JP3247811B2 - Pattern matching device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern matching apparatus, which uses a neural network to recognize handwritten characters, voice recognition, or information signals, etc., as a pattern. The present invention relates to a pattern matching apparatus that teaches an output and thereafter automatically performs a matching process.

[0002]

2. Description of the Related Art Conventionally, a learning operation of recognizing handwritten characters and inputting a pattern to be identified so as to obtain a control output corresponding to various information signals and correcting the corresponding output by comparing it with a teacher pattern has been repeated. Thereafter, there is known a pattern matching apparatus which automatically performs an identification by inputting an arbitrary identification pattern. Such a pattern matching apparatus is generally configured using a neural network configured by connecting a multiple-input one-output type element (unit) to another unit via a coupling element.

A neural network used in this type of pattern matching device is an information processing technology that imitates the information processing function of the cerebrum of an organism, and is useful in that it is easy to apply and can be applied to nonlinear problems. Has characteristics. Further, since the method is different from that of the conventional Neumann computer, application to a problem that has been difficult in the Neumann computer is expected.

The components of the neural network are shown in FIG.
A unit 1 corresponding to a nerve cell of an organism as shown in
The unit 1 includes a coupling element 2 corresponding to a nerve fiber, which connects an input part and an output part. Normally, effective information processing is not performed by using only a single unit. Therefore, a network is formed by appropriately connecting these units using a plurality of units and a coupling element.

[0005] The units are coupled via a coupling element of a plurality of other units. As shown in the following equation 11, the outputs of these units are multiplied by a weighting factor assigned to each coupling element, and then added. An input and output value of an input / output function having a difference between the input value and the threshold value as an independent variable is output, and becomes an input to another unit via an output coupling element.

[0006]

[Equation 11] Here, o _i : the output value of the i-th unit, f: the input / output function w _ij : the strength of the connection from the j-th unit to the i-th unit (coupling coefficient), o _j : the j-th unit The output value of the unit of θ _i is the threshold value of the i-th unit. Also, depending on the definition, Equation 11 is

(Equation 12) In some cases,

At this time, as the input / output function f, FIG.
A saturated function called a sigmoid function as shown in FIG. This function simulates the function of nerve cells of a living organism, and is a monotonically increasing function that takes a value from 0 to 1. When the input value is equal to the threshold value (or when the input value is equal to the threshold value and the absolute value is (When equal signs are different).

Next, useful information processing becomes possible by configuring a network using the units having the above functions. A typical network shown in FIG. 6 is considered. I have. FIG. 6A shows a hierarchical network in which 1 constitutes an input layer.
Units 2 and 3 are units constituting the first and second intermediate layers, respectively, and 4 is a unit constituting the output layer, and 5, 6, and 7 are coupling elements. Such a hierarchical network has a hierarchical structure in which units between different layers are connected but units in the same layer are not connected, and information is transmitted in one direction from an input unit to an output unit. Is done. It is a feed-forward network. This model is known as a cerebellar model, and is also an important network used in many fields such as putter matching and control.

In this model, a pattern or a time-series signal to be identified is given to the input layer. For example, when applied to putter matching such as recognition of a handwritten character, the input layer is provided with a handwritten character or a feature pattern obtained by performing some kind of feature extraction processing on the handwritten character. On the other hand, each unit in the output layer is set so as to give a determination result for a category to be learned. That is, if the output of the unit 1 is, for example, 0.5 or more, the neural network is set so that it can be considered that the given pattern has been identified as a category 1 pattern.

FIG. 6B shows a mutual connection type network, where 1 is a unit and 2 is a coupling element. As described above, in the interconnected network, all units are interconnected. It should be noted that, in addition to the above, there are networks having feedback and networks having a partial connection between the same layers.

Next, learning is performed by giving learning data to the network so that the network has useful functions as described above. Since there are many learning methods and types of networks, there are also many learning methods. Here is the most typical method at present. A method of learning the three-layer hierarchical neural network shown in FIG. 7 using a back propagation algorithm will be described as an example.

In the learning, first, a coupling coefficient and a threshold value are set to small values, an output when a learning pattern is given to the input layer is calculated, and the result and an appropriate output for the pattern are designated. This is performed by comparing the teacher pattern with the teacher pattern and changing the coupling coefficient and the threshold value until the output of the output layer approaches the teacher signal for the given learning pattern. The detailed steps of the learning are as follows.

(Step 1) A coupling coefficient matrix {W _ji }, {V _kj } and a threshold vector {θ _j }, {γ _k }, which determine the state of the network, are each initialized with a small random value. (Step 2) The first pattern is set as a learning pattern. (Step 3) Put the value of the learning pattern into the output {I _i } of the input layer unit, and use the coupling coefficient matrix {W _ji } from the input layer to the intermediate layer and the threshold θ _j of the intermediate layer unit j. , An input U _j to the intermediate layer unit j, and an output H _j of the intermediate layer unit j is obtained from the input U _j and the sigmoid function f shown in FIG.

(Equation 13)

(Step 4) Output ΔH of the intermediate layer unit
and _j}, by using the threshold value gamma _k of the coupling coefficient matrix {V _kj} and the output layer unit k from the intermediate layer to the output layer, prompted S _k to the output layer unit k, the input S _k The output O _k of the output layer unit k is obtained by the sigmoid function f.

[Equation 14]

(Step 5) The teacher signal T of the learning pattern
_From the difference between _K and the output O _k of the output layer, a coupling coefficient connected to the unit k of the output layer and an error δ _k with respect to the threshold value of the intermediate layer unit k are _obtained .

(Equation 15)

(Step 6) From the error δ _k , the coupling coefficient matrix {V _k } from the hidden layer to the output layer, and the output H _{j of the} hidden layer,
An error δ _j between the coupling coefficient connected to the intermediate layer unit j and the threshold value of the intermediate unit is obtained.

[0017] (Step 7) by adding the product of the output H _j and constants α error [delta] _k and the intermediate layer unit j in the output layer unit k obtained in step 5, the intermediate layer unit j in the output layer Modify the coupling coefficient V _kj leading to unit k. Also, by adding the product of the error δ _k and the constant β,
Modify the threshold value γ _k of the output layer unit k.

(Equation 16)

[0018] (Step 8) and the error [delta] _j in the intermediate layer unit j, by adding the output I _i and the product of the constant α of the input layer units i, an intermediate layer unit j from the input layer unit i
_Is corrected to the coupling coefficient W _ji leading to. Also, the threshold θ _j of the intermediate layer unit _j is corrected by adding the product of the error δ _j and the constant β.

[Equation 17]

(Step 9) The next pattern is set as a learning pattern. (Step 10) Return to Step 3 until the learning pattern ends. (Step 11) The number of times of learning is updated. (Step 12) If the number of repetitions of the learning is equal to or less than the limit number, the process returns to Step 2.

Through such learning, the neural network acquires the pattern matching ability, and can identify a pattern that has already been learned and an unlearned pattern belonging to a learned category.

[0021]

However, in a conventional pattern matching apparatus using neural network processing, when a pattern belonging to a category that has not been learned is identified, it is erroneously determined to be one of the learned categories. There are often cases. This problem is generally a problem in response to unlearned data.

Incidentally, the reason for the erroneous determination is that the conventional input / output function outputs a value of 1 or 0 regardless of the value when a value considerably larger or smaller than the threshold value is input. . It is considered that since a so-called saturated function is used, there is only an operation of simply setting a boundary between the learned categories.

An object of the present invention is to provide a pattern matching apparatus using neural network processing capable of appropriately identifying and outputting an unlearned pattern by paying attention to the above conventional problems. Even if the sum of the input signals multiplied by the weight coefficient in each unit is not only close to the threshold but also greatly deviated from the threshold, an appropriate output signal is output, thereby improving the accuracy of the pattern matching judgment. It is an object to provide a pattern matching device that can be improved.

[0024]

[0025]

[0026]

In order to achieve the above object, a pattern matching apparatus according to the present invention comprises a unit group for inputting a pattern to be identified and a unit group for outputting a pattern corresponding to the input pattern. And an intermediate unit group that connects the input unit group and the output unit group with a plurality of coupling elements, wherein each unit is obtained by multiplying the output from the preceding unit by a weighting factor assigned to each coupling element. Addition and input, calculation is performed by a calculation unit having an input / output function using the difference between these input values and the set threshold value as an independent variable, and input to another unit via a coupling element for output After repeating the output from the output unit group to correct the weight coefficient of the related unit by the teacher signal, pattern matching is performed. In the pattern matching apparatus, the input and output functions of the arithmetic unit in the unit, c
_{When 1} and c ₂ are coefficients and the threshold is θ,

(Equation 18) Or

[Equation 19] Is a differential function of x with respect to x

(Equation 20) It is configured so that

According to a second aspect of the present invention , when the input / output function of the input / output function operation unit uses c ₁ and c ₂ as coefficients and the threshold value as θ,

(Equation 21) Or

(Equation 22) Is a differential function of x with respect to x

(Equation 23) It was configured to be:

According to a third aspect of the present invention , when the input / output function of the operation unit in the unit is such that c is a coefficient and a threshold is θ,

(Equation 24) Or

(Equation 25) It was configured to be:

Further, in the fourth invention , in particular, when the input / output function of the operation unit in the unit has c as a coefficient and a threshold as θ,

(Equation 26) Or

[Equation 27] It was configured to be:

[0030]

According to the above arrangement, the input / output function of the unit takes the maximum value for the input value equal to the threshold value as shown in FIGS. For, the corresponding output does not exceed the maximum value. However, depending on the method of defining the threshold value, the maximum value is obtained when the input value is equal to the threshold value and the absolute value is different and the sign is different. The same applies to the following inventions.

The input / output function of the unit is shown in FIG.
A function as shown in (4) is used. In other words, the function takes the maximum value for an input value equal to the threshold value, and for a value different from the threshold value, the corresponding output does not exceed the maximum value, and is a function that does not take a negative value. is there.

The first invention embodies the above means, and an effective output is obtained for a threshold value and an input near the threshold value. A second invention is effective output can be obtained for an input range that is proportional to the threshold and the threshold (determined by c ₁ and c _2). According to the third aspect, an effective output is obtained for a threshold value and an input near the threshold value. By increasing n, the boundary between categories becomes steeper.
According to a fourth aspect of the present invention, a threshold value and a range (c
Is determined, the output is valid. By increasing n, the boundary between categories becomes steeper.

In these functions, the coefficients c, c ₁ , c ₂ and the threshold value θ can be constants or determined by learning according to the problem to be solved. In addition, when a constant is used, a value of 1 can be selected.

As described above, even when any of the input / output functions of the present invention is used, a valid output is obtained from the unit only when an input value close to the threshold value is given to the unit.
A small value is given as an output signal to an input signal far away from the threshold value. As a result, the judgment output for an unlearned pattern is improved, the erroneous judgment output can be significantly reduced, and the judgment accuracy of the pattern matching is improved.

[0035]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram illustrating an overall configuration of a pattern matching apparatus according to an embodiment and an explanatory diagram of a main configuration.
As shown in the figure, the apparatus has an image receiving apparatus 10 for capturing an image to be recognized such as characters or various signal waveforms to be subjected to pattern matching. The image receiving device 10 can directly display the recognition target image, and outputs the image display signal to the pattern input unit 12. The pattern input unit 12 divides an image display area of the image receiving apparatus 10 into a two-dimensional matrix, and separates individual image signals corresponding to each divided area into a subsequent stage in accordance with the magnitude of a signal amount, for example, signal intensity or density. It is to be output to the conversion processing unit 14.
Thus, the input image signal is output as a pattern signal.

A conversion processing unit 14 for inputting an output signal from the pattern input unit 12 includes an input layer 16 as a unit group to which a pattern to be identified is input, and a unit group to output a pattern corresponding to the input pattern. And an intermediate layer 18 as an intermediate unit group connecting the input layer 16 and the output layer 20 by a plurality of coupling elements.
It is composed of These input layer 16, intermediate layer 1
8 and the output layer 20 have a plurality of multi-input / one-output elements (units) 22 in each layer, and are configured by coupling the unit 22 of each layer to the unit 22 of another layer via the coupling element 24. It is configured to use the neural network obtained.

The components of each of the layers 16, 18, and 20 include a unit 22 corresponding to a nerve cell of an organism, and a unit 2
It comprises a coupling element 24 corresponding to a nerve fiber, which connects the input unit and the output unit of the second unit, and a plurality of units and these units are appropriately connected to each other using a coupling element to form a network. In the embodiment, a hierarchical neural network is adopted, and although units between different layers are connected, units of the same layer are not connected, and information is transmitted in one direction from an input unit to an output unit. In other words, it is a feed-forward network.

The unit 22 includes a plurality of lower units 22.
, And is configured to calculate based on the input / output function f shown in the following Expression 28 in the embodiment.

[Equation 28]

The input / output function f outputs a maximum value when the absolute value of the input value is equal to the threshold value, and outputs a positive output value not exceeding the maximum value for an input value different from the threshold value. Where c is a coefficient, θ is a threshold, and is a constant or a variable.

First, in the input layer 16, which is the first input unit of the conversion processing unit 14, the output calculation circuit calculates the output of each unit in the input layer based on the input pattern. Each unit 22 corresponds to each element of the pattern input unit 12 and 1
In this example, the input / output function is a linear function, and the output value of each unit in the input layer is equal to the value of the input pattern.

Next, in the input calculation circuit in the input portion to the intermediate layer 18, first, the output of the input layer 16 and the input layer 16
The input supplied to each unit of the intermediate layer 18 is calculated using the coupling coefficient between the intermediate layer 18 and. A plurality of output signals from a plurality of units 22 of the input layer 16 at the previous stage are input to each unit 22 of the intermediate layer 18 to become a multi-input signal.

(Equation 29) Is calculated based on Here, w _ij is the coupling strength (coupling coefficient) from the j-th unit to the i-th unit, θ _i is the threshold value of the i-th unit, and U _j is the j-th unit.
The output value of the th unit. Thereby, the output U of the unit 22 is multiplied by the weighting coefficient W assigned to each coupling element, and then added to the input. The difference between these input values and the threshold is used as an independent variable, The output value to the intermediate layer 18 is calculated by the input / output function f.

That is, in the intermediate layer 18, the output calculation circuit performs an operation based on the expression 28 based on the input value calculated by the above expression 29, from the plurality of units 22 in the intermediate layer 18 to the lower layer. Is calculated (output signal U). This is, specifically,

[Equation 30] Is calculated as Here, H _{j is} the output value of the j-th unit, U _{j is} the output value of the j-th unit, and θ _i is the threshold value of the i-th unit.

Further, the thus obtained intermediate layer 18
The output H _i is input to the output layer 20 via the coupling element 24 from, the input calculation circuit at the input portion to the output layer 20, first, the output H _i of the intermediate layer 18, the intermediate layer 1
The input supplied to each unit 22 of the output layer 20 is calculated using a coupling coefficient between the output layer 8 and the output layer 20. This arithmetic processing is performed in the same manner as in Expression 29 described above.

(Equation 31) Here, V _kj is a coupling coefficient from the intermediate layer to the output layer, γ _k is a threshold value γ _{k of} the output layer unit k, and S _k is an input value to the output layer unit k.

Next, the output value to the output layer 20 is calculated by the input / output function f based on the equation (28). That is, in the output layer 20, by the output calculation circuit,
Based on the input value calculated by the above equation (31), a calculation is performed based on the above equation (28), and a plurality of units 22 in the output layer 20 are calculated.
, An output value to a lower layer is calculated (output signal O). This is, specifically,

(Equation 32) Is calculated by

The output signal O thus obtained is output for each unit 22 of the output layer 20. This is basically output from the pattern output unit 26, and the output pattern matching the output signal is stored in the memory. 28, output as a corresponding pattern, and display and output the content on a display device 30 or the like.

In this embodiment, a learning function by back propagation is provided. For this reason, an error detection unit 32 for inputting an output signal from the output layer 20 of the conversion processing unit 14 is provided. Correction processing of the coupling coefficient of each coupling element 24 and the threshold value of the unit 22 of the conversion processing unit 14 is performed so as to eliminate the error, and the pattern matching capability for the input pattern is acquired. In this learning, first, the coupling coefficient and the threshold are set to small values, the output when the learning pattern is given to the input layer is calculated, and the result and the teacher pattern that specifies the appropriate output for the pattern are calculated. This is performed by changing the values of the coupling coefficient and the threshold value until the output of the output layer approaches the teacher signal for the given learning pattern. For this purpose, a teacher device 34 is provided, and a coupling coefficient connected to the unit 22 of the output layer 20 based on the difference between the output pattern from the output layer 20 and the teacher signal, and the intermediate layer 18.
The processing such as calculating the error with respect to the threshold value of the unit 22 is performed by the correction processing using the above-described formulas 15 to 17. This learning process is repeated a set number of times,
Try to gain learning skills. After this processing, the error detection unit 32 is bypassed, and the output of the conversion processing unit 14 is directly displayed on the display device 30 via the pattern output unit 26.

An embodiment in which the operation of identifying a pattern represented by a two-dimensional vector using the pattern matching apparatus having the above-described configuration will be described.

FIG. 3 shows the circuit configuration of the conversion processing unit 14 described above. Reference numeral 40 denotes an output calculation circuit of the input layer 16 which calculates the output of each unit 22 of the input layer based on the input pattern. In this example, since the input / output function is a linear function as generally used, the output value is equal to the value of the input pattern. Reference numeral 42 denotes an input calculation circuit of the intermediate layer 18, which calculates an input supplied to each unit 22 of the intermediate layer 18 by using an output of the input layer calculation circuit 40 and a coupling coefficient between the input layer 16 and the intermediate layer 18. Reference numeral 44 denotes an output calculation circuit of the intermediate layer 18, which calculates the output of the intermediate layer 18 using the input / output function of Expression 28 based on the input value calculated in 42. 46 is the output layer 2
The zero input calculation circuit calculates the input supplied to each unit 22 of the output layer 20 using the output of the intermediate layer output calculation circuit 44 and the coupling coefficient of the intermediate layer 18 and the output layer 20. 4
Reference numeral 8 denotes an output calculation circuit of the output layer 20, which calculates the output of the output layer 20 using the input / output function of Expression 28 based on the input values calculated by the four output layer input calculation circuit 46.

FIG. 4 shows two-dimensional vectors used for learning. Category 1 is indicated by a symbol “*”, and category 2 is indicated by a symbol “+”. The data tables belonging to the categories 1 and 2 and the test data are shown in the following table. In this data, data numbers 1 to 18 are learning data (odd numbers belong to category 1 and even numbers belong to category 2), and data numbers 19 to 29 are test data.

[0050]

[Table 1]

Using the apparatus according to the embodiment, the learning data (data number 1) of Table 1 belonging to each of the categories 1 and 2 in FIG.
After learning by inputting the non-learned data (data Nos. 19 to 29), it is determined whether the unlearned data belongs to which category. Similarly, a learning pattern corresponding to FIG. 4 is trained with corresponding data 1 to 18 using a network in which the input / output function of the output layer 20 is a sigmoid function as in the related art. Table 2 shows the results of inputting the data (data numbers 19 to 29) and testing which category the unlearned data belongs to.

[0052]

[Table 2]

When these results are compared, there is no great difference in the identification results for the learning patterns (numbers 1 to 18).
Patterns that do not belong to the learned category (numbers 19 to 2)
There is a difference in the identification result for 9). That is, in the conventional method, output values for these patterns are large. For example, if an output of 0.5 or more is obtained, it is determined that the pattern belongs to the set category, and data numbers 20, 22, 24 , 26, and 28 are identified as belonging to category 1, and data numbers 23, 2
Patterns 5, 27 and 29 are identified as belonging to category 2. Therefore, the validity rate of data that does not belong to the learned category is 9%. On the other hand, when the pattern matching device according to the present invention is used, the data number 2
8, an erroneous determination is made, but for other patterns, it is determined that the output value is less than 0.5 and does not belong to the learned category. Accordingly, the validity rate of data that does not belong to the learned category is 90%, which is higher than that of the conventional method.

[0054]

As described above, according to the pattern matching apparatus of the present invention, the sum of the input signals multiplied by the weight coefficient in each unit is not only close to the threshold but also greatly deviates from the threshold. Even if it is, an excellent output signal can be output and the effect of improving the accuracy of pattern matching can be improved. Without this, it is possible to remarkably reduce the conventional drawback of making an erroneous determination such as assigning a pattern belonging to an unlearned category to one of the learned patterns.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram of an embodiment of a pattern matching apparatus according to the present invention.

FIG. 2 is an example of an input / output function of the present invention.

FIG. 3 is a circuit configuration example of a conversion processing unit according to the embodiment of the present invention.

FIG. 4 is a diagram in which a two-dimensional pattern used for learning is plotted on xy coordinates.

FIG. 5 shows a schematic diagram of a neural network element (unit) and a sigmoid function which is a kind of input / output function used in this unit.

FIG. 6 is a schematic diagram of a configuration example of a network, in which (1) is a schematic diagram of a hierarchical network and (2) is a schematic diagram of an interconnected network.

FIG. 7 is a schematic image of a three-layer hierarchical network.

DESCRIPTION OF SYMBOLS 10 Image receiving device 12 Pattern input unit 14 Conversion processing unit 16 Input layer 18 Middle layer 20 Output layer 22 Unit 24 Coupling element 26 Pattern output unit 28 Memory 30 Display device 32 Error detection unit 34 Teacher device 40 Input layer output Calculation circuit 42 Middle layer input calculation circuit 44 Middle layer output calculation circuit 46 Output layer input calculation circuit 48 Output layer output calculation circuit

Continuation of the front page (56) References JP-A-5-197701 (JP, A) JP-A-4-311254 (JP, A) JP-A-7-210533 (JP, A) JP-A-7-44641 (JP) , A) Shigeki Nakauchi et al., "Color Learning Using Three-Layer Neural Networks and Its Internal Expression Analysis", Transactions of the Institute of Electronics, Information and Communication Engineers, Vol. J73-D-II, No. 8. The Institute of Electronics, Information and Communication Engineers, published by, issue date: August 25, 1990, pp. 1242-1248 (Patent Office CS DB Document No .: CNT199800681017) Kazuki Nakajima et al., "Graphical Recognition by Psychological Potential Field and Neural Network", IEICE Technical Report, Vol. 90, no. 491 (PRU90-145 to 156), The Institute of Electronics, Information and Communication Engineers, Issued on: March 20, 1991, pp. 79 to 86 (Patent Office CSDB Document Number: CS NT199900657010) (58) Int.Cl. ⁷ , DB name) G06N 1/00-7/00 G06G 7/60 JICST file (JOIS) CSDB (Japan Patent Office)

Claims (4)

(57) [Claims] 1. A unit group for inputting a pattern to be identified, a unit group for outputting a pattern corresponding to the input pattern, and an intermediate unit for connecting the input unit group and the output unit group by a plurality of coupling elements. Group, each unit is multiplied by an output from a unit at the preceding stage after multiplying a weighting factor assigned to each coupling element and then input, and a difference between these input values and a set threshold is calculated. The weighting factor of a unit operated by an operation unit having an input / output function as an independent variable, inputting to another unit via a coupling element for output, and outputting from the output unit group by a teacher signal In a pattern matching apparatus configured to perform pattern matching after repeating the correction of, the input / output function of the arithmetic unit in the unit , C _1, c ₂
Is a coefficient and the threshold is θ, Or Is a differential function of x with respect to x A pattern matching device, characterized in that: 2. A unit group for inputting a pattern to be identified, a unit group for outputting a pattern corresponding to the input pattern, and an intermediate unit for connecting the input unit group and the output unit group by a plurality of coupling elements. And each of the units has a weighting factor assigned to each coupling element multiplied by the output from the unit at the preceding stage and then added and input, and a difference between these input values and a set threshold value is calculated. The weighting factor of a unit operated by an operation unit having an input / output function as an independent variable, inputting to another unit via a coupling element for output, and outputting from the output unit group by a teacher signal In a pattern matching apparatus configured to perform pattern matching after repeating the correction of, the input / output function of the arithmetic unit in the unit , C _1, c ₂
Is a coefficient and the threshold is θ, Or Is a differential function of x with respect to x A pattern matching device, characterized in that: 3. A unit group for inputting a pattern to be identified, a unit group for outputting a pattern corresponding to the input pattern, and an intermediate unit for connecting the input unit group and the output unit group by a plurality of coupling elements. Group, each unit is multiplied by an output from a unit at the preceding stage after multiplying a weighting factor assigned to each coupling element and then input, and a difference between these input values and a set threshold is calculated. The weighting factor of a unit operated by an operation unit having an input / output function as an independent variable, inputting to another unit via a coupling element for output, and outputting from the output unit group by a teacher signal In a pattern matching apparatus configured to perform pattern matching after repeating the correction of, the input / output function of the arithmetic unit in the unit , And coefficient c, when the threshold theta, [Expression 7] Or A pattern matching device, characterized in that: 4. A unit group for inputting a pattern to be identified, a unit group for outputting a pattern corresponding to the input pattern, and an intermediate unit for connecting the input unit group and the output unit group by a plurality of coupling elements. Group, each unit is multiplied by an output from a unit at the preceding stage after multiplying a weighting factor assigned to each coupling element and then input, and a difference between these input values and a set threshold is calculated. The weighting factor of a unit operated by an operation unit having an input / output function as an independent variable, inputting to another unit via a coupling element for output, and outputting from the output unit group by a teacher signal In a pattern matching apparatus configured to perform pattern matching after repeating the correction of, the input / output function of the arithmetic unit in the unit , And coefficient c, when the threshold theta, [Expression 9] Or A pattern matching device, characterized in that: