JPS63265372A - Signal pattern identifying device - Google Patents

Abstract

PURPOSE:To facilitate the identification of a signal pattern by deciding the category of an input signal from an addition value. CONSTITUTION:A pattern obtained by converting an input signal vector is inputted to a category deciding device group 2. Then the categories of all prototypes that have large degrees of resemblance to each prototype stored in a prototype memory 13 and can be identified with those prototypes in the memory 13 are defined as the category value the input signal. Then the category of the input signal is decided from the addition value of those categories.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a signal pattern identification device that classifies input signals into a plurality of categories according to their properties.

[Conventional technology]

A perceptron is well known as a conventional example of this type of signal pattern identification device. A learning pattern recognition device typified by the persebtron has a normal vector of a hyperplane called a weight vector to classify patterns into categories, and divides the pattern along the hyperplane.

In addition, as a learning device with a memory function, research on adaptive pattern recognition by Suzuki and Arimoto of the Faculty of Engineering Science, Osaka University.
Universal Learning Machine (hereinafter abbreviated as ULM). ULM is a method in which presented patterns and their categories are appropriately memorized in the order in which they are presented so as to form a root-branch binary format, and the category of the stored pattern that is closest to the presented pattern is output.

[Problem that the invention seeks to solve]

In a learning pattern recognition device such as a perceptron, identification is impossible unless the pattern can be linearly separable in principle, that is, it can be clearly divided by a single hyperplane. As a countermeasure for cases where linear separation is not possible, there is a method of having multiple hyperplanes and cutting the pattern space into zig-2-zag sections by AND or OR, but this has problems with convergence, and It is necessary to decide whether to prepare Therefore, in order to identify a pattern space with a complex distribution,
Even when one hyperplane is sufficient, many hyperplanes are used, which poses a problem of inefficiency.

On the other hand, in ULM, the structure of stored binary books largely depends on the order in which they are presented. In addition, if the wrong education is carried out, there is a problem that it will remain as a binary book and affect the ability to discern.

Furthermore, ULM does not have an input signal conversion means and has the function of grouping bit patterns similar to the input signal into one category, but in general, signals have specific points that are the key to pattern distribution, and the input signal Even if they are similar overall, if the keys are different, the categories to which they belong may be very different.

That is, ULM is primarily intended for handwritten character recognition, and does not meet the requirements for other input signals in general.

The present invention solves these problems and provides a signal pattern identification device that can be easily applied to input signals with patterns that are not linearly separable, has high discrimination ability, and can classify any signal pattern into categories. The purpose is to

[Means for solving problems]

A signal pattern identification device according to the present invention includes an input signal converting means for converting an input signal vector into a pattern that emphasizes its characteristics, and a prototype memory storing a prototype as a representative pattern for the pattern obtained by the input signal converting means. a prototype application range storage means that stores the application range of each prototype, a prototype category storage means that stores the category to which each prototype belongs, and a pattern obtained from the input signal conversion means and each prototype storage means stored in the prototype storage means. Examine the degree of similarity with the prototype, find out the prototypes that fall within the prototype application range stored in the prototype application range storage means from the similarity, extract the categories to which these prototypes belong from the prototype category storage means, and calculate the category value of the input signal. It is characterized by comprising a category determining means to seek, an adding means for adding all the category values obtained by the category determining means, and a category determining means for determining the category of the input signal from the output value of the adding means. do.

[Effect]

The pattern obtained by converting the input signal vector is input to the category judgment means, and if the degree of similarity with each prototype stored in the prototype storage means is large, that is, the distance to the prototype is short, the pattern is classified as a prototype. They are considered to be the same. The categories of all prototypes that can be identified as the same are taken as the category values of the input signal, and the category of the input signal is determined from the summed value.

In this way, a pattern close to the prototype that falls within the scope of the prototype is determined to belong to the category to which the prototype belongs, and the category of the input signal is determined. Therefore, unlike the identification method using a hyperplane, input signals with patterns that cannot be linearly separated can be easily identified. Furthermore, the discrimination ability is easily increased by increasing the number of prototypes, and there is no adverse effect due to the order of presentation patterns as in the case of ULM. Furthermore, by converting an input signal into a pattern that emphasizes its characteristics and then processing it, it becomes possible to classify various input signals into categories.

〔Example〕

Hereinafter, the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram of a signal pattern identification device according to an embodiment of the present invention.

In FIG. 1, an input signal vector (components may be binary signals of 0/1 or analog signals) is input to an input signal converter 1, and is converted into a pattern with emphasized features.

Various specific configurations of the input signal converter 1 can be considered depending on the purpose, but in general, one that performs threshold processing of random connections as shown in FIG. 2 is suitable. The conversion shown in Figure 2 can be expressed as follows.

It is assumed that the pattern (vector) obtained by the input signal converter 1 is y, and y has 0 and 1 components.

This pattern is input to the category determiner group 2.

The category determiner group 2 determines the category of the pattern y from the input signal pattern converter 1 based on the contents of the prototype memory 13 , the prototype application range memory 14 , and the prototype category memory 15 .

That is, the category judge group 2 first selects the prototype (71, l-1,
2, . . . , n) and the pattern y (similarity). As an example of the distance, for example, the Hamming distance d
Take (yl, F). Next, this distance d is compared with the contents of the prototype application range storage 14 (the application range corresponding to 71 is set as al), and dO'1. F) <al
For yl, the category to which the prototype yt belongs is extracted from the prototype category storage 15. Let this extracted prototype category be al. Now, there are two prototype categories to which prototype yt belongs,
One side is 1. The other is represented by -1. That is, ai-1o
It is r-1.

The categories thus obtained by category judge group 2 are input to adder 3, and for each prototype Yl d (F
l, Y) <θI is added. The output value of this adder 3 is b-Σa1.

The output of the adder 3 is input to a category determiner 4.

The category determiner 4 determines that the input signal belongs to category 1 if the output value of the adder 3 is b>o, and determines that the input signal belongs to category 1 if b>0.

In the case of b-, it is determined that it does not belong to any category.

FIG. 3 shows an example of the relationship between the prototype and input signals in this embodiment. FIG. 4 shows an example of category determination when this embodiment is applied to the identification of input signals with patterns that cannot be linearly separable.

The basic configuration of the present invention has been described above, but in this embodiment, the learning function described below is further added.

In the present invention, a prototype may be appropriately provided in the initial state, but there is no need for a prototype from the beginning. Assuming that there are no prototypes in the initial state, when an input signal is presented, the category determiner group 2 first outputs information as to which prototype has fired. Category affiliation check mechanism 6 learns this information when all prototypes do not fire (we are currently considering the case where there are no prototypes, so naturally this corresponds to the case where all prototypes do not fire). Supplied to signal generator 5.

The output signal of the learning signal generator 5 is sent to the switch 9. In this case, the switch 9 receives a command to perform case A control from the learning signal generator 5. Under the control of case A, the prototype generation mechanism 10 generates the current input signal conversion pattern y.
is generated as a prototype, and it is stored in the prototype memory 13 t:,@. Furthermore, when a prototype is written in the prototype storage 13, the prototype application range storage 14, the prototype category, and the gory storage 1
5, information on the prototype scope and prototype category corresponding to the prototype is also written.

Generally, even if several prototypes are stored, if the category affiliation check mechanism 6 detects non-firing of all prototypes, the conversion pattern y of the input signal is newly stored in the prototype storage 13 using the same logic. Accordingly, new information is also stored in the prototype scope storage 14 and the prototype category storage 15.

Further, in this embodiment, a prototype forgetting function is also added. By category judge group 2, d(Yi, 7)<
When determining yl which is θ, the firing frequency of each prototype is checked by the prototype firing frequency checking mechanism 7. When the firing frequency of a prototype falls below a certain threshold, the prototype is removed from the contents of the prototype memory 13 by the prototype forgetting mechanism 8, as the prototype does not play an important role in signal identification, and at the same time, the prototype is removed from the contents of the prototype memory 13. Prototype coverage and prototype categories are also removed from the contents of prototype coverage store 14 and prototype category store 15.

Assuming that a pattern appears with equal probability over the entire range, the larger the prototype application range θl is, the more frequently it fires.

The actual firing frequency is the number of steps n.

If the number of firings is m, it is m/n. Therefore, a method can be considered in which the prototype is removed when the prototype firing frequency is (m/n)<k(θl). Here, k(θ
1) is a monotonically increasing function regarding θ1. This is achieved through flexible and narrow prototypes.
This means changing the frequency standard value.

The functions of the prototype firing frequency checking mechanism 7 and the prototype forgetting mechanism 8 will be explained in more detail with reference to FIG. In Figure 6, the number of steps (↓ number) is n-18, the number of firings is 5 for prototype ■, 1 for prototype ■, the area is 7 for prototype ■, 1 for prototype ■, and the total area is 20. It shows. In addition,
The above function corresponds to prototype area/total area.

In this example, for prototype ■ m/n =
5/18 (0,28), k (θl)−7/20(
0.35). Therefore, m/n<k(θ1), and prototype (3) will be removed. In addition, the prototype ■ is m / n -1/18 (
0,OG).

k (θ2)-1/20(0,05), m/n
>k(θ2), the prototype ■ will remain without being removed.

Note that if the pattern does not appear with equal probability in the entire range (for example, it often appears around prototype pt, but rarely appears around prototype P2), m/
It is desirable to take a reference value that also depends on the prototype itself, such as n<k(θ, p) (representing a p prototype).

By removing prototypes that are not very useful for signal discrimination as described above, that is, by performing prototype forgetting, the discrimination ability may decrease to a degree, but on the other hand, if an atypical prototype is memorized, it may be This has the effect of creating opportunities for changes to the prototype. FIG. 6 shows the manner in which the prototype is added and the prototype is forgotten as described above.

Next, a method of changing the scope of application of the prototype will be explained. The output of the category determiner 4 and the teacher signal are input to the learning signal generator 5. If the re-inputs match, the learning signal generator 5 does not generate any signal because it means that correct signal identification has been performed. In other words, the scope of application of the prototype is not changed by learning.

If the output of the category determiner 4 and the teacher signal do not match, the switch 9 performs case B control.

In case B control, among the prototype application ranges stored in the prototype application range storage 14 by the application range converter 11, the prototype category in the prototype fired in response to the input signal matches the category of the input signal. Reduce the scope of prototype application by one.

An example of this situation is shown in FIG.

If there is an input signal with a pattern matching the prototype, in addition to the above learning, the category converter 12 updates the stored contents of the prototype category storage 15. Now, if the factor that determines category a1 is τl, then
−1,0≦τ1≦1.0.

The category is determined by As an example, τl is updated as follows. If the n-th update is expressed as τl (n), τI
(o)-al (category when stored).

r 1(n) -a r 1(n)+ (1-a)
It becomes P-E. Here, P is the output category of the category determiner 4, E is 1 if the output category and the true category match, and -1 if they do not match, and α is a parameter of 0, α, and 1. be. If there is a lot of noise and there is a high possibility that the prototype category is incorrect, then α can be decreased, and if it is fairly reliable, α can be increased. The above learning is completed when all input signals can be correctly classified into categories.

Note that the present invention is not limited to the above embodiments,
Various modifications can be made as follows.

In the embodiment, the input signal converter 1 performs random connection threshold processing in order to convert the input signal vector into a pattern with emphasized features, but other conversion methods include, for example, φp(X)-ΩPI There is a transformation called Xi. φp is the component of the feature pattern, p is the mask vector (component is 0/1, p is different for each component), X is the input signal vector, Ω is AND, OR, E
It is a logical function such as xclusive OR. The Kono transformation method is suitable for inferring logical expressions that hold between input signals.

In addition, as another conversion method, input signal vector
(The components are analog, 0 x a≦x1≦b) and the luminance of each pixel of the image is taken as a component, and there is the following conversion. Here, φ is a transformation that assigns each component of the input signal vector X to each block (1 to Ω) of the number line shown in FIG. This conversion method is effective when performing category classification based on the rough shape of an image.

In addition, Hamming distance was used to examine the similarity between the pattern into which the input signal was converted and the prototype in the category determiner group 2, but basically any distance may be used as long as it satisfies the axiom of distance. In other words, the pattern is a1 the prototype is b
Then, is a general distance called the 9th power norm. If the pattern component is 0/1 and p-1, this corresponds to the Hamming distance. When set to p-1100, the distance focuses only on the part where the a and b components differ the most. Also, the larger p is, the more the difference between the corresponding components of a and b (lO') will have an effect, and the smaller p is, the more the different number of corresponding components of a and b (the number of components) will have an effect. .

In addition, the present invention can be implemented with various modifications without departing from the scope.

〔Effect of the invention〕

According to the present invention, a pattern obtained by converting an input signal vector so as to emphasize its features is manually inputted to the category judgment means, and a pattern that is highly similar to each prototype stored in the prototype storage means and is the same as the prototype is used. By using the categories of all visible prototypes as the category values of the input signal, and determining the category of the input signal from their summed values, it is possible to eliminate patterns that are not linearly separable, which is difficult with the hyperplane identification method. It is also possible to easily identify patterns for input signals, and the identification ability can be improved by increasing the number of prototypes.Also, by using the system of listing prototypes, it is possible to easily identify patterns, which was a problem with ULM. There is no adverse effect due to the order of patterns, and stable and highly reliable identification is possible.

Furthermore, since the input signal is processed after being converted into a pattern that emphasizes its characteristics, various input signals can be classified into categories.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a signal pattern identification device according to an embodiment of the present invention, FIG. 2 is a diagram for explaining a specific example of an input signal converter in the embodiment, and FIG. 3 is a diagram for explaining a specific example of an input signal converter in the embodiment. A diagram for explaining the actions of the category judge group and the adder, FIG. 4 is a diagram showing a specific example of category determination in the same embodiment, and FIG. 5 is a prototype firing frequency check mechanism and prototype forgetting mechanism in the same embodiment. 6 is a diagram for explaining the operation of changing the contents of the prototype storage device and the prototype application range storage device in the same embodiment. FIG. FIG. 8 is a diagram for explaining the operation, and FIG. 8 is a diagram for explaining another example of the input signal converter. l...input signal converter, 2...category judge group,
3...Adder, 4...Category determiner, 5...Learning signal generator, 6...Category affiliation determining mechanism, 7...
- Prototype firing frequency check mechanism, 8... Prototype forgetting mechanism, 9... Switcher, 10... Prototype generation mechanism, 11... Applicable range converter, 12...
Category converter, 13... Prototype memory, 14
. . . Prototype application range storage, 15 . . . Prototype category storage. Applicant's agent Patent attorney Takehiko Suzue Kumeno, No V Gun Figure 2 Figure 4 1θ1 1 2 3 4 ~------n-(b-a)
b-a Figure 8

Claims (6)

[Claims] (1) Input signal conversion means for converting an input signal vector into a pattern that emphasizes its features; prototype storage means for storing prototypes as representative patterns for the patterns obtained by the input signal conversion means; and application of each prototype. a prototype application range storage means that stores the range; a prototype category storage means that stores the category to which each prototype belongs; and a degree of similarity between the pattern obtained from the input signal conversion means and each prototype stored in the prototype storage means. , find the prototypes that fall within the prototype application range stored in the prototype application range storage means based on their similarity, and extract the categories to which these prototypes belong from the prototype category storage means to obtain the category value of the input signal. The apparatus is characterized by comprising a determining means, an adding means for adding all the category values obtained by the category determining means, and a category determining means for determining the category of the input signal from the output value of the adding means. Signal pattern identification device. (2) The prototype storage means, the prototype scope storage means, and the prototype category storage means are controlled based on a learning signal from the learning signal generation means that generates a learning signal from the category determined by the category determination means and the teacher signal. The signal pattern identification device according to claim 1, characterized in that: (3) The prototype storage means receives information provided via the learning signal generation means when the category affiliation check means detects that all prototypes have not fired based on the prototype firing information from the category judgment means. 3. The signal pattern identification device according to claim 2, wherein the signal pattern identification device stores a prototype generated based on. (4) The prototype application range storage means expands or contracts the stored prototype application range based on the learning signal from the learning signal generation means. Signal pattern identification device. (5) The signal pattern identification according to claim 2, wherein the prototype category storage means corrects errors in the stored prototype categories based on the learning signal from the learning signal generation means. Device. (6) The prototype storage means, the prototype application range storage means, and the prototype category storage means store the prototypes whose frequency is determined to be below the threshold value by the prototype firing frequency checking means based on the prototype firing information from the category determining means and the corresponding prototypes. 2. The signal pattern identification device according to claim 1, wherein the contents of the prototype application range and the prototype category are removed by the prototype forgetting means.