JPH05181993A - Nuro network - Google Patents

Abstract

PURPOSE: To change the electrical resistivity of mutually connected resistance elements by adjusting the quantity of light reaching a photoconductive phase from a light source by providing a photoconductive layer forming the resistance elements, the light source, and a liquid crystal cell matrix between the light source and photoconductive layer in a constitution which gives the mutual connection of variable resistors in a neural network. CONSTITUTION: The substrate 15 of a liquid crystal display panel 14 is made of glass or another appropriate transparent material and a polarizer layer 16 is provided below the substrate 15. A set of transparent plates 18 is positioned on the opposite surface of the substrate 15 at an interval of, for example, 15 m and the spaces between the plates 18 and substrate 15 are filled up with liquid crystals 20. A second polarizer layer 21 is provided on the upper surfaces of the plates 18. A neural network 23 and liquid crystal display panel 14 thus constituted can be stuck to each other with an adhesive layer 28 between layers 21 and 27. Light irradiates the area corresponding to the neural network and the resistivity of the area is set in accordance with the intensity of transmitted light, and then, the mutual connection between the row and column conductors 25 and 27 of the network is weighted.

Description

Detailed Description of the Invention

[0001]

FIELD OF THE INVENTION The present invention relates to neural networks, and more particularly to an arrangement for varying interconnection weights between neural network elements.

A typical neural network consists of an array of elements interconnected by conductive links. These elements interact by transmitting electrical signals through the connecting links. These connections are made by a resistor which may be a thin film resistor or the like. The weighting of the link is changed by adjusting the resistance of the link, and this adjustment can be made accordingly.

It is an object of the present invention to provide an arrangement for implementing this resistance adjustment.

That is, the present invention provides a variable resistance interconnect in a neural network, provided in a region of a photoconductive material forming a resistive element, a light source, and between the light source and the photoconductive region. It is characterized by comprising a light adjusting means for adjusting the amount of light reaching the region from the light source.

The light controlling means may be composed of a liquid crystal light valve, or may be composed of a film of photographic emulsion or acetate covered with a photomask.

The preferred photoconductive material is amorphous silicon.

Embodiments of the present invention will now be described with reference to the accompanying drawings for purposes of illustration only.

FIG. 1 is a schematic circuit diagram showing a main part of a single-level artificial neural network.

FIG. 2 is a schematic circuit diagram showing a main part of a multilevel artificial neural network.

FIG. 3 is a schematic plan view showing a structure forming a main part of a neural network.

FIG. 4 is a schematic cross-sectional view showing the first embodiment of the invention.

FIG. 5 is a schematic cross-sectional view showing a second embodiment of the invention.

As shown in FIG. 1, the cross section of the artificial neural network has row lines R1, 1 '; 2, 2'; 3, 3 '; 4,
4 '; ... n, n'and column lines C1, 2, 3, 4,
... consisting of h. The row / column lines have resistance elements G ⁺ ₁₁ , G ⁺ ₁₂ , ... G ⁺ _1h ; G ^- ₁₁ , G ^- ₁₂ having conductance.
_{^{··· G - 1h; G + 21}} , G + 22, ··· G + 2h; G - 21, G -
₂₂ ... G ^- _2h ; G ⁺ ₃₁ , G ⁺ ₃₂ , ... G ⁺ _3h ; G ^- ₃₁ ,
G ^- ₃₂ ... G ^- _3h ; G ⁺ ₄₁ , G ⁺ ₄₂ , ... G ⁺ _4h ;
G ^- ₄₁ , G ^- ₄₂ ... G ^- _4h ; G ⁺ _n1 , G ⁺ _n2 , ...
G ⁺ _nh ; G ^- _n1 , G ^- _n2 ... G ^- _nh .
If necessary, each voltage V ⁺ ₁ , V ^- ₁ ; V ⁺ ₂ , V ^- ₂ ; V ⁺ ₃ ,
From _n was added to the row line, each amplifier _{A 1, A 2, A 3} , A 4 ··· A h row line _{^{- ··· V + n, V;}} V -; - 3 V + 4, V 4 Receive output.

The current input I1 to the amplifier A1 at an arbitrary instant is given by the following equation. I ₁ = V ⁺ ₁ G ⁺ ₁₁ + V ^- ₁ G ^- ₁₁ + V ⁺ ₂ G ⁺ ₂₁ + V ^- ₂ G ^- ₂₁ +
... V ⁺ _n G ⁺ _n1 + V ^- _n G ^- _n1

Similarly, the current inputs I ₂ to I _h to the amplifiers A _{2 to} A _h are obtained by the following equations, respectively. I ₂ = V ⁺ ₁ G ⁺ ₁₂ + V ^- ₁ G ^- ₁₂ + V ⁺ ₂ G ⁺ ₂₂ + V ^- ₂ G ^- ₂₂ +
^{_{··· V + n G + n2 +}} V - n G - n2 I h = V + 1 G + 1h + V - 1 G - 1h + V + 2 G + 2h + V - 2 G - 2h +
... V ⁺ _n G ⁺ _nh + V ^- _n G ^- _nh

Therefore, the outputs of the amplifiers A ₁ -A _h depend on the row input voltage and the conductance value of the connecting link, ie the magnitude of the weight. Each amplifier is an operational amplifier that adds a supply current to its input, and has means for adding a non-linear function such as a sigmoid to its output. Each amplifier may be composed of a thin film transistor.

Further, the row line 0 is connected to a voltage source S which supplies a positive or negative voltage of a predetermined value. Resistive elements G _o1 , G _o2 , G _o3 connected between the row line 0 and the column conductor
..By adjusting the conductance value of G _oh
An offset current is supplied to the amplifiers A _{1 to} A _h .

By connecting the positive and negative row line pairs, the conductance value connected to the positive voltage source is increased and the conductance value connected to the negative voltage source is decreased, or vice versa. Use either positive or negative weight. Despite the very high resistance of the elements in the low conductance state, they supply a small current to each amplifier, which tends to reduce the effective difference between the on and off states of the row conductors. However, this small current is equalized by the corresponding high conductance element, but is offset by the current components of opposite polarity. Furthermore, both links connected between the common column lines with a pair of row lines (e.g., link G ⁺ ₁₁ and G ^- ₁₁₎
Is in the low conductance state, the generated off-state currents supplied to each amplifier will cancel.

FIG. 2 shows a cross section of a multilevel artificial neural network in which the above-mentioned interconnected single-level networks are formed in a stacked state. The input voltage of the second level network L ₂ is the first
Derived from the amplifier of level network L ₁ . Similarly, the third
The input voltage of the level network L ₃ is derived from the amplifier of the second level network L ₂ . The same applies below. In this case, the fourth level perceptron consists of an input network plane, an output network plane and two network planes hidden on both sides.

FIG. 3 illustrates an advantageous method of placing resistive elements between row and column lines. Parallel spaced metal row (or column) conductors 10 are disposed on one side of the amorphous silicon layer 11. On the opposite side of the layer is disposed a set of parallel spaced column (or row) conductors 12 formed from a transparent conductive material such as indium tin oxide (ITO).
The conductor 12 is orthogonal to the conductor 10. Illumination of the conductor 11 and the layer 11 between them reduces the resistivity of the photoconductive amorphous silicon to a level that depends on the intensity of the incident light. If the incident light intensity is constant, the resistivity settles down at this level. When the light source is turned off or the light is turned off, the resistivity of the layer 11 returns to its original level. Layer 11 may be a continuous layer or, if desired, may be patterned and divided into rectangular sections at the intersections of conductors 10,12.

In the case of the configuration 13 of the present invention in which variable resistance connection is made at the row / column conductor intersections, the light incident on the photoconductive layer is controlled by the liquid crystal cell. An outline of this configuration is shown in FIG. The substrate 15 of the liquid crystal display panel 14 is made of glass or another suitable transparent material, and the polarizer layer 1 is provided under the substrate 15.
6 is provided. On the opposite side of the substrate 15 is formed a set of transparent row conductors 17 (eg made of ITO). A transparent (eg glass) plate 18 with transparent column conductors underneath it is provided from the substrate 15 eg 1
The liquid crystal substance 20 is filled in the gap with a gap of 5 μm. The second polarizer layer 2 is formed on the upper surface of the plate 18.
1 is provided. Light from the backlight 22 passes through the cells of the liquid crystal display panel in the clear (transmissive) state, but is blocked by the cells in the dark (non-transmissive) state. This is determined by the drive signal applied to the row conductors 17 and the column conductors 19. Although the display panel 14 is a direct multiplex panel, an active matrix display composed of thin film drive transistors formed on the substrate 15 can be used instead. The column conductor may also be replaced by a continuous reference electrode.

When constructing the neural network 23, a row conductor 25, which may be made of, for example, titanium, gold, aluminum or the like, is provided on the base body 24. An amorphous silicon layer 26 is formed on the conductor 25, and a transparent (eg, ITO) layer is formed on the silicon layer 26.
Column conductor 27 is deposited.

The circuit network 23 and the liquid crystal display panel 14 having the above-mentioned structure are provided with the adhesive layer 2 between the layers 21 and 27.
It can be joined by 8. Alternatively, they may be integrated and held by an air layer provided between the layers 21 and 27 or another material layer having suitable optical characteristics. Spacer members (not shown) may be incorporated into layer 28. What is important in each case is the extremely precise alignment of the liquid crystal cell and the neural network cell.

The liquid crystal substance 20 may be, for example, a twisted nematic substance, a super twisted type or a ferroelectric type substance,
In both cases two polarizer layers 15 and 21 are required. Alternatively, guest / host materials may be used,
In this case, only one polarizer bistability is required. Also, in the case of dye phase change materials, no polarizer layer is required.

In another embodiment shown in FIG. 5, a substrate 24 and
An amorphous silicon layer 27 is interposed between the conductors 25 and 27 to form a neural network 23 '. A relatively thick dielectric layer 29 of, for example, polyamide is deposited on the conductor 27, and the column electrodes 19 of the liquid crystal panel are formed on this layer 29 so that the polarizer layer 21 and the substrate 18 are not used. Instead of the dielectric layer 29, a multilayer dielectric may be used. The liquid crystal material may be a guest / host material or a dye phase change material. In the latter case, the polarizer layer 16 can be omitted.

When the above embodiment is used, the individual cells of the liquid crystal display panel are arranged by appropriately addressing the row and column conductors so that the visible light from the backlight 22 is variably transmitted. To do it. Light hits the corresponding regions of the neural network such as regions 29, 30, 31 and the resistivity of these regions is set according to the intensity of the transmitted light, whereby the row / column conductors 25, 27 of the neural network are set. Set interconnection weights for.

The cells shown are very few, but in reality,
The number of cells in the neural network and the liquid crystal display panel is generally several thousand. For example, a 125 mm ² panel is a 1000 × 1000 cell matrix.

The above multilevel structure is shown in FIG. 4 or FIG.
Can be formed by interconnecting a plurality of the above configurations.

Instead of the above liquid crystal cell, a photographic emulsion or acetate film having a photomask provided by printing or other means may be used as a means for adjusting the amount of light hitting the photoconductive layer.

[Brief description of drawings]

FIG. 1 is a schematic circuit diagram showing a main part of a single-level artificial neural network.

FIG. 2 is a schematic circuit diagram showing a main part of a multilevel artificial neural network.

FIG. 3 is a schematic plan view showing a structure forming a main part of a neural network.

FIG. 4 is a schematic cross-sectional view showing the first embodiment of the invention.

FIG. 5 is a schematic cross-sectional view showing a second embodiment of the invention.

[Explanation of symbols]

 L1, L2, L3 network 10, 12 conductor 14 liquid crystal display panel 15, 18, 24 substrate 16, 21 polarizer layer 17, 19 conductor 20 liquid crystal substance 22 backlight 26 amorphous silicon layer

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] June 19, 1992

[Procedure Amendment 1]

[Document name to be corrected] Drawing

[Correction target item name] All drawings

[Correction method] Change

[Correction content]

[Fig. 2]

[Figure 3]

[Figure 1]

[Figure 4]

[Figure 5]

 ─────────────────────────────────────────────────── ─── Continuation of the front page (71) Applicant 592118217 William Ireland Milne UK CB 50 Beeby, Cambridge, Burwell, Knows Street 133 (72) Inventor Pierro Miglioreit United Kingdom E 18, London, South Wood Ford , Peel Road 67 (72) Inventor Simon Christopher John Garth UK Seabee 10 1 Es W, Essex, Saffron Walden, Aicleton, Mill Lane 25 (72) Inventor William Ireland Milne UK Seabee 50 Beeby, Cambridge, Burwell, Knows Street 133

Claims (9)

[Claims] 1. A structure for providing variable resistance interconnections in a neural network, the region of photoconductive material forming a resistive element, a light source, and being provided between the light source and the photoconductive region. A neural network comprising a light adjusting means for adjusting the amount of light reaching the region. 2. The neural network according to claim 1, wherein the light control means comprises a liquid crystal light valve. 3. The neural network according to claim 1, wherein the light control means is a photographic emulsion film provided with a photomask. 4. The neural network according to claim 1, wherein the light control means is an acetate film provided with a photomask. 5. The neural network according to claim 1, wherein the photoconductive substance is amorphous silicon. 6. A liquid crystal comprising said photoconductive layer forming a plurality of said resistive elements between the intersections of row and column conductors of a neural network, and said light modulating means being substantially aligned at said intersections. Neural network according to any one of claims 1 to 5, consisting of a matrix of cells. 7. The neural network according to claim 6, wherein the matrix of liquid crystal cells is an active matrix in which the cells are addressed via transistors. 8. The neural network of claim 6 or 7, comprising a relatively thick layer of transparent dielectric material that also acts as one of the two layers holding the liquid crystal material therebetween. 9. The neural network of claim 8, wherein the dielectric material is polyimide.