JPS62133589A - Pattern storing and discriminating device - Google Patents

Abstract

PURPOSE:To improve the separating function of a pattern by reinforcing the coupling between each sequential information processing layer, in accordance with input information, and a result of processing by an information processing layer provided on the way. CONSTITUTION:An information processing layer which has placed two- dimensionally plural pieces each of exciting property elements U and suppressing property elements V by mixing them is arranged successively, and between each of the information processing layers, a variable exciting property coupling (a) and a variable suppressing property coupling (b) are provided in the forward direction. Each information processing layer, for instance, each U element U2 in a cell surface of the second stage receives the coupling (a) from one group of U elements of U1 of its previous stage, but a V element V1 of the first stage also receives a fixed suppressing property coupling (c) from the same U element, and its output is provided to the element U2 by the coupling (b). Also, the coupling (a) and the (b) are reinforced in accordance with a rule for reinforcing the coupling.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention is capable of storing various shape patterns to be identified and identifying new human cover patterns with the memorized patterns. The present invention relates to a pattern storage and identification device applied to storage and identification.

[Conventional technology]

Conventionally, devices of this type require sophisticated pattern recognition functions. As for the sophistication of this pattern recognition function, Japanese Patent Application Laid-Open No. 51-352474 “
Examples include ``Self-Organization Formation Method'' and ``Self-Organization Memory Device'' published in Japanese Unexamined Patent Publication No. 58-221474. Japanese Patent Application Publication No. 1973-
In Publication No. 352474, the final output is obtained only by processing the output from each information processing layer in the forward direction! :4 has been completed. Furthermore, the information signals to be processed in each information processing layer are constructed in a sequential manner from the human-powered layer to the deeper layers.

In addition, in Japanese Patent Application Laid-Open No. 58-221474, by providing a feedback-type suppressive coupling, [11] a good pattern separation function is achieved even for similar patterns that have a lot of overlap with each other. It is possible to hold it by +iJ function.

[The problem that the invention attempts to solve]

However, this type of conventional example has the following problems. That is, in Japanese Patent Application Laid-Open No. 51-352474, a pattern 1. (If there were many similar patterns, there was a problem that both the new and old patterns would be considered the same.
In Japanese Patent No. 74, since it is necessary to provide a feedback suppressive coupling, there is a problem that the circuit becomes complex.

The present invention has been made in consideration of the above circumstances, and the (-
The first objective is to provide a l<turn memory identification device which can measure the direction of the pattern separation function, and which can simplify the configuration without the need to provide a feedback-type suppressive coupling or the like. .

[Means for solving problems]

An interest of the present invention is to improve pattern separation capabilities through improved self-assembly methods.

That is, the present invention allows various shape patterns to be identified to be
In a pattern storage identification device which is applied to the identification of all products requiring identification of a new human cover pattern to a stored pattern, 1. u
Vi, an information processing layer in which excitatory and inhibitory nonlinear pixel elements are mixed and arranged in a two-dimensional manner is sequentially arranged, and the information processing layer is sequentially arranged between each stroke of the information processing layer. By providing directional connections and inhibitory connections, the connections between the sequential information processing layers are self-organized in correspondence with input information and processing results in the intermediate information processing layers. It is something.

[Effect]

In the above-mentioned (13th invention, when storing several patterns in the self-organizing storage device), in the overlearning of each turn, all the different patterns are However, if they are the same or included in one or the other, among the remaining non-enhanced cell faces, the cell face with the cell with the maximum output is designated as the enhanced cell face of that pattern. This makes fa different from others,
We will strengthen the BL filter. Note that the advantages of strengthening fa and b1 are the same as in the conventional method.

〔Example] Hereinafter, details of the present invention will be explained with reference to Examples.

In general, Shenfeng t in a neural network that imitates the human cerebrum.
As a unit information processing element, that is, a neural element, as a cell model, 1. Non-linear processing of input analog signal
Alternatively, an analog element that generates a positive analog output signal, an analog threshold element that generates a binary output signal of 0 or 1, or the like is used. When cascading information processing layers in which such information processing elements are two-dimensionally arranged for pattern information processing in multiple stages and connecting the elements in each layer, it is necessary to connect each neural element to another layer. It does not give the possibility of combining with all neurotoxins,
The possibility of connection is provided only between neural elements within a limited range of minute regions.

This region is called the bondable region of the cell element, and the strength of bonding is differentiated between the center and the periphery even within the bondable region.

1-Avoid changing the strength of the 1111 combination of neural elements between the information 91j processing layer river L1 as described above according to the pattern of human input information, and self-organize the neural network by strengthening the so-called 'i combination strength. . Here, strengthening the bond strength means forming a new bond when there is no bond between elements and the bond strength is 0, and forming a new bond when there is already a bond between elements. It means further strengthening. Neural elements that make connections between each other include sensitizing elements, which have an agonizing effect on other elements by increasing their output, and sensitizing elements, which have an inhibitory effect on other elements by decreasing their output. Two types of suppressive elements are set.

In strengthening the connections between the neural elements r, for each layer, the neural elements that formed the maximum output within the connection i-+J function region, ryo, shi, showed the best response.
In other words, it follows the rule 1 that only the input terminal connection of the neurotoxin r that has formed the output that is most suitable for human power is strengthened only when human power is present.

FIG. 1 shows an example of a pattern storage identification device in which information processing layers in which information processing elements are two-dimensionally arranged for pattern information processing are arranged in cascade in multiple stages.

Uo is the human power layer, and U elements arranged in nx rows and ny columns are U. Expressed as (re). Ull~U1. is the cell surface of the first stage, and a bond is made between this U+, z(71) element r-, and fjJ is the previous stage U. Intralayer bond r + J function 6r1
Four U elements arranged within the area are expressed as Uo(@+u+).

1.1 here? -1-Kl, which is 9-(n)(,
ny), and 1 indicates the deviation of the two-dimensional coordinate from n.

Now, as shown in FIG. 1, each information processing layer, for example, each U element U21. (n) is a group of U elements of the preceding stage U, that is, IU 51 ("11
), U r (re+u), "', U, (re+t+)f
, respectively, from Kaka's Kokame connection fa2 (u), a,', (a), =-, a,,',
(a) As shown in the figure, the first stage V element V+(
9) The same U element (U, 1 (FLIvJ), TJ 32 (t'L+u>
, −+ , U I, kl(M”ll)1 receives a fixed inhibitory ♀, − combination C+ (u), and its output is U 2L in −L by the inhibitory connection b2 of III change.
(7t) has been added. These variable bonds a and b are strengthened according to the bond strengthening rules described above.

Here, the rules for strengthening F111 are concretely shown using FIG. 1 as an example, as follows.

The output of the entire U element near the second stage coordinates is I U 24 (n, +uA ), U
2. z(ftl + a), -.

U2. For example, the maximum value of these prime r is expressed as *(F
LI). However, Nel ko (1, 2, -, ni 2), n
*('Ih, + u.

Although we have shown the method for selecting the maximum value for one neighborhood in the neighborhood system that completely covers U2, the same operation is performed for the remaining neighborhoods to select the maximum value. By the way, due to the restriction that one maximum value cannot be selected for one cell surface, if two or more maximum values are selected in different neighborhoods of one cell surface, the maximum value among them is used as the true value. Considered the maximum value. Therefore, one maximum value of height is selected for each cell plane, and the maximum value of 1 is at most 1 in a certain neighborhood of the neighborhood system. The cells selected in this way are called reinforced cell surfaces. - The aforementioned maximum value selection method will be explained with reference to FIG.

In Figure 2, the second stage consists of 6 cell surfaces, and each cell surface has 4 cell surfaces.
Suppose that it is covered with near f'l (1, 2, 3, 4).

Then, the maximum value of neighborhood 1 is U2. In L (FLI),
The maximum value of neighborhood 2 is U 2,1 (n2), and the neighborhood f573
The maximum value of is U2,3 (desk 3), and the maximum value of neighborhood 4 is U2
Suppose that it was taken at j (re 4). Since U2 and U3 have two maximum values in different neighborhoods, the maximum value among them is selected. Therefore, the enhanced cell surface is 1.3.5. Therefore, filters a and b to be strengthened are ia2. a2. a2. -=, a2, b211・Llj
L・L Tow! Ia2. a2. a2. −． a2, b21; di m,
l+, i Kl, turtle 1 (a2.a2.a
2. -, a2, b211, E r, i l,
i kl, i L, and the electricity to be strengthened is, for example, Δa, (III) ccc, (u) ・U3
1(F'++u+)Δa2 (+u)ocCl (a)
・U'g(t'h+1lj)7.1 Δb2 y: Vl It was determined as follows.

Next, a method for selecting reinforced cell surfaces according to the present invention will be explained based on the pattern storage identification device shown in FIG.

Now, the 00th layer (these three bats, rAJ, rBJ.

Suppose that a and b are reinforced while cyclically presenting "C", and after the presentation of cold tofu H, the reinforced cell planes for rAJ and rBJ patterns are the same, or -noj becomes a shape that includes other Jj. . For example, suppose that in the second stage, the reinforced cell surface of rCJ is 1,2.5, the reinforced cell surface of rAJ is 2.3.6, and the reinforced cell surface of rBJ is 2,3. The reinforced cell surface of rBJ has three elements that are not included in the reinforced cell surface of "C", but it is completely included in the reinforced cell surface of rAJ. Therefore, the rBJ enhanced cell plane is selected from those other than 1.2, 3, and 5.6, and at that time, the cell plane having the maximum value among the remaining cell planes is specified. In this way, the pattern is presented cyclically.
(Each time 7 is presented, by comparing the enhanced cell surface II of each pattern for one cycle immediately before each stage,
As mentioned above, new reinforced cell surfaces are selected as necessary.

By creating the following (1), different reinforcing cell planes are selected for different patterns, a and b are strengthened accordingly, and finally each pattern is made into a different one. As a specific example, five number patterns rlJ, r2J.

When r3J, r4J, and r5J are cyclically presented in the U layer, FIGS. 3 and 4 show how the enhanced cell surface changes depending on the number of times of presentation>J<.

Here, Fig. 3 shows the conventional method, and Fig. 4 shows the present example)
This is the J method. The number of cell faces in the final stage is 24, and it can be seen from FIG. 3 that in the 15th presentation, "4" and "5" are output as the same cell face 2. On the other hand, in the method of this embodiment, as shown in FIG.
lJ, r2J, r3J.

r4J and r5J are respectively 18, 21, 20°19.3
It can be seen that the output is 1111.

FIG. 5 shows an example of an enhanced cell-to-cell selection device according to the conventional method, and FIG. 6 shows an example of an enhanced cell-plane selection device according to this embodiment. Since the function of the selection device in each stage is the same, the second stage will be described in connection with FIG.

First, the conventional method will be explained using FIG.

■ expresses the maximum value of each neighborhood on each cell surface of the U layer,
Store the maximum value and the coordinates at which it was detected in the specified location (■). The maximum value and maximum value detection location storage device (2) stores the maximum value and maximum value detection coordinates using the cell surface and nearby addresses as keys. When all values have been memorized (a), the primary enhanced cell surface selection device (①) detects the maximum value regarding the same neighborhood of each cell surface, and calculates the maximum value, the cell curve number that takes the maximum value, and Memorize cell plane coordinates. In the example in Figure 5,
Since there are four neighbors, 1.2, 3.4, the four values (maximum value, cell curve number and cell plane coordinates that take the maximum value) are stored.

For these stored data, check whether the same cell song number has been selected in the secondary reinforced cell song selection device (①), and if so, select the cell song number and cell song number that has the maximum value among them. Select only the corresponding in-cell coordinates and remove the rest.

In this way, the output of ■ is (reinforced cell track number and cell plane coordinates).

Next, an example using the enhanced cell surface selection device according to this embodiment shown in FIG. 6 will be described.

(2) to (q) are similar to the functions shown in FIG. Additional functions according to this embodiment are ■ to ■. ■ is a device that stores the maximum value and the maximum (l'i detected coordinates) on each cell surface. ■ is a storage area with (N-1) elements for the total number of patterns to be stored (From Φ■
The output to is the data of (enhanced cell track number, coordinates in the cell plane) at that stage. For example, the total number of patterns to be memorized is N or 3, among which rAJ and rBJ.

If it is "C", it is U in Figure 1. Three patterns are presented cyclically. That is, rAJ→rBJ −r
The time series of cJ - rAJ → rBJ → rcJ →... is U. The pattern data changes, and at some point Uo
If there is "C" in , the oldest content of the soft buffer in ■ is related to rAJ (reinforced cell song number, data of cell plane coordinate l, and the next oldest is related to rBJ.

Since the total number N is 3, only these two pieces of data are included. Therefore, I would like to explain ■, but ■ receives (reinforced cell surface number, cell quotient coordinates) from C), and ('D
In case 1〉, the set of (reinforced cell song numbers) received from cell plane number) (N-1 in total)
If it is not a subset for any of (individuals), - output (reinforced cell face number, cell inside location) received from C) to ■, and strengthen a and b with this.

In Case 2, if the conditions of Case 1 are not evaluated, the set of (reinforced cell song numbers) received from -〇) and each 61 of (1)
Find the union of the results of (reinforced cell number) in the region '1, and for the cell planes that are not included in it, calculate the cell plane number that takes the maximum value and the coordinates within the cell plane corresponding to this in ■. Determine from the stored data, and output the 1-reinforced cell surface number, cell internal coordinates), which is the union of this data and the data received from ■, to ■, and use this to strengthen a and b.

〔Effect of the invention〕

Below 1. According to the invention as detailed, different a.

b is strengthened (self-organized), and as a result, the ability to separate and memorize different patterns as different things becomes significantly higher than the conventional JJ method. In other words, it is possible to improve the pattern decomposition function.

Furthermore, since there is no need to provide a blJ51 type suppression coupling, etc., the structure can be simplified compared to Japanese Patent Application No. 58-221474.

[Brief explanation of drawings]

FIG. 1 is a schematic configuration diagram showing a pattern storage identification device 4 according to an embodiment of the present invention, FIG. 2 is a schematic diagram for explaining a conventional maximum value selection method, and FIG. 3 is a pattern classification diagram according to a conventional method. FIG. 4 is a schematic diagram showing an example of pattern separation capability according to the present embodiment; FIG.
The figure is an outline (13
It is a complete drawing. U3-・Manpower layer, U 1, 1- U l, Kl "'
1st stage cell surface, U2,1 to U2,8□...2nd stage cell surface, U3,1 to U3. Old... 3rd stage cell surface, a...
・+1■ Variable inhibitory connections, b... Variable inhibitory connections, C...
・Fixed inhibitory bond, U...inhibitory element, ■...inhibitory element.

Claims (1)

[Claims] A plurality of information processing layers in which a plurality of excitatory and inhibitory nonlinear threshold elements are mixed and arranged in two dimensions are sequentially arranged, and forward excitatory connections and inhibitory connections are established between the information processing layers. A pattern storage identification device characterized in that by providing connections, the connections between the successive information processing layers are self-organized in correspondence with input information and processing results in the intermediate information processing layers.