JPH05225163A - Neural network system and learning method for neural network - Google Patents

Abstract

PURPOSE:To improve the reliability and the performance of e neural network system by providing an output node value evaluating means and evaluating the output reliability of a neural network for each instance. CONSTITUTION:An unlearned instance deciding neural network 100 of an output node value evaluating part 100 serves as an evaluating neural network to perform the learning with use of the learned data equal to that of a question solving neural network 101 and also to perform the learning with the use of unlearned data. Then the different fixed output node value is outputted based on a fact whether the optional input node value to be inputted is identical with the learned data or the unlearned data. An input node value evaluating part 118 decides whether the optional input node value is sililar or not to the input node value of the learned data used for learning. Therefore it is possible to decide whether the reliable result can be acquired or not from the network 101 with use of the output value 116a and 116b.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network system for solving a problem by using a neural network which has learned the correspondence between given cases, and in particular, is the answer estimated by the neural network reliable? The present invention relates to a neural network system and a learning method for a neural network suitable for efficiently determining whether or not to make a decision.

[0002]

2. Description of the Related Art At present, a neural network can be easily learned by a learning algorithm called backpropagation, and is widely applied to pattern recognition, voice recognition synthesis, signal processing, knowledge processing and the like. There is. That is, in the neural network, the value of a target item that is related to them but whose relationship is not concrete is estimated based on some key items, but this estimation is performed accurately. In order to input the value of a known key item and the value of the target item corresponding to this key item to a neural network having an input node corresponding to the key item and an output node corresponding to the target item, The learning is performed as the node value and the teacher data value (learning data) of the output node corresponding thereto. In this way, in the neural network after sufficient learning, if the value of the key item which is the basis of estimation is given to the input node, the value of the output node at this time becomes the target item value to be estimated. Hereinafter, such a neural network will be described with reference to FIGS.

FIG. 8 is an explanatory diagram showing the structure of a neuron, which is a constituent element of a neural network. Neuron 8
Reference numeral 01 denotes some input terminals 802 for inputting input signals from the outside of the neural network or output terminals of other neurons, and output terminals for outputting output signals to the input terminals of other neurons or the outside. And a terminal 803. Each neuron has, as its own parameters, one bias value and a weight value corresponding to the number of input terminals connected to other neurons,
And one input / output characteristic function. The output signal of each neuron having such a configuration is determined by adding the value obtained by multiplying the value of each input terminal by each weight value and the bias value, and applying it to the input / output characteristic function. .. Then, each neuron learns a required input / output characteristic by adjusting a weight value and a bias value at the time of learning.

FIG. 9 is a block diagram showing the structure of a neural network using the neurons shown in FIG. A neuron that directly receives an input from the outside of the network is called an input node 901, and a vector of each value given to the input terminal of this input node 901 is called an input node value. A neuron that directly outputs to the outside of the network is called an output node 902, and this output node 902
The vector of each value appearing at the output terminal of is called the output node value. Furthermore, neurons other than the input node 901 and the output node 902 are called intermediate nodes 903, and the vector of each value appearing at the output terminal of this intermediate node 903 is called the intermediate node value. The application of a neural network is to give the input node value to the neural network, and sequentially output the values of the output terminals of the neurons with the same values of the input terminals until the output node value is determined. To obtain the output node value determined in this way as a result output value. The learning data consists of a pair of a teacher input value and a teacher output value. Then, when the teacher input value is used as the input node value and applied to the neural network, the value to be output as the output node value is the teacher output value. Further, learning is performed by applying the bias value of each neuron and the weight so that the output node value obtained by applying the teacher input value of the learning data to the network becomes a value close to the paired teacher output value. It is done by adjusting the values.
Incidentally, for a feedforward type neural network in which the connection between neurons does not form a loop, for example, an error back propagation method or the like is known as a specific method thereof.

FIG. 10 is a block diagram showing the structure of a processing system using the neural network shown in FIG.
The central processing unit 1001 executes programs such as conversion processing and data processing. The storage device 1002 stores data, the state of the neural network, various programs, and the like. Then, the input device 1003 inputs each data and inquires about each process. Further, the output device 1004 displays an inquiry about each process and an output result.

In the neural network system having the above-mentioned configuration, in order to guarantee that the correspondence between the key item and the target item obtained by the learned neural network gives a correct result even in an unknown case. In the past, a known case that was not used for learning is input to the neural network after learning, and the degree of agreement between the output value obtained as a result and the case value is examined. The method was used. Such conventional techniques include, for example, "Stock Market Prediction With Modular Neural Networks" by T, Kimoto, K, Asakawa, M, Yoda, M, Takeoka, Proceeding of IJCNN '.
90 (1990) Volume 1, pages 1 to 6 (Kimoto,
T., Asakawa, K., Yoda, M., Takeoka, M., "Stock Marke
t Prediction System with ModularNeural Networks "/
Proc. Of IJCNN '90, vol.1, pp.1-6, 1990), there is an example in which a neural network is evaluated using a correlation coefficient.

In the case of making a decision with a large risk based on the answer given by such a neural network, if the reliability of each case can be evaluated, it will be useful information for a person who makes a decision. .. That is, if all of the key items to be input are key items used for learning input, the output target item is sufficiently reliable, but some of the key items to be input are not used for learning. If it is a key item, the output target item is 100
% Not reliable. However, the conventional technique merely evaluates the statistical reliability of a set of cases, such as the error rate for all cases.

[0008]

The problem to be solved is that, in the technology related to the conventional neural network, the reliability of the value of each target item obtained corresponding to the value of each key item is That is, it cannot be evaluated and cannot be applied when making a determination that requires the reliability of an output item close to 100%. The object of the present invention is to solve these problems of the prior art and to enable evaluation of the reliability of the output of the neural network for each case.
Neural network system and neural network that can be usefully utilized and can improve system reliability and performance even when making a judgment with a large risk that requires a reliability close to 0%. Is to provide a learning method.

[0009]

In order to achieve the above object, a neural network system and a learning method for a neural network according to the present invention include (1) an input node consisting of a plurality of neurons, an intermediate node, and an output. A neural net composed of nodes is provided, and for this neural network, learning data is used in advance so that the correct output node value corresponding to the input node value input to the input node is output to the output node. In the neural network system that learns to, and outputs a more correct output node value corresponding to an arbitrary input node value, based on the arbitrary input node value input to the input node of the neural network, this arbitrary input node value Providing an output node value evaluation unit that evaluates the certainty of the output node value corresponding to the value It is characterized in. Also, (2)
In the neural network system according to (1) above, the output node value evaluation unit includes an input node, an intermediate node, and an input node, an intermediate node, and an output node corresponding to the input node, the intermediate node, and the output node of the neural network, respectively. , Learning using the same learning data as the neural network and learning using unlearned data different from this learning data, inputting any input node value input to the neural network, and inputting this arbitrary input data Equipped with an evaluation neural net that outputs different constant output node values depending on whether the value is learning data or unlearned data. Based on the output node value from this evaluation neural network, an arbitrary input node value can be set. It is characterized in that the certainty of the corresponding output node value is evaluated. (3) In the neural network system according to (1) or (2), the output node value evaluation unit is composed of an arbitrary number of neurons added to the output node of the neural network. , An output node for evaluation for performing learning of a neural network by learning data and learning of a neural net by unlearned data other than this learning data with different output node values, and an output node from this evaluation output node It is characterized in that the certainty of the output node value corresponding to an arbitrary input node value is evaluated based on the value. Also, (4)
In the neural network system according to (2) or (3) above, the output node value evaluation unit uses the learning data used for learning of the neural network to evaluate the evaluation neural network and the output node. The present invention is characterized by including an evaluation learning data generation unit that generates learning data used for learning and unlearned data. (5) In the neural network system according to any one of (1) to (4), the output node value evaluation unit uses an arbitrary input node value input to the input node of the neural network for learning. It is characterized by comprising an input node value evaluator for judging whether the input node value of the learning data is similar to the input node value and evaluating the certainty of the output node value with respect to an arbitrary input node value. .. (6) In the neural network system according to (5), the input node value evaluation unit includes a learning input node value storage unit that stores all components of all input node values of learning data used for learning. , An input node value discriminating unit for discriminating whether or not all components of an arbitrary input node value to the neural network correspond to the components stored in the learning input node value storing unit, It is characterized in that the certainty of the output node value with respect to an arbitrary input node value is evaluated based on the discrimination result of the value discriminating unit. (7) In the neural network system according to any one of (1) to (6), the output node value evaluation unit is an intermediate node corresponding to an arbitrary input node value input to an input node of the neural network. An intermediate node value evaluator that determines whether the value is similar to the intermediate node value at the time of learning using the learning data and evaluates the certainty of the output node value with respect to any input node value. It is characterized by having. Also, (8) above (7)
In the neural network system described in (1), the intermediate node value evaluation unit stores a learning intermediate node value storage unit that stores all intermediate node values at the time of learning using learning data, and an arbitrary input node value to the neural network. An intermediate node value discriminating unit for discriminating whether or not all the components of the corresponding intermediate node value correspond to the components stored in the learning intermediate node value storage unit, and the discrimination of the intermediate node value discriminating unit Based on the result, the certainty of the output node value with respect to the arbitrary input node value is evaluated. Also,
(9) In the neural network system according to any one of (1) to (8), based on the evaluation result of the uncertainty of the output node value by the output node value evaluation unit, a warning to the user is arbitrarily included in advance. It is characterized by providing an improper output time response processing unit that performs a predetermined corresponding internal process. (10) In the learning method of the neural network in the neural network system according to (4), the evaluation learning data generation unit generates learning data and unlearned data each having a constant value with different output node values. A learning method for a neural network, wherein each of the generated neural net for evaluation and the output node is learned by using learning data and unlearned data each having a different constant output node value.

[0010]

According to the present invention, when the neural network is applied, it is determined whether or not the output node value is as reliable as the value output based on the learning data. For example, apart from the neural net that actually solves the problem,
A neural network having the same structure as this neural network is constructed. Then, for all input node values of the same learning data, a certain fixed value (for example, "0.0") corresponds as the teacher data value of the output node, and in addition to the above learning data, the input node The newly constructed neural net is trained by the learning data in which another constant value (for example, “1.0”) corresponds to the input data that is the boundary of the appropriate space that can take values as the teacher data value. .. Then, at the time of estimation for solving the problem, the value of each key item is given to this new neural network as an input node value, and the output node value of this new neural network at this time (for example, “0.8 ",
"0.9" or the like) is the reliability of the estimated value obtained from the given key item value. Further, when the input node value given at the time of application exceeds a predetermined range, the output of the problem solving neural network may be considered unreliable. Alternatively, the learning data input value is applied to the problem solving neural network in which the learning has converged, and the statistical information of the intermediate node value at that time (for example, the maximum value and the minimum value, or the average value and the variance) is recorded. Every time, when applying the neural network, compare the value of the intermediate node value with the stored statistical information at the time of learning, and if it is out of the appropriate range, the output of the problem solving neural network for that input value is not reliable. You can also in this way,
Based on the input key item value, the degree of reliability of the output target item value can be determined, which is useful information for a user who makes a high-risk decision.

[0011]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a first embodiment of a configuration according to the present invention of a neural network system embodying the present invention. The neural network system of the first embodiment determines whether the problem solving neural network 101 for solving a given problem and the answer given by the problem solving neural network 101 are reliable. And an output node value evaluation unit 100 for doing so.
Further, the output node value evaluation unit 100 determines the unlearned case determination neural network 110 as the evaluation neural network of the present invention in order to determine whether the answer of the problem solving neural network 101 is highly reliable. An unlearned case determination teacher input / output value creation unit 109 as an evaluation learning data generation unit according to the present invention, an input node value evaluation unit 118 according to the present invention, and an intermediate node value evaluation unit 119,
An unlearned case discrimination teacher input value storage area 113 for storing learning data generated by the unlearned case discrimination teacher input / output value creation unit 109 and an unlearned case discrimination teacher output value storage area 114 are provided. is doing.

The problem solving neural net 101 and the unlearned case discrimination neural net 110 are provided with input value conversion units 102 and 111 and output value conversion units 103 and 112, respectively. 1
11 is a neural network 101 for problem solving at the time of learning.
When the unlearned case discrimination neural net 110 reads data from the problem solving teacher input value storage area 104 and the unlearned case discrimination teacher input value storage area 113, respectively, the values are read as the problem solving neural network. 101 and a conversion for normalizing into a range suitable for input to the unlearned case discrimination neural net 110. Also, when applied, the same conversion is performed on the applied input value 115 read by each neural network. Further, the output value conversion units 103 and 112 are configured such that, at the time of learning, the problem solving neural net 101 and the unlearned case discrimination neural network 110 respectively have a problem solving teacher output value storage area 105 and an unlearned case discrimination. Teacher output value storage area 11
When reading the data from No. 4, the value is converted into an appropriate range as an output value of the problem solving neural network 101 and the unlearned case discrimination neural network 110, and is converted. Also, when applied, inverse conversion is performed on the output node value output by each neural network.

The intermediate node value evaluation unit 119 stores the statistical information on the intermediate node value of the problem solving neural network 101 during learning as a learning intermediate node value statistical information storage area as a learning intermediate node value storage unit of the present invention. Based on the contents stored in the learning intermediate node value statistical information collection processing unit 106 stored in 107 and the learning intermediate node value statistical information storage area 107, the intermediate node values of the problem solving neural network 101 being applied are stored. It is composed of an intermediate node value check processing unit 108 as an intermediate node value discriminating means of the present invention for checking whether or not it is different from the statistical information. Also,
Although not shown in the figure, the input node value evaluation unit 118, like the intermediate node value evaluation unit 119, also has a learning input node value storage unit that stores statistical information related to the input node value being learned, and an application when applied. It is configured by an input node value determination unit that checks whether the input value 115 deviates from the stored statistical information and outputs the result. The neural network system of the first embodiment is provided with the unlearned case discrimination neural net 110, the input node value evaluation unit 118, and the intermediate node value evaluation unit 119 as the output node value evaluation unit. However, a configuration in which each is individually provided may be used. The operation of the neural network system of the first embodiment according to the present invention will be described below.

First, the operation of the unlearned case discrimination neural net 110 will be described. The problem solving neural network 101 includes a problem solving teacher input value storage area 104.
And the problem-solving teacher output value storage area 105, the corresponding learning data is read, the corresponding relationship is learned, and the applied input value 115 is applied to the input node and the result is output to the output node. To get The unlearned case discrimination teacher input / output value creation unit 109 creates learning data for the unlearned case discrimination neural network 110 by the procedure described later with reference to FIG. 113 and an unlearned case discrimination teacher output value storage area 11
Remember it in 4. The unlearned case discrimination neural network 110 includes a teacher input value storage area 113 for unlearned case discrimination.
The corresponding learning data is read from each of the unlearned case discrimination teacher output value storage area 114 and the correspondence relationship is learned. That is, the problem solving neural network 1
For all input node values of the learning data used by 01, a certain constant value (for example, “0.0”) is set as the teacher data value of the output node, The unlearned case discrimination neural net 110 is trained with the learning data corresponding to another constant value (for example, “1.0”) as the teacher data value for the input data at the boundary. The problem solving neural net 101 and the unlearned case discriminating neural net 110, which have been learned in this way, have the applied input value 1
15 by applying 15 to the corresponding input node, resulting output value 117 for applied input value 115,
A determination result 116a for determining whether or not the applied input value 115 is included in the learning data is output to each corresponding output node. Here, if the given input node value exists in the learning data, the unlearned case discrimination neural net 110 has a predetermined output node value (for example, “0.
0 ”), but data that is the boundary of a proper space that can be taken by input node values given as other learning data, has another output node value (for example,“ 1.0 ”). In other cases, for example, if the relationship between the respective values of the output nodes with respect to the input node is always continuous, the farther away the given input node value is from the input node value group of the learning data, the more The node value behaves so as to be close to the latter value (for example, "1.0"). In the unlearned case discrimination neural net 110, when the value input at the time of application already exists as learning data or there is a value close to the learning data, the output from the problem solving neural net 101 It can be determined that the value is an output result with high reliability. As a result, when the problem solving neural net 101 wants to estimate the value of the target item with respect to a certain value of the key item, the values of those key items are similarly learned to the unlearned case discrimination neural net 110 that has been learned. The output node value of the unlearned case discrimination neural net 110 at this time can be used as the input node value, and can be used as the reliability of the estimated value obtained from the given key item value.

Next, the operations of the intermediate node value evaluation unit 119 and the input node value evaluation unit 118 will be described. In the intermediate node value evaluation unit 119, the learning intermediate node value statistical information collection processing unit 106 stores statistical information about the intermediate node values of the problem solving neural net 101 during learning in the learning intermediate node value statistical information storage area 10.
7 and the intermediate node value check processing unit 108 applies the applied input value 115 to the input node of the problem solving neural network 101 at the time of application, and the intermediate node value statistical information at the time of learning. Storage area 107
It is checked whether or not it is different from the statistical information stored in 1., and the result is output to the determination result 116c. As a concrete learning intermediate node value statistical information processing, for example, the maximum value and the minimum value obtained at the time of learning for each intermediate node are stored, and at the time of the check processing, each intermediate node value is between the values. Check whether or not, and similarly, store the average value and the variance value,
It can be considered to check whether or not it is in a significant area. Further, also in the input node value evaluation unit 118, similar to the intermediate node value evaluation unit 119, statistical information regarding the input node value of the problem solving neural network 101 during learning is stored in a learning input node value storage (not shown). The input node value determination unit (not shown) stores the input node value at the time of application to check whether the input node value deviates from the stored statistical information, and outputs the result to the determination result 116b. As a result, if the input node value or intermediate node value at the time of application deviates from the distribution of the input node value or intermediate node value at the time of learning, the result of the output node value at that time will be As a result, high confidence can be disappointing.

As described above, the output value of the unlearned case discrimination neural net 110 and the intermediate node value check processing unit 1
08, and the input node value evaluation unit 11 for the input node values converted by the input value conversion units 102 and 11.
Using at least one of the output values of the check function of 8,
From the problem solving neural net 101, it is possible to determine whether or not a reliable result has been obtained.

FIG. 2 is a flow chart showing an embodiment of the processing operation according to the present invention of the unlearned case discrimination teacher input / output value creating section in FIG. The flowchart of the present embodiment shows a procedure for creating learning data for the unlearned case discrimination neural net 110 in FIG. 1. First, the problem solving teacher input value storage area 104 in FIG. The teacher input value is read (step 501), and the read data is written to the unlearned case discrimination teacher input value storage area 113 of FIG. 1 (step 502). Then, corresponding to each of the writing problem solving teacher input values, the unlearned case discrimination teacher output value storage area 11 of FIG.
"0.0" is written in 4 (step 503). Check whether all data has been read (step 50
4) If not completed, the process returns to step 501 and the above procedure is repeated. If the reading is completed in step 504, a value corresponding to the boundary of the input space is generated (step 505), and the generated value is written in the unlearned case discrimination teacher input value storage area 113 in FIG. 1 (step 506). ). Corresponding to the written value, “1.
"0" is written (step 507). It is checked whether or not an appropriate end condition is satisfied (step 508). If not satisfied, the process returns to step 505, and if the condition is satisfied, this procedure is ended. The termination condition of step 508 differs depending on the input space boundary data generation unit. For example, when the input boundary space is a multidimensional rectangular area, it is possible to sequentially take the grid points as the input boundary space data, and in that case, the end condition is to obtain all the grid points. Conceivable. Further, it may be possible to randomly select some of the grid points instead of generating all the grid points as the other input space boundary data generation unit. The termination condition in this case may be a condition that the number of obtained grid points exceeds a predetermined number.

FIG. 3 is an explanatory view showing an embodiment of the data structure in the problem solving teacher input value storage area and the problem solving teacher output value storage area in FIG. The problem solving teacher input / output value table 301 is stored in the problem solving teacher input value storage area 104 of FIG. 1, and includes an input value portion 302 including a plurality of fields corresponding to the problem solving teacher input value, and FIG.
Is stored in the problem solving teacher output value storage area 105 and is composed of an output value portion 303 including a plurality of fields corresponding to the problem solving teacher output values. At this time,
Each field group of the same record (No. 1 to n in the figure) has a problem solving neural net 10 shown in FIG.
When learning 1, the pair of input values (input value (1), input value (2), ...) And the output value (output value (1), output value (2)) which is the teacher value thereof. , ...).

FIG. 4 is an explanatory diagram showing an embodiment of the data structure in the unlearned case teacher input value storage area and unlearned case teacher output value storage area shown in FIG. The unlearned case discrimination teacher input / output value table 401 is stored in the unlearned case teacher input value storage area 113 of FIG. 1, and the input value portion 402 is composed of a plurality of fields corresponding to the unlearned case discrimination teacher input values. And the unlearned example teacher output value storage area 11 of FIG.
4 and a discriminant value 403 composed of a field corresponding to a teacher output value for discriminating an unlearned case. At this time, the same record (No1 to n + 2 ... in the figure)
Each of the field groups of ()) corresponds to a pair of input values (input value (1), input value (2), ...) When learning the unlearned case discrimination neural net 110 of FIG. , Represents the output value that is the teacher value. At this time, the value of the input value portion 402 composed of a plurality of fields corresponding to the unlearned case discrimination teacher input value is the same as that existing in the input value portion of the problem solving learning data (No. 1 to n in the figure). , There are two types of boundary values of the input space (Non + 1, n + 2, etc. in the figure), and the value of the discriminant value 403, which is a field corresponding to the unlearned case discrimination teacher output value, is
A value (“0.0”, “1.
0 ”).

FIG. 5 is a block diagram showing an embodiment of the configuration of a host system using the neural network system shown in FIG. In the neural network system of FIG. 1, when the obtained reliability is low, the host system of this embodiment gives a warning to the user and prompts the corresponding process. In this embodiment, the input value processing unit 503
Is applied manually by the user from the input terminal 505 or by sequentially reading data from the input data storage area 506 to obtain the applied input value 115. For example, the unlearned case discrimination as shown in FIG. A possible neural network system 501 is provided. The output value processing unit 504 outputs the result output value 11 output from the neural network system 501.
7 is displayed on the output terminal 507 or is sequentially written in the output data storage area 508. In this case, together with the result output value 117, the applied input value 115 at that time and the determination result 116 of the unlearned case determination may be displayed on the output terminal 507 or written in the output data storage area 508. The input / output control unit 502 is the inappropriate output handling unit of the present invention, and the determination result 11 of the unlearned case determination output by the neural network system 501.
Based on the value of 6, it is determined whether or not a reliable result output value 117 has been obtained for the given applied input value 115. When the result output value 117 for the applied input value 115 is determined to have low reliability based on the value of the determination result 116,
The input / output control unit 502 outputs the result to the output terminal 507 via the output value processing unit 504, and prompts the user to warn the output value processing unit 504. Through this, normal display is suppressed and writing to the output data storage area 508 is suppressed. Further, the input / output control unit 502 temporarily redoes the input from the user's input terminal 505 or suspends the reading from the input data storage area 506 via the input value processing unit 503, and the user Processing such as making an inquiry as to whether or not to continue the processing of is performed. The determination result 116 may be given as a binary value indicating whether the reliability is acceptable or not, and may be given as a degree. If the determination result 116 is given as a degree, the value is a predetermined constant value. The judgment is made depending on whether or not it exceeds. Further, as shown in FIG. 1, the unlearned case discrimination neural network 11 and the input node value evaluation unit 11 are also provided.
8. If some determination results by the intermediate node value evaluation unit 119 and the like are obtained and it is determined by one or more methods among them that it is not reliable, it can be determined that it is not reliable as a whole.

FIG. 6 is a block diagram showing a second embodiment of the configuration according to the present invention of the neural network system according to the present invention. The neural network system of the second embodiment, for example, evaluates the present invention by correcting the structure of the problem solving neural network 101 in FIG. 1 and determining whether or not the data input to the output node has been learned. An unlearned case discrimination output node (described as Ot _{2 in the} figure) 606 is added as an output node for use, and for a known input node value, a target value corresponding to the input node value and learned For a set of a specific value for data discrimination (for example, “0.0”) and the input node value of the appropriately generated input value proper range boundary value, the teacher value of the discrimination node is set to A value that can be distinguished from a specific value (for example,
By using a combination of "1.0") as learning data, the value of the unlearned case discrimination output node 606 newly increased when the neural network is applied is used as the reliability for the applied input value at that time. Is. The configuration and operation will be described below. In the neural network system of the second embodiment, the teacher input / output value creation unit 602 reads learning data from the problem solving teacher input value storage area 601 and the problem solving teacher output value storage area 603, and stores the teacher input value storage. Data is written to each of the area 604, the teacher output value storage area 605, and the unlearned case determination teacher output value storage area 610 in a correspondence relationship described later with reference to FIG. 7. The unlearned case discrimination node-attached neural network 611 reads data from the teacher input value storage area 604, the teacher output value storage area 605, and the unlearned case discrimination teacher output value storage area 610, respectively, and establishes the corresponding relationship. learn.
In this learning, especially, the unlearned case discrimination output node 6
For learning of 06, a set of a known input node value, a target value corresponding to the input node value, and a specific value (for example, “0.0”) for discriminating learned data is appropriate. For the input node value of the input value proper range boundary value generated in
The value of the teacher of the discrimination node is combined with a set of values (for example, “1.0”) that can be distinguished from the specific value described above and used as learning data.

The neural network 611 with an unlearned case discrimination node that has been learned in this way has the applied input value 6
12 is applied by inputting 12 into the corresponding input node group (indicated as In in the figure) 615 through the input value conversion unit 607, and the result output value 61 for the applied input value 612 is applied.
4 and a discrimination result 613 for discriminating whether or not the applied input value 612 is included in the learning data, respectively, as a problem solving output node group (described as Ot _{1 in the} figure) 616 and an unlearned case discrimination output node. From 606, the output value conversion unit 608,
It is output via the discriminant value converter 609. Thus, if the neural network 611 with an unlearned case discrimination node is used, the problem solving output node group 61 is generated based on the discrimination result 613 output from the unlearned case discrimination output node 606.
It is possible to determine whether or not the result output value 614 output from 6 has high reliability. The input value conversion unit 6
07 is a neural network 6 with an unlearned case discrimination node
When reading data from the teacher input value storage area 604 at the time of learning 11, the value is converted into a value corresponding to the unlearned case discrimination node-attached neural network 611. further,
Also when applied, the same conversion is performed on the applied input value 612 read by the neural network with unlearned case discrimination node 611. The output value conversion unit 608 reads the data from the teacher output value storage area 605 at the time of learning the neural network 611 with unlearned case discrimination node, and sets the value to a value corresponding to the neural network 611 with unlearned case discrimination node. Convert. Further, at the time of application, the value of the problem solving output node group 616 output by the neural network with unlearned case discrimination node 611 is inversely transformed to obtain a result output value 614. The discriminant value converter 609 reads the data from the unlearned case discriminant node neural network 611 when the data is read from the unlearned case discriminant teacher output value storage area 610 during learning of the unlearned case discriminant node neural network 611. Convert to the value corresponding to. When applying the unlearned case discrimination node-attached neural network 611, the unlearned case discrimination node-attached neural network 61 is used.
The value of the unlearned case discrimination output node 606 output by 1 is inversely transformed to obtain a discrimination result 613.

FIG. 7 is an explanatory diagram showing an embodiment of the data structure in the teacher input value storage area, the teacher output value storage area, and the unlearned example teacher output value storage area in FIG. A neural network teacher input / output value table 701 with an unlearned case discrimination node is stored in the teacher input value storage area 604 of FIG.
An input value portion 702 including a plurality of fields corresponding to the teacher input value, an output value portion 703 including a plurality of fields corresponding to the teacher output value stored in the teacher output value storage area 605 of FIG. The learning value is stored in the unlearned case discrimination teacher output value storage area 610, and is constituted by a discrimination value 704 including a field corresponding to the teacher discrimination value. At this time, each field group of the same record (No1 to n + 2 ... In the figure) becomes a pair of input values (inputs) when learning the unlearned case discrimination node-attached neural network 611 of FIG. Value (1), input value (2) ...),
The problem solving output node group 616 of FIG.
Output values (output value (1), output value (2) ...) of FIG.
Represents the output value of the unlearned case discrimination output node 606. It should be noted that the value of the input value portion 702 consisting of a plurality of fields corresponding to the teacher input value for determining an unlearned case is input in the input value portion 702 (No1 to n) of the learning data for problem solving. Boundary value of space (Non + 1, n +
2)), and the value of the discriminant value 704, which is a field corresponding to the output value of the unlearned case discrimination output node 606 in FIG. 6, can distinguish these two types of input values. It has a value (“0.0”, “1.0”).
Based on the neural network teacher input / output value table 701 with the unlearned case discrimination node, the neural network 611 with unlearned case discrimination node shown in FIG. 6 outputs the result output value corresponding to the input value when applied. It is possible to output information for determining the reliability of the.

As described above with reference to FIGS. 1 to 7, in the neural network system and the learning method of the neural network of this embodiment, for a certain input node value,
A neural net that outputs the distribution of the key items of the learning data and the information indicating the degree of similarity with the input node value is constructed. This allows the user to easily determine whether or not the result obtained by the neural network system is sufficiently reliable. When the result is that the reliability is low, the user can discard the predicted data, re-learn, or re-examine the input data. The present invention is
As noted in FIG. 1 and the like, it is not limited to the embodiment described with reference to FIGS. For example, the present invention is not limited to the one using the neural network having the hierarchical structure shown in this embodiment. Further, for example, in the evaluation of the output value based on the distribution of the input node values, it is possible to use the value before conversion in the input value conversion unit.

[0025]

According to the present invention, the reliability of the output of the neural network is evaluated for each case.
Even when making a judgment with a great risk that a reliability close to 0% is required, it can be beneficially used, and the reliability and performance of the neural network system can be improved.

[0026]

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of a configuration according to the present invention of a neural network system according to the present invention.

FIG. 2 is a flowchart showing an embodiment of the processing operation of the teacher input / output value creation unit for unlearned case discrimination in FIG. 1 according to the present invention.

FIG. 3 is an explanatory diagram showing an embodiment of a data structure in a problem solving teacher input value storage area and a problem solving teacher output value storage area in FIG. 1;

4 is an explanatory diagram showing an example of a data structure in a teacher input value storage area for an unlearned example and a teacher output value storage area for an unlearned example in FIG. 1. FIG.

5 is a block diagram showing an embodiment of a configuration of a host system using the neural network system in FIG.

FIG. 6 is a block diagram showing a second embodiment of the configuration of the neural network system according to the present invention according to the present invention.

FIG. 7 is an explanatory diagram showing an example of a data structure in the teacher input value storage area, the teacher output value storage area, and the unlearned example teacher output value storage area in FIG. 6;

FIG. 8 is an explanatory diagram showing a configuration of a neuron that is a component of a neural network.

9 is a block diagram showing a configuration of a neural network using the neurons in FIG.

10 is a block diagram showing a configuration of a processing system using the neural network in FIG.

[Explanation of symbols]

100 output node value evaluation unit 101 problem solving neural net 102 input value conversion unit 103 output value conversion unit 104 problem solving teacher input value storage area 105 problem solving teacher output value storage area 106 learning intermediate node value statistical information collection processing Part 107 Learning intermediate node value statistical information storage area 108 Intermediate node value check processing part 109 Unlearned case discrimination teacher input / output value creation section 110 Unlearned case discrimination neural net 111 Input value conversion section 112 Output value conversion section 113 Not Teacher input value storage area for learning case determination 114 Teacher output value storage area for unlearned case determination 115 Applied input value 116, 116a, 116b, 116c Discrimination result 117 Result output value 118 Input node value evaluation unit 119 Intermediate node value evaluation unit 301 Problem solving teacher input / output value table 302 Input value part 03 Output value part 401 Teacher learning input / output value table for unlearned case example 402 Input value part 403 Discriminant value 501 Neural network system 502 Input / output control part 503 Input value processing part 504 Output value processing part 505 Input terminal 506 Input data storage area 507 Output terminal 508 Output data storage area 601 Problem solving teacher input value storage area 602 Teacher input / output value creation unit 603 Problem solving teacher output value storage area 604 Teacher input value storage area 605 Teacher output value storage area 606 Unlearned case discrimination output Node 607 Input value conversion unit 608 Output value conversion unit 609 Discrimination value conversion unit 610 Teacher output value storage area for unlearned case discrimination 611 Neural net with unlearned case discrimination node 612 Applied input value 613 Discrimination result 614 Result output value 615 Input node Group 616 No output for problem solving Group 701 Neural network teacher input / output value table with unlearned case discrimination node 702 Input value portion 703 Output value portion 704 Discrimination value 801 Neuron 802 Input terminal 803 Output terminal 901 Input node 902 Output node 903 Intermediate node 1001 Central processing unit 1002 Storage device 1003 input device 1004 output device

Claims (10)

[Claims] 1. A neural network comprising an input node composed of a plurality of neurons, an intermediate node, and an output node, the learning network being used in advance for the neural network. A neural network system that performs learning so as to output the correct output node value corresponding to the input node value input to the node to the output node, and outputs a more correct output node value corresponding to any input node value. At
A neural network comprising an output node value evaluation means for evaluating the certainty of an output node value corresponding to the arbitrary input node value based on the arbitrary input node value input to the input node of the neural network. Net system. 2. The neural network system according to claim 1, wherein the output node value evaluation means is an input node, an intermediate node, and an input node, an intermediate node corresponding to each of the output nodes of the neural network,
And learning using the same learning data as that of the neural network and learning using unlearned data different from the learning data, and an arbitrary input node value input to the neural network is obtained. An input neural network for evaluation, which outputs different constant output node values depending on whether the arbitrary input data value is the learned data or the unlearned data, and the output node from the evaluation neural network is provided. A neural network system characterized in that the certainty of an output node value corresponding to the arbitrary input node value is evaluated based on the value. 3. The neural network system according to claim 1 or 2, wherein the output node value evaluation means is composed of an arbitrary number of neurons added to an output node of the neural network. Was
Learning the neural network with the learning data,
An output node for evaluation for performing learning of the neural network by unlearned data other than the learning data with different output node values, and the arbitrary input node based on the output node value from the output node for evaluation A neural network system characterized by evaluating the certainty of an output node value corresponding to a value. 4. The neural network system according to claim 2 or 3, wherein the output node value evaluation means uses the learning data used for learning of the neural network to perform the evaluation neural network. A neural net system comprising: a net; and an evaluation learning data generating means for generating learning data and unlearned data used for learning the output node. 5. The neural network system according to any one of claims 1 to 4, wherein the output node value evaluation means uses the learning as an arbitrary input node value input to an input node of the neural network. An input node value evaluating means for judging whether or not the input node value of the learning data used for the above is similar, and evaluating the certainty of the output node value with respect to the arbitrary input node value. Characteristic neural network system. 6. The neural network system according to claim 5, wherein the input node value evaluation means stores learning input node value storage means for storing all components of all input node values of learning data used for the learning. And an input node value discriminating means for discriminating whether or not all the components of an arbitrary input node value to the neural network correspond to the components stored in the learning input node value storing means, A neural network system characterized in that the certainty of an output node value with respect to the arbitrary input node value is evaluated based on the result of the determination by the input node value determining means. 7. The neural network system according to claim 1, wherein the output node value evaluation means is an intermediate node corresponding to an arbitrary input node value input to an input node of the neural network. An intermediate node value for judging the certainty of the output node value with respect to the arbitrary input node value by determining whether the node value is similar to the intermediate node value at the time of learning using the learning data. A neural network system comprising an evaluation means. 8. The neural network system according to claim 7, wherein the intermediate node value evaluation means stores learning intermediate node value storage means for storing all intermediate node values at the time of learning using the learning data. Intermediate node value discriminating means for discriminating whether or not all the components of the intermediate node value corresponding to any input node value to the neural network correspond to the components stored in the learning intermediate node value storing means; A neural network system, characterized by comprising: evaluating the certainty of the output node value with respect to the arbitrary input node value based on the determination result of the intermediate node value determining means. 9. The neural network system according to claim 1, further comprising a warning to the user based on the evaluation result of the uncertainty of the output node value by the output node value evaluating means. A neural network system characterized in that it is provided with an improper output response processing means for performing corresponding internal processing arbitrarily determined in advance. 10. A learning method for a neural network in a neural network system according to claim 4, wherein the evaluation learning data generating means comprises the learning data and the unlearned data each of which has a constant output node value. And the evaluation neural net and each of the output nodes are learned by using learning data and unlearned data composed of the different constant output node values. Learning method.