JPH0357555B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an associative memory device, and more particularly to a memorization section of an associative memory device that memorizes input data.

[Conventional technology]

Generally, in ordinary memories used in electronic computers, a method is adopted in which information stored therein is accessed by specifying an address.

[Problems to be Solved by the Invention] In such conventional storage devices, there is a problem in that if the contents of the address are destroyed, the contents of the stored data will become completely unrecognizable.
Also, regardless of the stored contents, it can only be accessed by address.

The present invention is intended to solve the above problems,
To provide a storage unit of an associative memory device which does not become impossible to read even if a part of the stored content is destroyed and allows access related to the stored content.

[Means for solving problems]

For this purpose, the storage unit of the associative memory device of the present invention is
A multiplexing system that multiplexes input or feedback signals and outputs the result, an expansion system that expands and outputs the input or feedback signal, and an output system that multiplies the output of the multiplexing system and the output of the expansion system. 1 arithmetic unit, a memory that stores two-dimensional information, performs addition and subtraction on the contents as necessary, and outputs the contents, and a second arithmetic unit that multiplies the output read from the memory and the output of the multiplexing system. an arithmetic unit, a demultiplexing system that demultiplexes and outputs the output from the second arithmetic unit, and a threshold memory that stores the output from the demultiplexing system while performing a threshold value operation and can be read out as necessary. and a feedback system that returns the output results to the input.

[Effect]

The storage unit of the content addressable memory device of the present invention multiplexes and expands the input or feedback signal, multiplies it in the first arithmetic unit, writes the multiplication result to the memory, and outputs the output from the memory and the multiplexing system. It is possible to memorize by demultiplexing the multiplication result obtained by multiplying the output of becomes.

〔Example〕

Examples will be described below based on the drawings.

First, the principle of the associative memory device according to the present invention will be explained.

In the present invention, an autocorrelation matrix is used as a method for realizing associative memory, and a memory matrix is formed by the autocorrelation of the contents to be stored.

The arithmetic processing at the time of memorization is expressed as follows.

M=Σχ·χ′ (1) Here, χ is an input vector representing the content to be stored, χ′ is a transposed vector of χ, and M is a storage matrix. That is, the autocorrelation of the content to be memorized is taken, and this is added many times to form the memory.

At the time of recollection, by operation with this memory matrix,
You can recall the whole from a part. The calculation process during recall can be expressed as the following formula.

y=φ(M・χ) ...(2) Here, y is the output vector, χ is the input data,
φ represents threshold operation. By operating M・χ, χ
Even if the data is incomplete due to some missing parts, as long as the memory matrix M is created by the calculation process of equation (1), the recalled data y will be the original data χ with the missing parts filled in. Data close to that can be obtained. Note that the noise portion is removed by detecting data above a predetermined level by operating the threshold value of φ.

Next, as a method when recollection using the memory matrix in equation (1) above does not provide sufficient separation, a sequential calculation method for forming a memory matrix with a higher degree of separation will be described.

M _o+1 = M _o + α [χ − φ {Σ (M _o + χ)}] χ′
...(3) Here, α is the learning gain, and Σ is an operator for matching the order with χ′ by taking the partial sum of M _o ·χ. n+1th memory matrix M _o+1
is the result recalled by M _o for the nth time φ
It is obtained by correcting M _o by multiplying the correlation between the error component at the time of recall, which is represented by the difference between {Σ(M _o · χ)} and χ, and the learning gain α. Note that the learning gain α is selected to a value such that M _o converges.

In this way, by performing the calculation process of equation (3) until M _o converges, a correlation matrix M with improved separability can be obtained,
A memory matrix is formed by the autocorrelation of the content you want to remember, and when making associations, you can recall the whole from a part by taking the correlation with this memory matrix.

FIG. 6 is a diagram for explaining the configuration and operation of a spatial light modulation tube, which is a basic component in an embodiment of the associative memory device according to the present invention, in which 1 indicates an input image;
2 is a lens, 3 is a photocathode, 4 is a microchannel plate, 5 is a mesh electrode, 6 is a crystal, 6
1 is a charge storage surface, 7 is a half mirror, 8 is monochromatic light, 9 is an analyzer, and 10 is an output image.

In the figure, an input image 1 incident on a photocathode 3 of a spatial light modulation tube via a lens 2 is converted into a photoelectron image. After this photoelectron image is multiplied by the microchannel plate 4, the charge storage surface 61 of the crystal 6
form a charge pattern. The electric field across the crystal 6 changes depending on the charge pattern, and the refractive index of the crystal 6 changes due to the Pockels effect.

Here, when the crystal 6 is irradiated with linearly polarized monochromatic light 8, the polarization state of the reflected light from the charge accumulation surface 61 has changed due to the birefringence of the crystal 6, so if it is passed through the analyzer 9, An output image 10 having a light intensity corresponding to the light intensity of the input image 1 is obtained.

Next, the main functions of such a spatial light modulation tube related to the present invention will be explained.

(a) Memory function The spatial light modulation tube has a memory function that retains the charge distribution on the surface of the electro-optic crystal for a long time. Since crystal 6 has a very high electrical resistance value,
The charge distribution on the crystal surface 61 can be maintained for several days or more.

(b) Subtraction function The spatial light modulation tube can selectively form a positive or negative charge distribution on the surface of the electro-optic crystal.
FIG. 7A is a graph showing secondary electron emission characteristics of an electro-optic crystal.

As shown in FIG. 7A, if the primary electron energy E incident on the charge storage surface 61 is smaller than the first crossover point E1 or larger than the second crossover point E2, the number of primary electrons is Since it is larger than the number of secondary electrons emitted at the crystal surface (δ
<1), the crystal surface is negatively charged. When the energy of the primary electron is between E1 and E2, the number of secondary electrons is 1.
Since the number of secondary electrons is greater than the number of secondary electrons (δ>1), the crystal surface becomes positively charged.

When accumulating charge in a crystal, whether to write with positive charge or negative charge depends on Vc and Vc shown in Figure 6.
This is done by controlling the voltage of Vb. Here, the subtraction function can be provided by two methods: first writing a negative voltage and then charging positively, or first charging positively and then writing negative charges. The amount of subtraction can be controlled in three ways:

That is, a method of changing the intensity of incident light during subtraction;
These methods include changing the duration of the voltage applied to the microchannel plate 4, and changing the voltage applied to the microchannel plate 4.

Although writing and erasing methods are well known, they will be explained using a positive charge image as an example with reference to FIGS. 7A and 7B. Note that the voltage Vc applied to the secondary electron collecting electrode is set at the second crossover point _E2 in FIG. 7A.

[Erase operation]

When the crystal back voltage Vb is set to a potential corresponding to the second crossover point _E2 , the crystal surface potential Vs is
The value is the potential E ₂ plus the potential increase of positive charges caused by writing. In the case of this surface potential, the energy of the incident primary electron is _greater than E2, so the secondary electron emission ratio δ<1, and the surface potential is
Negative charge is accumulated until E ₂ is reached. potential is E ₂
When δ=1, an equilibrium state is reached, and the surface charge becomes zero.

[Write operation]

The crystal back surface voltage Vb is set to a voltage E' that provides a sufficient dynamic range between the first crossover point _E1 and the second crossover point _E2 . At this time, the crystal surface potential Vs also becomes approximately E', so the energy of the incident primary electron is between E ₁ and E ₂ . Therefore, since the secondary electron emission ratio δ>1, a positive charge image is formed. The emitted secondary electrons are collected by a secondary electron collection electrode at a potential Vc higher than Vs.

(C) Real-time threshold operation function The spatial light modulation tube can perform real-time threshold operation according to the setting conditions of Vc and Vb shown in FIG. The mesh electrode 5 is provided near the crystal surface, and when it is set to a predetermined potential,
When sufficient electrons are supplied to the crystal surface, the crystal surface potential becomes the potential of the mesh electrode, and this potential becomes the crossover point. In other words, if the crystal surface potential is lower than the mesh electrode, the secondary electrons emitted from the crystal surface will be collected by the mesh electrode, so the secondary electron emission will increase relative to the incident electrons, and the crystal surface potential will rise. Conversely, if the crystal surface potential is higher than the mesh electrode potential, the number of incident electrons will be greater than the secondary electrons, causing the crystal surface potential to drop, and eventually become constant when the crystal surface potential becomes equal to the mesh electrode potential. Now mesh electrode 5
When the voltage Vc is kept low at about 0.1KV and Vb is lowered stepwise below 0.1KV, the potential of the crystal surface also decreases stepwise, and the charge accumulation surface 61 of the crystal 6
becomes a negative potential, and electrons no longer reach it, resulting in a so-called lockout state. However, when Vb is slowly lowered in a ramp-like manner, electrons are supplied to the crystal surface in areas where the intensity of the incident light is high and a large amount of electrons are supplied, and secondary electron emission increases, causing negative potential. However, in areas where the intensity of the incident light is low and the amount of supplied electrons is small, the supply of electrons cannot keep up with the potential drop, resulting in a negative potential and no electrons reaching the crystal surface. Therefore, in response to the intensity of light incident on the photocathode 3, there are areas where the crystal surface becomes negative potential and writing is not performed, and areas where electrons reach the crystal surface and writing occurs without the surface becoming negative potential. The part where is done and can be done,
As a result, threshold operation is performed depending on the intensity of the incident light.

(d) AND operation function based on real-time threshold operation As described above, the spatial light modulation tube has a real-time threshold operation function. Using this, do the following
It is possible to perform an AND operation. That is, the eighth
As shown in the figure, when the two inputs are superimposed,
The input light intensity has three levels: 0, 1, and 2. If a real-time threshold operation is performed at level 2 of these, the result will be as shown in FIG. 9, which means that an AND operation has been executed.

Next, other components necessary for the present invention will be explained. Note that the input image is assumed to be a 2×2 pattern.

Enlargement system As shown in Figure 10, a 2x2 input image is converted into a 4x4
to be enlarged and projected. In reality, this is done using a lens or the like.

Multiplexing System A 2×2 input image is repeatedly projected as shown in FIG. 11A to obtain a pattern as shown in FIG. 11B. Actually, this is realized by a lens array or the like.

Demultiplexing system As shown in Figure 12, a 4x4 input image is converted into a 2x2
Project it so that it is superimposed on the image. Actually, this is realized by a lens array or the like.

Next, a coding method that performs coding suitable for two-dimensional data using the aforementioned enlarged image, multiplexed image, and demultiplexed image will be explained using a 2×2 image.

The method of memorization and recollection that was separately filed by the present applicant will be explained.

The vector χ and its transposed vector χ′ are expressed as follows.

χ＝x ₁ x ₂ x ₃ x ₄ , χ′=(x ₁ , x ₂ , x ₃ , x ₄ ) In this case, the memorization is M=χ・χ′=x ₁ x ₂ x ₃ x ₄・( x ₁ , x ₂ , x ₃ , x ₄ ) = x ₁ x ₁ x ₁ x ₂ x ₁ x ₃ x 1 x ₄ x ₂ x ₁ x ₂ x ₂ x ₂ x ₃ x ₂ x ₄ x ₃ x _{1 x 3} x ₂ x ₃ x ₃ x 3 x _{4 x 4} x ₁ x ₄ x ₂ x ₄ x ₃ x ₄ x ₄ , while recollection is expressed as z′=M・y=x ₁ x ₁ x ₁ x ₂ x ₁ x ₃ x ₁ x ₄ x ₂ x ₁ x ₂ x ₂ x ₂ x _{3 x 2} x ₄ x ₃ x ₁ x ₃ x ₂ x ₃ x ₃ x ₃ x ₄ x ₄ x ₁ x ₄ x _{2 x 4} x ₃ x ₄ x ₄ y ₁ y ₂ y ₃ y ₄ = x ₁ x ₁ y ₁ x 1 x ₂ y ₂ x 1 x _{3 y 3 x 1} x ₄ y _{4 x 2 x 1} y ₁ x ₂ x ₂ y ₂ x ₂ x ₃ y ₃ x ₂ x ₄ y ₄ x ₃ x ₁ y ₁ x ₃ x ₂ y ₂ x 3 x ₃ y ₃ x ₃ x ₄ y _{4 x 4 x 1} y ₁ x ₄ x ₂ y ₂ x It is expressed as ₄ x ₃ y ₃ x ₄ x ₄ y ₄ .

Here, if we make z' one-dimensional and call it z, then z=x ₁ x ₁ y ₁ +x ₁ x ₂ y ₂ +x ₁ x ₃ y ₃ +x ₁ x ₄ y ₄ x ₂ x ₁ y ₁ +x ₂ x ₂ y ₂ +x ₂ x ₃ y ₃ +x ₂ x ₄ y ₄ x ₃ x ₁ y ₁ +x ₃ x ₂ y ₂ +x ₃ x ₃ y ₃ +x ₃ x ₄ y ₄ x ₄ x ₁ y ₁ +x ₄ x ₂ y ₂ +x ₄ x ₃ y ₃ + x ₄ x ₄ y ₄ .

However, in order to perform image processing with this method, it is necessary to convert the two-dimensional data of the image into one-dimensional data.

In contrast, in the present invention, memorization and recall are performed by the following calculations.

Now let χ be the content you want to remember, and χ = x ₁ x ₂ x ₃ x ₄ . Here, due to the magnification system, x mag=x ₁ x ₁ x ₂ x ₂ x 1 x ₁ x ₂ x ₂ x _{3 x 3 x 4} x ₄ x ₃ x ₃ x ₄ x ₄ Due to the multiplex system, x mlt We get =x ₁ x ₂ x ₁ x ₂ x ₃ x ₄ x ₃ x ₄ x ₁ x ₂ x ₁ x ₂ x ₃ x ₄ x ₃ x ₄ . At this time, if the multiplication of the elements of x mag and x mlt, x mag・x mlt, is M, then M=x mag・x mlt=x ₁ x ₁ x ₁ x ₂ x ₂ x ₁ x
₂ x ₂ x ₁ x ₃ x ₁ x ₄ x ₂ x ₃ x ₂ x ₄ x ₃ x ₁ x ₃ x ₂ x ₄ x _{1 x 4 x 2} x ₃ x ₃ x 3 x ₄ x ₄ x ₃ x _{4 x 4} (... multiplication between elements is called Hadamard product), and the same result as the one obtained by the matrix operation described above is obtained (however, the array is different),
It can be seen that it is possible to memorize by x mag x mlt. Also, let y mlt=y ₁ y ₂ y ₁ y ₂ y ₂ y ₄ y ₃ y ₄ y ₁ y ₂ y ₁ y ₂ y ₃ y ₄ y ₃ y ₄ , and multiply the elements of M and y mlt by M・y
If mlt is z', z'=M・y mlt = x ₁ x ₁ x ₁ x ₂ x ₂ x 1 x ₂ x ₂ x ₁ x ₃ x ₁ x ₄ x ₂ x ₃ x ₂ x ₄ x _{3 x 1} x ₃ x ₂ x ₄ x ₁ x ₄ x ₂ x ₃ x _{3 x 3 x 4} x ₄ x ₃ x ₄ x ₄ y ₁ y ₂ y ₁ y ₂ y ₃ y ₄ y ₃ y ₄ y ₁ y ₂ y ₁ y ₂ y ₃ y ₄ y ₃ y ₄ =x ₁ x ₁ y ₁ x 1 x ₂ y ₂ x ₂ x ₁ y ₁ x ₂ x ₂ y ₂ x ₁ x ₃ y _{3 x 1 x 4} y ₄ x ₂ x ₃ y ₃ x ₂ x ₄ y ₄ x ₃ x ₁ y ₁ x ₃ x ₂ y ₂ x ₄ x ₁ y ₁ x ₄ x 2 y ₂ x ₃ x ₃ y ₃ x ₃ x ₄ y ₄ x _{4 x 3} y ₃ x ₄ x ₄ y ₄ . Here, let z be the demultiplexing of z', then z=x ₁ x ₁ y ₁ +x ₁ x ₂ y ₂ +x ₁ x ₃ y ₃ +x ₁ x ₄ y ₄ x ₂ x ₁ y ₁
+x ₂ x ₂ y ₂ +x ₂ x ₃ y ₃ +x ₂ x ₄ y ₄ z=x ₁ x ₁ y ₁ +x ₁ x ₂ y ₂ +x ₁ x ₃ y ₃ +x ₁ x ₄ y ₄ x ₂ x ₁ y ₁
+x ₂ x ₂ y ₂ +x ₂ x ₃ y ₃ +x ₂ x ₄ y ₄ x ₃ x ₁ y ₁ +x ₃ x ₂ y ₂ +x ₃ x ₃ y ₃ +x ₃ x ₄ y ₄ x ₄ x ₁ y ₁ +x ₄ x ₂ y ₂ + x ₄ x
₃ y ₃ + x ₄ x ₄ y ₄ , which is the same result as the above matrix operation (however, the array is different), and M・y
It can be seen that a recalled pattern can be obtained by demultiplexing the mlt image.

In this way, memorization can be performed using the multiplexing system and expansion system, and recall can be performed using the demultiplexing system, so it is possible to realize associative memory entirely through two-dimensional processing. Become.

Next, the associative memory according to the present invention will be specifically explained. Note that the calculation is performed using the following equation, which is similar to equation (3) above.

M _o+1 = M _o +α [x mag − [φ{Σ(M _o・x mlt)}] mag] x mlt
...(4) Here, α is a learning gain, x mag is an enlarged image of x, x mlt is a multiplexed image of x, and Σ is a demultiplexed image. The n+1st memory matrix M _o+1 is the result of recollection by the nth M _o [φ{Σ(M _o ×
It can be obtained by correcting M _o by the product of the error component at the time of recall indicated by the difference between x mlt)} mag and x mag and x mlt multiplied by the learning gain α. As described above, the learning gain α is selected to a value such that M _o converges, and by performing this calculation process until M _o converges, M _o with an improved degree of separation can be obtained.

FIG. 1 is a block diagram showing the configuration of an associative memory device according to the present invention, in which 101 is an enlargement system;
2 is a multiplexing system, 103 is a first arithmetic unit, 104 is a memory, 105 is a second arithmetic unit, 106 is a demultiplexing system, and 107 is a threshold memory.

In the figure, as mentioned above, the enlarging system 101 converts, for example, a 4×4 input x into a 16×16 x mag
The multiplexing system 102 similarly creates 16×16 x mlt from the 4×4 input x. The first arithmetic unit 103 multiplies the elements of x mag and x mlt. The first memory 104 has x mag x
This is the memory for adding mlt and subtracting feedback values. The second arithmetic unit 105 multiplies the contents of the memory 104 by xmlt to create recall data. The demultiplexing system 106 converts 16×16 data into 4×
This is for converting into data of 4. The threshold value memory 107 is a memory that performs a threshold value operation, detects data above a predetermined level, and cuts out noise. The output y of this threshold memory is fed back to the input side through a feedback system.

Next, the operation during memorization will be explained according to equation (4). The flow of memorization is shown by thin lines in the figure.

First, an enlargement system 101 creates x mag from two-dimensional input data x to be memorized, and at the same time, a multiplex system 102 creates x mlt.

In the second arithmetic unit 105, the
M _o and x mlt from multiplexing system 102,
M _o ·x mlt is calculated, passed through the demultiplexing system 106, subjected to threshold value operation in the threshold value memory 107, and then stored in memory. Here, it is assumed that M _o is initially set to, for example, "0". Of course, “0”
It may be set to a value other than .

Using the created x mag and x mlt as input data, the first arithmetic unit 103 multiplies the elements x
Execute mag・x mlt, multiply by α and set memory to 10
Add to M _o stored in 4.

The result of
(M _o x)}.

The first arithmetic unit 103 calculates φ{Σ(M _o x)} x
mlt is calculated, multiplied by α, and subtracted from M _o in the memory 104.

x mag x mlt is multiplied by α and added to M _o , and then φ{Σ(M _o x)} x mlt
The result of multiplying α by α and subtracting _{it from M o becomes} M _o+1 .

The input data is memorized by performing the above operations until M _o converges. expect
Once M _o is obtained, processing is performed based on that value.

Next, the time of recall will be explained.

In this case, it is assumed that the memory matrix M has already been calculated and memorized. The flow during recall is shown by the thick line in the figure.

First, xmlt is created by the multiplexing system 102 from two-dimensional input data x to be recalled.

In the second arithmetic unit 105, the memory 104
Multiplication with M read from , M.

The output is read from the threshold memory 107, and if one recollection is insufficient, the output is returned to the input by the feedback system, recollection is performed, and desired data is read.

In this way, memorization and recall can be performed.

Next, an embodiment in which the associative memory device shown in FIG. 1 is constructed using a spatial light modulation tube will be described.

FIG. 2 is a diagram showing an example of an associative memory device using a complete optical system using four spatial light modulation tubes.
In the figure, 201 to 204 are spatial light modulation tubes, 211 is an enlarged imaging system, 212 is a multiplexing system, 213 is an inverse multiplexing system, 221 to 228 are half mirrors, 231
-238 are reflecting mirrors, and S1, S2, and S3 are shutters.

In the figure, the arithmetic units 103 and 105 in FIG. There is.

The operating method will be explained below in order. Note that the calculation is performed based on equation (4) as in FIG. 1.

First, input U is sent to magnifying imaging systems 201 and 2 through a half mirror 228 and a reflecting mirror 238, respectively.
02, create U mag and U mlt, and write them to the spatial light modulation tubes 201 and 202, respectively.

Open shutter Sl and half mirror 221, 2.
22, reflecting mirror 231, half mirror 223, 2
Spatial light modulation tube 202 by monochromatic light via 24
→203 and calculates M _o ·U mlt.The result is passed through a half mirror 225 and an inverse multiplex imaging system 213, and is then processed by real-time threshold operation.
Write to the spatial light modulation tube 204 while operating the threshold value.

Open the shutter S2 and similarly read out monochromatic light from the spatial light modulation tube 201→202 via the reflecting mirror 232, half mirrors 223 and 224 to calculate U mag and U mlt. 234,235
The signal is added to the spatial light modulation tube 203 via.

After erasing the spatial light modulation tube 201, shutter S3
The spatial light modulation tube 204 is read out using monochromatic light through the half mirror 221, the reflecting mirror 236, and the half mirror 226, and the half mirror 22
7. An enlarged image is formed by the enlarged imaging system 211 via the reflecting mirrors 237 and 238, and (M _o ·U mlt)mag is written into the spatial light modulation tube 201.

Open shutter S2 and spatial light modulation tube 201→
Reads as 202, (M _o・U mlt)mag・U
mlt is calculated and subtracted from the spatial light modulation tube 203. At this time, the result will not be smaller than 0 due to the effect of Mesh.

Memorization is completed by performing the above operations until convergence for each input pattern.

Next, the recollection operation will be explained in order.

First, Umlt is written from the input U to the spatial light modulation tube 212 via the half mirror 228 and the multiplex imaging system 212.

Open the shutter Sl and open the spatial light modulation tube 202→2.
03, calculates M _o ·U mlt, and writes the result to the spatial light modulation tube 204 through a demultiplexing imaging system 213 by real-time threshold operation.

The spatial light modulation tube 204 is read out, and the output is obtained from the half mirror 227. It is also possible to recall this output again by opening the shutter S3 and returning it through the feedback system.

FIG. 3 is a diagram showing an embodiment of a hybrid content addressable memory device in which a part of the content addressable memory device using the complete optical system of FIG. 2 is replaced with an electronic circuit. In the figure, 301 to 303 are spatial light modulation tubes, 311 is an enlarged imaging system, 312 is a multiplex imaging system, 313 is an inverse multiplex imaging system, 321 to 324 are half mirrors, 331
~335 is a reflecting mirror, 341 is a light receiving matrix,
351 is a CPU, 361 and 362 are pattern presentation devices, and S1 and S2 are shutters.

In this embodiment, interface with electronic circuits is facilitated by replacing the input section and feedback section with a CPU. Additionally, the number of spatial light modulation tubes used is reduced to three, resulting in a system that can be implemented more easily. The operating procedure is exactly the same as that shown in FIG.

As described above, in the systems in the two embodiments, the spatial light modulation tubes 201 and 202, 301 and 30
2 is used for multiplication, but in reality it is {0,
1}, so the logical operation AND
It is the same as the operation. Therefore, by using AND operation using real-time threshold operation for this part,
It is possible to reduce the number of spatial light modulation tubes by one.

FIG. 4 is a diagram showing an embodiment of an associative memory device using a complete optical system, which is realized by reducing the number of spatial light modulation tubes by one and using three by introducing a real-time threshold operation AND operation. In the figure, 401-40
2 is a spatial light modulation tube, 411 is an enlarged imaging system, 412
is a multiplex imaging system, 413 is an inverse multiplex imaging system, 421~
430 is a half mirror, 431 to 437 are reflecting mirrors, and S1, S2, S3, and S4 are shutters.

In the figure, a spatial light modulation tube 401 performs an AND operation, a spatial light modulation tube 402 functions as a memory, and a spatial light modulation tube 403 functions as a threshold memory.

First, memorization will be explained.

Open the shutter S1 and input U, and multiple images are formed by the multiplex imaging system 412, and the spatial light modulation tube 401
Write U mlt to .

Read out as spatial light modulation tube 401 → 402
M _o ·U mlt is calculated and written to the spatial light modulation tube 403 through a demultiplexing imaging system 413 by real-time threshold operation while performing threshold operation.

After erasing the spatial light modulation tube 401, the shutter S
1. Open S2, create multiple imaging and enlarged imaging, and use real-time threshold operation to calculate U mag and U
Executes mlt AND operation.

Read out the spatial light modulation tube 401, and
Add U mlt to spatial light modulation tube 402 .

After erasing the spatial light modulation tube 401, the shutter S
1. Open S4 and perform real-time threshold operation.
(M _o・U mlt) Execute the AND operation of mag and U mlt.

The spatial light modulation tube 401 is read out and subtracted from the spatial light modulation tube 402. At this time, the result is Mesh
does not become smaller than 0 due to the effect of

Memorization is completed by performing the above operations until convergence for each input pattern.

Next, recall will be explained.

Open shutter S1 and multiplex image U of input U.
mlt is written to the spatial light modulation tube 401.

Spatial light modulation tube 401 → 402 and readout,
M _o ·U mlt is calculated and the result is written to the spatial light modulation tube 403 through a demultiplexing imaging system 413 by real-time threshold operation.

The spatial light modulation tube 403 is read out and an output is obtained from the half mirror 425. It is also possible to re-recall this output by opening the shutters S1 and S3 and returning it through the feedback system.

FIG. 5 is a diagram showing an embodiment of a hybrid content addressable memory device in which a part of the content addressable memory device using the complete optical system of FIG. 4 is replaced with an electronic circuit. In the figure, 501 and 502 are spatial light modulation tubes, 511 is an enlarged imaging system, 512 is a multiplex imaging system, 513 is an inverse multiplex imaging system, 521 to 523 are half mirrors, 531
~534 is a reflecting mirror, 541 is a light receiving matrix,
551 is a CPU, and 561 and 562 are pattern presentation devices.

In this embodiment, the input part and the feedback part of the associative memory system shown in FIG. This is a hybrid associative memory system with one fewer modulation tube.

The operating procedure in that case is exactly the same as in the case of FIG.

In each of the above embodiments, the basic configuration of the present invention shown in FIG. Of course, it may be constructed using an electric circuit system.

〔Effect of the invention〕

As described above, according to the present invention, a part of the memory contents is stored by an associative memory device that finds information that is the same as reference input from the outside or that meets certain conditions and retrieves the remaining related parts. Even if the memory is damaged, reading is not completely impossible, and access related to the stored contents becomes possible, which is useful for information retrieval.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an associative memory device according to the present invention, and FIGS. 2, 3, 4, and 5 are examples of optical systems of the associative memory device of the present invention using spatial light modulation tubes. FIG. 6 is a diagram for explaining the configuration of a spatial light modulation tube, which is a basic component in an embodiment of the associative memory device according to the present invention, and FIG.
The figure is for explaining the secondary electron emission characteristics of the crystal surface and writing and erasing during spatial light modulation. Figures 8 and 9 are for explaining the AND operation using the real-time hard clip function of the spatial modulation tube. Figure 1
0 is a diagram for explaining an enlargement system, FIG. 11 is a diagram for explaining multiplexing, and FIG. 12 is a diagram for explaining a demultiplexing system. 1... Image, 2... Lens, 3... Photocathode, 4
...Microchannel plate, 5...Mesh electrode, 6...Crystal, 61...Charge storage surface, 7...
...half mirror, 8...monochromatic light, 9...analyzer,
10...Output light, 101...Enlargement system, 102...
...Multiplexing system, 103... First arithmetic unit, 104...
...Memory, 105...Second arithmetic unit, 106...
Demultiplexing system, 107... Threshold memory, 201-2
04...Spatial light modulation tube, 211...Enlarged imaging system,
212...Multiple imaging system, 213...Inverse multiple imaging system, 221-228...Half mirror, 231-
238...Reflector, S1, S2, S3...Shutter.

Claims (1)

[Claims] 1. A multiplexing system that multiplexes and outputs input or feedback signals, and an expansion system that magnifies and outputs input or feedback signals,
A first arithmetic unit that multiplies the output of the multiplexing system and the output of the expansion system and outputs the result, and a memory that stores two-dimensional information, performs addition and subtraction on the content as necessary, and outputs the content. , a second arithmetic unit that multiplies the readout output from the memory and the output of the multiplexing system;
A demultiplexing system that demultiplexes and outputs the output from the second arithmetic unit, a threshold memory that stores the output from the demultiplexing system while performing a threshold value operation, and can read out the output result when necessary. A memorization unit of an associative memory device equipped with a feedback system for returning input.