JPH07168799A - Learning device for neural network - Google Patents

Abstract

PURPOSE:To enhance the efficiency of learning by providing a sensitivity representative value calculation means which finds the representative value of the sensitivity of output data when micro change is applied to the input data of teacher data and a recognition result correct answer degree calculation means which calculates the correct answer degree of a recognition result, and judging whether or not the learning for each teacher data should be continued based on the output of these means. CONSTITUTION:The sensitivity representative value calculation means 1 finds the representative value of sensitivity of the output data of a neural network when the micro change is applied to the sensitivity for respective teacher data i.e., the input data of the teacher data. The recognition result correct answer degree calculation means 2 calculates the correct answer degree of the recognition result for respective teacher data. The correction answer degree can be calculated by comparing the actual output value of an output layer neuron with the output data of the teacher data. Furthermore, a learning condition judging means 3 judges whether or not the learning for respective teacher data should be continued based on the output of the sensitivity representative value calculation means 1 and that of the recognition result correct answer degree calculation means 2.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to pattern recognition,
The present invention relates to a neural network that performs character recognition, prediction, or the like, and more specifically, to a learning device for a neural network that gives teacher data to a neural network and updates internal weights of the neural network so that the teacher data can be correctly recognized.

[0002]

2. Description of the Related Art A neural network is generally composed of a plurality of neurons. Each of these neurons generally weights inputs from other neurons or external sources and then sums them to obtain an output y as a function of the sum x. As this function, a sigmoid function given by the following equation is generally used in many cases.

Y = 1 / (1 + exp (-x)) FIG. 26 shows the form of this sigmoid function. Since this function is a non-linear function having a saturation characteristic, the value of the output is close to 0 or 1 when the change of the output with respect to the change of the input, that is, the sensitivity of the input to the input is small. Such a property is not limited to the sigmoid function, and a non-linear function having a saturation characteristic has the same property.

FIG. 27 shows an example of a hierarchical neural network. In the figure, the network is composed of an input layer, an intermediate layer, and an output layer. Generally, each layer is composed of a plurality of neurons, and the neurons are completely connected between the layers. Generally, the intermediate layer is not limited to one layer, and there may be a plurality of layers.

When such a neural network is made to perform pattern recognition, a plurality of typical pattern feature values are selected and used as input data to the input layer neurons.
The output layer unit is applied to each of the patterns to be recognized, and for the correct pattern, teaching data is given such that the output value of the corresponding output layer neuron is 1, and the output value of the other output layer neurons is 0, and learning is performed. Let it be done.

After the learning of the teacher data is completed, input data of which pattern is unknown can be given to the input layer neuron and the input pattern can be judged by the output value of the output layer neuron.

Neural networks are often applied when the input / output relationship is not clear and it is difficult to represent them by mathematical expressions. Regarding the learning of such a neural network, in Japanese Patent Application Laid-Open No. 3-118658 (the title of the invention: a function analysis device of a learning device) previously filed, the neural network corresponding to the weight of the connection determined by learning is described. A functional analysis device capable of estimating the input / output relationship has been disclosed. In the case of a three-layer hierarchical neural network, if the weight matrix of each neuron is W _n and the representative value of the sigmoid function is S _n , the influence matrix is O = W ₀ S ₀ W ₁ It becomes S ₁ , and the input / output relationship is estimated by the magnitude of this influence.

Further, in Japanese Patent Application No. 4-66200 (Title of Invention: Learning Method of Neural Network), all sample data are not necessarily learned in a neural network for category recognition such as pattern recognition. A learning method of a neural network that speeds up learning by learning is disclosed.

Further, the following method has been used as a method for deleting unnecessary input items from the input items input to the input layer of the neural network. In the case where the neural network is a three-layer hierarchical neural network in which the number of neurons in the input layer is m, the number of neurons in the intermediate layer is 1, and the number of neurons in the output layer is n, the weight from the i-th input layer neuron to the k-th intermediate layer neuron is W ₁ ( i, k), intermediate layer k
If the weight from the n-th neuron to the j-th neuron in the output layer is W ₂ (k, j), the inverse mapping R from the input i to the output j
_ij is _calculated by the following formula.

[0010]

[Equation 1]

Conventionally, a value such as this inverse mapping is used as an input / output relation value, and an input item i having a small absolute value of the input / output relation value is judged as an unnecessary input item and the input item is deleted. .

[0012]

Conventionally, when the learning should be terminated when the neural network is made to perform learning, for example, when the number of times of learning exceeds a certain set value, or the recognition error is a constant value. The judgment was made in the following cases or using a combination of both.

Further, generally, learning is performed assuming that all the teacher data is correct, but some actual data include inconsistencies. In such a case, it takes a very long time to learn the teacher data. There was a problem. In addition, there is a problem that even if the teacher data is correctly recognized, the correct recognition is not always possible when a slight change is given to the teacher data.

Since the input items of the neural network are determined by human experience and intuition, not all input items used for learning have a great influence on the output value. Therefore, in order to learn the minimum error with the minimum required number of input items, it is necessary to determine unnecessary input items and reduce the number of input items.

However, in the case of deleting an input item by using the inverse mapping, the input item i having a small absolute value of the input / output relation value is judged as an unnecessary input item and the input item i is deleted. . Therefore, since the input value is large even though the input / output relation value is small, there is a problem that the input item having a large influence on the output is deleted. On the contrary, since the input value is small even though the input / output relation value is large, there is a problem that the input item having a small influence on the output remains without being deleted.

Further, when it is difficult to express the input / output relationship of the neural network by a mathematical expression, the user could not understand the input / output relationship. The present invention is
An object of the present invention is to provide a neural network learning device for excluding inappropriate teaching data during learning to improve learning efficiency and for constructing a neural network resistant to noise added to an input.

Further, the present invention provides a learning device of a neural network for deleting unnecessary input items, and also provides a user with a display of input / output relations in an easy-to-understand manner to provide a material for determining input item deletion. Moreover, it is to provide materials for supporting the judgment of the learning situation.

[0018]

FIG. 1 is a block diagram showing a first principle configuration of the present invention. FIG. 1 is a block diagram showing the principle configuration of a learning device for a neural network that gives training data to the neural network and updates the internal weights of the neural network so that the training data can be correctly recognized.

In FIG. 1, the sensitivity representative value calculation means 1 calculates the sensitivity to each teacher data, that is, the representative value of the sensitivity of the output data of the neural network when the input data of the teacher data is slightly changed.
As the representative value of the sensitivity for one teacher data, for example, the maximum sensitivity of the outputs of the output layer neurons when each input data to the input layer neurons is slightly changed is used.

The recognition result correctness degree calculating means 2 calculates the correctness degree of the recognition result for each teacher data. This correctness degree is calculated, for example, by comparing the actual output value of the output layer neuron with the output data of the teacher data.

Further, the learning situation judging means 3 judges whether or not learning should be continued for each teacher data based on the output of the sensitivity representative value calculating means 1 and the output of the recognition result correct answer degree calculating means 2. In the determination, for example, when the values of the representative value of the sensitivity and the correct answer degree are normalized in the range of 0 to 1,
The judgment is performed with 0.5 as the respective boundary value. For example, when the degree of correct answer is 0.5 or more and the representative value of sensitivity is 0.5 or less, it is assumed that the learning for the teacher data has been sufficiently performed, and the learning for the teacher data is terminated.

FIG. 2 is a block diagram showing a second principle configuration of the present invention. FIG. 1 is a block diagram showing the principle configuration of a learning device for a neural network that gives training data to the neural network and updates the internal weights of the neural network so that the training data can be correctly recognized.

In FIG. 2, the reference input value recognizing means 4 receives input reference data and outputs a recognition result for the input data. Reference input value recognition means 4
Is composed of a neural network having the same layer structure as the neural network, and the same weight as the weight is given to each neuron for recognition.

The input / output relationship calculation means 5 is a reference input data, a teacher data, a recognition result of the teacher data input to the neural network, and a reference input value recognition means 4.
The change amount of the output value with respect to the change amount from the reference input value of the teacher data is calculated from the recognition result for the reference data output from the neural network and the weight of the neural network.

The input item determination means 6 determines whether or not to delete the input item based on the amount of change in the output calculated by the input / output relationship calculation means 5. The input item deleting means 7 deletes the data corresponding to the input item of the teacher data when the judgment result of the input item judging means 6, that is, when it is judged that the input item can be deleted.

The input / output relationship display means 8 receives information input / output to / from the input / output relationship calculation means 5, that is, reference input data, teacher data, recognition results of the teacher data input to the neural network, and reference input. The recognition result for the reference data output from the value recognition means 4, the weight of the neural network, or the change amount of the output value with respect to the change amount from the reference input value of the teacher data is displayed, and the input item is displayed to the user. Prompt for deletion decision.

The deletion item input means 9 is for inputting an input item to be deleted, and the user inputs / outputs the relation display means 8.
Refer to the displayed contents of and enter. Input item deleting means 7
Also deletes the data corresponding to the input item to be deleted, which is input by the deletion item input means 9, from the teacher data.

Further, in claim 10, the effective range calculation means refers to each teacher data to obtain and output an effective range for each input item. The data changing means changes at least one of the input items of the teacher data to data within the valid range. The data changing means generally changes the data to the upper limit value or the lower limit value of the effective range, but it is also possible to change the data to any value within the effective range.

The output relation display means displays the recognition result obtained by inputting the changed data to the neural network. Further, in claim 12, the input / output sign relation calculation means calculates and outputs the sign relation between the input and the output from the teacher data and the weight of the neural network.

The qualitative relation input means is for inputting a qualitative relation which is a relation between an input and an output peculiar to the neural network, and the user inputs a value peculiar to the neural network.

The learning status determining means compares the code relationship calculated by the input / output code relationship calculating means with the qualitative relationship input by the qualitative relationship input means to determine the learning status. The learning status display means displays the determination result from the learning status determination means.

[0032]

According to the present invention, it is determined during learning of the neural network whether or not it is necessary to continue learning for each teacher data. As described with reference to FIG. 14, when the output of the output layer neuron is close to 0 or 1, that is, when the sensitivity is low, the output does not change much even if the input is changed, that is, learning does not progress further. At this time, if the recognition is performed almost correctly, the learning for the teacher data may be finished.

That is, the teacher data being learned can be classified into four using the sensitivity of the teacher data and the correct answer degree of the recognition result. In the first case, the sensitivity is low and the accuracy is high. Such teacher data has no room for further learning, and can be judged to have been completely learned.

In the second case, the sensitivity is high and the correct answer is high. In such a case, the output value greatly changes due to the change in the input data, and it is determined that the learning should be continued.

In the third case, the sensitivity is low and the correct answer degree is low. It shows that even if learning does not progress further on such data, correct results are not obtained, that is, it is inconsistent with the already trained data, and such data is trained further. It should not be done and you need to check for inconsistencies.

The fourth case is a case where the sensitivity is high and the correct answer degree is low. In such a case, it indicates that learning is insufficient, and there is a possibility that future learning will reduce the sensitivity and improve the correct answer rate.

As described above, according to the present invention, by using the sensitivity of the teacher data and the correct answer degree of the recognition result, whether or not the learning should be continued with respect to the teacher data, that is, the learning status of the neural network is determined. You can do it.

Further, in the present invention, the necessity of the input item of the teacher data input to the neural network is determined. Reference input data, output data which is a recognition result for the reference input data, teacher data,
The absolute value of the change amount of the output value with respect to the change amount from the reference input value of each teacher data for each input item calculated from the recognition result of the teacher data input to the neural network and the weight of the neural network is the value. Is large, it has a large effect on the output, so it is regarded as a necessary input item. On the contrary, if the absolute value of the change amount of the output value is small, the influence on the output is small, so that the input item is regarded as unnecessary and the data corresponding to the input item of the teacher data is deleted.

As a method of judging deletion, if the value obtained by adding the maximum absolute value of the change amount of the output value and the maximum actual error of the output is within a predetermined allowable error range, the input is made. It is determined that the item can be deleted, the data corresponding to the input item of the teacher data is automatically deleted, and the teacher data is reorganized. Further, if the value obtained by adding the maximum absolute value of the change amount of the output value and the maximum actual error of the output is larger than the predetermined allowable error, it is determined that the input item cannot be deleted. It

As another deletion determination method, the amount of change in the output value is displayed to the user and it is determined whether the user deletes the input item. When the user determines to delete the input item, an instruction to delete the input item is issued, the data corresponding to the input item of the teacher data is deleted, and the teacher data is reorganized.

Further, in the present invention, the effective range is obtained for each input item from the teacher data used for learning of the neural network, and the data corresponding to one input item of the teacher data is set to, for example, the upper limit value of the effective range. The lower limit value is replaced and input to the neural network for recognition. The user can easily understand the effect of each input item on the output by performing recognition after replacing the data for each teacher data and creating and displaying a graph of the recognition result for each upper limit value or lower limit value. You will be able to.

Further, in the present invention, the learning is inconsistent by calculating the qualitative relationship such as the inverse mapping obtained from the sensitivity of the neural network and the weight of the connection and comparing it with the qualitative relationship input by the user. It is determined whether or not there is.

[0043]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 3 is a block diagram of the structure of a first embodiment of a learning device for a neural network according to the present invention. In the figure, the learning device stores a teacher data storage unit 11 for storing teacher data, a learning control unit 12 for controlling learning of a neural network using the teacher data, a layer structure of the neural network, a weight of connection between neurons, and the like. Correctness degree calculation for calculating the correctness degree of the recognition result using the teacher data stored in the neural network storage unit 13 and the teacher data storage unit 11 and the recognition result output from the neural network storage unit 13 for the teacher data Part 14, giving a slight change to the input data of the teacher data,
It is determined whether the learning of the teacher data should be continued using the outputs of the sensitivity calculation unit 15 that calculates the sensitivity to the teacher data using the output of the corresponding neural network, the correct answer degree calculation unit 14, and the sensitivity calculation unit 15. And a learning data changing unit 17 for changing learning data based on the judgment result of the learning condition judging unit 16 and notifying the learning data to the teacher data storage unit 11 and the learning control unit 12. ing.

If the sensitivity is low and the correct answer degree of the recognition result is high, the learning situation judging section 16 judges that the learning for the teacher data should be ended, and the sensitivity is small,
Data having a low correct answer in the recognition result is regarded as having a contradiction, and is transmitted to the learning data changing unit 17 so as to be omitted from the learning target. In other cases, it is determined that the learning is to be continued, and the learning control unit 12 passes through the learning data changing unit 17.
To continue learning the teacher data.

FIG. 4 is a block diagram of the structure of a second embodiment of the learning device for the neural network. Comparing FIG. 3 with FIG. 3 showing the first embodiment, it is different only in that a learning status display unit 18 is used instead of the learning status determination unit 16. The learning status display unit, instead of the learning status determination unit, displays the relationship between the sensitivity of the teacher data and the correct answer degree of the recognition result for the teacher data using, for example, a graph, and the learning data is learned. It assists the user in determining whether or not to continue. The user determines the learning status in accordance with the display on the learning status display unit, and controls the learning data changing unit 17 to change the learning data or to continue the learning of the teacher data.

FIG. 5 is a block diagram of the structure of the third embodiment of the learning device for the neural network. Comparing this figure with the first embodiment, an unlearned data changing unit 20 is provided in place of the learning data changing unit 17, and newly unlearned data, that is, data stored in the teacher data storage unit 11. An unlearned data storage unit 19 for storing unlearned data different from the above is provided, and the correct answer degree calculation unit 14 calculates the correct answer degree of the recognition result of the neural network for the unlearned data instead of the teacher data, and the sensitivity calculation. The difference is that the unit 15 similarly calculates the sensitivity to unlearned data, and the learning status determination unit determines the validity of unlearned data. Similar to the second embodiment of FIG. 4, it is possible to provide a learning status display unit instead of the learning status determination unit 16 to assist the user in determining the learning status.

Next, a method of sensitivity calculation for determining the learning situation, that is, for determining the necessity of continuing learning with respect to certain teacher data will be generally described. After learning once or a plurality of times, the output value and sensitivity are obtained for each learned teacher data. Where x1, ...
.., xm are input values to the input layer neurons, y1, ..
,, yn are the output values from the output layer neurons, f1, ...
,, fn are input / output related functions, s11, ..., snm
Is the sensitivity, and ε is the perturbation applied to the input to obtain the sensitivity.

Y1 = f1 (x1, ..., xm): yn = fn (x1, ..., xm) s11 = (f1 (x1 + ε, ..., xm) -f1 (x1, ・ ・ ・ ・ ・ ・, xm)) / ε: s1m = (f1 (x1, ・ ・ ・ ・ ・ ・, xm + ε) -f1 (x1, ・ ・ ・ ・ ・ ・, xm)) / ε s21 = (f2 (x1 + ε, ・ ・ ・ ・ ・ ・, xm) -f2 (x1, ・ ・ ・ ・ ・ ・, xm)) / ε s2m ＝ (f2 (x1, ・ ・ ・ ・ ・ ・, xm + ε) -f2 (x1, ... ・ ・ ・, xm)) / ε :: sn1 ＝ (fn (x1 + ε, ・ ・ ・ ・ ・ ・, xm) -fn (x1, ・ ・ ・ ・ ・ ・, xm)) / ε: snm = (fn (x1, ... ・ ・ ・, xm + ε) -fn (x1, ... ・ ・ ・, xm)) / ε In the present embodiment, these sensitivities s₁₁, S₁₂・ ・ ・ S_nmNou
The one with the maximum absolute value is obtained as the representative value of sensitivity. Was
However, it is also possible to use the average or median as the representative value.
Is. Note that the input / output relationship for obtaining these sensitivities
Function f of₁, F ₂・ ・ ・ F_nThe formula is not always clear
Depending on the value of the input, the output layer neuron
It indicates that the output value of the input is determined.

FIG. 6 is a flow chart of sensitivity calculation processing for teacher data. When the process is started in the figure, first, in step S21, the maximum absolute value of the network-specific sensitivity to the teacher data is obtained. This maximum value is determined by the network and does not depend on input data. How to obtain this maximum value will be described later.

Subsequently, in S22, the minute change to be applied to each input variable in the input data, that is, the magnitude of the perturbation is determined, one teacher data is selected in S23, and the network for the data selected in S24 is selected. Output is required.

Next, one input variable is selected from a plurality of input variables in S25, a perturbation is added to the value of the selected input variable in S26, and the output of the network for the resulting input data is obtained, and in S27. The output value obtained in S24 described above is subtracted from the obtained output value, and the difference is divided by the magnitude of the perturbation to obtain the sensitivity of each output with respect to the selected input. The process from S25 to S27 is repeated for all of the plurality of input variables, and when the sensitivities to all the input variables are obtained, the representative value of the sensitivity is obtained in S28. Here, it is assumed that the one having the largest absolute value is selected. Finally, the representative values of the sensitivities obtained for all the input variables in S29 are divided by the maximum absolute value of the sensitivity peculiar to the network obtained in S21, and normalized to the range of 0 to 1, S23 to S
The processes up to 29 are repeated for all the teacher data.

Next, the maximum absolute value of the sensitivity peculiar to the network will be described. Here, the maximum value of the absolute value of the inherent sensitivity is obtained in order to use this maximum value in the above-mentioned normalization. The maximum sensitivity depends on the configuration of the neural network. The two-layer network shown in FIG. 7 will be described first.

Sensitivity means that for input x and output y

[0054]

[Equation 2]

(Y (x0 + ε) −y (x
0)) / ε. In the case of a two-layer structure, if the sigmoid function is the transfer function,

[0056]

[Equation 3]

Therefore, the maximum absolute value of sensitivity is (1 /
4) • max | W1 (i, j) |. Next, for the three-layer network shown in FIG. 7, the sensitivity is obtained as follows.

[0058]

[Equation 4]

(The first term on the right side represents the sum of only positive terms, and the second term represents the sum of only negative terms.) 0 <f '(uj) ≤1 / 4

[0060]

[Equation 5]

Similarly,

[0062]

[Equation 6]

It is Therefore, the maximum absolute sensitivity is

[0064]

[Equation 7]

The maximum of all i and j should be obtained. This value is determined by the network and does not depend on the input. The calculation of the sensitivity of the teacher data will be described in more detail using a specific example. FIG. 9 is a specific example of teacher data. In the figure, there are four types of teacher data patterns A to D, _three inputs x ₁ to x 3 and two outputs z ₁ and z ₂ . As the perturbation value added to the input variable, it is necessary to select a small value that does not have much effect on the entire teacher data, and here it is set to 0.01. Then, the sensitivity value having the maximum absolute value is obtained as a representative value of the sensitivity with respect to the teacher data, and the representative value in the range of 0 to 1 is obtained by dividing by the maximum value peculiar to the network.

The sensitivity calculation for the teacher data of FIG. 9 is performed by the following procedures 1) to 5). 1) Find the maximum absolute value of the sensitivity peculiar to the network. This calculation method is as described in FIGS. 2) Of the four target patterns, first pay attention to the pattern A, and output f1 (x1, x2, x
3) and f2 (x1, x2, x3) are calculated. 3) Add ε to x1 and output f1 (x1 + ε, x2, x
3), f2 (x1 + ε, x2, x3) is obtained. Similarly, for x2 and x3, f1 (x1, x2 + ε, x
3), f2 (x1, x2 + ε, x3), f1 (x1, x
2, x3 + ε), f2 (x1, x2, x3 + ε). At this time, the sensitivity s (A, i,
j) is determined as follows. s (A, 1,1) = (f1 (x1 + ε, x2, x3) −
f1 (x1, x2, x3)) / ε s (A, 1, 2) = (f1 (x1, x2 + ε, x3) −
f1 (x1, x2, x3)) / ε s (A, 1, 3) = (f1 (x1, x2, x3 + ε) −
f1 (x1, x2, x3)) / ε s (A, 2, 1) = (f2 (x1 + ε, x2, x3) −
f2 (x1, x2, x3)) / ε s (A, 2, 2) = (f2 (x1, x2 + ε, x3) −
f2 (x1, x2, x3)) / ε s (A, 2, 3) = (f2 (x1, x2, x3 + ε) −
f2 (x1, x2, x3)) / ε 4) In order to obtain the representative value of the sensitivity of the pattern from the six values, the one with the maximum absolute value is extracted. This absolute value is the representative value of the sensitivity of pattern A in this network. In other words, this corresponds to finding the maximum value among the influences of the minute fluctuation of the input of the pattern A on the output. 5) Similarly, for the remaining patterns, a representative value of sensitivity is obtained, divided by the absolute value of sensitivity unique to the network, and normalized.

FIG. 10 shows representative values of sensitivity before normalization and representative values after normalization for the four types of patterns A to D in FIG. Here, the maximum absolute sensitivity of the network is 1.34659.

Next, how to obtain the correct answer degree with respect to the recognition result of the teacher data will be described. The number of output layer neurons is n
And the actual output value for the teacher data is y ₁ ~
If y _n and the output values of the teacher data are z ₁ to z _n , the correct answer degree of the recognition result for the teacher data is given by the following equation.

[0069]

[Equation 8]

[0070] The output value z _{i (i} = 1,2 of the output value y _i and the teacher data of this correct value of the degree of actual neuron, ...
, N) becomes 1 only when they completely match, and approaches 0 as the actual output value and the output value of the teacher data deviate. The accuracy of the correct answer may be calculated by the following formula.

[0071]

[Equation 9]

The values of these expressions are also in the range between 0 and 1, and the closer the recognition result is, the closer to 1. It is possible to use other calculation formulas that have a magnitude relationship with goodness or badness of the recognition result, not necessarily in the range of 0 to 1.

FIG. 11 is an explanatory diagram of a method of determining the necessity of learning the teacher data based on the relationship between the sensitivity of the teacher data and the degree of correct answer of the recognition result. As described above, both the sensitivity and the degree of correct answer are normalized in the range of 0 to 1, and in general, 0.5 can be considered as the boundary value of the judgment criterion. Regarding the correct answer, it is considered that the recognition target by the neural network is pattern recognition, and an error of maximum 0.5 is allowed for each output. Therefore, 0.5 is used as the criterion for determining the correct answer.

As described above, the teacher data having a low sensitivity and a high degree of correct answer, that is, the data is judged to have been learned, and conversely, the teacher data having a high sensitivity and a small degree of correct answer is unlearned, that is, learning should be continued. Judged as data.
On the other hand, it is determined that the data having a small sensitivity and the accuracy of correct answers are inconsistent with other teacher data, and the data having a high sensitivity and the accuracy of correct answers are determined to require further learning.

FIG. 12 is a diagram showing specific sensitivities and correct answers to teacher data. In the figure, four types of patterns E to H are shown as teacher data, and the sensitivity after normalization and the correct answer degree of the recognition result are shown. As described above, since the boundary values of both determinations are 0.5, the pattern E is classified into the data of FIG. 10, that is, the data with contradiction, the pattern F is the unlearned data, the pattern G is the learning required data, and the pattern H. Is classified as trained data.

FIG. 13 is an example of the display result of the learning status display section 18 in the second embodiment of FIG. In the figure, the locus of the relationship between the sensitivity during learning and the degree of correct answer for the teacher data 1 and the teacher data 2 is given. With respect to the teacher data 1, the sensitivity decreases as the learning progresses, and the correct answer degree increases, so that it is determined that the learning is appropriately performed. On the other hand, with respect to the teacher data 2, the sensitivity decreases as the learning progresses, but the correct answer degree of the recognition result also decreases, and it is determined that there is a contradiction with other teacher data, and the learning is continued. It is judged to be unsuitable. Displaying such a learning status can assist the user in making a decision.

Although the sensitivity and the degree of correct answer of 0.5 are used as the criteria for the learning situation determination, the criteria for this determination can be changed if necessary. For example, when high accuracy is required for output, the criterion for the correct answer is
If it is larger than 0.5, and if a large error is likely to occur in the input, it is appropriate to make the sensitivity judgment standard larger than 0.5.

When the number of neurons in the input layer is large, that is, when the number of input variables is large, it is appropriate to set the sensitivity criterion to be smaller than 0.5. The reason is that the maximum value of the above-mentioned network-specific sensitivity is required on the condition that all inputs have a certain value. Therefore, in general, as the number of inputs increases, a value closer to that maximum value is realized. This is because it becomes difficult.

Next, a method of determining the consistency between unlearned data and teacher data in the third embodiment shown in FIG. 5 will be described. As to the unlearned data, as described with reference to FIG. 11, the data having low sensitivity and low correct answer degree are determined to be inconsistent with the teacher data stored in the teacher data storage unit 11. Here, the determination of the consistency with the teacher data will be described using the specific example shown in FIG.

In FIG. 14, four patterns of rising 1, 2 and falling 1, 2 are stored in the teacher data storage unit 11, and it is assumed that learning has already been performed. A ~
It is assumed that the four patterns of D are unlearned data stored in the unlearned data storage unit 19, and the sensitivity and the correct answer degree after normalization of these four patterns are calculated as shown in FIG. And

In FIG. 14, the pattern A has the contradiction of FIG. 11, that is, it is classified as data. That is, although this pattern is very close to the pattern of rising 1 when viewed as input data, the output data is the same as that of falling 1 or 2, which is inconsistent with the teacher data. Patterns B and C are classified as unlearned, and it is determined that they should be added as teacher data to be learned. The pattern D is classified into learned data, and it is determined that it is not necessary to add this pattern as teacher data. FIG. 15 is a block diagram showing the configuration of a fourth embodiment of the learning device for the neural network of the present invention.
In the figure, the learning device stores a teacher data storage unit 31 that stores teacher data, a teacher data learning unit 32 that controls learning of a neural network using the teacher data, a layer structure of the neural network, a weight of connection between neurons, and the like. In the neural network storage unit 33, the teacher input value recognition unit 34 and the teacher input value recognition unit 34 that recognize the teacher data stored in the teacher data storage unit 31 by the layer structure of the neural network storage unit 33 and the weight of the connection. A recognition result storage unit 35 that stores the recognition result,
A reference input value storage unit 36 that stores a reference input value, a reference input value that recognizes the reference input value data stored in the reference input value storage unit 36 based on the layer structure of the neural network storage unit 33 and the weight of the connection. The recognition unit 37, the reference output value storage unit 38 that stores the recognition result in the reference input value recognition unit 37, the teacher data stored in the teacher data storage unit 31, the reference input value data stored in the reference input value storage unit 36, The recognition result stored in the recognition result storage unit 35, the recognition result stored in the reference output value storage unit 38, the layer structure of the neural network storage unit 33, and the change amount of the output value with respect to the change amount of the input value depending on the connection weight. An input / output relationship calculation unit 39 that calculates a relationship, an input item deletion determination unit 40 that determines whether or not an input item is an item to be deleted from the calculation result of the input / output relationship calculation unit 39, The input item deletion unit 41 deletes input items based on the determination result of the input item deletion determination unit 40, and the teacher data reorganization unit 42 reorganizes the teacher data deleted by the input item deletion unit 41. There is.

The reference input value data stored in the reference input value storage unit 36 is reference data for each teacher data input to the teacher data storage unit 31, and is an average value of each teacher data, or These are teacher data and the like used in the previous learning.

The reference input value recognizing unit 37 has a layer structure of a neural network similar to that of the teacher data learning unit 32, and the weight of the connection obtained as a result of the learning in the teacher data learning unit 32 is directly input layer neurons. It is given as a weight between the intermediate layer neurons, the intermediate layer to the intermediate layer, or the intermediate layer to the output layer. The reference input value recognition unit 37 obtains and outputs a reference output value by using the reference input value data and the weight of the connection obtained as a result of the learning in the teacher data learning unit 32.

Next, a method of calculating the amount of change in the output value in the input / output relation calculating unit 39 and a method of judging the deletion of the input item in the input item deletion judging unit 40 will be described below with reference to the flowchart of FIG. However, it is assumed that the neural networks in the teacher input value recognition unit 34 and the reference input value recognition unit 37 are composed of three layers of the number of input layer neurons m, the number of intermediate layer neurons 1, and the number of output layer neurons n.

First, in step S61, the inverse mapping of the input layer neuron i and the output layer neuron j of the neural network is obtained. To obtain the inverse mapping of the output for all input items, the calculation of the inverse mapping R _ij is repeated (m × n) times. Further, if the inverse mapping for a specific input item is to be obtained, it is only necessary to obtain n inverse mappings for one input item. Also, without using the inverse mapping formula,
The maximum absolute value of the sensitivity for each input item described above may be used. That is, the sensitivity of the output layer neuron to the input layer neuron is obtained for each teacher data, and the maximum absolute value of the obtained sensitivity is set to the maximum absolute value of the output layer neuron to the input layer neuron. All combinations (m × from input layer neurons to output layer neurons)
For n), the maximum absolute value of sensitivity is obtained and used instead of the above-mentioned inverse mapping.

Next, in step S62, the reference input value data I _{01 to} I _0m stored in the reference input value storage unit 36.
Is _supplied to the reference input value recognition unit 37, reference output data O _{01 to} O _0n are obtained, and the reference output value storage unit 38 stores them.
The reference input value recognition unit 37 has the same layer structure as the neural network of the teacher input value recognition unit 34, and the weight at the present time learned by the teacher input value recognition unit 34 (information stored in the neural network storage unit 33). ) Is used to calculate. However, the reference input value recognition unit 37 calculates and outputs output data from the reference input value data and the weight data stored in the neural network storage unit 33 without performing learning.

Subsequently, in step S63, the input values I _{11 to} I _1m, which are the first teacher data, to the output values O _{11 to} O.
_{Calculate 1n} . The input value is input to the teacher input value recognition unit 34, and the teacher input value recognition unit 34 performs learning so that the output data has an appropriate value. Therefore, the weight used by the reference input value recognition unit 37 in the previous step S62 is updated every time the teacher input value recognition unit 34 learns. When the learning of the input value is completed, the output value is transferred to the recognition result storage unit 3
Store in 5.

Next, in step S64, the input data I _1i, output data O _1j corresponding to the input data I _1i
Then, the change amount V _ij is calculated from the following equation.

[0089]

[Equation 10]

The input / output relation calculation unit 39 repeats the calculation of the variation amount V _ij for one teacher data (m × n) times. In the equation for _{obtaining the} change amount V _ij , the denominator is the product of the difference between the reference input value data and the input data of the teacher data multiplied by the inverse mapping, that is, the sum of the change amount of the output with respect to the change amount of the input is the output layer neuron j. Seeking for.
The numerator is obtained by multiplying the difference between the reference input value data and the input data of the teacher data by the inverse mapping, that is, the input layer neuron i
The amount of change in the output of the output layer neuron j with respect to the amount of change in the input is calculated. Therefore, it can be said that the change amount V _ij represents how much the change amount of the input data of the input item i affects the change of the output of the output layer neuron j.

The above steps S62 to S64 are executed as follows.
Repeat for the number of teacher data. Next, the input / output relationship calculation unit 39 selects the maximum value max of each of the change amounts V _ij of the teacher data obtained in steps S62 to S64.
(V _ij ) is calculated (step S65).

Next, the input / output relation calculation unit 39 obtains an error E _j of the output j in learning (difference between the output value of the teacher data and the output value after learning) for each teacher data, and the error E _j.
The maximum value of is max (E _j ) (step S66).

Finally, in step S67, it is determined whether or not the input item i should be deleted. When the allowable error of the output j is L _j , the input item deletion determination unit 40 determines the input item i if all the outputs j (1 ≦ j ≦ n) for the input item i satisfy the input / output relationship of the following equation. It is determined to be deleted, and it is determined that the input item i is not deleted if even one is not satisfied.

[0094]

[Equation 11]

If the input item deletion determination unit 40 determines that the input item i is to be deleted, the input item deletion unit 41 deletes the input item i, and the teacher data reorganization unit 42 causes the teacher data other than the input item i of the teacher data to be deleted. Reorganize the data to generate new teacher data. The reorganized teacher data is stored in the teacher data storage unit 31, and the teacher input value recognition unit 34 relearns based on the reorganized teacher data.

FIG. 21 is a diagram showing an example of teacher data. As input data of the teacher data, the temperature is measured three times at regular intervals to obtain temperature 1, temperature 2, and temperature 3. However, as the measured temperature, for example, a value obtained by normalizing the maximum temperature as “1” and the minimum temperature as “0” is used. Further, as output data of the teacher data, whether the measured temperature changes of the temperature 1, the temperature 2, and the temperature 3 have an increasing tendency, a decreasing tendency, or no change is indicated by "1.
21 are added to the teacher data as 0 ”,“ 0.0 ”, and“ 0.5. ”As the teacher data, 11 kinds of data 1 to 11 are shown in FIG.
The respective average values of the temperature 3 and the output data of the teacher data are shown last.

FIG. 22 shows a table created by obtaining the output, the error, and the change amount of the recognition result obtained by using the teacher data shown in FIG. In this embodiment, the maximum change amount max (V ₁₁ ) from the input item 1 to the output is “3.547842”, and the maximum change amount max (V ₂₁ ) from the input item 2 to the output is “0.035236”. The maximum value max (V ₃₁ ) of the change amount from the item 3 to the output is “3.51256”, and the change amount of the input item 2 is the minimum. Therefore, input item 2
Is selected for deletion. Moreover, the maximum value max (E ₁ ) of the error is “0.022505”. Since the output data of the teacher data is of three types of "1.0", "0.0", and "0.5", if 1/2 of the minimum interval of each value is defined as the allowable error L ₁ , It can be set to L ₁ = 0.25.
Therefore, the input / output relation value is max (V ₂₁ ) + max (E ₁ ) = 0.057741, which is smaller than the allowable error L ₁ , and therefore the input item 2 (temperature 2) can be deleted.

FIG. 16 is a block diagram of the fifth embodiment of the learning device of the neural network. 4th figure
15 is different from that of FIG. 15 in that the input item deletion determination unit 40 is replaced with the input / output relationship display unit 43, and the deletion item input unit 44 is added. The input / output relationship display unit 43 displays the input / output data of the input / output relationship calculation unit 39 in place of the input item deletion determination unit 40, for example, using a graph as shown in FIG. 23, and deletes the input item. It assists the user in determining whether or not to do it. The user determines the deletion of the input item in response to the display of the input / output relationship display unit 43, and inputs the input item to be deleted through the deletion item input unit 44. The subsequent processing is similar to that of the fourth embodiment shown in FIG.

FIG. 23 shows a display example of the input / output relation display section 43. FIG. 23A shows the recognition result of the data 11 as a tendency (output) and the reference output value as a tendency (standard output). It is displayed as a band graph. Also, FIG.
In (b), the input value and change amount of the teacher data of the data 11 are displayed, and the change amounts of temperature 1, temperature 2, and temperature 3 are displayed as band graphs. Therefore, the user can understand how much the change amount of the input data of each input item (the difference between the reference input and the input value) influences the change of the output by referring to the change amount. Furthermore, by referring to the change amount of each teacher data, it is possible to find an input item having a small change amount of the output value and easily determine the deletion of the input item.

FIG. 17 is a block diagram of the structure of the sixth embodiment of the learning device for the neural network. Figure 5
Compared with the embodiment described above, instead of the teacher data storage unit 31 and the teacher input value recognition unit 34, the input data storage unit 45 that stores the input data at the time of recognition, and the input data stored in the input data storage unit 45 An input data recognition unit 46 that recognizes the layer structure of the neural network storage unit 33 and the weight of connection is used. Instead of the reference input value storage unit 36, the reference input value recognition unit 37, and the reference output value storage unit 38,
The reference input / output value storage unit 47 is used, and the teacher data learning unit 32, the input item deletion unit 41, the teacher data reorganization unit 42,
The difference is that the deleted item input section 44 is deleted.

The reference input / output value storage section 47 stores a reference input value and a reference output value. The reference input value is the average value of the input data stored in the input data storage unit 45,
The characteristic representative value or the input data used for the calculation of the change amount of the previous output value is used. Further, as the reference output value, an output value at the time of recognition based on the layer structure and the connection weight stored in the neural network storage unit 33 is calculated in advance and stored in the reference input / output value storage unit 47.

The input data recognition section 46 recognizes the input data stored in the input data storage section 45 by the layer structure of the neural network storage section 33 and the weight of the connection, but since the learning is not performed, the weight is not updated. Not performed.

The input / output relation calculation unit 39 calculates the change amount of the output value based on the input data, the reference input value, and the reference output value stored in the input data storage unit 45. The calculation method of the variation of the output value is obtained by the same method as that described in the fourth embodiment. The calculated change amount value of the output value is displayed by the input / output relationship display unit 43 in a format that the user can understand, for example, using a graph as shown in FIG. Further, by updating the input data stored in the input data storage unit 45 and inputting it to the input data recognition unit 46, a new change amount value of the output value is calculated and displayed on the graph. It is possible to know which input value change causes the change in the output value of the neural network from the graph and to understand the input-output correspondence.

In the fifth embodiment, learning is performed every time the teacher data is input, and the weight between the neurons is changed.
In the sixth embodiment, input data is given to a neural network for which learning has already been completed, and the amount of change in output value is obtained. Therefore, by comparing the input / output correspondence with the already learned neural network with the experience and knowledge of the user, it is determined whether the input / output relationship is correct, that is, whether the neural network is effective or not. Is done.

FIG. 18 is a block diagram showing the configuration of the seventh embodiment of the learning device for the neural network. In the figure, the learning device is a teacher data storage unit 3 for storing teacher data.
1. A teacher data learning unit 32 that controls learning of a neural network using teacher data, a layer structure of the neural network, a neural network storage unit 33 that stores connection weights between neurons, etc., valid for each input item of teacher data Effective range calculation unit 4 for finding the range
8. One of the input items of the teacher data from the teacher data storage unit 31 is replaced with the upper limit value and the lower limit value of the effective range output from the effective range calculation unit 48, and the teacher data is stored in the layer of the neural network storage unit 33. The recognition unit 49 is configured to perform recognition based on the configuration and the weight of connection, and an output relationship display unit 50 that displays the output from the recognition unit 49 using a graph as shown in FIG. 24, for example.

The effective range calculating section 48 determines that the effective range is
For example, the maximum value and the minimum value for each input item are obtained from the teacher data and output. The recognition unit 49 uses the teacher data storage unit 3
The data of the input item i of the teacher data from 1 is replaced with the maximum value or the minimum value of the input item i obtained by the effective range calculation unit 48 for recognition, and the output value and the minimum value using the maximum value are used. The output value that was output. Further, the recognition unit 49 performs recognition without replacing the input item i and outputs the recognition result as an actual value. Similarly, if the data of the input item i is replaced with respect to each teacher data or if recognition is performed without replacement, the output value using the maximum value and the output value using the minimum value for each teacher data. , Three kinds of output values of the actual value can be obtained. If the output of the neural network is a monotone increase in which the output value increases with an increase in the input value, or a monotone decrease in which the output value decreases with a decrease in the input value, the output relation display unit 50
When three kinds of output values for each teacher data are arranged and displayed in order, each output value is displayed with a certain width as shown in FIG. The user can know the influence on the output of the input item i by referring to the output graph. That is, if the width of the graph is large, it can be determined that the input item i has a large influence on the output, and conversely, if the width of the graph is small, it can be determined that the input item i has a small influence on the output.

FIG. 19 is a block diagram showing the structure of an eighth embodiment of the learning device for the neural network. In the figure, the learning device is a teacher data storage unit 3 for storing teacher data.
1. A teacher data learning unit 32 that controls learning of a neural network using teacher data, a layer structure of the neural network, a neural network storage unit 33 that stores the weights of connections between neurons, etc. The output sign relation calculation unit 51, the qualitative relation input unit 52 for inputting the qualitative relation between input and output, the output from the input / output sign relation calculation unit 51 and the output from the qualitative relation input unit 52 are compared to determine the learning status. Learning status determination unit 5 for making a determination
3. The learning status display unit 54 displays the output from the learning status determination unit 53.

The input / output code relation calculation unit 51 uses the input item i
And the output j are represented by a sign, for example, the inverse mapping described above and the maximum value of the sensitivity described in the fourth embodiment are calculated. When it is clear that the correlation between the input item i and the output j is either a positive correlation or a negative correlation,
The user inputs the code of the correlation from the qualitative relation input unit 52.

The learning status determination unit 53 includes a qualitative relationship input unit 5 and the code relationship calculated by the input / output code relationship calculation unit 51.
The codes input from 2 are compared, and if the codes are the same, there is no contradiction in learning, and if the signs are opposite, it is determined that there is a contradiction in learning. Alternatively, even if the signs are the same, if the absolute value is small, it is possible to determine that the learning is inconsistent. The learning status display unit 54 displays, for example, the code relationship calculated by the input / output code relationship calculation unit 51, the code input from the qualitative relationship input unit 52, and the determination result.

FIG. 25 shows a display example of the learning status display section 54, which relates to the prediction of the pollution degree of the tunnel. A neural network that predicts the degree of air pollution in the tunnel is constructed from the ventilation volume of the tunnel and the traffic volume of the vehicle. It is clear from the user's experience that the larger the ventilation volume of the tunnel, the smaller the pollution degree, and the larger the traffic volume of the vehicle, the larger the pollution degree.

Therefore, a qualitative relationship is established that the relationship between the ventilation volume of the tunnel and the pollution degree has a negative correlation, and the relationship between the traffic volume of the vehicle and the pollution degree has a positive correlation. Then, the input / output relation value, for example, the value of the inverse mapping is obtained. Since the qualitative relationship of ventilation volume is "-", it is judged that the learning is not performed correctly if the value of the inverse mapping is positive, and the learning is correct if the value of the inverse mapping is negative. It is determined that it is being performed.

Further, since the qualitative relation of the traffic volume is "+", if the value of the inverse mapping is a positive number, it is judged that the learning is correctly performed, and the value of the inverse mapping is a negative number. If so, it is determined that the learning is not performed correctly. If it is determined that the learning is not performed correctly, it is possible that the number of times of learning is insufficient or the teacher data has an error.

[0113]

As described in detail above, according to the present invention, it is possible to determine whether or not to continue learning of the teacher data during learning of the neural network, thus shortening the time required for learning. Can be made. Further, it is possible to construct a neural network that is resistant to a slight change in input, and this greatly contributes to improvement in practicability in application fields such as pattern recognition by a neural network.

Further, according to the present invention, the change amount is obtained based on the change of the output with respect to the change of the input, and the unnecessary input is deleted based on the input / output relation value calculated from the change amount. Can be simplified. Therefore, the amount of calculation and the amount of storage during operation are reduced, and the cost of hardware and the calculation time can be shortened.

Furthermore, by displaying the value of the amount of change, the user can understand the structure of the neural network, the system with excellent maintainability can be constructed, and the accuracy of the neural network can be improved. Become.

[Brief description of drawings]

FIG. 1 is a block diagram (part 1) of the principle configuration of the present invention.

FIG. 2 is a block diagram (part 2) of the principle configuration of the present invention.

FIG. 3 is a configuration block diagram of a first embodiment of a learning device for a neural network.

FIG. 4 is a configuration block diagram of a second embodiment of a learning device for a neural network.

FIG. 5 is a configuration block diagram of a third embodiment of a learning device for a neural network.

FIG. 6 is a flowchart of a representative value calculation process of sensitivity of teacher data.

FIG. 7 is a diagram showing a structure of a two-layer neural network.

FIG. 8 is a diagram showing a structure of a three-layer neural network.

FIG. 9 is a diagram showing a specific example of teacher data.

10 is a diagram showing a representative value of sensitivity with respect to the teacher data of FIG.

FIG. 11 is a diagram illustrating a method of determining the necessity of continuing learning based on the relationship between the sensitivity to teacher data and the degree of correct answer.

FIG. 12 is a diagram showing an example of a representative value of sensitivity for teacher data and a correct answer degree of a recognition result.

FIG. 13 is a diagram showing a display example of a learning status display unit.

FIG. 14 is a diagram for explaining determination of the consistency between learned teacher data and unlearned data.

FIG. 15 is a configuration block diagram of a learning device for a neural network according to a fourth embodiment.

FIG. 16 is a configuration block diagram of a fifth embodiment of a learning device for a neural network.

FIG. 17 is a configuration block diagram of a sixth embodiment of a learning device for a neural network.

FIG. 18 is a configuration block diagram of a seventh embodiment of a learning device for a neural network.

FIG. 19 is a configuration block diagram of an eighth embodiment of a learning device for a neural network.

FIG. 20 is a flowchart of an input item deletion determination method.

FIG. 21 is a diagram showing an example of teacher data.

FIG. 22 is a diagram showing an output, an error, and a change amount for each teacher data.

FIG. 23 is a diagram showing a display example of an input / output relation display section.

FIG. 24 is a diagram showing an effective range of the seventh embodiment.

FIG. 25 is a diagram showing prediction of a pollution degree of a tunnel.

FIG. 26 is a diagram showing an input / output relationship of neurons.

FIG. 27 is a diagram showing an example of a hierarchical neural network.

[Explanation of symbols]

 1 sensitivity representative value calculation means 2 recognition result correct answer degree calculation means 3 learning status determination means 4 standard input value recognition means 5 input / output relation calculation means 6 input item determination means 7 input item deletion means 8 input / output relation display means 9 deleted item input Means 11 Teacher data storage unit 12 Learning control unit 13 Neural network storage unit 14 Correct answer degree calculation unit 15 Sensitivity calculation unit 16 Learning status determination unit 17 Learning data change unit 18 Learning status display unit 19 Unlearned data storage unit 20 Unlearned data change Unit 31 Teacher data storage unit 32 Teacher data learning unit 33 Neural network storage unit 34 Teacher input value recognition unit 34 35 Recognition result storage unit 36 Reference input value storage unit 37 Reference input value recognition unit 38 Reference output value storage unit 39 Input / output relation Calculation part 40 Input item deletion judgment part 41 Input item deletion part 42 Teacher data reorganization Part 43 Input / output relation display unit 44 Deleted item input unit 45 Input data storage unit 46 Input data recognition unit 47 Reference input / output value storage unit 48 Effective range calculation unit 49 Recognition unit 50 Output relation display unit

Claims (12)

[Claims] 1. A learning device for a neural network, which applies training data to a neural network and updates internal weights of the neural network so that the training data can be correctly recognized. Sensitivity representative value calculation means for obtaining a representative value of sensitivity of output data of the neural network when data is slightly changed, and recognition result correct answer degree calculation for calculating the correct answer degree of the neural network recognition result for each of the teacher data. Means for learning, and based on the output of the sensitivity representative value calculating means and the output of the recognition result correct answer degree calculating means, learning status judging means for judging whether or not learning should be continued for each of the teacher data. A learning device for a characteristic neural network. 2. An unlearned data storage unit for storing unlearned data different from the teacher data, the output of the sensitivity representative value calculation unit and the output of the recognition result correct answer degree calculation unit for the unlearned data. The learning device for a neural network according to claim 1, wherein the learning status determination means determines the consistency between the unlearned data and the teacher data based on the following. 3. A learning device of a neural network for giving teacher data to a neural network and updating internal weights of the neural network so that the teacher data can be correctly recognized. Sensitivity representative value calculation means for obtaining a representative value of sensitivity of output data of the neural network when data is slightly changed, and recognition result correct answer degree calculation for calculating the correct answer degree of the neural network recognition result for each of the teacher data. Means, and learning status display means for displaying the relationship between the sensitivity representative value output by the sensitivity representative value calculation means and the correct answer degree output by the recognition result correct answer degree calculation means. Learning device. 4. An unlearned data storage unit for storing unlearned data different from the teacher data, wherein the learning status display unit outputs the sensitivity to the unlearned data by the sensitivity representative value calculation unit. The neural network learning apparatus according to claim 3, wherein the relationship between the representative value and the correct answer degree output by the recognition result correct answer degree calculating means is displayed. 5. A learning device for a neural network, which applies training data to a neural network and updates internal weights of the neural network so that the training data can be correctly recognized, recognizes reference input data, and outputs the result. Reference input value recognition means for outputting, reference input data, the teacher data, recognition results of the teacher data input to the neural network, recognition results output from the reference input value recognition means, and the neural network. Based on the weight of the network, the input / output relation calculation means for calculating the change amount of the output value with respect to the change amount from the input data which is the reference of the teacher data, and And an input item determining means for determining whether or not to delete the input item, Learning device neural network and having an input item deleting means for deleting the data corresponding to the input item of the teacher data based on the determination result. 6. The learning device for a neural network according to claim 5, wherein the reference input value recognizing means is a neural network having the same layer configuration and weight as the neural network. 7. A learning device for a neural network, which applies teaching data to a neural network and updates internal weights of the neural network so that the teaching data can be correctly recognized, recognizes reference input data, and Reference input value recognition means for outputting, reference input data, the teacher data, recognition results of the teacher data input to the neural network, recognition results output from the reference input value recognition means, and the neural network. An input / output relation calculation means for calculating the change amount of the output value with respect to the change amount from the input data which is the reference of the teacher data from the weight of the network, and an input / output for displaying the information input / output to / from the input / output relation calculation means Relationship display means, deletion item input means for inputting input items to be deleted, and deletion item input Learning device neural network and having an input item deleting means for deleting the data corresponding to the input item of the teacher data based on the input stage, the. 8. A learning device for a neural network, which applies teaching data to a neural network and updates internal weights of the neural network so that the teaching data can be correctly recognized, input data, the input data to the neural network. An input / output relation calculating means for calculating a change amount of the output value with respect to a change amount of the teacher data from the reference input data, based on the recognition result of the input data, the reference input / output data, and the weight of the neural network; An input / output relation display means for displaying information input / output to / from the input / output relation calculation means, and a learning device for a neural network. 9. The learning device for a neural network according to claim 8, wherein when the input data is input to the neural network for recognition, the recognition is performed without updating the weight. 10. A learning device for a neural network, which applies teacher data to a neural network and updates internal weights of the neural network so that the teacher data can be correctly recognized. Effective range calculating means for obtaining an effective range, data changing means for changing at least one of the input item data of the teacher data to data within the effective range, and the changed data are input to the neural network. A learning device for a neural network, comprising: output relation display means for displaying a recognition result obtained by the above. 11. The neural network learning apparatus according to claim 10, wherein the data changing unit changes the upper limit value and the lower limit value of the effective range. 12. A learning device for a neural network, which applies training data to a neural network and updates internal weights of the neural network so that the training data can be correctly recognized, wherein the learning data is input from the training data and the weight of the neural network. Input / output code relationship calculating means for calculating a code relationship between the input and output, qualitative relationship input means for inputting a qualitative relationship which is a relationship between the input and output peculiar to the neural network, and the input / output code A learning status judging means for judging a learning status by comparing the sign relation calculated by the relation calculating means with the qualitative relation; and a learning status displaying means for displaying a judgment result of the learning status. Neural network learning device.