JPH0659764A - Optical neuro arithmetic element - Google Patents

Abstract

(57) [Abstract] [Purpose] A versatile, small-sized integrated device that can perform operations in parallel. A light emitting section 2, a liquid crystal spatial light modulating element section 3 and a light receiving section 4 are adhered in this order, and the light emitting section 2 and the liquid crystal spatial light modulating element section 3 and the liquid crystal spatial light modulating element section 3 are attached. There is no gap between and the light receiving part 4. Further, since the liquid crystal spatial light modulation element section 3 is provided with the heat resistant monolithic polarizing layers 11 and 20, it is thermally strong. Further, the liquid crystal spatial light modulation element section 3 has matrix-like regions for adjusting the amount of transmitted light vertically and horizontally, and the amount of transmitted light in each region is independently adjusted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an optical neuro arithmetic element mainly for neural networks used in the fields of information processing systems, equipment control and the like.

[0002]

2. Description of the Related Art The above-mentioned neural network imitates information processing in the brain of an organism, and is extremely useful in the areas such as pattern recognition and association of voices, characters, images, etc. which conventional Neumann computers are not good at. It has demonstrated its ability and has been actively researched and applied in recent years.

FIG. 3 is a model diagram of one unit, that is, a neuron that constitutes a neural network, which corresponds to one brain cell of an organism. The operation of the neuron will be described below with reference to FIG. Neuron _i
(I = 1 to m) is usually treated as generating only one output signal from a plurality of input signals. The n input signals I _j (j = 1 to n, each I _j has a value in the range of 0 ≦ I _j ≦ 1) are multiplied by the value of the weight W _ij and then transmitted to the neuron _i . The total sum X _i of the signals transmitted to the neuron _i is expressed by the following equation 1.

[0004] _{X i = W i1 I 1 +} W i2 I 2 + ··· + W in-1 I n-1 + W in I n ... (1) A sum X _i of the output Y _i input signals from neuron _i
And is expressed by the following two equations.

Y _i = f _i (X _i ) ... (2) The above f is a response function (value in the range of 0 ≦ f ≦ 1), and corresponds to, for example, 1 / (1 + EXP (−X _i ).

A neural network is constructed by connecting a large number of neurons as shown in FIG. 3, and an output signal of a certain neuron becomes an input signal of a large number of other neurons. Further, the neural network has its processing content determined by the coupling between the neurons, the weight W, and the response function f, and the weight W is sequentially changed so that a desired output is derived with respect to the input signal. Learning is carried out by going, and it grows to the state where desired processing is performed. In addition, generally, the larger the number of neurons and the more the connections between neurons, the more complicated the processing becomes.

Initially, such a neural network was realized by a kind of computer simulation in which the connection between neurons and the output operation of neurons were sequentially calculated by a general-purpose computer.

However, when the number of neurons in the neural network and the number of connections between the neurons are increased in order to meet the demand for more complicated processing, the amount of calculation naturally becomes enormous, and in terms of processing speed. There was a very problem.

Therefore, it has been proposed to perform operations in parallel using optical hardware. Fig. 4 shows its configuration. The electrical input signal group is converted into light intensity by a light emitting element array 24 including, for example, five LEDs (light emitting diodes), LDs (laser diodes), etc., and these are converted to a vertical direction (or a horizontal direction) by a lens system 25 on the input side. ), And the light is incident on the planar optical mask 26. The optical mask 26 is divided into 5 × 5 = 25 regions, and each region has a unique transmittance. The transmittance of each region corresponds to the weight W of the neuron, and the difference in the transmittance of light becomes the magnification of the input signal. The light whose intensity is changed by the optical mask 26 passes through the output-side lens system 27 which is divided into five in the direction perpendicular to the input-side lens system 25, and is also composed of five light-detecting light-receiving elements. Collected in the light receiving element array 28, a signal corresponding to each total light receiving amount is extracted from each light receiving element. This corresponds to obtaining the sum of products for each neuron after the above-mentioned input signals I _j are multiplied by the respective weights W _ij .

With this system, the advantage of using light is that, compared with the case where the neural network is sequentially processed by the computer for each unit, the advantage of using light is that the load for multiple units is applied to multiple input signals at the same time. Thus, the speed is improved in that the processing of multiplying the sum and the sum is performed at once and in parallel.

[0011]

However, in the conventional configuration as shown in FIG. 4, a space for expanding light is required in the lenses such as the lens systems 25 and 27 or an optical system having a similar function, so that the calculation is performed. There is a problem that the structure of the device becomes large, and there has been a demand for downsizing the device and making it an integrated element.

On the other hand, for the optical mask 26 of FIG. 4, it is most convenient to use a flat plate whose transmittance in each region is fixed, for example, an ND (neutral density) filter having a density distribution. However, in that case, the transmittance once determined, that is, the load cannot be changed, as a matter of course. In other words, the function of neuro operation is 1
It is limited in type and lacks versatility. Therefore, when the neural network required by the user is created, the same system as the processing system to be created is simulated by an ordinary computer, and a transmission distribution is replaced based on the weight W _ij obtained by neuro learning. Will be done. When this is integrated into an element, the element formation process is performed after the learned load is determined, so it takes time from the ordering to the completion.

On the contrary, if the transmittance of the optical mask is variable for each minute region, the user can purchase the device in the form of a finished product and independently learn the neuron, which is versatile and has the above-mentioned fixed transmittance. Time problems are also eliminated as compared to the case of the mask. It is considered that the most popular liquid crystal element at present is used for controlling the transmittance. However, the liquid crystal display element needs to use a polarizing plate together due to its mechanism, but the conventional polarizing plate has a drawback that it is weak against heat, and an optical neuro arithmetic device capable of performing arithmetic operations in parallel is manufactured by an IC manufacturing technique. It has been a serious obstacle to the demand for miniaturization and integration into integrated devices.

The present invention has been made in view of the above-mentioned problems of the prior art, and provides a versatile, small-sized integrated optical device capable of performing arithmetic operations in parallel. The purpose is to do.

[0015]

In the optical neuro arithmetic element of the present invention, a light emitting portion in which a plurality of linear light sources whose light amounts are independently controlled are provided in parallel, and a matrix area for adjusting the transmitted light amount. And the amount of transmitted light in each region is independently adjusted, and the light emitted from each light source of the light emitting unit is received and modulated in the region of one vertical column.
In addition, a liquid crystal spatial light modulator having a heat-resistant monolithic polarizing layer and a plurality of linear photodetectors for receiving transmitted light from one horizontal row of the matrix-like region are provided in parallel, and each linear photodetector is detected. A light receiving section for detecting modulated light transmitted through the liquid crystal spatial light modulating element section, and the light emitting section and the light receiving section are provided so as to sandwich and contact the liquid crystal spatial light modulating element section. The above object is achieved thereby.

[0016]

In the present invention, the light emitting portion, the liquid crystal spatial light modulating element portion and the light receiving portion are provided in this order, and between the light emitting portion and the liquid crystal spatial light modulating element portion, and the liquid crystal spatial light modulating element portion. There is no gap between and the light receiving part. Further, since the liquid crystal spatial light modulation element section is provided with a heat resistant monolithic polarizing layer, it is thermally strong. Further, the liquid crystal spatial light modulation element section is provided with matrix-like regions for adjusting the amount of transmitted light vertically and horizontally, and the amount of transmitted light in each region is adjusted independently.

[0017]

EXAMPLES Examples of the present invention will be specifically described below.

FIG. 1 is a perspective view showing the structure of the optical neuro-computation element of this embodiment, and FIG. 2 is a sectional view of the optical neuro-computation element. In this optical neuro arithmetic element, a light emitting section 2, a liquid crystal spatial light modulating element section 3 and a light receiving section 4 are provided between two glass substrates 1 and 5, and a plurality of electrodes 9, 14, 1 are provided outside.
5 and 23 are formed.

The light emitting portion 2 is formed by arranging linear ELs, and the insulated ZnS layer 7 is provided with an ITO (transparent electrode) film 6,
It has a structure sandwiched between 8. The ITO film 6 has a linear shape, and the electrodes 9 connected to the ITO films 6 are
The ZnS layer 7 can emit light linearly by applying a voltage from the outside.

A liquid crystal spatial light modulation element section 3 is provided on the light emitting section 2. The liquid crystal spatial light modulation element section 3 includes, in order from the light emitting section 2 side, a transparent insulating film 10, a monolithic polarizing plate 11, a transparent insulating film 12 and a TFT (thin film transistor).
The array 13 has a first laminated body in which the ITO film 18 and the transparent insulating film 1 are provided above the first laminated body.
It has the 2nd laminated body which consists of the monolithic polarizing plate 20 pinched | interposed by 9 and 21. The first laminated body and the second laminated body are bonded to each other with a constant spacing maintained by a spacer 17 interposed therebetween, and between the first laminated body and the second laminated body, A liquid crystal layer 16 made of a ferroelectric liquid crystal is enclosed in a portion surrounded by the spacer 17. In addition,
An alignment film (not shown) is formed on the upper and lower surfaces sandwiching the liquid crystal layer 16, and the alignment direction of the liquid crystal layer 16 is determined by this alignment film.

The monolithic polarizing plate 11 is a metal thin film having slits. For example, a thin metal film of Al, Cr or the like is formed on the transparent insulating film 10, and thin linear slits with an interval narrower than half the wavelength of the light emitted from the light emitting unit 2 are formed on the entire surface thereof. It is performed by forming by line lithography technology. In the monolithic polarizing plate 11 having such slits, the electric vector component in the direction perpendicular to the extending direction of each line of the slits is passed.

The TFT array 13 provided above the monolithic polarizing plate 11 is for driving the liquid crystal layer 16 in a matrix, and is arranged in a matrix. The TFT array 13 is driven by a voltage applied from the outside of the liquid crystal spatial light modulation element section 3 through electrodes 14 and 15 formed outside the element, and has a light transmittance determined based on the orientation angle of the liquid crystal layer 16. Control is performed by dividing into a number of matrix-shaped regions. That is, the liquid crystal spatial light modulation element unit 3
Is provided with matrix-like regions for adjusting the amount of transmitted light in the vertical and horizontal directions, and the amount of transmitted light in each region can be independently adjusted according to the energization state to the electrodes 14 and 15.

The monolithic polarizing plate 20 provided above the liquid crystal layer 16 is manufactured by the same manufacturing method as the monolithic polarizing plate 11 provided below the liquid crystal layer 16, but the polarization direction thereof is the polarization direction of the monolithic polarizing plate 11. And are vertical.

The light receiving portion 4 provided on the liquid crystal spatial light modulation element portion 3 is formed by arranging linear photodiodes 22 made of amorphous silicon side by side. One end of an electrode 23 is electrically connected to each photodiode 22, and each electrode 23 takes out a current according to the amount of light received by the photodiode 22.

The optical neuro arithmetic element having the above structure is manufactured as follows. First, the liquid crystal spatial light modulation element portion 3 portion below the liquid crystal layer 16 and the light emitting portion 2 are formed on the glass substrate 1, while the liquid crystal spatial light modulation element portion 3 portion above the liquid crystal layer 16 is formed. The light receiving part 4 and the glass substrate 5 are formed.

Next, the facing surfaces of the glass substrates 1 and 5 are subjected to an alignment treatment for determining the alignment angle of the liquid crystal, and then the glass substrates 1 and 5 are bonded with a spacer 17 at a constant interval. A liquid crystal layer 16 is formed by enclosing a ferroelectric liquid crystal capable of high-speed operation in a space surrounded by both glass substrates 1 and 5 and a spacer 17.

In the manufacturing process, since the monolithic polarizing plates 11 and 20 are formed of a metal thin film such as Al or Cr, even when heat is generated during the formation of the TFT array 13 or in the subsequent element manufacturing stage. The predetermined function can be achieved without being destroyed.

Next, the operation contents of the optical neuro-computation element of the present invention will be described.

The input signal is input to the electrode 9 of the optical neuro-computation element in the form of an electric signal to cause the linear EL of the light emitting section 2 to emit light. The light emitted from the light emitting section 2 is applied to the liquid crystal spatial light modulation element section 3 to undergo intensity modulation. This intensity modulation is performed by matrix driving the TFT array 13 from the electrodes 14 and 15 and changing the light transmittance of each matrix-shaped region, which corresponds to the control of the weight of the neuron.

The intensity-modulated light is emitted from the linear E of the light emitting section 2.
The light is incident on the light receiving unit 4 having a linear photodiode array arranged in a direction orthogonal to L. Each linear photodiode of the light receiving section 4 outputs a current according to the total amount of received light received. The signal related to this current is taken out to the outside through the electrode 23 and subjected to threshold value processing, for example, by a threshold value processing circuit provided outside, and becomes a neuron output signal. Regarding this output, for example, when an operational amplifier is used and a current (or voltage) above a certain level is reached, a voltage signal may be output from the operational amplifier.

Therefore, according to the present invention, there is no need to provide a space for expanding the width of the light between the light emitting section and the liquid crystal spatial light modulating element section and between the liquid crystal spatial light modulating element section and the light receiving section. The element can be miniaturized. Moreover, since a monolithic polarizing plate made of heat-resistant metal is used, it is possible to use IC manufacturing technology, and it is possible to integrate an optical neuro arithmetic element using liquid crystal for transmittance control (load expression). Become.

Further, the weight application in the neural network can be changed to the control of the liquid crystal. This means that not only the learning function of the neural network is realized, but also multiple types of neuro computations can be performed by a single element by time-division processing, and the user can complete the optical neuro computation element after completion. This means that the optimal learning can be performed by directly incorporating it into the target system, which means that double cost reduction can be achieved.

Furthermore, in order to control the load change, it is necessary to use an ordinary computer system, but in the present technology, the final output obtained by the neural network is processed by an ordinary computer or a microcomputer. Therefore, it is not necessary to consider the cost increase due to the load control. If it is used in a simple device that requires only a single function, it is of course advantageous in terms of cost to use an optical mask with a fixed load, but for such a small-scale operation. In most cases, it is processed by a microphone computer instead of being converted into an element. On the other hand, in the optical neuro arithmetic element of the present invention, by performing time-division processing of the input, output, and weight signals, a large-scale neural network in which a larger number of neurons than the optical neuro arithmetic element has is virtualized. It is possible to configure and multiple neural networks.

Although the threshold value processing circuit is provided outside the optical neuro-computation element in the above embodiment, the present invention is not limited to this, and it goes without saying that it may be incorporated in the optical neuro-computation element.

[0035]

According to the present invention, since no space is required between the light emitting portion and the liquid crystal spatial light modulating element portion and between the liquid crystal spatial light modulating element portion and the light receiving portion, it is possible to form an integrated element. Can be miniaturized. Further, since the heat-resistant monolithic polarizing layer is used for the polarizing plate required for the liquid crystal spatial light modulation element portion, it is thermally strong. Further, the liquid crystal spatial light modulation element section is provided with matrix-shaped regions for adjusting the amount of transmitted light in the vertical and horizontal directions, and the amount of transmitted light in each region is adjusted independently, so that it is versatile and performs operations in parallel. be able to

[Brief description of drawings]

FIG. 1 is a perspective view showing an optical neuro-computation element of this embodiment.

FIG. 2 is a cross-sectional view showing an optical neuro arithmetic element of this embodiment.

FIG. 3 is a model diagram of neurons forming a neural network.

FIG. 4 is a perspective view showing a configuration of a conventional optical neuro arithmetic device.

[Explanation of symbols]

 1 glass substrate 2 light emitting part 3 liquid crystal spatial light modulation element part 4 light receiving part 5 glass substrate 6 ITO film 7 ZnS layer 8 ITO film 9 electrode 10 transparent insulating film 11 monolithic polarizing plate 12 transparent insulating film 13 TFT (thin film transistor) array 14 electrode 15 electrode 16 liquid crystal layer 17 spacer 18 ITO film 19 transparent insulating film 20 monolithic polarizing plate 21 transparent insulating film 22 photodiode 23 electrode

Claims (1)

[Claims] 1. A light-emitting unit in which a plurality of linear light sources whose light amounts are independently controlled are arranged in parallel, and matrix-shaped regions for adjusting the amount of transmitted light are provided vertically and horizontally, and the amount of transmitted light in each region is independent. And a liquid crystal spatial light modulation element section having a heat-resistant monolithic polarizing layer, which is configured to receive and modulate light emitted from each light source of the light emitting section in a vertical one-row region. , A plurality of linear photodetectors for receiving transmitted light from one horizontal row of the matrix-like region are provided in parallel, and each linear photodetector detects the modulated light transmitted through the liquid crystal spatial light modulation element section. An optical neuro arithmetic element comprising: a light receiving section for detecting; and the light emitting section and the light receiving section provided with the liquid crystal spatial light modulation element section sandwiched and in contact with each other.