JPH04328691A - Optical neural processor - Google Patents

Abstract

PURPOSE:To simplify an optical system and to improve the performance by providing a time series optical weighting means and a light accumulating threshold processing means and performing the weighting, adding and threshold- processing of an arbitrary number with one optical system. CONSTITUTION:A time series weighting means is constituted of a programmable weight generating part 1 and a code converting part 2 and the different weighting is performed to the time series. An accumulation type threshold processing part 3 is constituted by using a space light modulator consisting of the ferroelectric positive liquid crystal. For a constant determined from the intensity and wavelength time of the light made incident from within the limited time for the writing input of the space light modulator, the action of threshold, that is, the threshold processing is performed. Here, plural types of the coupling realized optically by the programmable weight generating part 1 and the code converting part 2, that is, the light of the weighting between neurones in the neural net is inputted to the part 3, an output 0(t+1) is outputted, held at a latch 4 and used for the next input.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] The present invention relates to an optical neural processor that performs neural network processing operations using light as a medium, and in particular, an optical neural processor suitable for simplifying the configuration of an optical system and improving processing performance. It concerns processors.

0002

2. Description of the Related Art In recent years, research and development have been carried out on processing devices that use networks that imitate the human brain, that is, neural networks. Furthermore, research is progressing on optical neural processors that control neural network processing using light as a medium. Regarding such an optical neural processor, as explained below, there is a paper published in the Department of Electronics, Information and Communication Studies journal "D-II Vol. J73-D-I
I No. 8 (1990), pages 1346 to 1353, there is a ``Light Side Suppression Neural Network.''

FIG. 4 is a block diagram showing the configuration of an optical side suppression neural network using a conventional optical neural processor. This light-side suppression neural network applies positive and negative weights to the branched input patterns using light as a medium, and then compares the combined results with a predetermined threshold to determine the input pattern. The corresponding recognition result is output, and the input I(0) 41 is electrically or optically branched at a branching section 42, and the branched input at the branching section 42 is multiplied by a positive weight and a negative weight, respectively. Function processing units 43 and 44, an inversion unit 45 that inverts the inputs multiplied by negative weights in the weight function processing unit 44 to positive values, and the respective inputs multiplied by these positive weights and negative weights are added. an optical merging section 46; a threshold processing section 47 that performs output control based on the addition result at the optical merging section 46;
, a latch 48 that holds outputs O(t) and O(t+1). It should be noted that here, the positive weights act to emphasize so-called neurons so that they are excited together, and the negative weights act competitively to suppress the excitation of other neurons.

The operation of this light-side suppression neural network will be explained below. First, the input I(0) 41 is input to the branch unit 42
One is multiplied by a positive weight W+ based on the intensity of the light in a weighting function processing section 43, and the other is multiplied by a negative weight W+ based on the intensity of the light in a weighting function processing section 44. -, respectively, Iexc(=ΣO(t
)・w+) and Iinh(=ΣO(t)・w−). Note that since negative values cannot be expressed by light intensity, the negative weighted output is made into a positive value by adding a certain value to the negative value so that it appears to be a pseudo negative value. This operation is reversal in the reversing unit 45, and from Iinh to (1-Iinh)
I am getting . Next, at the optical merging section 46, Iexc and (1-Iinh) are added, and Iexc-Iinh+1
The threshold processing unit 47 performs threshold processing on this,
Obtain output O(t+1). If pure subtraction is possible, then
You can perform thresholding by running Iexc-Iinh, where you can raise the threshold by one level and run Iexc-Iinh.
−Iinh+1 is subjected to threshold processing by Iex
The result is the same as performing threshold processing on c-Iinh, and there is no problem. Input I(0) 41 is at time t=0
For example, if the output O(1) is obtained by this input, this is temporarily stored in the latch 48, and the output O(1) read from the latch 48 is newly read out.
The same process as above is used to obtain the output O(2). If the error between this output O(2) and the output O(1) is small, it indicates that the recognition result is sufficient, so repeat this process several times until the output becomes steady. , finish this process and get the final output.

Next, the configuration of the optical system of such an optical neural processor/neural network will be explained using a more specific example. 5 and 6 are block diagrams showing the configuration of a two-dimensional suppression neural network using a conventional optical neural processor. In particular, the configuration of the optical system is shown in detail, and this two-dimensional side suppression neural network is
Using each processing element shown in the side suppression neural network in Fig. 4, weighting is performed using light as a medium for two-dimensional input such as figures, and recognition processing of the input is performed. 0) A branching unit 52 (shown in FIG. 6) that branches 51, and a positive weighting unit 53 (shown in FIG.
) and the negative weighting unit 54, and the subtraction unit 56 which combines the outputs from the positive weighting unit 53 and the negative weighting unit 54 using light as a medium and adds them. =Iexc-Iinh+1) into an electrical signal, and then performs threshold processing and output control using a CCD (C
It is comprised of a latch 58 that latches the output from a charge coupled device (charge coupled device) camera 57 and a CCD camera and sends it out as the next input.

[0006] Here, the positive weighting section 53, the negative weighting section 54, and the subtracting section 56 are each constituted by an optical system including a lens and are independent systems. That is, the positive weighting section 53 and the negative weighting section 54 each include a lamp as a light source of readout light, two lenses, and an aperture (denoted as A in the figure) that blocks light from the lamp. , two polarizers (denoted as POL in the figure, Polarizer) that polarize light,
A spatial light modulator (indicated as SLM in the figure, Spatia
(Light Modulation) and a mask that is made of a pinhole array or the like and uses diffraction to weight the light modulated by the spatial light modulator, and receives a negative feedback optical signal and a positive feedback optical signal, respectively. Create an optical signal. Note that the two-dimensional suppression neural network shown in this figure is provided with a light receiver such as a CCD camera that converts these feedback optical signals into electrical signals and outputs them. Note that the polarizer after the mask in the negative weighting section 54 performs inversion processing of the negative weighting. Further, the subtraction unit 56 includes a plurality of lenses, a polarizer (indicated as POL(1) to (4) in the figure),
a spatial light modulator (denoted as SLM (1) and (2) in the figure) that optically modulates the readout light based on the outputs from the respective light receivers of the positive weighting section 53 and the negative weighting section 54;
It is composed of a mirror (denoted as M in the figure) and a half mirror (denoted as H.M. in the figure), and combines the negative feedback optical signal created by the positive weighting section 53 and the negative weighting section 54 with the positive feedback optical signal. subtract the feedback optical signal. In this way, with the configuration of independent optical systems, using light as a medium, the positive weighting section 53 creates a positive feedback optical signal, the negative weighting section 54 creates a negative feedback optical signal, and the subtracting section 56 creates a positive feedback optical signal. , add them. C
The CD camera 57 receives the addition result (Y=I
exc-Iinh+1) is converted into an electrical signal, subjected to threshold processing, and output. Output O from CCD camera 57 (
t) is held in latch #2 of latch 58 and input to branch 52 as the next input. Then, similar weighting and addition processing is performed, and as a result, the output O(t)
is held in latch #1, and the processing result is finally output when the error between the outputs O(t) and O(t+1) held in latch #1 and latch #2 becomes smaller.

[0007]

The conventional optical neural processor of the optical neural network shown in FIGS. 4 to 6 has the following problems. For example, Figures 5 and 6
Weighting to achieve the desired processing requires two optical systems, a positive weighting section 53 and a negative weighting section 54, and an optical system for adding up the results obtained by these optical systems. system, resulting in a complex configuration using a total of three optical systems. Further, weighting for realizing desired processing is specifically implemented using, for example, a pinhole array. In this case, the actually required coupling may not necessarily be obtained with a single pinhole array, but rather multiple pinhole arrays (usually each with different weighting) are used, each with different weights. In some cases, the combined weights are superimposed to obtain a combination. In such a case, it is necessary to configure an optical system using a pinhole array that performs multiple branches at the branching section 42 and weights differently at each branch destination, resulting in an extremely complicated optical system. turn into. Furthermore, threshold processing in conventional optical neural processors involves converting an optical signal into an electrical signal using a CCD camera or the like, and then performing electrical processing, making the control process complicated. Further, in the subtraction unit 56, positive and negative addition is performed optically in order to perform subtraction, but it is difficult to miniaturize by using a laser for the readout light. This is because the subtraction unit 56 will cause interference if a recent semiconductor laser with high spatial coherence, single wavelength, and large emission output is used. For this reason, the configuration of the optical system is restricted. Furthermore, the conventional technology requires two latches to hold the output. This is because the processing result Y is obtained in real time after outputting O(t) from latch #2, so it is necessary to latch the output O(t) and output O(t+1) separately. . In other words, if there is only one latch, the processing result will be obtained immediately, and it will not be possible to store it unless the previously latched information is erased. However, when it is erased, the original input is erased, so the output is This is because you will not be able to obtain it.

The problem to be solved is that in conventional optical neural processors, each process such as weighting, addition, and threshold processing is performed by each processing element alone, and the configuration including the optical system is complicated. This is the point where it becomes. The purpose of the present invention is to provide an optical neural processor that solves these problems of the prior art, simplifies the configuration of the optical system, improves processing accuracy and speed, and makes it possible to reduce the size and cost of the device. It is to be.

[0009]

[Means for Solving the Problems] In order to achieve the above object, the optical neural processor of the present invention has the following features: (1) Using light as a medium, weighting an input with a plurality of values, and input weighted with the plurality of values. The optical neural processor controls the output corresponding to the input based on the comparison between the added input value and a predetermined threshold value. The present invention is characterized by the provision of a time series weighting section that applies weights using light that changes serially as a medium. (2) In the optical neural processor described in (1) above, the time-series optical weighting section sequentially receives and accumulates the light output corresponding to the input weighted in time-series with each value, The present invention is characterized in that it includes a light accumulation threshold processing section that controls the output using the light corresponding to the input as a medium based on a comparison between the accumulated light and a predetermined threshold value.

0010

[Operation] In the present invention, a time series weighting section is provided,
Weighting of input using light as a medium is performed using weights that change over time. This allows an arbitrary number of weights to be applied with one optical system. Further, a light accumulation threshold processing section is provided to sequentially accumulate the respective weighted lights and control the output corresponding to the input based on the accumulation result. As a result, the weighting, addition, and threshold processing processed by the neural network can be performed in one optical system, achieving completeness of the system and simplifying the configuration of the optical system of the optical neural processor. can.

[0011]

Embodiments Hereinafter, embodiments of the present invention will be explained in detail with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the configuration of an optical side suppression neural network using an optical neural processor according to the present invention. In particular, the present invention is applied to the light-side suppression neural network shown in FIG. The light-side suppression neural network of this embodiment uses weighting functions W+ and W- to multiply input I(0).
A programmable weight generation unit 1 that generates , a code conversion unit 2 that controls inversion/non-inversion of the output multiplied by the weight function in the programmable weight generation unit 1 , and an optical addition based on the weight function from the code conversion unit 2 - Consists of an accumulation-type threshold processing unit 3 that performs subtraction and threshold processing, and a latch 4 that holds the processing results from the accumulation-type threshold processing unit 3. Figure 4
The difference from the conventional light-side suppression neural network shown in FIG. In other words, a code converter 2 that can program whether or not to invert the output multiplied by the weighting function.
and that the optical addition/subtraction and threshold processing are executed by the storage type threshold processing section 3, which is a single unit. The programmable weight generation section 1 and the code conversion section 2 constitute the time series weighting means of the present invention, and weight the time series differently. The storage type threshold processing section 3 corresponds to the optical storage threshold processing means of the present invention, and is, for example, a spatial light modulator made of a ferroelectric positive liquid crystal having a GaAs (gallium arsenide) pin structure on the input surface. (SLM) provides the following features. Its characteristics are that for the write input of the spatial light modulator, the constant determined from the intensity, wavelength, time, etc. of the light incident within a finite time,
This is a threshold operation, that is, threshold processing. From now on, such a spatial light converter,
In particular, it will be described as AT-SLM. In this way, the light-side suppression neural network of this embodiment uses multiple types of connections optically realized by the programmable weight generation section 1 and the code conversion section 2, that is, the light-side suppression neural network that uses the light weighting between neurons in the neural network. , the output O(t+1) based on the threshold operation of the AT-SLM is output as a processing result, or is held in the latch 4 and used as the next input. It is characterized by its use.

The operation when using a combination of diffraction by pinholes as the weighting function W will be described below. The programmable weight generator 1 can generate only one type of weight function W at a certain time, but can generate different weight functions W at different times. Specifically, for example, the pinhole array used in the prior art is replaced with a spatial light modulator such as a liquid crystal panel, and different weights are generated in time series. That is, if the wavelength of the readout light is λ, the radius of the pinhole is a, and the distance from the pinhole to the light-receiving surface is L, then the optical weight W(r) obtained at a point r from the center of the light-receiving surface is is w(r
)={(πa2÷λL)2}×{2J1(2πar÷λ
L)
÷(2πar÷λ
L)}2

...(1)
J1: A first-order Bessel function. Here, by changing the radius a using a spatial light modulator such as a liquid crystal panel, or by changing the wavelength of the light source or, in some cases, the distance L, various weight w(r) profiles ( Synapse connection distribution creation) becomes possible, and weights w(r) obtained by integrating these can be used. That is, in the conventional technology, if the system was constructed without using merging and branching, only the weight w(r) expressed by the above equation (1) was obtained, but by applying the present invention, Then, with only a slight change in the system, w(r)=Σ{(πa2÷λL)2}×{2J
1 (2πar÷λL)

÷(2πar÷λL)}2

・
...(2) can be obtained. The summation (Σ) is performed by the accumulation-type threshold processing unit 3 on the weight function w generated by the programmable weight generation unit 1 during the process ready state (AT-SLM write timing).

[0013] Furthermore, by using the code conversion unit 2 in combination with the programmable weight generation unit 1, the signal multiplied by the weight function is branched into either inversion or non-inversion processing, and a positive feedback signal Iexc is generated. , a negative feedback signal 1-Iinh is obtained. In other words, instead of creating weighting functions spatially separately and adding them as in the prior art, it is possible to generate a positive feedback optical signal at a certain timing and a negative feedback optical signal at another timing. These signals are input to the accumulation type threshold processing section 3. This eliminates the need for two weight function generators used in the prior art, and also eliminates the need for a merging processing element that optically adds the positive and negative weights. In this way, the negative part is added to the above-mentioned equation (2) indicating the positive part, and weighting according to the following equation becomes possible. w(r)=Σ{(πa2÷λL)2}×{2J
1 (2πar÷λL)

÷(2πar÷λL)}2
+Σ{1-(πa2÷λL)2}×{2J1(2πa
r÷λL)

÷(2πar÷λL)}2

・・・
- (3) Note that the sum (Σ) here is the same as in equation (2), and the accumulation type threshold processing unit 3 is in the processing ready state (AT-S
This is performed regarding the weight function w generated by the programmable weight generation unit 1 during the LM write timing). Then, the accumulation-type threshold processing section 3 sequentially accumulates signals using light as a medium that have been weighted differently in time series, and performs threshold processing on the total sum (Iexc-Iinh+1).

[0014] In this way, the light-side suppression neural network of this embodiment can perform the weight creation and addition operations, which were conventionally performed separately in a spatially divided system, in a time-series manner. . As a result, the optical system of the conventional two-dimensional side suppression neural network, which has an extremely complicated configuration as shown in FIGS. 5 and 6, can be simplified as shown in FIG. 2 below. Can be done. FIG. 2 is a block diagram showing an embodiment of the configuration of a two-dimensional suppression neural network using an optical neural processor according to the present invention. In particular, the present invention is applied to the two-dimensional suppression neural network shown in FIGS. 5 and 6. The two-dimensional suppression neural network of this embodiment includes a programmable mask 21 corresponding to the programmable weight generation section 1 in FIG. 1, and a code conversion section 2 in FIG.
The programmable weighting circuit 20 of the present invention includes a polarization modulator 22 corresponding to the storage type threshold processing section 3 of FIG. , the optical signal from the programmable weighting circuit 20 is transmitted to the lens L (3)
, L(4) and converts it into an electrical signal.
The configuration includes a camera 28 and a latch 29 that holds the electrical signal converted by the CCD camera 28 for use as the next input to the programmable weighting circuit 20.

As described above, in the programmable weighting circuit 20, the programmable mask 21 corresponds to the programmable weight generation section 1 in FIG. 1, and the polarization modulator 22 corresponds to the code conversion section 2 in FIG. The mask 21 and the polarization modulator 22 constitute a time-series weighting means of the present invention, which performs positive/negative and various weighting function w(r) profiles, and corresponds to the accumulation type threshold processing section 3 of FIG. The AT-SLM 23 and the polarizing beam splitter 27 constitute the light accumulation threshold processing means of the present invention, and perform threshold processing to control the output. The programmable weighting circuit 20 includes a conventional lamp, lenses L(1), L(2), aperture A (indicated only by A in the figure), polarizers 24, 25 (in the figure, POL(1),
POL (2)), a spatial light modulator 26 (denoted as SLM in the figure) made of a liquid crystal cell, etc. Then, the output of the light that has entered the spatial light modulator 26 from the lamp via the lens L (1), the aperture A, the lens L (2), and the polarizer 25 is input to the input I (0) or the output O. (t). Up to this point, the operation is the same as that of the conventional technology.

Furthermore, a programmable weighting circuit 20
Now, the programmable mask 21 gives weights that change over time to the modulated light from the spatial light modulator 26. That is, for example, the weights generated using the diffraction of light by a pinhole array or the like are changed by a method such as changing the radius of the pinhole in time series. The polarization modulator 22 also changes its polarization characteristics in time series in synchronization with the operation of the programmable mask 21, and performs negative weight inversion processing. In this way, with one optical system,
Any type of weighting can be done. After this,
Accumulation of weightings and threshold processing are performed, but a possible coping method is to provide a memory corresponding to each type of weighting and to accumulate weightings that change over time. However, this method requires parallel memories and addition processing of parallel information between memories, making the system complicated. In order to solve such a problem, in this embodiment, the AT-SLM 23 is used to perform weighting accumulation and threshold processing. That is, light weighted by the programmable mask 21 and the polarization modulator 22 is received by the AT-SLM 23 via the polarizer 24. As explained in FIG. 1, this AT-SLM 23, for example,
By using a spatial light modulator made of ferroelectric positive liquid crystal with a GaAs (gallium arsenide) pin structure on the input surface, the intensity, wavelength, and
A threshold operation, that is, threshold processing is performed on a constant determined from time or the like. Through this operation, the AT-SLM 23 sequentially accumulates the weighted light and controls the readout output based on the comparison with the threshold value. That is, the read light source (Read
out Light) from the AT-SLM 23 via the polarizing beam splitter 28,
Output. This output light (Y) is transmitted through lenses L(3) and L(
4), it is input to the CCD camera 28 as the output of Y=(Iexc-Iinh+1), held in the latch 29, and output to the programmable weighting circuit 20 as the next input O(t). In this way, the two-dimensional suppression neural network provided with the programmable weighting circuit 20 of this embodiment performs weighting and addition to input, threshold processing, etc. with one optical system, and is superior to conventional technology. , the configuration of the optical system is significantly simplified.

The advantages of the light-side suppression neural network of this embodiment will be explained below in comparison with the prior art. (1) Programmability can be achieved with a simple system:
When performing multiple weightings, the conventional method is to perform each weighting process, optically combine them, take the sum for positive and negative values, obtain the output, and input it to threshold processing. Instead, the optical system was extremely complicated. However, in the configuration of this embodiment, the storage process and addition process are optically performed by the AT-SLM 23, so the configuration of the optical system is simplified. (2) Threshold processing can also be easily realized: In conventional technology, threshold processing can be performed by using the threshold characteristics of a normal spatial light modulator or by using C
This was done by converting it into an electrical signal using a CD camera, etc., and then processing it electrically. In this embodiment, by using the AT-SLM 23, threshold processing can also be realized with the same element as optical addition. (3) A single optical system performs positive and negative optical analog addition or simple subtraction, simplifying the system: In the prior art, optical subtraction is called optical pseudo-subtraction, and positive signal and a pseudo negative optical signal are created in separate optical systems and added optically to perform positive and negative optical analog addition. The difference in the generation of positive and negative signals depended on the position of the analyzer. For example, if the polarizer and analyzer are orthogonal, a positive optical signal will be generated, and if they are parallel, a negative optical signal will be generated. In this example, a programmable polarization modulation element is used instead of a fixed polarization element in the analyzer, and by changing the modulation of the element in the same optical system, it can produce either a positive signal or a negative signal. Can be done. In other words, the signals are switched in time series such as a positive signal at a certain timing and a negative signal at another timing. For example, a liquid crystal cell is used as such a polarization modulation element. Then, after this, with AT-SLM23, simply
Since it is sufficient to add the positive and negative signals to realize pseudo-subtraction, a single optical system can be used to combine the weighting function multiplication processing section and the
An addition processing section and a threshold processing section can be configured. (4) It becomes possible to easily use a coherent laser as a processing light source: In conventional technology, optical addition is performed to perform subtraction, but in this case, the laser causes interference. Therefore, it is difficult to use it as a readout light, and the configuration of the optical system is restricted. In this embodiment, since the light intensity is taken in for a single light beam, there is no problem of interference caused by merging a plurality of light beams, and a laser can be easily used. (5) AT-SLM also functions as a latch, so
Reducing the number of external latches: In the prior art, two external latches were required, as shown in FIGS. 5 and 6. In this embodiment, the AT-SLM 23 is, for example,
The input surface has a PIN structure of GaAs (gallium arsenide), and the optical modulation section is made of ferroelectric liquid crystal (FLC), and the memory property of this ferroelectric liquid crystal allows it to function as a latch. Therefore, only one latch needs to be provided externally. Note that if a ferroelectric liquid crystal is used as the spatial light modulator 206 in FIG. 2 for converting input into optical information, ferroelectric liquid crystals have memory properties, This eliminates the need for latches and reduces the number of parts.

FIG. 3 is a block diagram showing an embodiment of the configuration of a general optical neural network using an optical neural processor according to the present invention. Although it is time-series, a programmable optical weight generation unit 31 (denoted as W in the figure) creates a weight function, and optically multiplies the input by this weight function. And A.T.
- The SLM 32 performs time-series addition (accumulation) and threshold processing, and also functions as a latch for optical signals. If necessary, the system can be completed with a latch 4 provided externally. This configuration can be applied to various optical neural networks or optical neural processors, including the side suppression type described in detail above. For example, processing is performed by inputting an optical signal modulated by a spatial light modulator or an optical signal weighted by a holographic optical element to the AT-SLM, instead of weighting by various types of diffraction.

As explained above with reference to FIGS. 1 to 3, when constructing a two-dimensional side suppression optical neural network or a general optical neural processor/neural network, the optical neural processor of this embodiment is used. By using this, the optical system becomes extremely simple. Since the adjustment of the optical system in an optical neural processor is generally delicate, simplification of the optical system is extremely important from the viewpoint of configuring an actual optical neural processor. That is, by simplifying the optical system, the optical neural processor can be made more precise, faster, smaller, and less expensive. Moreover, since the optical neural processor of this embodiment can program various weighting functions, although it is a simple optical system, programmability can be greatly improved and it can correspond to various processing contents.

[0020]

According to the present invention, it is possible to simplify the configuration of the optical system of an optical neural processor, improve processing accuracy and speed, and reduce the size and cost of the optical neural network.

[0021]

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the configuration of an optical side suppression neural network using an optical neural processor according to the present invention.

FIG. 2 is a block diagram showing an embodiment of the configuration of a two-dimensional suppression neural network using an optical neural processor according to the present invention;

FIG. 3 is a block diagram showing an embodiment of the configuration of a general optical neural network using an optical neural processor according to the present invention;

FIG. 4 is a block diagram showing the configuration of an optical side suppression neural network using a conventional optical neural processor.

FIG. 5 is a block diagram showing part of the configuration of a two-dimensional side suppression neural network using a conventional optical neural processor.

FIG. 6 is a block diagram showing part of the configuration of a two-dimensional suppression neural network using a conventional optical neural processor.

[Explanation of symbols]

1 Programmable weight generation unit 2 Code conversion unit 3 Accumulation threshold processing unit 4 Latch 20 Programmable weighting circuit 21 Programmable mask 22 Polarization modulator 23 AT-SLM 24, 25 Polarizer 26 Spatial light modulator 27 Polarization beam splitter 28 CCD camera 29 , 30 latch 31 programmable optical weight generator 32 A
T-SLM 41 Input I(0) 42 Branching units 43, 44 Weighting function processing unit 45 Inverting unit 46 Optical merging unit 47 Threshold processing unit 48 Latch 51 Input I(0) 52 Branching unit 53 Positive weighting unit 54 Negative Weighting section 55 Inversion section 56 Subtraction section 57 CCD camera 58 Latch

Claims (2)

[Claims] Claim 1: Using light as a medium, weighting an input with a plurality of values, adding the input weighted with the plurality of values, and comparing the added input value with a predetermined threshold value. In an optical neural processor that controls an output corresponding to the above input based on the comparison, a time series weighting means is provided for weighting the above input with a predetermined value using light as a medium that changes in a time series. An optical neural processor featuring 2. The optical neural processor according to claim 1, wherein the time-series optical weighting means sequentially receives and accumulates the light output corresponding to the input weighted in time-series with each value. . An optical neural processor comprising a light accumulation threshold processing means for controlling an output using light corresponding to the input as a medium based on a comparison between the accumulated light and a predetermined threshold.