JPH02256027A - Neuronetwork circuit - Google Patents

Abstract

PURPOSE:To provide the circuit which has a learning function and with which a threshold processing function is attained by hardware and a gradation structure is easily attained by connecting a part (PC part) having a photoconductivity to a thin-film field effect type transistor (TFT) driving a liquid crystal. CONSTITUTION:The PC part has an electrostatic capacity and acts as a capacitor when this part is not irradiated with light. The part loses its function and the electric charges charged therein are annihilated when this part is irradiated with the light of a certain specified quantity or above. The gate voltage can, therefore, be diminished by irradiating the PC part with the light when the TFT is in a conducting state while the charged PC part is kept connected to a gate electrode. The breaking of the conducting state of the TFT is consequently possible. Namely, the threshold processing to incident light is executed. Not only the computing time is shortened by the adoption of the hardware for the threshold processing but also the degree of the convergence of the learning and the operating capacity are improved as compared to the case of learning only the synaps bonding strength.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a neural network circuit that performs input/output operations similar to those of the nervous system, such as pattern recognition, associative memory, and parallel arithmetic processing.

Conventional technology Currently, the development of hardware for the neural functions of living organisms, that is, neurocomputers, is actively being carried out.
There is no hardware version of the learning function and threshold processing function, and these processes depend on a conventional Phono-Neumann type computer. Further, at present, attempts have been made to implement only the neural circuit model consisting of 27I in the input layer and the output layer into hardware, and there has been no report on the implementation of the hierarchical structure model shown in FIG. 8 into hardware.

Problems to be Solved by the Invention Neurocomputers configured with hardware perform learning and threshold processing sequentially and serially using computers, and due to calculation time problems, no one on a scale that can be put to practical use has yet been reported. do not have. In order to put neurocomputers into practical use, it is necessary to perform more complex calculations, increase the number of neurons, and improve learning functions. To achieve this, it is necessary to implement learning functions and threshold processing functions into hardware. .

In order to improve the abilities such as association, ambiguity processing, and pattern identification, which are the characteristics of Mata and Neurocombination,
It is effective to make the network circuits have a hierarchical structure. However, it has been difficult to convert the hierarchical structure into hardware due to problems such as the large number of wires between neurons. Therefore, the neurocomputers currently implemented in hardware are based on a simple model consisting of two layers, an input layer and an output layer, and cannot be put to practical use.

The present invention is intended to solve the above-mentioned conventional problems, and provides a neural network circuit that has a learning function, implements a threshold processing function in hardware, and can easily realize a hierarchical structure. The purpose is to

Means for Solving the Problems In order to achieve the above object, a neural network circuit of the present invention includes a liquid crystal layer, a photoconductive part, and a field effect transistor, and the field effect transistor has a photoconductive part. is characterized in that it includes at least a spatial light modulation element that modulates the electric field intensity applied to the liquid crystal layer by irradiating light onto a portion to which the liquid crystal layer is electrically connected and has photoconductivity. .

A thin film field effect transistor (hereinafter abbreviated as TPT) that drives the working liquid crystal has a photoconductive portion (hereinafter abbreviated as TPT).
(abbreviated as PC portion) (for example, the PC portion is connected to the gate electrode). When the PC part is not irradiated with light, it has a capacitance and functions as a capacitor, but when it is irradiated with light exceeding a certain amount of light, it loses its function and the electric charge that was charged disappears. Therefore, when the TPT is in a conductive state with a charged PC part connected to the gate electrode, the gate voltage can be reduced by irradiating the PC part with light, so that the conductive state of the TFT can be cut off. In other words, threshold processing can be performed on incident light.

Furthermore, if multiple PC parts are connected in series and these PC parts are charged with gate voltage, the gate voltage changes depending on the number of PC parts that are irradiated with light, so the operating state of the TFT can be changed. . That is, threshold processing can be performed for multiple inputs. Further, this threshold value can be controlled by changing the gate voltage that charges the PC portion.

If the threshold processing function of neurons is realized using such a spatial light modulation element, threshold processing that has traditionally been performed using computers can be performed simultaneously and at high speed. Furthermore, since the threshold value can be arbitrarily set by changing the voltage applied to the spatial light modulation element, efficient operation of the neural network circuit can be achieved by controlling the threshold value. In other words, by implementing threshold processing in hardware, not only can calculation time be shortened, but also the degree of convergence of learning and operational performance can be improved compared to the case where only the synaptic connection strength is learned.

Further, when a synaptic connection is formed using this spatial light modulation element, the synaptic connection strength changes with respect to the intensity of the incident signal light, so that unsupervised learning becomes possible.

Furthermore, since this spatial light modulator can be used as a transmissive type, it is possible to construct a hierarchical structure model simply by overlapping them, making it possible to realize more advanced functions.

Embodiments An embodiment of the present invention will be described with reference to the drawings.

FIG. 1 shows an example of a neural network circuit according to an embodiment of the present invention. FIG. 1(a) schematically shows the configuration of a neural network circuit, and FIG. 1(b) and FIG. 1(c) are equivalent circuit diagrams of a spatial light modulation element used as a threshold element. The input pattern display element 101 corresponds to the input layer, the threshold value element 102 corresponds to the output layer, and a synaptic connection therebetween is formed by a synaptic connection element 103.

The input cover pattern display element 101 is a two-dimensional array of light emitting diodes, electroluminescent (EL) elements, a-Si:H or polycrystalline S! ) It consists of an active matrix type liquid crystal cell (hereinafter abbreviated as AM-LC cell) using a transistor array as an active element and a 90° twisted nematic liquid crystal, and a liquid crystal display element that combines a fluorescent lamp and a surface diffuser plate. This is a device that provides human input signals to the circuit.

The threshold element 102 is constructed by arranging a plurality of spatial light modulation elements in a matrix, which are represented by the equivalent circuits of FIGS. 1(b) and 1(c). (b) shows two TPTs (TF
Tl, TPT2), a liquid crystal layer (expressed as CLC in terms of capacitance), and a photoconductive layer in the PC portion.

The photoconductive layer is connected between the gate electrode of TPT2 and the ground, and acts as a dielectric when no light is irradiated, so it is expressed as capacitance Cpc. The operation of the spatial light modulation element will be explained using this circuit diagram.

Va is applied to the gate electrode of TFTl, and the photoconductive layer Cp
c is charged and TFTI becomes conductive. When charging of Cpc is completed, VG is decreased to cut off conduction of TFTI.

At this time, the gate voltage of TPT2 is approximately equal to Vp, and T
PT2 becomes conductive and AC voltage V is applied to the liquid crystal layer CLC. Next, the photoconductive layer CPC is irradiated with light, and when the light intensity becomes large enough to sufficiently reduce the resistivity of the photoconductive layer, the charges stored in the photoconductive layer are neutralized, and the gate voltage of TPT2 becomes Decrease. Therefore, conduction of TPT2 is cut off, and AC voltage V is no longer applied to liquid crystal layer CLC. In this way, the voltage applied to the liquid crystal layer can be controlled by light irradiation, and the device can be operated as a spatial light modulator having a threshold processing function. The circuit (C) consists of one TPT, a photoconductive layer Cpc, and a liquid crystal layer CLC, and the photoconductive layer is TPT.
The liquid crystal layer is connected to the source electrode of No. 3 in series. T.P.
When voltage Va is applied to the gate electrode of T3, TPT3 becomes conductive, and AC voltage V is applied to the photoconductive layer and the liquid crystal layer. However, since C8C is as small as Cic or smaller, (2) mainly affects Cpc in the photoconductive layer and does not affect much in the liquid crystal layer CLC. Next, the photoconductive layer Cpc
When the photoconductive layer is irradiated with light and the resistivity of the photoconductive layer is made sufficiently small, V is applied to CLc. Therefore, the voltage applied to the liquid crystal layer can be changed by light irradiation. As in the case (b), it is possible to change the voltage applied to the liquid crystal layer.

The material used for the photoconductive layer of the spatial light modulator operates as a dielectric in the dark, but loses its dielectric properties due to photoconductivity when irradiated with light. For example, Cd Sr
Cd T el Cd S el Z n Sr
Z nS IBI G aAL G aN+
G a P+ G a A I As.

Compound semiconductors such as InP, S el S e T e
l A Amorphous semiconductor such as sSe, S 1+ G
el S 1 +-*Cx5S I I-xG el
u Gel-xcx (O<X<1) polycrystalline or amorphous semiconductor, and (1) phthalocyanine pigment (P
c) For example, metal-free P CI X P c (X
=Cul N I+ C01T t O+ M
g+ S t (OH)t, etc.), AlClPcC
1, Ti0CIPcC1゜InClPcC1, InCl
Pc, InBrPcBr, etc. (2) Monoazo dye,
Azo dyes such as disazo dyes, (3) penylene pigments such as penylene acid anhydride and penylene acid imide,
(4) Indigoid dye, (5) Quinacridone pigment,
(6) polycyclic quinones such as anthraquinones and pyrenequinones, (7) cyanine dyes, (8) xanthine dyes, (9) charge transfer complexes such as PVK/TNF,
(10) eutectic complexes formed from biryllium salt dyes and polycarbonate resin; (11) organic semiconductors such as azulenium salt compounds.

In addition, amorphous S l+ G e S 11-x
c+uS s + -x G ex+ G e +-
8C, below C, a-8i, a-Ge, a Si+-8
CXI a -S 11-x G ex, aGe+-
*When abbreviated as Cx) is used in the photoconductive layer 103, hydrogen or a halogen element may be included, and oxygen or nitrogen may be included to reduce the dielectric constant and increase the resistivity. P-type impurities B and A are used to control resistivity.
elements such as I, Ga, or P+, which is an n-type impurity.
Elements such as As and Sb may be added. By stacking amorphous materials doped with impurities in this way, p/n+p/l
The dielectric constant may be controlled by forming a junction such as +1/np/i/n and forming a depletion layer within the photoconductive layer.

In addition to such an amorphous material, the dielectric constant of the photoconductive layer may be controlled by laminating two or more of the above-mentioned materials to form a heterojunction. Moreover, the film thickness of the photoconductive layer is 0.01 to 1
00 μm is desirable.

Figure 1 (d) and (e) are respectively Figure 1 (b) and (c).
FIG. 2 is an equivalent circuit diagram in which the equivalent circuit shown in FIG. 1 is changed to one in which a plurality of CPcs are connected in series. In this case, depending on the number of photoconductive layers Cpck (k: LL...+n) to be irradiated with light, the TPT (TPT2 in FIG.
The conduction state of PT3) can be changed.

The synapse coupling element 103 has a function of changing the transmittance, polarization, refractive index, etc. of light or converting the wavelength, and for example, the spatial light modulation element used in the AM-LCD or the threshold element 102 can be used. I can do it.

Specific examples will be described below with reference to the drawings.

Example 1 In the example of the neural network circuit of the present invention shown in FIG. 1(a), the structural cross-sectional view in FIG.
The spatial light modulator represented by the plan view of threshold element +0
2 was used to construct an optical associative memory.

The equivalent circuit of this spatial light modulator is shown in FIG. 1(d). This equivalent circuit is basically the same as that shown in FIG. 1(b), except that a plurality of photoconductive layers are connected in series. In other words, the structure is such that the gate voltage of the TPT changes depending on the number of photoconductive layers that are irradiated with light. This spatial light modulator is made of hydrogenated amorphous silicon (hereinafter referred to as a
-9i:H) was used as the semiconductor layer 201, and two TPTs were formed on one glass substrate 202.

To form TPT, first, a gate electrode 203 is formed on a glass substrate 202 using, for example, Cr, and then a gate insulating film 20 is formed.
4. Semiconductor layer 20! , the semiconductor protective layer 205 is exposed to plasma C
It was formed and patterned using the VD method. After interposing an n-type semiconductor layer 206 to improve ohmic properties,
A source electrode 207 and a drain electrode 208 were formed at once using A1, for example, to produce a TPT. Next, transparent electrode 2
09 is formed of, for example, ITO, and the gate electrode 203 is connected through a hole made in the gate insulating film 204 by etching.

For example, a-Sl:H was formed and patterned as a photoconductive layer 210 thereon, and a transparent electrode 211 was further formed to serve as an electrode of the photoconductive layer 210. Next, pixel electrodes 2+3 for applying an electric field to the liquid crystals 2!2 were formed of transparent electrodes such as ITO. After that, an alignment film 214 was applied and rubbed. The other glass substrate 2!5 is provided with a counter electrode 216 and a light shielding layer 217, and similarly coated with an alignment film 218 and subjected to a rubbing process, but this rubbing process is performed in a direction that is approximately 80° different from that of the previous substrate. Ta. Then, twisted nematic liquid crystal 212 was sealed between both substrates, and polarizing plates 219 were placed in front and behind the substrates. This has a transmission type structure, and the output is given as a transmission image.

In addition, as shown in the plan view of FIG. 2(b), one TPT
The number of photoconductive layers 2!0 connected in series to the gate electrode was 16, and the spatial light modulators were arranged in a 4x4 matrix.

The input cover turn display element 101 is an AM-LC cell that combines a fluorescent lamp and a surface diffuser (however, these AM-LC
The number of pixels of the C cell was 400×400 (=IGOOOO). The light emitted from this AM-LC cell was 400 to G50 nm. On the other hand, the synaptic coupling element 103 used an AM-LC cell using two color polarizing plates with parallel polarization directions. The transmission characteristics of this color polarizing plate in the polarization direction and viewing angle direction are as shown in FIG. 2(C). Furthermore, on the opposite side of the spatial light modulation element used for the threshold element 102 to the synaptic coupling element +03, a color filter is placed that does not transmit light with a wavelength shorter than λ1. The transmitted light from the element 102 becomes output light subjected to threshold processing at a wavelength longer than λ1.

The three letters A, J, and O of the alphabet with a matrix size of 4×4 as shown in FIG. 3 were input to this neural network circuit, learned, and stored in the synaptic coupling element +03. During learning, when we controlled the threshold value according to the input pattern, it converged more than twice as fast as when it was not used, and even when inputting an incomplete pattern as shown in Figure 4. I got a complete output result. We also confirmed that the output (recall results) can be displayed in real time and can be obtained more quickly than when using a computer.

Example 2 A hierarchical neural network circuit as shown in FIG. 5(a) was constructed.

First, the configuration of this circuit will be explained. The input display element 501 is an array of light emitting diodes arranged in two dimensions.
, λ2. It has different emission wavelength components of λ3 (however, λ1, λ2, and λ3). Synaptic coupling element 502.50
3 are both constructed of AM-LC with a pixel count of 512×512. The two polarizers used in the AM-LC of the synaptic coupling element 502 are color polarizers whose polarization directions are parallel, and the transmission characteristics in the direction perpendicular to the polarization direction are shown in FIG. 5(b). It's something like this. Threshold chain element 5
The structure of 04 is represented by the equivalent circuit shown in Fig. 1(e), and as shown in Fig. 5(C), photoconductive layers are arranged in a matrix of 16xlG, which is further arranged in a matrix of IGxlG. be. The transmitted light of the threshold chain element 504 passes through an optical filter 505 having transmission characteristics as shown in FIG. 5(d), and enters the synaptic coupling element 503. The two polarizing plates used in the AM-LC of the synaptic coupling element 503 are color polarizers as in the synaptic coupling element 502, and the transmission characteristics in the direction perpendicular to the polarization direction are shown in FIG. 5(e). It is as shown in . The structure of the threshold chain element 506 is a 4x4 structure as shown in FIG. 1(e).
Furthermore, 4 photoconductive layers were arranged in a matrix of
They are arranged in a x4 matrix. The transmitted light of the threshold chain element 506 passes through an optical filter 508 having transmission characteristics as shown in FIG. 5(f), and an output pattern of wavelength λ3 is obtained. Here, the photoconductive layer used in the threshold chain element 504 absorbs λ1 and absorbs λ2. It is a-5IC:H which does not absorb much light of λ3, and the threshold chain element 5
The photoconductive layer used for 0[i absorbs λ2 and absorbs λ3
a-Sl:I(, which does not absorb much of It has a hierarchical structure.

In this circuit, four pieces of input information (X) displayed in a matrix of IGxlG as shown in FIG.
were made to memorize and learn patterns. In learning, output patterns are determined for each of the four input patterns, and the synaptic coupling elements 502 and 50 are connected to each other in order to obtain the correct output pattern.
3 AM-, LC display signals and threshold chain element 50
This was done by controlling the voltage of 4.508 to change the threshold value.

As a result, learning converged more than three times faster than when the threshold value was not changed, and it was confirmed that threshold processing could be performed much faster than when using a computer.

Also, as shown in Fig. 7, 1/8 of the pattern is added to this circuit.
When looking at the output results for the hidden input, the recognition rate was approximately 88%.

Furthermore, even if the X-trix size of the AM-LC and the spatial light modulator in this neural network circuit is increased in this inverted position, the same effect can be obtained;
The recognition rate can be further improved by increasing the number of pairs of LC and spatial light modulator to increase the number of layers.

Example 3 In the neural network circuit of Example 1, a spatial light modulation element represented by the equivalent circuit of FIG. 1(b) or (c) was used instead of the AM-LC as the synaptic coupling element 103. However, the photoconductive layer of this spatial light modulator is a-SI.
C:H was used. Further, a-Sl2H was used for the photoconductive layer constituting the spatial light modulator of the threshold chain element 102, and was arranged in a matrix as shown in FIG. 2(b). The input display element 101 uses a light emitting diode array (matrix size 4×4) whose emission wavelength includes wavelength λ as light that is well absorbed by a-SIC:H and wavelength λ2 as light that is well absorbed by a-Sl:H. did.

This neural network circuit uses the output from a light sensor that responds to 16 types of single wavelength light as input information, and -
For example, unsupervised learning was performed by inputting input information of sunlight, fluorescent light, incandescent light bulb, and candle flame into this neural network circuit and manipulating the threshold value. As a result, it was possible to identify 100%, and the degree of convergence of learning was 1.5 to 1.5% higher than when no learning was performed on the threshold value.
The song is twice as fast.

In this way, a neural network circuit with excellent operational ability was obtained even with unsupervised learning.

Effects of the Invention According to the present invention, a neural network circuit is obtained which has a learning function, has a threshold processing function in hardware, has excellent learning convergence and operation ability, and can easily construct a hierarchical structure. It will be done.

[Brief explanation of drawings]

FIG. 1(a) is a schematic diagram of the configuration of a neural network circuit in an embodiment of the present invention, and FIG. 1(b) and (C)
. (d) and (e) are equivalent circuit diagrams of a spatial light modulation element used in a neural network circuit according to an embodiment of the present invention;
a) and (b) are a cross-sectional view and a plan view, respectively, of a spatial light modulation element used in a neural network circuit according to an embodiment of the present invention, and FIG. 2(C) is a neural network according to an embodiment of the present invention. Graphs 3 and 4 showing the transmission characteristics of the color polarizing plate used in the circuit are graphs showing the character-filled cover patterns stored in the neural network circuit of one embodiment of the present invention and their imperfections used for recall, respectively. Figure 5 (a) is a diagram showing the pattern, and Figure 5 (a) is a schematic diagram of the configuration of a neural network circuit according to an embodiment of the present invention.
), (e), and (f) are a transmission characteristic diagram of a color polarizing plate, a top view of a spatial light modulation element, a transmission characteristic diagram of a color filter, and a color polarization diagram, respectively, used in a neural network circuit according to an embodiment of the present invention. The transmission characteristic diagram of the board, the transmission characteristic diagram of the color filter, and Figures 6 and 7 are diagrams showing an example of the pattern stored in the neural network circuit and an incomplete pattern used for recall, respectively. Figure 8 is the diagram showing the neural network circuit. FIG. 2 is a model diagram showing an example of a hierarchical structure of a network. 101... Input pattern display element, 102... Threshold element, 103... Synapse coupling element, 501...
- Input display element, 502, 503... Synaptic coupling element, 504, 508... Threshold element, 505, 50
7...Optical filter. Name of agent: Patent attorney Shigetaka Awano (Figure 1 (OL)). Mezzo station shrimp (α) AI? +t λ3 1 length (input〈λ2〈space 3) No. ■ (C) Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. Fig. 3 (nδ1) Fig. Fig. Fig. 1 (C) Fig. Fig. Fig. 2) 7r, = /113 N

Claims (4)

[Claims] (1) Comprising a liquid crystal layer, a photoconductive part, and a field effect transistor, the photoconductive part is electrically connected to the field effect transistor, and the photoconductive part has the photoconductivity. A neural network circuit comprising at least a spatial light modulation element that modulates the electric field intensity applied to the liquid crystal layer by irradiating a portion with light. (2) The neural network circuit according to claim 1, wherein the neural network circuit includes a plurality of spatial light modulation elements, and the photoconductive portions of each of the spatial light modulation elements are made of different materials. (3) The neural network circuit according to claim 1, wherein the photoconductive portion includes a depletion layer therein. (4) The neural network circuit according to claim 1, characterized in that there is a plurality of photoconductive parts.