror 21 shown in FIG. Light beam spread and intensity distribution can be achieved.

As a result, by feeding back the output excitation signal light and the other output light, the inhibition signal light, to the optical processing unit via the optical wiring, the space of lateral inhibitory synaptic connections as shown in Fig. 6 is created. It is possible to distribute the light input to the slit to multiple neurons according to the target intensity distribution.

FIG. 14 is a characteristic diagram showing a function that synthesizes the spatial intensity distribution of the synaptic connections of the lateral inhibitory type shown in FIG. 6.

When the spatial intensity distribution of lateral inhibitory synaptic connections is synthesized by the above method, a function curve shown in FIG. 14 is obtained. In addition,
For the optical wiring, it is also possible to use spatial wiring using the lens diameter or a bundle of multiple fibers.

〔Effect of the invention〕

As explained above, according to the present invention, since it has a self-organizing function that could not be realized with conventional computers, it is self-adaptive and can perform intelligent information processing such as complex pattern recognition. Parallelism.

It can be performed optically by taking advantage of the incoherence and broadband properties of the optical path. As a result, the present invention can be expected to be widely applied to various expert systems that can only be achieved using AI using large-scale computers.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an optical neural computer showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of calculating temporal changes in the output signals of neurons at each position, and FIG. 3 is a diagram showing the interaction between neurons. A model diagram showing connections. Figure 4 is a diagram of a neural network with a layered structure. Figure 5 is a diagram of a one-dimensional neural network. Figure 6 is a spatial distribution diagram of each synaptic connection strength. Figure 7 is a sigmoid diagram. FIG. 8 is a specific configuration diagram of the optical calculation unit in FIG. 1, and FIG. 9 is an output obtained by the calculation method of the present invention, which is spatially 1 of the separated exclusive OR
A diagram showing the output for each pixel; Figure 10 is a diagram of a mask with slits used as the calculation kernel in Figure 8;
Figure 11 is a configuration diagram of the optical distributor in Figure 1, Figures 12 and 13 are diagrams showing the slit and the spread of the light beam, respectively, and Figure 14 is a composite of the spatial intensity distribution of synaptic connections in Figure 6. FIG. 1.2.8: Neuron, 3: Synaptic connection of feedback, 4: Synaptic connection from other neurons or synaptic connection with input layer, 5: Synaptic connection of output to other neurons, 6, 7: Neuron # 1 to #2,
Synaptic connection strength of neurons #2 to #1, 9: synaptic connection, 10: input signal, 11: output signal, 12: optical operation unit, 13: signal input section, 16: excitation signal ON
and an optical splitter for the output signal from the optical operation unit when the suppression signal is OFF, 18: optical connection connecting the optical splitter and the signal input section, 19: optical input/reflection type spatial light modulator, 2
0: Birefringent plate, 21: Half mirror 22: Computation kernel,
23: Inlet cover turn A. 24: Input cover turn B, 25: Output light, 26: Readout laser light, 27: Output light with excitation signal ON and inhibition signal OFF, 28: Output light with excitation signal OFF and inhibition signal ON, 29: Block light 1 opaque plate, 30: area occupied by pixels, 31; slit through which the reflected light from the light input/reflection spatial light modulator passes through the birefringent plate and is spatially separated, 33: input light; 34: Mirror, 35, 36
: slit, 37: output light. ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓

Claims (2)

[Claims] (1) Each of the plurality of neurons constituting a computer processing device is a processor with nonlinear input/output characteristics of multiple inputs and one output, and has a feedback loop to itself, an output line to other neurons, and a It is interconnected with multiple input lines from neurons, and in response to sensory input received from the outside world, neurons are excited together through positive synaptic connections, and negative synaptic connections suppress the excitation of other neurons. In a neural computer that performs parallel distributed processing by performing competitive actions and outputting responses to the outside world, each of the neurons described above is an optical neuron using an optical element, and the optical neurons are arranged on a plurality of two-dimensional planes. an optical logical operation unit, and a synaptic connection between the optical neurons is composed of an optical branching element that branches an optical signal and an optical connection that transmits the optical signal from the optical branching element. neural computer. (2) In the optical logic operation unit, each optical neuron receives optical input signals from a plurality of other optical neurons inputted via optical connections, and creates positive synaptic connections based on the received optical input signals. An excitation signal to perform this and an inhibitory signal to perform negative synaptic connection are generated separately, and the difference between the two signals is optically calculated for each optical neuron. Signal ON and suppression signal OFF
2. The optical neuron according to claim 1, wherein only the output signal in the case of 1 is separated, and the output optical signal is extracted and fed back via an optical connection as an input signal to other plurality of optical neurons. Computer.