JP3467914B2 - Hierarchical pattern classifier - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention generates and stores a hierarchical structure between items to be stored based on the similarities between the items to be stored, and systematically and efficiently uses the hierarchical structure. The present invention relates to a storage device capable of dynamically and flexibly retrieving stored items according to a purpose.

[0002]

2. Description of the Related Art In recent years, as a storage device, not only a device that searches a memory of an explicitly designated address or a memory whose contents completely match a search key at high speed and accurately, but also a search key is a storage item. Even if they do not completely match, a device that allows associative and flexible retrieval according to the contents of the stored items, and if the stored items are given, the hierarchical structure between the items is automatically set based on the mutual similarity between the items. There is a demand for an apparatus that can generate and store the hierarchical structure, display the hierarchical structure to a user, and systematically and efficiently search for the purpose by utilizing the hierarchical structure.

The following is a conventional hierarchical concept inheritance associative memory device (Hirahara, Oka, Kindo: "Cascade associative memory model for storing patterns having a hierarchical correlation", a communication technique, NC94.
-58, 1994). FIG. 6 shows a conventional hierarchical concept inheritance associative memory device. In FIG.
61 is the concept formation associative memory weight Wij held therein.
The operation is defined by (i = 1 to D, j = 1 to D), the concept formation associative memory signal X (t) at time t is updated by using the external input signal S as a trigger, and X (t-1) and X (t-1) Reference numeral 62 denotes a concept formation associative memory unit that outputs a stable concept formation associative memory signal X (= X (t-1) = X (t)) when (t) becomes equal. Associative memory weight W1ij (i = 1 to D, j = 1 to D)
The operation is specified by the external input signal S and the stable concept formation associative memory signal X output from the concept formation associative memory unit 61 is used as a trigger to update the concept inheritance associative memory signal X1 (t).
When 1 (t-1) and X1 (t) are equal, it is a concept inheritance associative memory unit that outputs a stable concept inheritance associative memory signal X1 (= X1 (t-1) = X1 (t)), and 63 is Stable concept formation associative memory signal X output from the concept formation associative storage unit 61 and concept succession associative storage unit 62
Output from the stable concept inherited associative memory signal X1
Is an output signal generation unit for generating.

Here, the number of dimensions of the external input signal S is D
The value of each element Si (i = 1 to D) of the external input signal S is ± 1. Further, for simplicity, the number of cascade levels will be described as two.

First, a method of generating P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2) for performing learning will be described. Prior to generation of the learning pattern S ^uv (u = 1 to P1, v = 1 to P2), P1 parent patterns S ^u (u = 1 to P1) are generated.
Each element S ^u i (i = 1 to D) of the parent pattern S ^u (u = 1 to P1) takes a value of ± 1 with the following probability Pr1 (S ^u i).

[0006]

[Equation 1]

Here, a (-1 <a <1) is the bias of the parent pattern. Here, if the average of x is represented by E [x], the bias of the parent pattern is defined by a = E [S ^u i]. Also, δ [x]
Takes a value of 1 when x = 0, and takes a value of 0 otherwise.

Next, P2 learning patterns S ^uv (v = 1 to P2) are generated for each parent pattern S ^u . Learning pattern S ^uv
Each element S ^uv i (i = 1 to D) of is ± 1 with the following probability Pr2 (S ^uv i)
Takes the value of.

[0009]

[Equation 2]

Here, b (0 <b <1) is a correlation parameter indicating the strength of the correlation between the parent pattern and the learning pattern.

The operation in the learning mode of the conventional hierarchical concept inheritance associative memory device will be described by using the learning pattern thus generated.

In the learning mode, the concept formation associative memory device 6
1 is P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2)
Received, the concept formation associative memory weight Wi
j (i = 1 to D, j = 1 to D)

[0013]

[Equation 3]

Update as follows. Here, d1 (Wij) may be a constant or a function of the concept-forming associative memory weight Wij, and takes a value from 0 to 1. For example, d1 (Wij) is

[0015]

[Equation 4]

Can be written as Similarly, d2 (Wij) may be a constant or a function of the concept formation associative memory weight Wij, for example,

[0017]

[Equation 5]

Can be written as According to these update formulas, P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2) are learned. By this learning, the learning pattern S ^uv (u = 1 ~
P1, v = 1~P2) be a pattern having a structure made with a correlation with the parent pattern S ^{u (u} = 1~P1), almost to its parent pattern S ^{u (u} = 1~P1) equal pattern S ^{'u (u} = 1~P
1) is stored in the concept formation associated storage unit 61. After that, the pattern S ′ ^u (u
= 1 ~ P1) is called a concept. From this, the concept
When the learning pattern S ^uv (v = 1 to P2) close to S ′ ^u (u = 1 to P1) is input to the concept formation associative storage unit 61 as the external input signal S, the concept formation associative storage unit 61 determines the concept formation. The associative memory signal X (t) is continuously updated, and when X (t-1) and X (t) become equal, the stable concept formation associative memory signal X (= X (t-1) = X (t)) Output. At this time, the stable concept associative memory signal X is a certain concept S ' ^u
Becomes For simplicity, hereinafter, the parent pattern S ^u and the concept S ′ ^u are equal (S ^u i = S ′ ^u i, u = 1 to P1, i = 1 to D), and the concept S
It is represented by ^u (u = 1 to P1).

Next, the concept inheritance associative storage unit 62
Pattern S^uvAnd its learning pattern S^uvThe concept formation associative memory
It is the stable concept associative memory signal X when input to the unit 61.
Concept S ^uAnd the concept inheritance association held inside
Memory weight W1ij (i = 1 to D, j = 1 to D)

[0020]

[Equation 6]

Update as follows. From this update formula of the concept inheritance associative memory weight W1ij, the concept inheritance associative memory unit 62 determines S ′ ^uv.
Inheritance pattern S'having i (= S ^uv iS ^u i, i = 1 to D) as an element
You can see that you remember ^uv . Concept S ^u and learning patterns
For correlation with S ^uv is very strong, inheritance pattern S 'many elements of ^uv is to take 1, inheritance pattern S' ^uv becomes sparse pattern.

The operation in the recall mode of the hierarchical concept inheritance associative memory device having the weight thus set will be described below.

In the recall mode, the external input signal S is input to the concept formation associative storage unit 61. This input is used as a trigger, and the concept formation associative storage unit 61 stores the concept formation associative storage signal X according to the concept formation associative storage weight Wij held therein.
Update (t) according to the following formula.

[0024]

[Equation 7]

[0025]

[Equation 8]

[0026]

[Equation 9]

[0027]

[Equation 10]

Here, Xi (t) represents the i-th element of the concept formation associative memory signal X (t) at time t, and Ui (t) is the i-th concept formation potential signal U (t) at time t. Indicates the element of. Here, when the concept forming associative memory signal X (t) is stabilized (X (t-1) = X (t)), the concept forming associative memory unit 61 determines that the stable concept forming associative memory signal X (= X (t -1) = X (t)) is output. The stable concept forming associative memory signal X becomes the concept S ^u most similar to the external input signal S.

When the concept formation associative storage unit 61 is stabilized,
The stable concept formation associative memory signal X and the external input signal S are input to the concept succession associative memory unit 62. Here, it is assumed that the stable concept forming associative memory signal X becomes the concept S ^u . Using these inputs as triggers, the concept inheritance associative memory unit 62 updates the concept succession associative memory signal X1 (t) according to the following formula in accordance with the concept inheritance associative memory weight W1ij held therein.

[0030]

[Equation 11]

[0031]

[Equation 12]

[0032]

[Equation 13]

Here, U1i (t) indicates the i-th element of the concept inheritance potential signal U1 (t) at time t. When the concept inherited associative memory signal X1 (t) is stabilized by this operation, the concept inherited associative memory unit 62 outputs the stable concept inherited associative memory signal X1 (= X1 (t-1) = X1 (t)). That is, the succession pattern S ′ ^uv corresponding to the learning pattern S ^uv similar to the external input signal S is recalled in the concept succession associative storage unit 62.

As described above, the concept inheritance associative storage unit 62 stores the inheritance pattern ^S'uv which is a sparse pattern. It is known that the storage capacity increases dramatically when the stored pattern is sparse, and therefore the concept inheritance associative storage unit 62 has a very large storage capacity.

When the concept formation associative memory unit 61 outputs the stable concept formation associative memory signal X and the concept inheritance associative memory unit 62 outputs the stable concept succession associative memory signal X1, the output signal generation unit 63 produces the following formula: Calculate the output signal R.

[0036]

[Equation 14]

Here, Ri represents the i-th element of the output signal R. In this way, the output signal generation unit 63 uses the external input signal
A learning pattern S ^uv similar to S can be recalled.

[0038]

However, in the above-mentioned conventional hierarchical concept inheritance associative storage device, since the inheritance patterns are stored in the concept succession associative storage unit 62 in a uniformly overlapped form, the storage items are stored in the same manner. Since the relationship, that is, the hierarchical structure cannot be extracted and the information about the classification of the pattern such as which category the input pattern belongs to cannot be obtained, the problem that the relationship between the memorized items cannot be provided to the user. However, it has a problem that it does not have a sufficient function to search systematically, efficiently, and according to the purpose.

The present invention is to solve the above-mentioned conventional problems. Based on the similarities between given memory items to be stored, a hierarchical structure between them is generated and stored, and the hierarchical structure is used. An object of the present invention is to provide a hierarchical pattern classification device, a simplified hierarchical pattern classification device, and a prioritized simplified hierarchical pattern classification device that can systematically, efficiently, and flexibly search for memory items according to the purpose. To do.

[0040]

To achieve this object, a hierarchical pattern classifying apparatus of the present invention includes a concept formation associative storage unit which receives an external input signal and outputs a stable concept formation associative storage signal, and a stable concept formation associative storage unit. It becomes operable by receiving a category detection unit for classifying the formed associative memory signal and an output signal from the category detection unit, and outputs a stable concept inherited associative memory signal in response to an external input signal and a stable concept formation associative memory signal. A plurality of concept inheritance associative memory units, a plurality of subcategory detection units for classifying stable concept inheritance associative memory signals, and an output signal generation for receiving stable concept formation associative memory signals and stable concept inheritance associative memory signals to generate output signals And a part.

[0041]

With this configuration, in the hierarchical pattern classifying apparatus of the present invention, inheritance patterns belonging to the same category (see the first embodiment) are stored in the same concept inheritance associative storage unit.
It is possible to explicitly take out the relationship between the memory items, and obtain information on the classification of the pattern such as which category the input pattern belongs to. That is,
Based on the similarities between the given items (learning patterns) to be stored, a hierarchical structure is generated and stored between these items, and the stored items are systematically and flexibly searched using the hierarchical structure. be able to. Further, by displaying the generated hierarchical structure, the user can easily understand the relationship between the memory items.

[0042]

【Example】

(Embodiment 1) Hereinafter, a first embodiment (hierarchical pattern classification device) of the present invention will be described with reference to the drawings.

In FIG. 1, 11 is an operation defined by a concept forming associative memory weight Wij (i = 1 to D, j = 1 to D, D is a dimension) held therein, and triggers an external input signal S. The concept formation associative memory signal X (t) is updated as
When X (t) becomes equal, stable concept formation associative memory signal X (=
X (t-1) = X (t)) is a concept formation associative memory unit,
An active value 12 is a weighted sum of the stable concept formation associated memory signal X output from the concept formation associated memory unit 11 and the category detection weights Cij (i = 1 to N, j = 1 to D) held therein. Is a category detection unit configured to output 1 from at least one cell, and 13i (i = 1 to N) is a category detection unit 12.
Can be operated only when cell i of 1 outputs 1.
Concept inheritance associative memory weight Wijk (j = 1
~ D, k = 1 ~ D), the operation is defined by external input signal
S and the concept inheritance associative memory signal Xi triggered by the stable concept formation associative memory signal X output from the concept formation associative memory unit 11.
(t) is updated, and when the Xi (t-1) and Xi (t) become equal, the stable concept inheritance associative memory signal Xi (= Xi (t-1) = Xi (t)) is output. The storage unit 14i (i = 1 to N) is a stable concept succession associative memory signal Xi output by the concept succession associative memory unit 13i and a subcategory detection weight held therein.
It consists of Ni cells i (i = 1 to Ni) that have the weighted sum of Cijk (j = 1 to Ni, k = 1 to D) as an active value, and at least one cell outputs 1 The sub-category detection unit 15 receives the stable concept formation associative storage signal X output from the concept formation associative storage unit 11 and the stable concept succession associative storage signal Xi output from the concept succession associated storage unit 13i. It is an output signal generation unit for generating.

In the present embodiment, the number of layers of the hierarchical pattern classification device is described as 2 for simplicity, but this limitation is not an essential problem, and the hierarchical pattern classification having three or more layers. It can be easily extended to the device. The number of dimensions of the external input signal S is D, and the value of each element Si (i = 1 to D) of the external input signal S is ± 1.

First, a method of generating P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2) for performing learning will be described. Prior to generation of the learning pattern S ^uv (u = 1 to P1, v = 1 to P2), P1 parent patterns S ^u (u = 1 to P1) are generated.
Each element S ^u i (i = 1 to D) of the parent pattern S ^u (u = 1 to P1) takes a value of ± 1 with the following probability Pr1 (S ^u i).

[0046]

[Equation 15]

Here, a (-1 <a <1) is the bias of the parent pattern and is defined by E [S ^u i] = a (E is an average value). Also, δ
[x] takes a value of 1 when x = 0, and takes a value of 0 otherwise.

Next, P2 learning patterns S ^uv (v = 1 to P2) are generated for each parent pattern S ^u . Learning pattern S ^uv
Each element S ^uv i (i = 1 to D) of is ± 1 with the following probability Pr2 (S ^uv i)
Takes the value of.

[0049]

[Equation 16]

Here, b (0 <b <1) is a correlation parameter between the parent pattern and the learning pattern. The operation in the learning mode of the hierarchical pattern classification device in FIG. 1 will be described using the learning pattern generated in this way.

In the learning mode, the concept formation associative memory device 1
1 is P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2)
Received, the concept formation associative memory weight Wi
j (i = 1 to D, j = 1 to D)

[0052]

[Equation 17]

Update as follows. Here, d1 (Wij) may be a constant or a function of the concept-forming associative memory weight Wij, and takes a value from 0 to 1. For example, d1 (Wij) is

[0054]

[Equation 18]

Can be written as Similarly, d2 (Wij) may be a constant or a function of the concept formation associative memory weight Wij, for example,

[0056]

[Formula 19]

Can be written as According to these update formulas, P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2) are learned. By this learning, the learning pattern S ^uv (u = 1 ~
P1, v = 1~P2) be a pattern having a structure made with a correlation with the parent pattern S ^{u (u} = 1~P1), almost to its parent pattern S ^{u (u} = 1~P1) equal pattern S ^{'u (u} = 1~P
1) is stored in the concept formation associative storage unit 11. Thereafter, the pattern S ′ ^u (u
= 1 ~ P1) is called a concept. This gives the concept S '
^{When the} learning pattern S ^uv close to ^u (u = 1 to P1) is input to the concept formation associative storage unit 11 as the external input signal S, the concept formation associative storage unit 11 updates the concept formation associative storage signal X (t). The stable concept formation associative memory signal X (= X (t-1) = X (t)) is output when X (t-1) becomes equal to X (t). The stable concept associative memory signal X at this time becomes the concept ^S'u . For simplicity, the parent pattern S ^u and the concept ^S'u are equal (S ^u i = S ' ^u i
, u = 1 to P1, i = 1 to D), and the concept is represented by S ^u (u = 1 to P1).

When the learning of the concept formation associative storage unit 11 is finished, the learning of the category detection unit 12 is started.

First, the category detection weight Cij (i = 1 to N, j = 1 to D) of the category detection unit 12 is initialized to a sufficiently small random number. In addition, P1 × P2 learning patterns S ^uv (u = 1
~ P1, v = 1 ~ P2) by adding a sufficiently small random number to the average pattern of the category detection unit 12 of the category detection weight Ci
By initializing j (i = 1 to N, j = 1 to D), the learning speed and learning result can be improved. Next, the learning pattern S ^uv is input to the concept formation association memory unit 11, and the concept formation association memory unit 11 outputs the stable concept formation association memory signal X,
That is, the concept S ^u is output. The category detection unit 12 will learn this concept S ^u . In this case, the category detector 12 calculates the weighted sum, and the activity value ^u i of the cell i of the category detector 12 is

[0060]

[Equation 20]

It becomes Where hi is the category detector 1
It is a variable threshold value of cell i of 2. Based on this activation value, the output value oi of the cell i of the category detection unit 12 is

[0062]

[Equation 21]

It becomes Here, c (0 <c ≦ 1) is a constant. From this, it can be seen that in the category detection unit 12, the cell having the maximum active value and the cell having the active value equivalent to the maximum active value output 1. When the output value of the cell is determined, the category detection weight Cij (i
= 1 to N, j = 1 to D) is updated according to the following weight update formula.

[0064]

[Equation 22]

Here, A is a sufficiently small positive real number.
From this, it can be seen that the category detection unit 12 updates only the weight that is combined with the cell that outputs 1. Further, the cell threshold of the category detection unit 12 is updated based on the following formula.

[0066]

[Equation 23]

Here, A1 is a sufficiently small positive real number of 1 or less. All learning patterns S ^uv (u = 1
~ P1, v = 1 to P2) is sufficiently repeated to complete the learning of the category detection unit 12.

When the learning of the category detector 12 is completed, the concept inheritance associative memory 13i (i = 1 to N) is learned.

Consider learning the learning pattern S ^uv . First, the learning pattern S ^uv is the concept formation associative storage unit 11
, And the concept formation associative memory unit 11 outputs the stable concept formation associative memory signal X, that is, the concept S ^u . It is assumed that the category detection unit 12 has learned this concept S ^u and that the cell i of the category detection unit 12 outputs 1. At this time, the concept inheritance associative storage unit 13i connected to the cell i of the category detection unit 12 becomes operable. In this case, the concept inheritance associative memory unit 13i receives the learning pattern S ^{u v} and the concept S ^u which is the stable concept associative memory signal X when the learning pattern S ^uv is input to the concept formation associative memory unit 11, Concept inheritance associative memory weight Wijk held in it
(J = 1 to D, k = 1 to D)

[0070]

[Equation 24]

Update as follows. From this update formula of the concept inheritance associative memory weight Wijk, the concept inheritance associative memory unit 13i determines S ′.
Inheritance pattern with ^uv j (= S ^uv jS ^u j, j = 1 to D) as elements
You can see that you remember ^S'uv . The correlation between the concept S ^u j and the learning pattern S ^uv j is so strong that the inheritance pattern ^S'uv
Many elements of S will take 1 and inheritance pattern ^S'uv
Becomes a sparse pattern. By performing such learning for all learning patterns S ^uv (u = 1 to P1, v = 1 to P2), the learning of the concept inheritance associative storage unit 13i ends.

In this way, when the external input signal S similar to the learning pattern S ^uv is input to the hierarchical pattern classifier, the concept formation associative memory unit 11 outputs the concept S ^u as the stable concept formation associative memory signal X. The concept inheritance associative storage unit 13
i is the inheritance pattern S'as the stable concept inherited associative memory signal Xi.
Output ^uv .

Next, learning of the subcategory detecting section 14i will be described. This learning is performed by the concept inheritance association memory unit 13.
Although the learning can be performed simultaneously with the learning of i, the following description will be made assuming that the learning of the concept inheritance associative storage unit 13i is completed.

First, the subcategory detection weight Cijk (j = 1 to Ni, k = 1 to D) of the subcategory detection unit 14i is initialized to a sufficiently small random number. Here, the learning pattern S
Think about learning ^uv . First, the learning pattern
S ^uv is input to the concept formation associated memory unit 11, and the concept formation associated memory unit 11 outputs the stable concept formation associated memory signal X, that is, the concept S ^u . It is assumed that the category detection unit 12 has learned this concept S ^u and that the cell i of the category detection unit 12 outputs 1. At this time, the concept inheritance associative storage unit 13i connected to the cell i of the category detection unit 12
Becomes an operable state, and the concept inheritance associative memory unit 13i outputs a succession pattern S ′ ^uv as a stable concept succession associative memory signal Xi. The subcategory detection unit 14i learns this inheritance pattern ^S'uv . In this case, the subcategory detection unit 14i calculates the weighted sum, and the activity value ^u ij of the cell j of the subcategory detection unit 14i is

[0075]

[Equation 25]

It becomes Here, hij is a variable threshold value of the cell j of the subcategory detection unit 14i. Based on this active value, the output value oij of the cell j of the subcategory detection unit 14i is

[0077]

[Equation 26]

It becomes Here, ci (0 <ci ≦ 1) is a constant. From this, it can be seen that in the subcategory detection unit 14i, a cell having an activity value of ci times or more the maximum activity value outputs 1. When the cell output value is determined, the subcategory detection weight Cijk (j = 1
~ Ni, k = 1 to D) are updated according to the following weight update formula.

[0079]

[Equation 27]

Here, Ai is a sufficiently small positive constant.
From this, it can be seen that the sub-category detection unit 14i updates only the weight that is combined with the cell that outputs 1. In addition, the cell threshold of the subcategory detection unit 14i is updated based on the following formula.

[0081]

[Equation 28]

Here, A2i is a sufficiently small positive real number of 1 or less. All learning patterns S ^uv (u = 1 ~
The learning of the subcategory detection unit 14i is completed by sufficiently repeating P1, v = 1 to P2).

The operation in the recall mode of the hierarchical pattern classifying apparatus having the weight thus set will be described below.

In the recall mode, the external input signal S is input to the concept formation associative storage unit 11. This input is used as a trigger, and the concept formation associative storage unit 11 stores the concept formation associative storage signal X according to the concept formation associative storage weight Wij held therein.
Update (t) according to the following formula.

[0085]

[Equation 29]

[0086]

[Equation 30]

[0087]

[Equation 31]

[0088]

[Equation 32]

Here, Xi (t) represents the i-th element of the concept formation associative memory signal X (t) at time t, and Ui (t) is the i-th element of the concept formation potential signal U (t) at time t. Indicates the element of. Here, when the concept formation associated memory signal X (t) is stabilized, the concept formation associated memory unit 11 outputs the stable concept formation associated memory signal X (= X (t-1) = X (t)). This stable concept formation associative memory signal X is the concept most similar to the external input signal S.
It can be expected that S ^u (u = 1 to P1). Hereinafter, it is assumed that the stable concept formation associative memory signal X when the external input signal S is input to the concept formation associative memory unit 11 becomes the concept S ^u .

Next, the stable concept formation associative memory signal X output from the concept formation associative storage unit 11, that is, the concept S ^u, is input to the category detection unit 12, and (Equation 20) and (Equation 2)
The output value of the cell of the category detection unit 12 is calculated from 1). Here, it is assumed that the cell i of the category detection unit 12 outputs 1.

At this time, the cell i of the category detection unit 12
The concept inheritance associative memory unit 13i connected to the active state becomes operable, and the external input signal S and the stable concept formation associative memory signal X output from the concept formation associative memory unit 11, ie, the concept S ^u
Is input to the concept inheritance associative storage unit 13i. With these inputs as triggers, the concept inheritance associative memory unit 13i updates the concept succession associative memory signal Xi (t) according to the following formula in accordance with the concept inheritance associative memory weight Wijk held therein.

[0092]

[Expression 33]

[0093]

[Equation 34]

[0094]

[Equation 35]

Here, Uij (t) indicates the j-th element of the concept inheritance potential signal Uij (t) at time t. When the concept inherited associative memory signal Xi (t) is stabilized by this operation, the concept inherited associative memory unit 13i outputs the stable concept inherited associative memory signal Xi (= Xi (t-1) = Xi (t)). That is, the succession pattern S ′ ^uv corresponding to the learning pattern S ^uv similar to the external input signal S is recalled in the concept succession associative storage unit 13i.

[0096] Thus, the concept inherited associative memory unit 13i stores a inheritance pattern S ^'uv a sparse pattern. It is known that the storage capacity increases dramatically when the stored pattern is sparse, and therefore the concept inheritance associative storage unit 13i has a very large storage capacity.

[0097] Next, stable concept inherited associative memory signal Xi to output concept inherited associative memory unit 13i, that is, the concept S ^'uv is input to the sub-category detection unit 14i, the sub from (Equation 25) and (Expression 26) The output value of the cell of the category detection unit 14i is calculated, and 1 of the subcategory detection unit 14i is calculated.
One cell outputs 1.

From this fact, which category the external input signal S belongs to in the category detecting section 12, that is, the first hierarchical level, and which subcategory detecting section 13i, that is, the subcategory in the second hierarchical level, belong to. Can be judged.

Further, the output signal generation unit 15 outputs the stable concept formation associative storage signal X output from the concept formation associative storage unit 11, that is, the concept S ^u, and the stable concept succession associative storage signal output from the concept succession associative storage unit 13i. By receiving Xi, that is, the inheritance pattern S ′ ^uv , the output signal R is calculated based on the following equation.

[0100]

[Equation 36]

Here, Ri represents the i-th element of the output signal R. In this way, the output signal generator 15 generates the learning pattern S ^uv most similar to the external input signal S.

Further, by expanding the function of the category detection unit 12, it is possible to speed up learning and retrieval and reduce memory. For example, the learning pattern sub-set i that gives the output 1 to the cell i is stored in each cell i of the category detection unit 12. When the number of elements of this learning pattern sub-set i is 0, it is realized by deleting the concept inheritance associative storage unit 13i and the sub-category detection unit 14i related to the cell i of the category detection unit 12.

Furthermore, the D-dimensional learning pattern S ^uv has T partial learning patterns S ^uv j (j = 1 to T) (partial learning pattern
Let the dimension of S ^uv j be D j. ), The dimension of the inheritance pattern S ′ ^uv stored in the concept inheritance associative storage unit 13i can be reduced by expanding the function of the category detection unit 12. For example, the category detection unit 12
The learning pattern sub-set i that gives the output 1 to the cell i is stored in each cell i of. If a partial learning pattern of all learning patterns included in this learning pattern sub-set i is equal to all learning patterns included in the learning pattern sub-set i, the partial learning pattern is deleted from each learning pattern (that is, Delete the partial inheritance pattern corresponding to the partial learning pattern), category detection unit 1
The dimension reduction of the inheritance pattern can be realized by making the concept inheritance associative storage unit 13i and the subcategory detection unit 14i on and after the second learn.

Similarly, if the function of the subcategory detecting unit 14i is expanded as described above, the dimension of the inheritance pattern can be further reduced.

In the present embodiment, the number of layers of the hierarchical pattern classifying device has been described as 2 for the sake of simplicity. However, this limitation is not an essential problem, and a hierarchical pattern classifying device having three or more hierarchical layers can be used. Can also be easily extended. FIG. 2 shows a schematic diagram (excluding the output generation unit) of the hierarchical pattern classification device when the number of layers is three. In FIG. 2, reference numeral 21 denotes a module A in which the concept formation associative storage unit 11 and the category detection unit 12 are integrated, and 22i (i = 1 to 1).
N) is a concept formation associative storage unit for succession patterns (the operation is the same as the concept formation associative storage unit 11, and the concept formation associative storage unit 11 learns the learning pattern itself,
The inheritance pattern concept formation associative storage unit learns the inheritance pattern of the learning pattern) and the category detection unit (the operation is the same as that of the category detection unit 12 and learns the stable pattern output from the inheritance pattern concept formation associative storage unit). 23i is a module B in which
j (j = 1 to Ni) is a module C in which the concept inheritance associative storage unit 12i and the subcategory detection unit 12i are integrated. In FIG. 2, for simplicity, the number of input lines from the module A21 to the module B22i is one.
Actually, two kinds of signals, that is, a stable concept formation associative memory signal output from the concept formation associative storage unit inside the module A21 and an output signal from the cell of the category detection unit are module B2.
It will enter 2i. Similarly, although there is one input line from the module B22i to the module C23ij, in reality, from the stable concept formation associative storage signal output from the inheritance pattern concept formation associative storage unit inside the module B22i and the cell of the category detection unit. Two types of signals of the output signals of the above will enter the module C23ij. In this way, it is possible to easily expand to a hierarchical pattern classification device having three or more hierarchical layers.

As described above, according to this embodiment, since the inheritance patterns belonging to the same category (see Embodiment 1) are stored in the same concept inheritance storage unit, it is possible to explicitly take out the relationship between the stored items. Therefore, it is possible to obtain information about the classification of patterns such as which category the input pattern belongs to. That is, based on the mutual similarity between given items (learning patterns) to be stored, a hierarchical structure between those items is generated and stored, and the hierarchical structure is used to systematically, efficiently and flexibly. You can search the memory. Further, by displaying the generated hierarchical structure, the user can easily understand the relationship between the memory items.

(Embodiment 2) A second embodiment (simple hierarchical pattern classification apparatus) of the present invention will be described below with reference to the drawings.

In FIG. 3, reference numeral 31 denotes an external input signal S (each element is represented by Si (i = 1 to D)) and a category detection weight Cij (i = 1 to N, j = 1 to 1) held therein. D is a category detection unit that is composed of N cells i (i = 1 to N) having a weighted sum with D) as an active value, and at least one cell outputs 1 and 31i (i = 1 Up to N) can be operated only when the cell i of the category detection unit 31 outputs 1, and the external category input signal S and the sub category detection weight Cijk (j = 1 to Ni, k = 1) held therein. To D) is a sub-category detection unit that is composed of Ni cells that have a weighted sum total as an active value, and at least one cell outputs 1 and 32 is a category detection unit 31.
Is initialized by giving a sufficiently small random number to the category detection weight Cij (i = 1 to N, j = 1 to D) of the category detection weight Cij (i = 1 to
N, j = 1 to D) in consideration of the subcategory detection unit 31i
(I = 1 to N) sub-category detection weight Cijk (j = 1 to N)
i, k = 1 to D) is an initial weight setting unit for initializing 3
An output signal generation unit 3 receives the subcategory detection weights Cijk (j = 1 to Ni, k = 1 to D) of the subcategory detection unit 31i (i = 1 to N) and generates an output signal R.

In this embodiment, the number of layers of the simplified hierarchical pattern classifier is set to 2 for the sake of simplicity.
This limitation is not an essential problem and can be easily extended to a simplified hierarchical pattern classification device having three or more hierarchical layers (described later).

The learning operation of the two-layer simple hierarchical pattern classifying apparatus configured as described above will be described. In the learning mode, learning of the category detection unit 31 starts.

First, the category detection weight Cij (i = 1 to N, j = 1 to D) of the category detection unit 31 is set to the initial weight setting unit 32.
Is initialized to a sufficiently small random number. Also, P1 ×
By adding a sufficiently small random number to the average pattern of the P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2), the category detection weight Cij (i = 1 to N, j = 1 of the category detection unit 31. It is possible to improve the learning speed and the learning result by initializing (~ D). Next, P1 × P2 learning patterns S ^uv (u
= 1 to P1, v = 1 to P2) are sequentially input to the category detection unit 31, and learning is sequentially performed. Consider a case where the learning pattern S ^uv is input to the category detection unit 31. In this case, the category detector 31 calculates the weighted sum, and the activity value ui of the cell i of the category detector 31 is

[0112]

[Equation 37]

It becomes: Where hi is the category detection unit 3
This is a variable threshold value for cell i of 1. Based on this active value, the output value oi of the cell i of the category detection unit 31 is

[0114]

[Equation 38]

It becomes: c is a positive constant of 1 or less. From this, it can be seen that in the category detection unit 31, the cell having the maximum active value and the cell having the active value equivalent to that output 1. When the output value of the cell is determined, the category detection weight Cij (i = 1 to N, j = 1 to
D) is updated according to the following weight update formula.

[0116]

[Formula 39]

Here, A is a sufficiently small positive real number.
From this, it can be seen that the category detection unit 31 updates only the weight that is combined with the cell that outputs 1. Further, the cell threshold of the category detection unit 31 is updated based on the following formula.

[0118]

[Formula 40]

Here, A1 is a sufficiently small positive real number of 1 or less. This kind of learning is performed with P1 × P2 learning patterns S ^uv
By sufficiently repeating (u = 1 to P1, v = 1 to P2), the learning of the category detection unit 31 ends.

When the learning of the category detecting unit 31 is completed, the sub-category detecting unit 31i (i = 1 to N) is learned.

First, the subcategory detection weights Cijk (j = 1 to Ni, k = 1 to D) of the subcategory detection unit 31i are initialized by the initial weight setting unit 32 as follows.

[0122]

[Formula 41]

Here, α is a sufficiently small random number. As a result, the sub-category detection weights Cijk (j = 1 to Ni, k = 1 to D) of the sub-category detection unit 31i are distributed around the category detection weights Cij (j = 1 to D) of the category detection unit 31. Will be done.

When the initial setting of the subcategory detection weights Cijk (j = 1 to Ni, k = 1 to D) of the subcategory detection unit 31i is completed, P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 ~
P2) is learned. Now, ^let us consider the learning of the learning pattern S ^uv . In this case, first, the learning pattern S ^uv
Is input to the category detection unit 31. The category detection unit 31 calculates the output value of the cell based on (Equation 37) and (Equation 38). Here, a cell that outputs 1 when the learning pattern S ^uv is input to the category detection unit 31 is set as a cell i. As a result, the subcategory detection unit 31i becomes operable, and the subcategory detection unit 31i
^{Learns the} learning pattern S ^uv . In this case, the subcategory detection unit 31i calculates the weighted sum, and the activity value uij of the cell j of the subcategory detection unit 31i is

[0125]

[Equation 42]

[0126] Here, hij is a variable threshold value of the cell j of the subcategory detection unit 31i. Based on this activation value, the output value oij of the cell j of the subcategory detection unit 31i is

[0127]

[Equation 43]

It becomes: ci is a positive constant of 1 or less. From this, it can be seen that in the subcategory detection unit 31i, a cell having an activity value of ci times or more the maximum activity value outputs 1. When the output value of the cell is determined, the subcategory detection weight Cijk (j = 1 to Ni, k = of the subcategory detection unit 31i
1 to D) are updated according to the following weight update formula.

[0129]

[Equation 44]

Here, Ai is a sufficiently small positive constant.
From this, it can be seen that the sub-category detection unit 31i updates only the weight that is combined with the cell that outputs 1. Further, the cell threshold of the subcategory detection unit 31i is updated based on the following formula.

[0131]

[Equation 45]

Here, A2i is a sufficiently small positive real number of 1 or less. This kind of learning is performed with P1 × P2 learning patterns S ^uv
By sufficiently repeating (u = 1 to P1, v = 1 to P2), the learning of the subcategory detection unit 31i ends.

The operation in the recall mode of the simplified hierarchical pattern classification device having the weight set in this way will be described below.

In the recollection mode, the external input signal S is input to the category detection unit 31, and (Equation 37) and (Equation 3)
The output value of the cell of the category detection unit 31 is calculated from 8). Here, it is assumed that the cell i of the category detection unit 31 outputs 1.

In this case, the subcategory detecting unit 31i becomes operable, and the external input signal S is input to the subcategory detecting unit 31i. In the subcategory detection unit 31i, the output value of the cell of the subcategory detection unit 31i is calculated from (Equation 42) and (Equation 43), and one cell of the subcategory detection unit 31i outputs 1.

From this fact, which category the external input signal S belongs to in the category detecting section 31, that is, the first hierarchical level, and which subcategory detecting section 31i, that is, the subcategory in the second hierarchical level, belong to. Can be judged.

If the cell j of the subcategory detecting section 31i outputs 1 when the external input signal S is input, the output signal generating section 33 outputs the output signal R

[0138]

[Equation 46]

Is output. In this way, the simplified hierarchical pattern classifier has a learning pattern S ^uv similar to the external input signal S.
Can also be recalled.

Further, by expanding the function of the category detection unit 31, it is possible to speed up learning and retrieval and reduce memory. For example, the learning pattern sub-set i that gives the output 1 to the cell i is stored in each cell i of the category detection unit 31. When the number of elements of this learning pattern sub-set i is 0, it is realized by deleting the sub-category detecting unit 31i related to the cell i of the category detecting unit 31.

Further, the D-dimensional learning pattern S ^uv has T partial learning patterns S ^uv j (j = 1 to T) (partial learning pattern
Let the dimension of S ^uv j be D j. ), The dimension of the learning pattern S ^uv can be reduced by expanding the function of the category detection unit 31. For example, the learning pattern sub-set i that gives the output 1 to the cell i is stored in each cell i of the category detection unit 31. When a partial learning pattern of all learning patterns included in this learning pattern sub-set i is equal to all learning patterns included in the learning pattern sub-set i, the partial learning pattern is deleted from each learning pattern, and the partial learning pattern is deleted. The pattern is detected by the sub-category detection unit 31 after the category detection unit 31
It is possible to reduce the dimension of the learning pattern by making i learn.

Similarly, if the function of the subcategory detecting unit 12i is expanded as described above, the dimension of the learning pattern can be further reduced.

In this embodiment, the number of layers of the simplified hierarchical pattern classification device is set to 2 for simplicity. However, this limitation is not an essential problem, and the simplified hierarchical pattern classification having three or more hierarchical numbers. It can be easily extended to the device. FIG. 4 shows a simplified diagram (excluding the output generation unit and the initial weight setting unit) of the simplified hierarchical pattern classification device when the number of layers is three. In FIG. 4, 41 is the same as the category detection unit 31, 41i (i = 1 to N) is the same as the subcategory detection unit 31i, and 41ij (j = 1 to Ni) is the same as the subcategory detection unit 31i. It is a sub-2 category detection unit having a function. In this way, it is possible to easily extend to a simplified hierarchical pattern classification device having three or more hierarchical layers.

As described above, according to the present embodiment, the hierarchical structure between the stored items (learning patterns) is generated and stored based on the similarities between the stored items (learning patterns), and the hierarchical structure is used. Then, the memory items can be searched systematically, efficiently and flexibly according to the purpose. Also, by displaying the generated hierarchical structure, the user can easily understand the relationship between the stored items. Further, compared to the hierarchical pattern classifying apparatus of the first embodiment, the hierarchical structure can be generated and the learning pattern can be searched at a higher speed, and the memory can be implemented with a smaller memory.

(Embodiment 3) A third embodiment of the present invention (prioritized simplified hierarchical pattern classification apparatus) will be described below with reference to the drawings.

In FIG. 5, reference numeral 51 controls the category detection section priority signal pi (i = 1 to T) and the subcategory detection section priority signal pij (i = 1 to N, j = 1 to T). A priority control unit 52 includes an external input signal S composed of T attribute signals Si (i = 1 to T) and category detection weights Cijk (i = 1 to N, j =) held therein. 1 to T, k = 1 to Dj) N number of cells i (i = 1, i) having a weighted sum as an active value
˜N), and at least one or more cells output 1 as a category detection unit, 52i (i = 1 to N)
Becomes operable only when the cell i of the category detection unit 52 outputs 1, and the external category signal S and the sub category detection weight Cijkl (j = 1 to Ni, k = 1) held therein.
Up to T, l = 1 to Dk) N with active weighted sum
The sub-category detection unit is composed of i cells and at least one cell outputs 1 and 53 is a category detection weight Cijk (i = 1 to N, of the category detection unit 52).
(j = 1 to T, k = 1 to Dj) is initialized by giving a sufficiently small random number, and the category detection weights Cijk (i = 1 to N, j = 1 to T, of the category detection unit 52 after learning). Considering k = 1 to Dj), the subcategory detection weight Cijkl (j = 1 to Ni, k = 1 to T, l = 1 to
Dk) is an initial weight setting unit for initializing the sub category detection weight Cijkl (j = 1 to Ni, k = 1 to T, l = 1 to 54) of the sub category detection unit 52i (i = 1 to N). Dk) is an output signal generation unit that generates an output signal R.

In the present embodiment, the number of layers of the prioritized simplified hierarchical pattern classification device is described as 2 for simplification, but this limitation is not an essential problem and the number of layers is 3 or more. It can be easily extended to a simplified hierarchical pattern classification device with priorities (see Example 2).
It is also assumed that the external input signal S is composed of T attribute signals Si (i = 1 to T), and each learning pattern S ^uv is similarly T attribute signals S ^uv i (i = 1 to T). T) (that is, can be divided into partial learning patterns). Further, the dimension of the attribute signal Si is represented by Di. Accordingly, the dimension D of the external input signal S can be expressed as ΣDi.

The operation of the two-layered prioritized simplified hierarchical pattern classification apparatus configured as described above will be described.

First, the priority control section 51 receives the category detection section priority signal pi (i = 1 to T) from the user (pi ≧
0) is stored, and the priority signal pi for category detection unit pi (i = 1
~ T) is

[0150]

[Equation 47]

It is updated according to. Next, the operation in the learning mode will be described. In the learning mode, the learning of the category detection unit 52 starts.

First, the category detection weight Cijk (i = 1 to N, j = 1 to T, k = 1 to Dj) of the category detection unit 52 is initialized to a sufficiently small random number by the initial weight setting unit 53.
Also, P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2)
The category detection weight Cijk (i = 1 to N, j = of the category detection unit 52 by adding a sufficiently small random number to the average pattern of
By initializing 1 to T, k = 1 to Dj), the learning speed and the learning result can be improved. Next, P1 × P2 learning patterns S ^uv (u = 1 to P1, v = 1 to P2) are sequentially input to the category detection unit 52, and learning is sequentially performed. Consider a case where the learning pattern S ^uv is input to the category detection unit 52. In this case, the category detector 52 calculates the weighted sum, and the activity value ui of the cell i of the category detector 52 is

[0153]

[Equation 48]

It becomes: Where hi is the category detection unit 5
It is a variable threshold value of cell i of 2. Based on this activation value, the output value oi of the cell i of the category detection unit 52 is

[0155]

[Equation 49]

[0156] Here, c is a positive constant of 1 or less. From this, it can be seen that in the category detection unit 52, the cell having the maximum active value and the cell having the active value equivalent to that output 1. When the output value of the cell is determined,
Category detection weight of the category detection unit 52 Cijk (i = 1
~ N, j = 1 to T, k = 1 to Dj) are updated according to the following weight update formulas.

[0157]

[Equation 50]

Here, A is a sufficiently small positive real number.
From this, it can be seen that the category detection unit 52 updates only the weight that is combined with the cell that outputs 1. Further, it can be seen that the attribute signal S ^uv j having a larger category detection unit priority signal p j (j = 1 to T) is preferentially learned. Further, the cell threshold of the category detection unit 52 is updated based on the following formula.

[0159]

[Equation 51]

Here, A1 is a sufficiently small positive real number of 1 or less. This kind of learning is performed with P1 × P2 learning patterns S ^uv
By sufficiently repeating (u = 1 to P1, v = 1 to P2), the learning of the category detection unit 52 ends.

When the learning of the category detector 52 is completed, the priority controller 51 sets the subcategory detector priority signal pij (i = 1 to N, j = 1 to T) as follows. Consider setting the priority signal pij (j = 1 to T) for the subcategory detecting unit of the subcategory detecting unit 52i. If the priority signal pj for category detection section is pj <1,

[0162]

[Equation 52]

Set as follows. If the priority signal pj for the category detection unit is pj = 1, the priority control unit 51 creates a set SETi of learning patterns that gives an output 1 to the cell i of the category detection unit 52, and all the sets included in the set SETi. If the attribute signals S ^uv j of the learning pattern S ^uv are equal,

[0164]

[Equation 53]

Otherwise,

[0166]

[Equation 54]

Is set. Next, according to the following formula,
The subcategory detection unit priority signal pij (j = 1 to T) is normalized.

[0168]

[Equation 55]

After the sub-category detecting unit priority signal pij (i = 1 to N, j = 1 to T) is set in this way, the sub-category detecting unit 52i (i = 1 to N) is set. Learning is done.

First, the subcategory detection weight Cijkl (j = 1 to Ni, k = 1 to T, l = 1 to D of the subcategory detection unit 52i).
k) is initialized by the initial weight setting unit 53 as follows.

[0171]

[Equation 56]

Here, α is a sufficiently small random number. Accordingly, the subcategory detection weight Cijkl (j = 1 to Ni, k = 1 to T, l = 1 to Dk) of the subcategory detection unit 52i is the category detection weight Cijk (i = 1 to N, of the category detection unit 52). k =
The distribution is centered on 1 to T, l = 1 to Dk).

When the initial setting of the subcategory detection weight Cijkl (j = 1 to Ni, k = 1 to T, l = 1 to Dk) of the subcategory detection unit 52i is completed, P1 × P2 learning patterns S ^uv ( u = 1 ~
Learning of P1, v = 1 to P2) is performed. Where the learning pattern
Think about learning S ^uv . In this case, first, the learning pattern S ^uv is input to the category detection unit 52. The category detection unit 52 calculates the output value of the cell based on (Equation 48) and (Equation 49). Where the learning pattern
A cell that outputs 1 when S ^uv is input to the category detection unit 52 is set as a cell i. As a result, the sub-category detection unit 52i becomes operable, and only the sub-category detection unit 52i learns the learning pattern S ^uv . In this case, the weighted sum is calculated in the subcategory detecting unit 52i, and the activation value of the cell j of the subcategory detecting unit 52i is calculated.
uij

[0174]

[Equation 57]

It becomes: Here, hij is a variable threshold value of the cell j of the subcategory detection unit 52i. Based on this active value, the output value oij of the cell j of the subcategory detection unit 52i is

[0176]

[Equation 58]

It becomes: ci is a positive constant of 1 or less. From this, it is understood that in the subcategory detection unit 52i, 1 is output to the cell having the maximum active value and the cell having the active value equivalent to the maximum active value. When the output value of the cell is determined, the subcategory detection weight Cijkl (j = 1 to Ni, k = 1 to T, l = 1 to Dk) of the subcategory detection unit 52i is updated according to the following weight update formula.

[0178]

[Equation 59]

Here, Ai is a sufficiently small positive constant.
From this, it can be seen that the sub-category detection unit 52i updates only the weight that is combined with the cell that outputs 1. Furthermore, the priority signal pik for the sub-category detector
It can be seen that the attribute signal S ^uv k having a larger (k = 1 to T) is prioritized for learning. Also, the subcategory detection unit 52i
The cell threshold of is updated based on the following formula.

[0180]

[Equation 60]

Here, A2i is a sufficiently small positive real number of 1 or less. This kind of learning is performed with P1 × P2 learning patterns S ^uv
By sufficiently repeating (u = 1 to P1, v = 1 to P2), the learning of the subcategory detection unit 52i ends.

The operation in the recall mode of the priority-ordered simplified hierarchical pattern classification device having the weight thus set will be described below.

In the recall mode, the external input signal S is input to the category detection unit 52, and (Equation 48) and (Equation 4)
The output value of the cell of the category detection unit 52 is calculated from 9). Here, it is assumed that the cell i of the category detection unit 52 outputs 1.

In this case, the subcategory detecting section 52i becomes operable, and the external input signal S is input to the subcategory detecting section 52i. In the subcategory detection unit 52i, the output value of the cell of the subcategory detection unit 52i is calculated from (Equation 57) and (Equation 58), and one cell of the subcategory detection unit 52i outputs 1.

From this fact, which category the external input signal S belongs to in the category detecting section 52, that is, the first hierarchical level, and which subcategory detecting section 52i, that is, which subcategory in the second hierarchical level, belongs. Can be judged.

When the cell j of the subcategory detecting section 52i outputs 1 when the external input signal S is input, the output signal generating section 54 outputs the output signal R

[0187]

[Equation 61]

Is output. In this way, the prioritized simplified hierarchical pattern classifier can also recall a learning pattern S ^uv similar to the external input signal S.

Further, by expanding the function of the category detection unit 52, learning and retrieval can be speeded up and memory can be reduced. For example, the learning pattern sub-set i that gives the output 1 to the cell i is stored in each cell i of the category detection unit 52. When the number of elements of this learning pattern sub-set i is 0, it can be realized by deleting the sub-category detecting unit 52i related to the cell i of the category detecting unit 52.

Furthermore, a priority signal for category detection section
pj (j = 1 to T) and priority signal pi for subcategory detector
From k (i = 1 to N, k = 1 to T), the dimension of the learning pattern S ^uv can be reduced as described below, and learning and search can be speeded up and memory can be reduced.

When the category detection section priority signal pj is 0, the attribute signal Sj of the external input signal S does not affect the operation of the category detection section 52 in the learning mode and the recall mode (Equation 48). Recognize. Similarly, when the subcategory detection unit priority signal pik of the subcategory detection unit 52i (i = 1 to N) is 0, the external input signal S
It can be seen from (Equation 57) that the attribute signal Sk of does not affect the operation of the subcategory detection unit 52i in the learning mode and the recall mode.

From this, when the category detection section priority signal pj is 0, the attribute signal Sj of the external input signal S
Detection weight of the category detection unit 52 to be combined with
Cijk (i = 1 to N, k = 1 to Dj) can be removed, and learning of the category detection unit 52, speedup of search, and reduction of memory can be realized. Similarly, the subcategory detection unit 5
When the priority signal pik for the sub-category detecting unit of 2i (i = 1 to N) is 0, the sub-category detecting weight Cijkl (j = 1 of the sub-category detecting unit 52i that is combined with the attribute signal Sk of the external input signal S). .About.Ni, l = 1 to Dk) can be removed, and learning of the subcategory detection unit 52i and search can be speeded up and memory can be reduced.

As described above, according to this embodiment, a hierarchical structure is generated and stored between the stored items (learning patterns) based on the similarities between the given stored items (learning patterns) and the priority signal from the user. However, by utilizing the hierarchical structure, it is possible to systematically, efficiently, and flexibly search the memory items according to the purpose. Also, by displaying the generated hierarchical structure, the user can easily understand the relationship between the stored items. Further, compared with the hierarchical pattern classification device of the first embodiment, the hierarchical structure can be generated and the learning pattern can be searched at a higher speed, and the device can be implemented with less memory.

[0194]

As described above, according to the present invention, the operation is defined by the concept-forming associative memory weight Wij (i = 1 to D, j = 1 to D) held inside, and the external input signal S is triggered. The concept formation associative memory signal X (t) is updated as, and when X (t-1) and X (t) become equal, the stable concept formation associative memory signal X (= X (t-1) = X
(t)), and the stable concept formation associative storage signal X output by the concept formation associative storage unit 11.
And the category detection weights Cij (i = 1 ~
N, j = 1 to D) and a category detection unit 12 that is composed of N cells i (i = 1 to N) having a weighted sum as an active value, and at least one cell outputs 1 , The operation is enabled only when the cell i of the category detection unit 12 outputs 1, and the operation is defined by the concept inheritance associative memory weight Wijk (j = 1 to D, k = 1 to D) held therein. Then, the concept inherited associative memory signal Xi (t) is updated by using the stable externally formed signal S and the stable concept formation associative memory signal X output from the concept formation associative memory unit 11 as a trigger, and Xi (t-1) and Xi (t) are updated. When are equal, the stable concept inherited associative memory signal Xi (= Xi (t-1) = X
concept inheritance associative storage unit 13i (i = 1 to
N) and the stable concept inheritance associative memory signal Xi output by the concept inheritance associative memory unit 13i and the subcategory detection weights Cijk (j = 1 to Ni, k = 1 to D) held therein. Ni cells i having i as an activation value (i = 1 to Ni)
A sub-category detection unit 14i (i = 1 to N) in which at least one cell outputs 1 and a stable concept formation associated memory signal output by the concept formation associated memory unit 11.
By providing X and the output signal generation unit 15 that receives the stable concept inheritance associative memory signal Xi output from the concept inheritance associative memory unit 13i and generates the output signal R, inheritance patterns belonging to the same category are the same concept inheritance memory unit. Since it is stored in
It is possible to obtain information regarding the classification of patterns such as which category the input pattern belongs to. That is, based on the similarities between given items to be stored (learning patterns), a hierarchical structure is generated and stored between these items, and the hierarchical items are used to systematically and flexibly store the stored items. You can search. Also, by displaying the generated hierarchical structure, the user can easily understand the relationship between the stored items.

[Brief description of drawings]

FIG. 1 is a block connection diagram of a hierarchical pattern classification device (two layers) according to a first embodiment of the present invention.

FIG. 2 is a block connection diagram of a hierarchical pattern classification device (three layers) according to the first embodiment of the present invention.

FIG. 3 is a block connection diagram of a simplified hierarchical pattern classification device (two layers) according to a second embodiment of the present invention.

FIG. 4 is a block connection diagram of a simplified hierarchical pattern classification device (three layers) according to a second embodiment of the present invention.

FIG. 5 is a block connection diagram of a prioritized simplified hierarchical pattern classification device according to a third embodiment of the present invention.

FIG. 6 is a block connection diagram of a conventional hierarchical concept inheritance associative memory device.

[Explanation of symbols]

11 Concept formation associative memory 12 Category detection section 13i Concept inheritance associative memory 14i Subcategory detection unit 15 Output signal generator 21 Module A 22i module B 23ij module C 31 Category detection section 31i Subcategory detection unit 32 initial weight setting section 33 Output signal generator 41 Category Detector 42i Subcategory detection unit 43ij Sub 2 category detector 51 Priority control unit 52 Category detection section 52i Subcategory detection unit 53 initial weight setting section 54 Output signal generator 61 Concept formation associative memory 62 Concept Succession Associative Memory 63 Output signal generator

Continuation of the front page (56) References JP-A-6-12396 (JP, A) JP-A-4-31979 (JP, A) JP-A-8-115386 (JP, A) Akiyoshi Ito, et al., "Hierarchical Dictionary Construction Method for Registering Character Types in Multiple Clusters in Printed Kanji Recognition System ", IEICE Transactions, Japan, The Institute of Electronics, Information and Communication Engineers, June 25, 1995, Vol. J78-D-II, No. 6, pp. 896-905 Yoshihiro Kojima et al., "Multi-font kanji recognition by structured neural network (NARA)", IEICE technical report, Japan, The Institute of Electronics, Information and Communication Engineers, September 20, 1991, Vol. 91, No. 231 (NC91-30 to 47), pp. 9-16 Makoto Hirahara, et al., "Cascade associative memory model for storing patterns with hierarchical correlation", IEICE Technical Report, Japan, The Institute of Electronics, Information and Communication Engineers, December 16, 1994. Sun, Vol. 94, No. 417 (NC94-51 to 59), pp. 57-64 M.I. V. TSODYKS, "Hier archical associative ve memory in neural network network with lo w activity level" Moden Physics Letters B, February 1990, No. Vol. 4, pp. 259-256, JST Material No .: T0431A, ISSN: 0217-9849 J. P. Sutton, et. a. , "Hierarchical model of memory and memory loss", Journal of Physics A, December 1988, Vol. 21, No. 23, pp. 4443-4454, JST Material No .: D 0390B (58) Fields investigated (Int.Cl. ⁷ , DB name) G06N 1/00-7/08 G06G 7/60 G06F 17/30 JST file (JOIS) CSDB (Japan (National Patent Office)

Claims (5)

(57) [Claims] 1. An operation is defined by a concept-forming associative memory weight held inside, a concept-forming associative memory signal is updated by using an external input signal as a trigger, and a stable concept-forming associative memory signal is stabilized. From the N (N is an integer of 1 or more) cells having an active value of the concept formation associative memory unit that outputs as a signal, and the weighted sum of the stable concept formation associative memory signal and the category detection weight held inside An operation is defined by a first hierarchy having a category detection unit configured to output 1 from at least one cell, and an operation is defined by a concept inheritance associative memory weight internally held, and an external input signal and the stable concept formation. The concept inherited associative memory signal is updated by using the associative memory signal as a trigger, and the stable concept inherited associative memory signal is output as a stable concept inherited associative memory signal N N number of concept inheritance associative memory units, and Ni having Ni as the active value (Ni is the sum of weighted sums of the stable concept inheritance associative memory signals output from each concept inheritance associative memory unit and the internally stored subcategory detection weights). A second hierarchy having N sub-category detectors each having at least one or more cells outputting 1 and at least one or more cells, the stable concept forming associative memory signal, and the stable concept inheritance. An output signal generation unit that receives an associative memory signal and generates an output signal. 2. The hierarchical pattern classification device according to claim 1, wherein the number of layers is three or more. 3. Each cell i (i is an integer of 1 or more and Ni or less) of the category detection unit stores a learning pattern which gives an output 1 to the cell i as a learning pattern sub-set i, and the number of elements of the learning pattern sub-set i 3. If the value is 0, the hierarchical pattern classifying apparatus according to claim 1 or 2, wherein the concept inheritance associative storage unit and the subcategory detection unit related to the cell i of the category detection unit are deleted. 4. Each cell i (i is an integer of 1 or more and Ni or less) of the category detection unit stores a learning pattern that gives an output 1 to the cell i as a learning pattern sub-set i, and all the learning patterns in the learning pattern sub-set i are stored. 4. The hierarchical pattern classification according to claim 1, wherein a partial learning pattern common to the learning patterns is detected, and the partial learning patterns are deleted from all the learning patterns in the learning pattern sub-set i. apparatus. 5. Each cell i (i is an integer of 1 or more and Ni or less) of the subcategory detection unit stores a learning pattern that gives an output 1 to the cell i as a learning pattern subset i, and 5. The hierarchical pattern according to claim 1, wherein a partial learning pattern common to all the learning patterns is detected and the partial learning pattern is deleted from all the learning patterns in the learning pattern sub-set i. Classifier.