JPH10187651A - Learning type recognition and judgement device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To shorten the learning time and to improve the recognition performance of the learning type recognition and judgement device which learns and recognizes inputted pattern data by automatically removing learning patterns which are not necessary for learning. SOLUTION: For example, a group reversion degree selected by a 2nd group selection part 1007 is large, in-group similarity which corresponds to the group reversion degree and is calculated by a similarity calculation part 1004 is large, and although it is decided that the reliability of a recognition result is very high, the recognition result and a tutor signal are different, namely, the tutor signal is misadded or a signal which is very difficult to discriminate is inputted. A 1st learning decision part 1019 automatically detects the said state and makes a learning control part 1013 make no load change. Consequently, the learning time is shortened and the recognition performance is prevented from decreasing.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning-type recognition / judgment device for learning and recognizing input pattern data.

[0002]

2. Description of the Related Art Conventionally, input pattern data is first subjected to a large classification to select a category group to which the input data belongs, and then to a sub-classification in the selected category group, thereby performing a classification of the input data. An example of a learning-type recognition / judgment device that performs recognition is described in, for example, IEICE Transactions D-II Vol.J75-D-II No.3 pp 545-553 “Large-Scale Neural Network“ CombNET-II ”” It is shown.

FIG. 10 is a block diagram showing the configuration of a conventional learning type recognition / judgment device. Reference numeral 101 denotes a large classification unit which classifies an input pattern signal into various categories. Reference numeral 102 denotes a sub-classification unit, which sub-classifies the input pattern signal in each category group. A group selection unit 103 selects a plurality of category groups from the output values of the large classification unit 101 (hereinafter, referred to as “fitness”).
104, a sub-classification unit input signal selection unit;
The sub-classification unit 102 for inputting an input pattern signal is selected based on the group selection information obtained in Step 3. 105
Denotes an identification unit that identifies an input pattern signal based on the degree of conformity of the category group selected by the group selection unit 103 and the output value of the subclassification unit 102.

In the large classification unit 101, an input unit 106 inputs an input pattern signal. 10
A multi-input / one-output signal processing unit 7 calculates the degree of conformity of each category group to the input pattern signal.

[0005] In the sub-classification unit 102, an input unit 108 inputs an input pattern signal output from the sub-classification unit input signal selection unit 104. A multi-input / one-output signal processing unit 109 is connected to the lower-layer input unit 108 or the multi-input / one-output signal processing unit 109.
Is multiplied by a weighting factor which is a degree of the connection, and the sum is summed and output. Here, the plurality of multi-input / one-output signal processing units have a layered structure and are connected to a network such that there is no mutual connection in each layer and the signal propagates only to the upper layer. The degree of similarity to each category in is determined. Reference numeral 110 denotes a maximum value selection unit which selects the maximum value from the outputs of the plurality of multi-input / output signal processing units in the uppermost layer.

[0006] In the identification unit 105, a similarity calculation unit 111 is used to determine the similarity of each category from the fitness of the category group selected by the group selection unit 103 and the output value of the subclassification unit 102 corresponding to the category group. Calculate the degree.
A category identification unit 112 identifies the input pattern signal by obtaining the maximum value of the similarity of each category obtained from the similarity calculation unit 111.

The operation of the conventional learning type recognition / judgment device configured as described above will be described below. An input pattern signal X composed of n pieces of feature data of a recognition target is
First, it is input to the input unit 106 of the large classification unit 101. _{X = (x 1, x 2} , ..., x n) ··· (1) Input unit 106 is equal to n pieces prepared wherein the number of data of the pattern data, each feature data x _i is input to the corresponding It is input to 106. Each multi-input first output signal processing section 107 of the major classification section 101, the weighting coefficient _{v ij (1 ≦ i ≦ m} r is the degree of coupling between the input x _j of the input section 101 connected thereto; the m _r Number of category groups, 1 ≦ j
.Ltoreq.n), calculate the sum, and then calculate the sum of the input pattern signal X and each of the multi-input / one-output signal processing units 1.
07 of the weighting coefficient norm of the vector _{V i | X |, | V} i |
The result of dividing by the product of is output. V _i = (v _i1 , v _i2 ,..., V _in ) (2) That is, the output value sim (X, V _i ) of the multi-input one-output signal processing unit 107 having the weight coefficient vector V _i is: _{sim (X, V i) =} (X · V i) / (| X || V i |) ··· (3) _{However, X · V i = Σ j} (x j × v ij) | X | = (Σx _j ² ) ^0.5 | V _i | = (Σv _ij ² ) ^0.5

Note that the weight coefficient vector V _i is
It is designed in advance so that a multi-input one-output signal processing unit determined for similar input pattern signals generates the maximum output.

According to the prior art, these weight coefficient vectors V _i are designed by the following method. First, in the first step, the input pattern signal X for weight coefficient vector design
Is input, V _C having the largest sim (X, V _i ) is obtained (at this time, X is optimally matched to V _C ).
Move _C closer to X. When an input pattern signal that optimally matches one weight coefficient vector becomes equal to or more than a certain number, the area assigned to the weight coefficient vector is divided into two. In the second step, V _i that optimally matches all input pattern signals for weight coefficient vector design is obtained,
Find out if it has changed from the previous time. And, if there is a change, to correct the V _i. At this time, similarly to the first process, the division of the weight coefficient vector is also performed. This is repeated until the correction and division of the weight coefficient vector are eliminated.

[0010] By designing the weight coefficient vector in this manner, the input pattern signals are roughly classified into a plurality of category groups. Then, the output value of each multi-input / one-output signal processing unit 107 is output to the group selection unit 103 as the degree of conformity of each category group to the input pattern signal X.

The group selecting unit 103 selects an arbitrary number of category groups in descending order of the degree of matching obtained by the large classifying unit 101.
It outputs group selection information indicating which category group has been selected and the degree of matching corresponding thereto. Based on the group selection information obtained from group selection section 103, sub-classified input signal selection section 10
4 selects the sub-classification unit 102 for inputting the input pattern signal, and outputs the input pattern signal to these sub-classification units 102.

In each of the sub-classification units 102 corresponding to the category group selected by the group selection unit 103 (that is, the sub-classification unit 102 to which the input pattern signal is input from the sub-classification input signal selection unit 104), The input pattern signal X is input to the unit 108. The input unit 108 is equal to the N prepared wherein the number of data of the pattern signal, each feature data x _i is inputted to the input unit 108, respectively corresponding.
Each multi-input one-output signal processing unit 109 of the sub-classification unit 102
The output of the lower-layer input unit 108 or the multi-input / one-output signal processing unit 109 connected thereto is multiplied by a weighting factor indicating the degree of connection thereof, and the sum is converted by a threshold function. Is output to the upper layer. Here, the multi-input one-output signal processing unit 10 in the uppermost layer of each sub-classification unit 102
9 is set to the same number as the number of categories of the pattern data included in each category group, and each multi-input / one-output signal processing unit 109 in the uppermost layer corresponds to each of these categories. The maximum value selecting section 80 selects the largest one of the output values of each multi-input one-output signal processing section 109 of the uppermost layer, the category corresponding to the multi-input one-output signal processing section 109, and the maximum output value Is output.

The weight coefficient of each multi-input / one-output signal processing unit 109 is based on the multi-input / one-output signal processing of the uppermost layer corresponding to each category for the input pattern signal having each category in the category group. The unit 109 has been learned in advance so as to generate the maximum output.

More specifically, such a weight coefficient learning method is performed by a learning algorithm called an error back propagation method. For the error backpropagation method, for example, DE Ru
"L" by melhart, GE Hinton and RJ Williams
earning Representations byBack-Propagating Error
s, "Nature, vol. 323, pp. 533-536, Oct. 9, 1986.

Hereinafter, an outline of the error back propagation method will be described. First, a pattern signal X for weight coefficient learning is input to the input unit 108 of the fine classification unit 102. As described above, each multi-input / one-output signal processing unit 109 includes a lower-layer input unit 108 connected thereto, or an output of the multi-input / one-output signal processing unit 109 and a weight coefficient indicating a degree of connection thereof. Is converted by a threshold function, and the value is output to the upper layer. Here, the error E with all of the output signal o _k and the desired output signal t _k multiple-input first output signal processing section 109 of the top layer (referred to as a teacher signal) is obtained as equation (4) . _{E = 0. 5 Σ p Σ} k (t k -o k) 2 ··· (4) However, sigma _p is the total sum related to the number of patterns of the teacher signal. The purpose of the learning is to determine the value of the weighting coefficient that minimizes the error E, and the value of 10
The change amount △ w _ij of the weighting coefficient of No. 9 is calculated based on Expression (5). Δw _ij = −ε∂E / ∂w _ij (5) where ε is a positive constant called a learning rate. By repeating the updating of the weight coefficient based on the equation (5) every time a learning pattern signal is input,
The error E can be reduced. When the error E becomes sufficiently small, the learning is terminated on the assumption that the output signal has become sufficiently close to the desired value.

According to such a weight coefficient learning method,
With respect to an input pattern signal having each category in the category group, the multi-input / one-output signal processing unit 109 of the highest layer corresponding to each category can generate the maximum output. Therefore, by selecting the one that generates the maximum output from the plurality of multi-input one-output signal processing units 109 in the highest layer by the maximum value selection unit 80, each of the sub-classification units Can recognize the category of the input pattern signal.

In the identification section 105, first, the similarity calculation section 1
In 11, the degree of conformity of the category group selected by the group selection unit 103 and the sub-classification unit 102 corresponding to the category group
The similarity of each category obtained by the sub-classification unit 102 is calculated from the output values of the sub-categories using Expression (6), and these similarities are output to the category identification unit 112. (Similarity) = (fitness) ^a (output value) ^b (6) where a and b are real constants.

Finally, the category identifying section 112 compares the similarities of the respective categories obtained from the similarity calculating section 111, and outputs a category corresponding to the maximum similarity among the categories as an identification result. However, the error back-propagation method requires a teacher signal to determine a category into which a pattern signal for learning is classified.
Humans give by hand. Therefore, there are cases where the teacher signal has an error, and learning such a pattern signal leads to an increase in the learning time or a result that the learning does not converge, which eventually leads to a reduction in recognition performance. This situation occurs not only when there is an error in the teacher signal, but also when learning a pattern signal in which personal characteristics appear remarkably, such as handwritten characters, to put very similar pattern signals into completely different categories. It also occurs when classifying or completely different pattern signals are classified into the same category.

In order to eliminate such a problem, Japanese Patent Laid-Open No. 2-235170 discloses a method in which only a pattern signal effective for learning is selected and used in advance by a statistical method. However, no consideration is given to a method of automatically removing a learning pattern signal that adversely affects learning and recognition performance during learning.

[0020]

However, in the conventional learning-type recognition / judgment device, it is necessary to provide a teacher signal for deciding into which category a pattern signal for learning is to be classified. Give in. Therefore, there are cases where the teacher signal has an error, and learning such a pattern signal leads to an increase in the learning time or a result that the learning does not converge, which eventually leads to a reduction in recognition performance. This situation occurs not only when there is an error in the teacher signal, but also when learning a pattern signal in which personal characteristics appear remarkably, such as handwritten characters, to put very similar pattern signals into completely different categories. It also occurs when classifying or completely different pattern signals are classified into the same category.

The present invention solves such a conventional problem, and automatically removes a learning pattern unsuitable for learning by using information such as a group membership degree and a category similarity degree, thereby shortening learning time and reducing recognition performance. It is an object of the present invention to provide a learning-type recognition / judgment device capable of improving the learning.

[0022]

To achieve the above object, a first aspect of the present invention is a group in which a plurality of group reference pattern signals representing a group of categories consisting of a set of similar patterns are stored. A dictionary, a fuzzy large classification unit that calculates a group belonging degree, which is a degree to which the input pattern signal belongs to each category group, using the group reference pattern signal; and Load change for controlling a bit shift amount of a transmission amount changing unit of a similarity calculation unit described later according to an output of a similarity calculation unit for calculating an intra-group similarity which is a similarity degree and a second group selection unit described later. A plurality of sub-classification units comprising a quantity control unit, a first group selection unit for selecting at least one category group from the group membership degree, and the intra-group similarity obtained by the sub-classification unit. Group membership An identification signal weighting unit, a category similarity calculation unit that calculates a category similarity that is a degree belonging to each category from an output of the identification signal weighting unit, and at least a larger one of the calculated category similarities. A category selection unit that selects two category similarities and the category; a rejection threshold value determination unit that determines a rejection threshold value based on a group belonging degree selected by the first group selection unit and an output of the category selection unit Department and
A reject threshold control unit that controls a reject determination unit according to a reject threshold value determined by the reject threshold determination unit; and a reject threshold value according to an output of the reject threshold control unit. A reject determination unit that outputs a recognition result of an input pattern signal or a reject signal indicating that recognition is not possible according to a comparison result with a ratio of two category similarities output from the category selection unit from the larger one, A first teacher signal generator for generating a teacher signal necessary for learning of the sub-classifier, a category information storage unit storing category information belonging to each sub-class, and a category of the input pattern exists according to the category information. A second group selecting unit that selects at least one group belonging degree and a category group from a group having a large group belonging degree in the sub-classifying unit; It is determined whether or not learning of the input pattern signal is possible based on the group belonging degree selected by the selecting unit and the intra-group similarity calculated by the similarity calculating unit. If it is determined that the input pattern signal is inappropriate for learning, a learning suppression signal is output. A learning determination unit, a comparison result of a category selected by the category selection unit with a category of an input pattern output from the first teacher signal generation unit, and learning output from the first learning determination unit. A learning control unit that controls learning of the sub-classification unit according to a suppression signal, wherein the similarity calculation unit includes a plurality of layers, and a lowest layer includes a plurality of third units. The recognition unit, the lowermost layer is constituted by a plurality of second unit recognition units, and the other layers are constituted by a plurality of first unit recognition units, wherein the first unit recognition unit comprises:
A signal input unit, a quantizer that performs quantization in accordance with an output of an input signal, an adjacent interval selection unit that selects a quantization interval adjacent to the quantization interval from a quantization result of the input signal, A weight table for storing a quantization section to which the signal belongs and a weight value for a quantization section adjacent to the quantization section; a path input section having a single or a plurality of path input terminals; A route output unit having a route output terminal, and a connection strength between a route input terminal of the route input unit and a route output terminal of the route output unit is changed by changing a setting position of a load applied to a route according to a quantization result. The second unit recognition unit includes a signal input unit, a quantizer that performs quantization in accordance with an output of the input signal, and a single or a plurality of path input terminals. With a path input unit and a single Is a path output unit having a plurality of path output terminals, a transmission amount changing unit that shifts a value transmitted to the path input terminal of the path input unit by n bits, and the path input terminal according to an output of the transmission amount changing unit. And a learning unit for increasing the connection strength between the path output terminal of the path output unit indicated by the output value of the quantizer and the third unit recognition unit. It has an adder for adding, and a threshold processing unit for performing threshold processing on an output signal of the adder. As a result, a learning-type recognition / judgment device that automatically removes a learning pattern unsuitable for learning to reduce learning time and improve recognition performance is obtained.

According to a second aspect of the present invention, a second learning determining unit for outputting a learning suppression signal in accordance with the category similarity output from the category selecting unit is provided in place of the first learning determining unit. This is a learning-type recognition determination device. As a result, a learning-type recognition / judgment device that automatically removes a learning pattern unsuitable for learning to reduce learning time and improve recognition performance is obtained.

According to a third aspect of the present invention, in place of the first learning determination section, the group membership degree and the category group selected by the second group selection section and the category information stored in the category information storage section are stored. Is a learning-type recognition / determination apparatus including a third learning determination unit that outputs a learning inappropriate signal in accordance with the determination. This allows
A learning-type recognition / judgment device that automatically removes a learning pattern unsuitable for learning to reduce learning time and improve recognition performance is obtained.

[0025]

BEST MODE FOR CARRYING OUT THE INVENTION
A group dictionary in which a plurality of group reference pattern signals representing a group of similar pattern groups are stored, and a group membership in which an input pattern signal belongs to each category group using the group reference pattern signal. A fuzzy large classifying unit for calculating a degree, a similarity calculating unit for calculating an intra-group similarity which is a degree of similarity of the input pattern signal to each category included in the category group, and a second group selecting unit to be described later. A plurality of sub-classification units each including a load change amount control unit that controls a bit shift amount of a transmission amount change unit of a similarity calculation unit described later according to an output; and at least one category group is selected from the group membership degrees A first group selection unit, an identification signal weighting unit that weights the intra-group similarity obtained by the sub-classification unit with the selected group membership, and a category based on the output of the identification signal weighting unit. And category similarity calculating section for calculating category similarity is the degree to which they belong the calculated from the direction of the category similarity greater least two categories similarity between the category selection section for selecting the category,
A reject threshold determining unit that determines a reject threshold from the group membership selected by the first group selecting unit and an output of the category selecting unit; and a reject threshold determined by the reject threshold determining unit. A reject threshold control unit that controls a reject determination unit according to a threshold value; and a reject threshold according to an output of the reject threshold control unit and a larger one output from the category selector. A reject determination unit that outputs a recognition result of an input pattern signal or a reject signal indicating that recognition is not possible according to a comparison result with a category similarity ratio, and generates a teacher signal necessary for learning of the fine classification unit. A first teacher signal generation unit, a category information storage unit storing category information belonging to each sub-classification unit, and a category of the input pattern based on the category information. A second group selecting unit that selects at least one group belonging degree and a category group from a group having a large group belonging degree among the sub-classifying units in which the group exists, and a group belonging selected by the second group selecting unit. Determine whether the input pattern signal can be learned from the similarity calculated in the group and the similarity calculated by the similarity calculator,
A first learning determining unit that outputs a learning suppression signal when determining that the learning pattern is inappropriate for learning; a category of the input pattern output from the first teacher signal generating unit and a category selected by the category selecting unit; In a learning-type recognition / determination apparatus including a comparison result and a learning control unit that controls learning of the sub-classification unit according to a learning suppression signal output from the first learning determination unit, the similarity calculation unit includes: The lowermost layer is constituted by a plurality of third unit recognition units, the lowermost layer is constituted by a plurality of second unit recognition units, and the other layers are constituted by a plurality of first unit recognition units. The first unit recognition unit includes a signal input unit, a quantizer that performs quantization in accordance with an output of an input signal, and a quantization interval adjacent to the quantization interval based on a quantization result of the input signal. Adjacent section to select A selection section, a quantization section to which the input signal belongs, and a load table storing weight values for quantization sections adjacent to the quantization section, a path input section having a single or a plurality of path input terminals, A path output section having a single or a plurality of path output terminals, and a path input terminal of the path input section and a path output terminal of the path output section by changing a set position of a load applied to the path according to a quantization result. A second unit recognition unit, a signal input unit, a quantizer that performs quantization in accordance with an output of the input signal, a single unit or a unit. A route input unit having a plurality of route input terminals, a route output unit having a single or a plurality of route output terminals, a transmission amount changing unit for shifting a value transmitted to the route input terminal of the route input unit by n bits; , To the output of the transmission amount changing unit A learning unit for increasing the connection strength between the path input terminal and the path output terminal of the path output unit indicated by the output value of the quantizer. The third unit recognition unit includes a plurality of path input terminals. A learning-type recognition / judgment device comprising: an adder for adding an input signal from a terminal; and a threshold value processing unit for performing a threshold process on an output signal of the adder. In the unit, from the group belonging degree selected in the second group belonging degree and the intra-group similarity calculated by the similarity calculating unit, the reliability of the recognition result is determined to be very high, When the recognition result and the teacher signal are different, that is, when a teacher signal addition error or a very difficult-to-identify signal is input, the detection is automatically detected and the learning control unit performs control so as not to change the weight. By reducing the learning time It has the effect of realizing and preventing a decrease in recognition performance.

According to a second aspect of the present invention, in place of the first learning determination section, a second learning determination section that outputs a learning suppression signal according to a category similarity output from a category selection section is provided. 2. The learning-type recognition / determination apparatus according to claim 1, wherein the second learning determination unit determines that the reliability of the identification result is extremely high based on the category similarity selected by the category selection unit. Nevertheless, if the recognition result is different from the teacher signal, that is, if a teacher signal addition error or a very difficult-to-identify signal is input, it is automatically detected, and the learning controller performs weight change. By controlling so as not to reduce the learning time, the learning time can be shortened and the recognition performance can be prevented from lowering.

According to a third aspect of the present invention, in place of the first learning judging section, the degree of group belonging selected by the second group selecting section and the category information stored in the category information storing section are used. 2. The learning-type recognition / determination apparatus according to claim 1, further comprising a third learning determination unit that outputs a learning inappropriate signal by the third learning determination unit. If there is a contradiction that the category to which the input pattern stored in the category information storage unit does not belong is included, that is, if a teacher signal addition error or a very difficult-to-identify signal is input, large classification is performed. In this case, the automatic detection is performed at the stage, and the control is performed so that the load is not changed in the learning control unit without performing the sub-classification, thereby shortening the learning time and preventing the recognition performance from lowering.

According to a fourth aspect of the present invention, there is provided a learning intensity control unit for changing the learning intensity of the similarity calculation unit in accordance with the output of the second group selection unit, instead of the load change amount control unit. A learning-type recognition / judgment device according to claim 1, which has the same effect as described above.

The invention according to claim 5 of the present invention is characterized in that the learning intensity control unit changes the number of times of learning of the corresponding sub-classification unit according to the output of the second group selection unit. The learning type recognition determination device described above can contribute to shortening of learning time and improvement of recognition performance.

According to a sixth aspect of the present invention, the fuzzy large classification unit calculates a distance between the input pattern signal and all group reference pattern signals stored in the group dictionary. And a plurality of dividers for calculating the reciprocal of the output of the distance calculator, an adder for adding each output of the divider, and multiplying the output of the adder and the output of the distance calculator. 4. The learning-type recognition / determination apparatus according to claim 1, further comprising: a plurality of multipliers; and a plurality of dividers for calculating a reciprocal of an output of the multiplier. Can be reduced and the recognition performance can be improved.

According to a seventh aspect of the present invention, in the third unit, it is desirable that the similarity calculating section outputs the output of the first teacher signal generating section as the largest value among the sub-classifying sections. The output of the teacher signal conversion unit is input to a signal input unit of a second unit recognition unit located in the lowermost layer and a teacher signal conversion unit that converts the number into a recognition unit number. The learning-type recognition / determination apparatus according to claim 1, 2, or 3, which can contribute to shortening of learning time and improvement of recognition performance.

The invention according to claim 8 of the present invention is characterized in that the reject threshold value determining section determines the reject threshold value according to the category selected by the category selecting section. Alternatively, the learning type recognition judging device described in 3 can contribute to shortening of learning time and improvement of recognition performance.

According to a ninth aspect of the present invention, the reject threshold determining unit determines the reject threshold according to the category group selected by the first group selecting unit. The learning-type recognition / judgment device according to item 1, 2, or 3, which can contribute to a reduction in learning time and an improvement in recognition performance.

According to a tenth aspect of the present invention, the reject threshold determining unit determines the reject threshold according to the category selected by the category selecting unit and the category group selected by the first group selecting unit. 4. The learning-type recognition / determination apparatus according to claim 1, wherein the learning time is determined, and the learning time can be reduced and the recognition performance can be improved.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In general, the pattern data input to the learning-type recognition / determination device includes a time-series pattern such as voice, a spatial pattern such as characters and images, and the like.
In the present invention, any pattern data may be used.

(Embodiment 1) FIG. 1 is a block diagram showing a learning type recognition judging device according to a first embodiment of the present invention. In FIG. 1, reference numeral 1001 denotes a group dictionary which stores a plurality of group reference pattern signals representing a category group composed of a set of similar patterns of input pattern signals. Reference numeral 1002 denotes a fuzzy large classification unit which calculates a group belonging degree, which is a degree to which the input pattern signal belongs to each category group, using the group reference pattern signal stored in the group dictionary 1001. 100
Reference numeral 3 denotes a sub-classifying unit, which is a similarity calculating unit 1004 for calculating an intra-group similarity, which is a degree that the input pattern signal is similar to each category included in the category group, and an output of a second group selecting unit described later. The load change amount control unit 1005 controls the load change amount of the similarity calculation unit in accordance with. 1006
Is a first group selection unit, which selects a plurality of category groups based on the degree of group membership. Reference numeral 1007 denotes a second group selection unit. At least one of the sub-classification units in which the category of the input pattern signal exists according to the category information stored in the category information storage unit 1012 described later, in descending order of group membership. This is for selecting the degree of belonging and the group of categories.

Reference numeral 1008 denotes a discrimination signal load unit, which converts the intra-group similarity obtained by the sub-classification unit 1003 into the first group selection unit 10.
06 is weighted by the degree of group membership of the category group selected. In the identification signal loading section 1008, 1
Reference numeral 009 denotes a multiplier that multiplies the group membership of the category group selected by the first group selector 1006 by the intra-group similarity obtained from the subclassification unit 1003 corresponding to the membership. Reference numeral 1015 denotes a category similarity calculation unit that adds the multiplication result for each category to calculate a sum (category similarity). Reference numeral 1010 denotes a category selection unit which selects at least two category similarities and their category numbers in descending order of the category similarity, and also assigns the category number of the category similarity having the largest category similarity to the input pattern signal. This is the identification result. Reference numeral 1016 denotes a reject threshold value determination unit which determines a reject threshold value according to the category group selected by the first group selection unit 1006.
Reference numeral 1017 denotes a reject threshold control unit that controls a reject determination unit described later according to the reject threshold. Reference numeral 1018 denotes a reject threshold value judging unit which obtains a ratio of the second category similarity to the largest category similarity from the largest category similarity, and outputs a reject signal when the ratio is equal to or larger than the reject threshold. If it is less than the reject threshold, the category number having the highest category similarity is output as the recognition result of the input pattern signal.

Reference numeral 1011 denotes a first teacher signal generator.
It generates a necessary teacher signal at the time of updating the load (at the time of learning) of each of the sub-classifying units 1003. Reference numeral 1012 denotes a category information storage unit that stores category information belonging to each of the sub-classification units 1003. Reference numeral 1019 denotes a first learning determination unit which determines whether or not the input pattern signal can be learned based on the group belonging degree selected by the second group selection unit 1007 and the intra-group similarity calculated by the similarity calculation unit 1004. If it is determined that it is inappropriate for learning, a learning suppression signal is output. Reference numeral 1013 denotes a learning control unit, which is a comparison result between the classification result of the category selection unit 1010 and the category of the input pattern output from the first teacher signal generation unit 1011 and output from the first learning determination unit 1019. The presence or absence of a load value change in the load change amount control unit 1005 is controlled according to the learning suppression signal.

FIG. 2 is a block diagram specifically showing a configuration of the fuzzy large classification unit 1002. An input unit 21 inputs an input pattern signal. Reference numeral 22 denotes a distance calculation unit that calculates the distance between all group reference pattern signals in the group dictionary 1001 and the input pattern signals. Reference numeral 23 denotes a divider for calculating the reciprocal of the output of the distance calculator 22. Reference numeral 24 denotes an adder for calculating the sum of outputs from the respective dividers 23. A multiplier 25 multiplies the output of the adder 24 by the output of the distance calculator 22. Reference numeral 26 denotes a divider.
The reciprocal of the output of 5 is calculated.

FIG. 3 shows an example of the first unit recognition unit k used in the similarity calculating section 1004 shown in FIG. 6 of the learning type recognition judging device according to the present embodiment.
In the figure, reference numeral 31 denotes a signal input unit for inputting feature data to be recognized and recognized, which is input via a signal input terminal 31a, to a quantizer 32. The quantizer 32 quantizes the input feature data, and inputs a value indicating the quantized level to the path load unit 33c and the adjacent section selection unit 34. The adjacent section selection section 34 selects a quantization section adjacent to the quantization section of the input signal (feature data) and outputs the selected quantization section to the path load section 33c. 33a0 is a path input terminal, 33b1 to 33b5
Is a path output terminal, which connects these terminals to each other when a network is constructed by combining unit recognition units. The path load section 33c is provided for the path input section 3
Load 33c1 connecting between 3a and route output unit 33b
To 33c5, and load values stored in a load table 33d to be described later, according to the quantization result.
And a load setting unit 33c0 set to 33c5. In the load table 33d, a load value for the quantization section to which the input signal belongs and a load value for the adjacent quantization section are stored. In the path load section 33c, the load setting section 33c0 performs a calculation based on the quantization result. , The weight value is set at a position corresponding to each quantization section. The loads 33c1 to 33c5 weight the route signal input from the route input unit, and the route output unit 33b outputs the weighted route signal to the route output terminals 33b1 to 33b5.

FIG. 4 shows an example of the second unit recognition unit p used for the preceding layer at the lowest layer of the network in the similarity calculating section 1004 shown in FIG. 6 of the learning type recognition judging apparatus according to the present embodiment. It is shown. The first shown in FIG.
, A signal input unit 41, a quantizer 42, a path input unit 43a having one path input terminal 43a1, and three path output terminals 43b1 to 43b4.
3b3, a route output unit 43b, and a route input unit 43a.
, A transmission amount changing unit 44 for shifting the value transmitted to the path input terminal 43a1 by n bits, and a path load unit 43c. The loads 43c1 to 43c3 are weights to be added to the route output signals output to the route output terminals 43b1 to 43b3 of the route output unit 43b. In the load change process, the learning unit 43c0 uses the weight according to the output value of the transmission amount change unit. The load between the path input terminal 43a1 and the path output terminal indicated by the output value of the quantizer 42 is updated. In the forward processing, no signal is input to the quantizer 42, the weights 43c1 to 43c3 weight the signal output from the transmission amount changing unit, and the route output unit 43b converts the weighted route signal into a route. Output terminals 43b1 to 43b
Output to 3.

FIG. 5 shows an example of the third unit recognition unit q used for the lowermost layer of the network in the similarity calculating section 1004 shown in FIG. 6 of the learning type recognition judging device according to the present embodiment. is there. 53a0 is an adder for adding input signals from a plurality of path input terminals, and 53b0 is
This is a threshold value processor that performs threshold processing on the path signal.

FIG. 6 shows an example of the similarity calculating section 1004 of the learning type recognition judging device according to the first embodiment. In the present embodiment, three types of categories are recognized based on two types of two pieces of feature data (first feature data and second feature data). It has a tree branch structure, and a network is configured so that one tree branch structure has four layers. The unit recognition units k11 to k12 and k21 to k30 forming the first and second layers are, for example,
The unit recognition units p31 to p350 forming the third layer using the first unit recognition unit shown in FIG. 3 use, for example, the second unit recognition unit shown in FIG. Constituent unit recognition units q41, q42,
For q43, the third unit recognition unit shown in FIG. 5 is used. The signal input terminal of the first unit recognition unit forming the first and second layers in each tree branch structure inputs the characteristic data to be determined, and the signal of the second unit recognition unit forming the third layer The input terminal receives a teacher signal of input data only during the load change process.

In FIG. 6, reference numeral 60 denotes an operation mode selection unit for switching the operation mode by outputting a load change and forward processing mode signal, and 61 denotes a teacher signal output from the first teacher signal generation unit 1011, This is a teacher signal conversion unit that converts a corresponding unit recognition unit number of the last layer.

The operation of the learning type recognition / judgment device configured as described above will be described below. N of the object to be recognized
First, an input pattern signal X composed of pieces of feature data is
It is input to the fuzzy large classification unit 1002. X = (x ₁ , x ₂ ,..., X _n ) (7) In the fuzzy large classification unit 1002, first, the input unit 21 inputs the input pattern signal X, and the X distance calculation units 22
Is output. Each distance calculation unit 22 includes a group dictionary 1001
The group reference pattern signal V _i (1 ≦ i ≦ r; r is the number of group reference pattern signals, that is, the number of category groups) representative of each category group stored in is stored in the equation (8). a divider 2 which calculates the distance d _i between X and V _i, corresponding respectively showing
3, and output to the multiplier 25. d _i = {X−V _i } ^{2 / (f−1)} (1 ≦ i ≦ r) (8) where f is a real number satisfying f> 1.

[0046] Each of the divider 23, the reciprocal of the distance d _i is calculated, and its output to the adder 24. Adder 24
Calculates the sum of the outputs of all dividers 23 and outputs the output to r multipliers 25. Each multiplier 25 multiplies the output of the corresponding distance calculator 22 and the output of the adder 24 and inputs the output to the corresponding divider 26. Each divider 26 calculates the reciprocal of the output of the corresponding multiplier 25. Finally, in the fuzzy large classification unit 1002, the output of each divider 26 is set to the first group selection unit 1006 as the group membership degree μ _i (1 ≦ i ≦ r) of each category group with respect to the input pattern signal X. Is output. That is, the degree of group membership μ _i (1 ≦ i ≦ r) of each category group is

[0047]

(Equation 1) Can be expressed as

The group reference pattern signal representative of each category group stored in the group dictionary 1001 is prepared in advance.
A conventional clustering method, for example, the K-means algorithm and the Isodata algorithm shown in "Pattern Information Processing" (Corona) edited by IEICE Makoto Nagao, and Y. Linde, A. Buzo, and RMGray " An Algo
rithmfor Vector Quantizer design, "IEEE Trans. Com
mun., COM-28, 1, pp. 84-95, Jan. 1980
It is designed using the BG algorithm.

Hereinafter, a method of designing the group dictionary 1001 using the K-means algorithm will be briefly described. (1) r sets (where r is a predetermined number of category groups) from a set of pattern signals for group dictionary design of a recognition target
Are appropriately selected, and these are set as r group reference pattern signals V _i (1 ≦ i ≦ r). (2) For all group dictionary design pattern signals X,
In each case, V _i that minimizes the distance d _i shown in equation (10) is obtained. At this time, X is a category group S _i (1 ≦ i
≤ r). d _i = {X−V _i } (10) (3) The average value of the pattern signals X belonging to each S _i is obtained,
This is defined as V _i ′. (4) If satisfied V _i '= V _i is for all i, store the group reference pattern signal V _i at this time to the group dictionary 1001. Otherwise, V _i ′ is set as a new group reference pattern signal V _i and the process returns to (2).

By designing the group reference pattern signal in this way, all the pattern signals can be divided into some similar pattern signal subsets (category groups). Note that the Isodata algorithm and the LBG algorithm are basically almost the same as the K-means algorithm.

Returning to FIG. 1, the first group selecting unit 1006 selects a plurality of category groups in descending order of the group belonging degree obtained by the fuzzy large classifying unit 1002 (except for the selected group belonging degree). 0 ") is output to the identification signal loading unit 1008. As a method of selecting a category group, a category group having a group membership degree equal to or higher than a certain threshold value may be selected.

The similarity calculating section 1004 of each sub-classifying section 1003 corresponding to the category group selected by the first group selecting section 1006 calculates the similarity as follows.
Here, the similarity calculation unit will be described as an example in which the input pattern signal is four-dimensional vector data composed of first feature data (two-dimensional) and second feature data (two-dimensional) and is classified into three types of categories. Will be described.

As a path signal to the path input terminals of the unit recognition units k11 to k12 of the first layer of each tree branch structure of each similarity calculation unit 1004, 1 is always given. First, a forward processing mode signal is output from the operation mode selection unit 60 to the teacher signal conversion unit 61. The input pattern signal is divided into first feature data and second feature data, and is input to the signal input terminals 31a of the first and second unit recognition units, respectively.

Each unit recognition unit of the first layer quantizes the first feature data by the quantizer 32, and outputs the quantization result to the load setting unit 33c0 and the adjacent section selecting unit 34. The adjacent section selection unit 34 selects a quantization section adjacent to the quantization section of the input signal (feature data), and outputs it to the path load unit 33c. The load setting unit 33c0 reads the load value corresponding to the preset input data and the load value of the adjacent section from the load table 33d, and sets the corresponding load value at the corresponding quantization level position. In this way, a plurality of paths are selected in one unit recognition unit, and a value obtained by multiplying the path signal by the load value is sent to the path input terminals of the unit recognition units k21 to k30 in the second layer.

The second feature data is input to the signal input terminal of the unit recognition unit in the second layer (in this case, two second feature data are k21 to k21, respectively). k2
5. Input to k26 to k30. ). A plurality of path signals are selected in the same manner as in the case of the first layer, and a value obtained by multiplying the path signal by the load value is sent to the path input terminals of the unit recognition units p31 to p350 of the third layer.

In the forward processing mode, no teacher input signal is input to the signal input terminals of the unit recognition units p31 to p350 of the third layer. Therefore, the load is not changed, the state of the load at the time of learning is held, and the path input signal of each unit recognition unit is multiplied by these loads to obtain the path input terminals of all the unit recognition units on the fourth layer. What,
A route signal is sent.

The adder 5 of each unit recognition unit of the fourth layer
3a0 adds the input path signals. The threshold value processor 53b0 performs a threshold value process on this signal and outputs the signal to a path output terminal. Here, a sigmoid function, a step function, or the like can be used as the function for performing the threshold processing.

In this manner, based on the characteristic data of the input pattern signal, the output of the unit recognition unit of the final layer based on the weight value at that time is calculated. This output value is output to the identification signal load unit 1008 and the first learning determination unit 1019 as an intra-group similarity vector of each category with respect to the input pattern signal X.

On the other hand, in the identification signal loading section 1008, first, the multiplier 1009 calculates the group membership degree μ _i of the category group selected by the first group selection section 1006 and the subclassification section 1003 corresponding to the category group. Are multiplied by the intra-group similarity vector (ν ₁ , ν ₂ ,..., Ν _c ) obtained from the above, and outputs the outputs to the category similarity calculation unit 1015. That is,
The multiplier 1009 calculates (the number of group reference pattern signals × the total number of category reference pattern signals in each sub-classification unit).
And a group membership degree μ _p (1 ≦ p ≦ r; r is the number of category groups) of a certain category group p selected by the first group selection unit 1006, and a detail corresponding to the category group p. Within-group similarity ν of a certain category q obtained from the classification unit 1003
The output value ξp _q of the multiplier 1009 to which p _q (1 ≦ q ≦ c; c is the number of categories belonging to the category group) is expressed as follows: ξp _q = μ _p · νp _q (13)

In the category similarity calculation unit 1015, all
The output value of the multiplier 1009 is
The sum of the selected output values is calculated as the category similarity r._s(1 ≦ s
≤N _CN_CIs the number of categories)
Output to the selection unit 1010. In the category selection section 1010
Corresponds to the highest similarity in the category similarity.
To the learning control unit 1013 as the identification result.
Output.

The first teacher signal calculation unit 1011 generates a category number corresponding to the category of the pattern X in synchronization with the input pattern signal X, and generates a similarity calculation unit 1004 between the learning control unit 1013 and the fine classification unit 1003. Output to Second
The group selection unit 1007 selects a plurality of category groups from the sub-classification unit in which the category of the input pattern exists based on the category information stored in the category information storage unit 1012 from the group having a higher degree of group membership, and performs the first learning. Output to the determination unit 1019.

The first learning judging unit 1019 calculates the similarity calculating unit 100 corresponding to the maximum group belonging degree corresponding to the maximum group belonging degree which is larger than a certain threshold value.
If the maximum value of the intra-group similarity calculated in step 4 is larger than a certain threshold value, a learning suppression signal is output to the learning control unit 1013.

The learning control unit 1013 compares the classification result of the input pattern signal X calculated as described above with the category of the input pattern signal X. If the comparison results are equal, and if the comparison results are different, When the suppression signal is output, the next input pattern is learned without performing the load update of each sub-classification unit 1003, and when the comparison result is different and the learning suppression signal is not output, Control is performed so that the load value of each sub-classification unit 1003 is updated. When the load value is changed, the second group selecting unit 1007 determines, based on the category information stored in the category information storing unit 1012, a plurality of sub-classifying units in which the category of the input pattern exists, starting from the group having the highest group membership. Are output to the load change amount control unit 1005.

The load change amount control unit 1005 updates the load value of the similarity calculation unit 1004 corresponding to the selected group belonging degree as follows. In FIG. 6, first, a load change processing mode signal is output from the operation mode selection unit 60 to the teacher signal conversion unit 61. The teacher signal conversion unit 61 sets the category number output from the first teacher signal generation unit 1011 to the number of the unit recognition unit of the last layer in the selected category group, that is, q41, q42, or q43. The signal is converted into a signal indicating whether it is desirable to output a large value, and the signal is input to the signal input terminal 4 of the unit recognition unit in the third layer.
1a. The quantizers 42 of the unit recognition units p31 to p350 are set such that, for example, when q41 has the largest output, the quantization level value is a value corresponding to the path leading to the unit recognition unit q41.

The load change amount control unit 1005 changes the bit shift amount of the transmission amount change unit 44 according to the value of the selected group belonging degree. For example, among the sub-classification units selected by the second group selection unit, the bit-shift amount of the sub-classification unit with the largest group membership is 0, the bit-shift amount of the second sub-classification unit is 1, and so on. Change the bit shift amount.

The learning unit 43c0 is connected to the route input unit 43a.
The load value 43c1 which is the connection strength between the path input terminal 43a1 of the path output terminal 43b1 (here, the path 43b1 connected to the unit recognition unit q41) indicated by the output of the quantizer 42 is transmitted from the transmission amount changing section. Increase by the output value.

As described above, for all the input data other than the data determined to be unsuitable for learning, the identification result of the category selection unit 1010 and the first teacher signal generation unit 1011
By repeating the above-described learning operation until the teacher signals output from the learning patterns match, the learning-type recognition determination device can recognize the input pattern.

Next, the operation of the learning type recognition judging device at the time of the recognition operation will be described. Category similarity calculator 10
The operations up to 15 are the same as the above-described forward operation of the learning operation, and thus will not be described.

The category selecting section 1010 selects at least two category similarities and their category numbers in descending order of the category similarity, and selects the reject threshold value determining section 1
016, the reject determination unit 1018 outputs the category having the highest category similarity to the learning control unit 101 as the classification result of the input pattern signal.
Output to 3.

In the reject threshold value determination unit 1016,
The reject threshold is determined according to the category group selected by the first group selection unit 1006, and is output to the reject threshold control unit 1017. For example, when a category group including categories having a high degree of similarity is selected, the reject threshold is increased to facilitate rejection.

In reject threshold control section 1017,
Reject determination unit 10 according to the reject threshold
18 is controlled.

The reject determination unit 1018 obtains the ratio of the second category similarity to the largest category similarity from the one with the largest category similarity, and indicates that recognition is not possible if the ratio is equal to or larger than the reject threshold. And outputting a category number corresponding to the category similarity having the highest category similarity as a recognition result of the input pattern signal when the input pattern signal is less than the reject threshold.

The same applies to a case where a learning intensity control unit 1014 described later is used instead of the load change amount control unit 1005.

As described above, according to the first embodiment,
In the first learning determination unit 1019, the second group selection unit 1
If the degree of group belonging selected in 007 is large and the degree of similarity in group calculated by the similarity calculating unit 1004 corresponding to the degree of group belonging is large, it is determined that the reliability of the recognition result is very high. Regardless, when the recognition result is different from the teacher signal, that is, when a teacher signal addition error or a very difficult-to-identify signal is input, the learning control unit 1013 automatically detects the load change. By performing the control so that the learning is not performed, it is possible to reduce the learning time and prevent a reduction in recognition performance.

(Embodiment 2) FIG. 7 is a block diagram showing a learning type recognition judging apparatus according to a second embodiment of the present invention. 7, a group dictionary 1001, a fuzzy large classification unit 1002, a fine classification unit 1003, a similarity calculation unit 1004, a load change amount control unit 1005, a first group selection unit 1006, a second group selection unit 1007, an identification signal Load part 1
008, multiplier 1009, category identification unit 1010, first teacher signal generation unit 1011, category information storage unit 10
12, learning control unit 1013, category similarity calculation unit 10
15, reject threshold determination section 1016, reject threshold control section 1017, reject determination section 1018
Is the same as that of the first invention described above. The configuration different from FIG. 1 is that the second learning determination unit 10 that outputs a learning suppression signal in accordance with the category similarity value selected by the category selection unit 1015 instead of the first learning determination unit 1019
20 is provided.

The operation of the learning-type recognition / judgment device configured as described above will be described below. As in the first embodiment, the classification result of the input pattern is output from the category selection unit 1015. The category selection unit 10
If the maximum value of the category similarity output from 15 is larger than a certain threshold, the second learning determination unit 1020 outputs a learning suppression signal to the learning control unit 1013.

The subsequent operation and recognition operation are the same as in the first embodiment, and will not be described. Note that, instead of the load change amount control unit 1005, a learning intensity control unit 101 described later is used.
The same applies to the case where No. 4 is used.

As described above, according to the second embodiment,
In the second learning determination unit 1021, the category selection unit 1
015, the category similarity of the identification result is large, and the reliability of the identification result is determined to be very high, but the recognition result is different from the teacher signal. When a mistake or a signal that is very difficult to identify is input, it is automatically detected, and the learning control unit 1013 controls the load not to be changed, thereby shortening the learning time and reducing the recognition performance. It is possible to prevent.

(Embodiment 3) FIG. 8 is a block diagram showing a learning type recognition judging device according to a third embodiment of the present invention. 8, a group dictionary 1001, a fuzzy large classification unit 1002, a fine classification unit 1003, a similarity calculation unit 1004, a load change amount control unit 1005, a first group selection unit 1006, a second group selection unit 1007, an identification signal Load part 1
008, multiplier 1009, category identification unit 1010, first teacher signal generation unit 1011, category information storage unit 10
12, learning control unit 1013, category similarity calculation unit 10
15, reject threshold determination section 1016, reject threshold control section 1017, reject determination section 1018
Is similar to that of the first embodiment described above. 1 is different from the first learning determination unit 1019 in that learning suppression is performed in accordance with the degree of group belonging selected by the second group selection unit 1007 and the category information stored in the category information storage unit 1012. Third learning determination unit 10 that outputs a signal
21 is provided.

The operation of the learning type recognition / judgment device configured as described above will be described below. As in the first embodiment, the input pattern signal X is first input to the fuzzy large classification unit 1002. Second group selector 10
In step 07, a plurality of category groups are selected from the sub-classification units having the categories of the input pattern in descending order of the degree of belonging, based on the category information stored in the category information storage unit 1012, and the third learning determination unit 1021 Output. The third learning determination unit 1021 assigns category information to a category group having a maximum group membership degree value of a category group output from the second group selection unit that is greater than a certain threshold value and has the maximum group membership degree value. If there is no category included in the category information stored in the storage unit 1012, a learning inappropriate signal is output to the learning control unit 1013. When the learning suppression signal is output, the learning control unit 1013 controls each of the sub-classification units 1013.
Control is performed so that the load update of No. 03 is not performed, and learning of the next input pattern is performed.

The subsequent operation and recognition operation are the same as in the first embodiment, and will not be described. Note that, instead of the load change amount control unit 1005, a learning intensity control unit 101 described later is used.
The same applies to the case where No. 4 is used.

As described above, according to the third embodiment,
In the third learning determination unit 1021, the second group selection unit 1
If there is a contradiction that the category group having a high degree of group membership selected in 007 does not include the category to which the input pattern stored in the category information storage unit 1012 belongs, that is, a teacher signal addition error or an emergency When a signal that is difficult to identify is input, the detection is automatically performed at the stage when the large classification is performed, and the learning control unit 1013 performs control so that the load is not changed without performing the fine classification, thereby shortening the learning time. This makes it possible to prevent a decrease in recognition performance.

(Embodiment 4) FIG. 9 shows an example of a learning intensity control unit according to a fourth embodiment of the present invention. The fourth embodiment is different from the sub-classifying unit 1003 in FIG.
A learning degree control unit 1014 that changes the learning intensity of the similarity calculation unit 1004 in accordance with the output of the second group selection unit 1007 instead of the load change amount control unit 1005 of FIG. In FIG. 9, reference numeral 141 denotes a reference load value update number setting unit for setting a reference value of the load value update number, and 142 denotes a reference value of the load value update number output from the reference load value update number setting unit and the second value. A multiplier for multiplying the reference value by the group membership degree corresponding to the group membership degree selected by the group selection unit 1007, 143 is a similarity calculation unit 1 based on the multiplication result.
004 is a load update number control unit that controls the number of load updates by the learning device 43c0.

If the result of identification of the input pattern is output from the category selection unit and the result of comparison with the teacher signal is different, the second group selection unit 1007 inputs the result based on the category information stored in the category information storage unit 1012. A plurality of group membership degrees are selected from the sub-classification units having the pattern categories in descending order of the group membership degrees, and output to the learning intensity control unit 1014 of the corresponding fine classification unit.

Each learning intensity control unit 1014 weights the product by multiplying the reference value of the number of load value updates set in advance by the reference load value update number setting unit 141 by the group membership degree. Output to the unit 143. The load update number control unit 143 repeats the load update process by the update number output from the multiplier 142.

As described above, according to the fourth embodiment,
A learning degree control unit 1014 that changes the learning intensity of the similarity calculation unit 1004 according to the output of the second group selection unit 1007 instead of the load update number control unit 1005 of the fine classification unit 1003
Therefore, the present embodiment has the same effect as the above embodiments.

[0087]

According to the first aspect of the present invention, in the first learning determination section, the degree of group membership selected by the second group selection section is large, and The intra-group similarity calculated by the similarity calculating unit corresponding to the degree is large,
Even if it is determined that the reliability of the recognition result is very high, the recognition result and the teacher signal are different,
In other words, when the addition error of the teacher signal or a signal that is very difficult to identify is input, the detection is automatically performed, and the learning control unit performs control so as not to change the load.
It is possible to shorten the learning time and prevent a decrease in recognition performance.

According to the second aspect of the present invention, in the second learning determining section, the category similarity of the classification result output from the category selecting section is large, and the reliability of the classification result is high. If the recognition result is different from the teacher signal even though it is determined to be very high, that is, if a teacher signal addition error or a very difficult-to-identify signal is input, it is automatically detected. However, by controlling the learning control unit not to change the load, it is possible to shorten the learning time and prevent a reduction in recognition performance.

According to the third aspect of the present invention, the third learning determining section stores the category information in the category group selected by the second group selecting section and having a high degree of group belonging. If there is an inconsistency that the category to which the input pattern stored in the section belongs is not included, that is, if a teacher signal addition error or a very difficult to identify signal is input, By automatically detecting, performing no sub-classification, and controlling the learning control unit so as not to change the load, it is possible to shorten the learning time and prevent the recognition performance from lowering.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a learning-type recognition / determination apparatus according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a fuzzy large classification unit of the learning-type recognition / determination apparatus according to the first embodiment;

FIG. 3 is a block diagram illustrating a configuration of a first unit recognition unit in a similarity calculation unit of the learning-type recognition / determination apparatus according to the first embodiment;

FIG. 4 is a block diagram illustrating a configuration of a second unit recognition unit in a similarity calculation unit of the learning-type recognition / determination apparatus according to the first embodiment;

FIG. 5 is a block diagram illustrating a configuration of a third unit recognition unit in a similarity calculation unit of the learning-type recognition / determination apparatus according to the first embodiment;

FIG. 6 is a block diagram illustrating a configuration of a similarity calculation unit of the learning-type recognition / determination apparatus according to the first embodiment;

FIG. 7 is a block diagram illustrating a configuration of a learning-type recognition determination device according to a second embodiment of the present invention.

FIG. 8 is a block diagram illustrating a configuration of a learning-type recognition / determination apparatus according to a third embodiment of the present invention.

FIG. 9 is a block diagram illustrating a configuration of a learning intensity control unit of a learning-type recognition / determination apparatus according to a fourth embodiment of the present invention.

FIG. 10 is a block diagram showing a configuration of a conventional learning-type recognition / judgment device.

[Explanation of symbols]

 1001 Group dictionary 1002 Fuzzy large classification unit 1003 Fine classification unit 1004 Similarity calculation unit 1005 Load change amount control unit 1006 First group selection unit 1007 Second group selection unit 1008 Identification signal load unit 1009 Multiplier 1010 Category selection unit 1011 First teacher signal generation unit 1012 Category information storage unit 1013 Learning control unit 1014 Learning intensity control unit 1015 Category similarity calculation unit 1016 Reject threshold determination unit 1017 Reject threshold control unit 1018 Reject determination unit 1019 First learning Determination unit 1020 second learning determination unit 1021 third learning determination unit k first unit recognition unit p second unit recognition unit g third unit recognition unit

Claims (10)

[Claims] 1. A group dictionary in which a plurality of group reference pattern signals representative of a group of groups of similar patterns are stored, and an input pattern signal belongs to each category group by using the group reference pattern signal. A fuzzy large classifier for calculating a degree of group membership, a degree of similarity, a similarity calculator for calculating an intra-group similarity that is a degree of similarity of the input pattern signal to each category included in the category group, and a second group A plurality of sub-classification units each including a load change amount control unit that controls a bit shift amount of the transmission amount change unit of the similarity calculation unit in accordance with an output of the selection unit; and at least one category group based on the group membership degree. A first group selection unit to be selected, an identification signal weighting unit that weights the intra-group similarity obtained by the sub-classification unit by the selected group membership degree, and a category from the output of the identification signal weighting unit. Genus A category similarity calculating unit that calculates a category similarity that is a degree of performing, a category selecting unit that selects at least two category similarities and the category from the larger of the calculated category similarities, and the first A reject threshold determining unit for determining a reject threshold from the group membership selected by the group selecting unit and the output of the category selecting unit; and a reject threshold determined by the reject threshold determining unit. A reject threshold control section for controlling the reject determination section, and a reject threshold value in accordance with an output of the reject threshold control section and the two category similarities output from the category selection section. A reject that outputs a recognition result of the input pattern signal or a reject signal indicating that the input pattern signal cannot be recognized according to the comparison result with the ratio A classifying unit, a first teacher signal generating unit for generating a teacher signal necessary for learning of the subclassification unit, a category information storage unit storing category information belonging to each subclassification unit, and an input pattern based on the category information. A second group selecting unit that selects at least one group belonging degree and a category group from a group having a large group belonging degree among the sub-classifying units in which the category exists, and a group selected by the second group selecting unit. Determine the necessity of learning the input pattern signal from the degree of belonging and the intra-group similarity calculated by the similarity calculation unit,
A first learning determination unit that outputs a learning suppression signal; a comparison result between a category selected by the category selection unit and a category of an input pattern output from the first teacher signal generation unit; A learning control unit that controls learning of the fine classification unit according to a learning suppression signal output from a determination unit, wherein the similarity calculation unit includes a plurality of layers, Is constituted by a plurality of third unit recognition units, a plurality of second unit recognition units on the lowermost layer, and a plurality of first unit recognition units on the other layers, wherein the first unit recognition unit is A signal input unit, a quantizer that performs quantization in accordance with the output of the input signal, an adjacent interval selection unit that selects a quantization interval adjacent to the quantization interval from a quantization result of the input signal, Attributes of input signal A load table for storing a quantization section and a weight value for a quantization section adjacent to the quantization section, a path input section having a single or a plurality of path input terminals, and a single or a plurality of path outputs. A path output unit having a terminal, and a path for changing a connection strength between a path input terminal of the path input unit and a path output terminal of the path output unit by changing a setting position of a load applied to the path according to a quantization result. A load unit, wherein the second unit recognition unit includes a signal input unit, a quantizer that performs quantization in accordance with an output of the input signal, and a path having a single or a plurality of path input terminals. An input unit, a path output unit having a single or a plurality of path output terminals, a transmission amount changing unit that shifts a value transmitted to the path input terminal of the path input unit by n bits, According to the path input terminal and front A learning unit configured to increase a connection strength between the path output terminal of the path output unit indicated by the output value of the quantizer and the path output terminal; A learning type recognition / judgment device comprising an adder and a threshold value processing unit for performing threshold processing on an output signal of the adder. 2. The learning type according to claim 1, further comprising a second learning determining unit that outputs a learning suppression signal in accordance with the category similarity output from the category selecting unit, instead of the first learning determining unit. Recognition determination device. 3. A third learning unit that outputs a learning inappropriate signal in accordance with the degree of group membership selected by the second group selection unit and the category information stored in the category information storage unit, instead of the first learning determination unit. The learning-type recognition / judgment device according to claim 1, further comprising a learning judgment unit. 4. A learning intensity control unit for changing the learning intensity of the similarity calculation unit in accordance with the output of the second group selection unit, instead of the load change amount control unit. Learning type recognition judgment device. 5. The learning-type recognition / judgment device according to claim 4, wherein the learning strength control unit changes the number of times of learning of the corresponding sub-classification unit according to the output of the second group selection unit. 6. A fuzzy large classification unit includes: a plurality of distance calculation units for calculating distances between an input pattern signal and all group reference pattern signals stored in a group dictionary; and an output of the distance calculation unit. A plurality of dividers for calculating a reciprocal; an adder for adding each output of the divider; a plurality of multipliers for multiplying an output of the adder and an output of the distance calculation unit; 4. The learning type recognition / judgment device according to claim 1, further comprising a plurality of dividers for calculating the reciprocal of the output of the device. 7. A teacher signal conversion unit for converting an output of the first teacher signal generation unit into a number of a third unit recognition unit which desirably outputs the largest value among the sub-classification units. 4. The learning according to claim 1, wherein an output of the teacher signal conversion unit is input to a signal input unit of a unit and a second unit recognition unit located in the lowermost layer. Type recognition judgment device. 8. The learning-type recognition / judgment device according to claim 1, wherein the reject threshold determining unit determines the reject threshold according to the category selected by the category selecting unit. . 9. The learning method according to claim 1, wherein the reject threshold value determining unit determines the reject threshold value according to the category group selected by the first group selecting unit. Type recognition judgment device. 10. The reject threshold determining unit determines a reject threshold according to a category selected by the category selecting unit and a category group selected by the first group selecting unit. Item 3. The learning-type recognition / judgment device according to item 1, 2 or 3.