JP2002358499A - Allocation device for neural network tutor signal and storage medium stored with program thereof - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a tutor signal allocating device which can efficiently perform tutor signal allocating operation. SOLUTION: A main neural network means which is made to learn with basic learning data and one or more subordinate neural network means which are made to learn by using mutually different tutor signal arrangements obtained by converting the arrangement of basic tuner signals for the basic learning input data are connected in parallel and tutor signal allocation and discrimination to generated input data is carried out by using at least output matching state detection between the main output signal sent out by the main neural network means and the conversion output obtained through an output converting means which reverse-converts the tutor signal arrangement conversion rule of output signals of the subordinate neural network means and tutor signal matching state detection between the tutor signal corresponding to the basic learning input data on which the generated input data are based and the main and conversion output signals, thereby allocating the tutor signals.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for allocating a teacher signal to input data for neural network processing which can be applied to communication network fault processing systems such as the Internet, and fields such as pattern recognition, voice recognition and image recognition. It relates to a method and a system.

[0002]

2. Description of the Related Art Conventional neural networks include various neural networks such as a multilayer (hierarchical) neural network and an interconnected neural network. In particular, multilayer (hierarchical) neural networks that have been learned using learning data including learning input data and teacher signals corresponding to the classification categories have been widely and widely used. As processes related to many neural networks up to now, there are a neural network learning process using learning data and an execution process performed using a connection weight coefficient after learning.

Here, a learning process will be briefly described by taking a multilayer neural network as an example. The learning input data to the multilayer neural network is composed of a vector having elements. After input to each unit of the corresponding input layer of the neural network, each unit of the intermediate layer (hidden unit) is further weighted and output. Is done. In the hidden layer, the sum of weighted outputs from each unit in the input layer is used as an input, and after subtracting a bias, it is output via a function having a nonlinear input / output characteristic called a sigmoid function. In the output layer, after the same input / output processing as that of the intermediate layer is performed, each output layer unit sends a corresponding output signal. In some cases, these output layer output signals are converted into binary output signals via a binarization threshold circuit and transmitted.

[0004] As a connection weight coefficient learning algorithm of the multilayer neural network, for example, there is a back propagation algorithm. After initial values are set using random numbers or the like for the connection weight coefficients, a back propagation algorithm is performed using a teacher signal corresponding to a prepared classification category and learning input data input to the input layer. Based on the above, adaptively update the connection weight coefficient so that the average error power between the output signal and the teacher signal is minimized,
Learning is performed so that an output signal as close as possible to the teacher signal is transmitted from the output layer. Here, when the average error power becomes equal to or smaller than a predetermined threshold, the learning is terminated assuming that the convergence has been achieved. Alternatively, 2 for all learning input data
The learning ends when the value output signal becomes the same as the corresponding binary teacher signal and becomes the correct answer output signal.

Next, an execution processing neural network is constructed by using the connection weighting factors obtained by learning (here, biases are also included in the connection weighting factors and handled), and the test input data prepared in advance and the Using test data consisting of the desired teacher signal corresponding to that,
For example, a test process for evaluating a generalization rate indicating a ratio of the number of output signals that are correct to the number of test input data is performed.
When a test input data is input, a desired output signal, that is, a case where a correct output signal is transmitted, and a case where an incorrect output signal is output,
That is, an output signal of an incorrect answer may be transmitted. In particular, when the neural network converges in a local minimum state as a result of learning, an output signal of an incorrect answer is easily transmitted, and the generalization rate is not so high.

[0006] After the above-mentioned learning processing or further test processing, in the execution processing in the pattern recognition, the network failure detection processing, and the like, an execution processing neural network in which the connection weight coefficient obtained by the learning processing is set is used. , To obtain an output signal for unknown input data. Here, if the learning input data used in the learning process is small, and the convergence to the local minimum is also low and the generalization rate is inferior,
A desired output signal cannot be obtained even for input data in the vicinity of the learning input data, and many incorrect output signals are generated.

Accordingly, it is necessary to collect as much input data as possible in advance and use it as learning data or test data. However, in general, it is often difficult to sufficiently collect training data and test data in advance.

Further, in order to realize a high generalization ability and a high-performance category separation ability, when a large amount of input data is collected, the work of assigning a teacher signal corresponding to a classification category to the input data becomes enormous, and When allocation is performed in a tentative manner, allocation errors are likely to occur. Furthermore, it is not clear what kind of property the collected input data is, that is, whether the data is close to or separated from the category separation boundary area. Even if the data is simply collected and learned, it is not always necessary to improve the performance. However, the amount of calculation becomes enormous, and in some cases, it is impossible to converge. In addition, the learning data may be in an overfitting state, resulting in performance degradation.

FIG. 2 shows a configuration example of a conventional teacher signal allocating method. These configurations and their operations will be briefly described.

In this configuration, the input data generating means 37 generates input data, and the data display input processing means 40
Display to the neural network designer via. On the other hand, the neural network designer inputs a correct teacher signal for the given generated input data,
These are generated data and stored in the generated data storage means 39.

Here, in the input data generating means 37,
When the generated input data is generated, in order to check whether the same generated input data already exists in the generated data storage means 39, the existing input data detecting means 38 detects the same generated input data. If present,
Discard the generated input data. As described above, the simple destruction processing of the duplicate generation data is performed.

In addition, input data actually generated is collected, the data is displayed to a neural network designer, and a corresponding teacher signal is allocated. The task of assigning teacher signals to a teacher is enormous. Until now, there has been no method for systematically and successfully collecting learning data for correctly forming a category separation boundary, and it is not clear how to select learning data from a huge amount of collected data. Therefore, the cross-validation method (Computer Syste
ms That Learn by Sholom M. Weiss and Casimir A. Ku
likowski, Morgan Kaufmann Publishers, Inc)
Learning data is created by various combinations of data and learning data is selected based on experience.
For this reason, problems such as the formation of incomplete separation boundaries depending on the learning data and overfitting due to learning are likely to occur, and learning is performed so that all correct output signals are sent out, achieving high generalization ability. Is difficult to do.

On the other hand, mechanically generating neighboring input data from basic learning input data and allocating teacher signals by simple visual observation requires a great deal of labor when the amount of generated input data to be generated is large. Mistakes also occur. Furthermore, it is not known which generated input data is effective and effective for learning and testing, respectively. If all generated input data is used for learning, the amount of calculation processing becomes enormous, and unnecessary generated input data is learned. There are many drawbacks, such as

[0014]

As described above, according to the prior art, when constructing a high-performance neural network, the neural network designer needs a great deal of preparation for collecting learning data and test data and allocating teacher signals. It is an amount of work, requires time and effort, and is easy to make mistakes in allocation. In particular, since the influence of the input data on the category separation boundary is not known at all and the teacher signal is mechanically assigned, it is possible to efficiently select the learning data and test data effective for forming the category separation boundary. No, it is necessary to collect more extensive data than necessary.
In addition, as a result of learning and testing by combining these data in various ways, such as the cross-validation method, the amount of arithmetic processing for learning and testing becomes enormous and a neural network with high-performance category separation capability There are drawbacks such as difficulty in obtaining a network.

Therefore, in practical systems such as a large-scale pattern recognition system and a network fault detection and search system, it is possible to efficiently prepare and use learning data and test data for realizing a high-performance neural network. Can not. Further, there is a drawback that performance improvement and generalization ability cannot be effectively improved according to the use environment of the execution processing neural network.

An object of the present invention is to provide a method and a system for solving these problems and a storage medium storing the program.

[0017]

The main method for solving the problem and a system using the same will be described below.

As a first solving means, basic learning input data and a teacher signal corresponding to a classification category are used as basic learning data, and learning is performed using the basic learning data and a different initial value of a connection weight coefficient. Two or more neural network means, input data generating means for generating generated input data based on the basic learning input data, and inputting the generated input data, which are sent from the parallel connected neural network means. Output coincidence detecting means for detecting an output coincidence state between output signals and transmitting an output coincidence signal, and detecting an output coincidence state between the teacher signal and the output signal corresponding to basic learning input data on which the generated input data is based. Teacher signal matching state detecting means for detecting a teacher signal matching state signal and sending a teacher signal matching state signal, and at least the output matching state signal At least teacher signal assignment identifying means for identifying the teacher signal coincidence state signal for assignment of the teacher signal to the generated input data, and allocating the teacher signal to the generated input data based on at least the result of the identification. The problem is solved by a neural network teacher signal allocating device.

As a second solving means, the main neural network means which uses the basic learning input data and the teacher signal corresponding to the classification category as basic learning data and learns with the basic learning data and sends out a main output signal is different from the main neural network means. Transforming the teacher signal arrangement of the basic learning data according to a conversion rule and learning conversion learning data in which a teacher signal is assigned to the basic learning input data; and learning one or more sub-neural network means connected in parallel with the main neural network means An input data generating means for generating generated input data based on the basic learning input data; and an output signal of the sub-neural network means, a teacher signal arrangement of the conversion learning data and a teacher signal arrangement of the basic learning data. Output conversion means for performing conversion according to an inverse conversion rule for converting and outputting a converted output signal; Output matching state detecting means for inputting data to the neural network means connected in parallel, detecting an output matching state between the obtained main output signal and the converted output signal, and sending out an output matching state signal; Teacher signal matching state detecting means for detecting a teacher signal matching state between the teacher signal corresponding to the basic learning input data on which the input data is based and the main and converted output signals and transmitting a teacher signal matching state signal; At least teacher signal assignment identifying means for identifying the output match state signal and the teacher signal match state signal for assignment of the teacher signal to the generated input data, and at least based on the result of the identification, The problem is solved by a neural network teacher signal allocating apparatus characterized by allocating a teacher signal to data.

As a third solving means, the neural network means sets different initial values of connection weights and performs learning using the basic learning input data. The problem is solved by the described neural network teacher signal assignment device.

According to a fourth aspect of the present invention, the neural network means sets the same initial value of the connection weight coefficient and performs learning using the basic learning input data. The problem is solved by the described neural network teacher signal assignment device.

As a fifth solution means, an output coincidence state signal indicating that all the output signals inputted by the output coincidence state detection means coincide with each other; The teacher signal corresponding to the basic learning input data on which the generated input data is input and the teacher signal matching state signal indicating that the output signal matches are output by the teacher signal assignment identifying means. And inputting the teacher signal to the generated input data by the neural network teacher signal allocating apparatus according to any one of the first, second, third and fourth means. Resolve.

According to a sixth aspect of the present invention, there is provided a neural network teacher signal allocating apparatus according to the fifth aspect, wherein the generated data is stored as at least third generated data.

As a seventh solving means, an output matching state signal indicating that at least one of the output signals inputted in the output matching state detecting means is different, and A teacher signal matching state signal indicating that at least one of the input output signals matches the teacher signal corresponding to the basic learning input data on which the generated input data is based; When input by the means, it is identified, a first teacher signal assignment request signal is generated, and further, any of the output signals corresponds to the basic learning input data on which the generated input data is based. When a teacher signal matching state signal indicating that there is no match with the teacher signal is input to the teacher signal assignment identifying means, it is identified and a second teacher signal assignment request signal is output. The neural network according to any one of the first, second, third, fourth, fifth and sixth means, wherein at least the transmitting means allocates a teacher signal to the generated input data based on information input from the outside. The problem is solved by a network teacher signal assignment device.

According to an eighth aspect of the present invention, the generated data to which the teacher signal is allocated is defined as first generated data based on the first teacher signal allocation request signal, and the second teacher signal allocation request signal is defined as the first generated data. Based on the above, the generated data to which the teacher signal has been allocated is stored as the second generated data.

As a ninth solution means, when the generated input data generated by the input data generating means matches the existing generated input data of the stored existing generated data, the existing input data detection signal is sent to the input data detecting signal. An existing input data detecting means for transmitting to the teacher signal assignment identifying means is provided. When the existing input data detecting signal is identified by the teacher signal assignment identifying means, a teacher signal assignment reconfirmation request signal is sent out, and a correct The neural network teacher signal according to any one of the sixth and eighth solving means, wherein when the teacher signal is input, the teacher signal is assigned to the existing generated input data, and the existing generated data is rewritten and stored. Solved by the assigning device.

Further, the problem is solved by a program storage medium storing a program for realizing the above solution.

In the neural network teacher signal allocating apparatus of the present invention, a neural network trained using basic learning data composed of teacher signals corresponding to representative and characteristic basic learning input data in order to separate categories of input data. The means and one or more neural network means, which are obtained by converting the teacher signal arrangement in the basic learning data and are trained using the conversion learning data having different teacher signal arrangements, are connected in parallel, and A category separation boundary plane slightly deviated from data is formed, and detection of input data close to the separation boundary area and estimation of correct or incorrect answer of an output signal are used for teacher signal allocation.

That is, based on the basic learning input data, for example, generation input data (neighboring input data) having a short distance is generated and input, and the main output signal from the neural network means learned using the basic learning data and Between the conversion rule obtained by converting the output signal of the neural network means learned using the conversion learning data having different teacher signal arrangements by the teacher signal arrangement conversion and the inverse conversion rule in the reverse direction. Output matching state detection is performed. As a result, the positional relationship of the generated input data on the equivalent category separation boundary surface, that is, information on whether the generated input data is around or separated from the separation boundary region is obtained. Further, by detecting a second matching state between the main output signal and the converted output signal corresponding to the generated input data and the teacher signal (reference teacher signal) corresponding to the basic learning input data, it is possible to estimate the correct or incorrect answer of each output signal. In this system, a teacher signal assignment mode is identified based on these pieces of information, and teacher signal automatic assignment, teacher signal confirmation assignment, or teacher signal inquiry assignment is performed.

If all of the main and converted output signals match and are also the same as the reference teacher signal, the generated input data departs from the separation boundary area and can be regarded as a correct answer output signal, and the output signal is used as the teacher signal. A teacher signal is automatically assigned. Also,
If all of the main and converted output signals do not match, and one of them is the same as the reference teacher signal, the generated input data is close to the separation boundary area, and either the main or converted output signal can be correct. Since the output signal is high, the output signal is set as a candidate for teacher signal assignment, and teacher signal confirmation assignment for requesting a neural network designer to confirm is performed.

Further, if the main and converted output signals do not all match, and also the reference teacher signal does not match, the generated input data is close to the separation boundary area and the formation of the separation boundary is insufficient, and all the output Because the signal is likely to be wrong,
A teacher signal inquiry assignment for inquiring a designer of a correct teacher signal is performed. As a result, the generated input data and the assigned teacher signal are stored as generated data (neighbor data).

With such a neural network teacher signal allocating device, teacher signal allocation can be performed almost automatically from basic learning input data to short-range generated input data. The amount of work can be significantly reduced. That is, a teacher signal input request is made to the neural network designer only when a teacher signal is confirmed or an inquiry is made, and the teacher signal is automatically allocated in other cases. Is greatly reduced. Further, by comparing the existing generated data with the generated data, and displaying the output signal and the reference teacher signal, the input data on the separation boundary surface can be grasped, and the teacher signal assignment error can be eliminated.

Further, with respect to the generated data systematically collected as described above, when there is a teacher signal error in the confirmation assignment, the query assignment, or further in the reconfirmation mode, the generated data is automatically set as learning data. By using the generated data of the assignment as test data, it is possible to efficiently form a more accurate category separation boundary surface, and to realize a neural network means excellent in category separation characteristics and generalization ability.

[0034]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The configuration and operation of a neural network teacher signal allocating apparatus according to a first embodiment of the present invention will be described below in detail. Here, for the sake of simplicity, a case where a binary teacher signal is used will be described as an example. However, in the present invention, as the teacher signal,
The present invention is not limited to the binary teacher signal, and may be a multi-value or continuous value. Also, any structure may be used as long as it is a neural network that learns using a teacher signal. A case in which short-range generated input data is used as nearby input data based on basic learning input data will be specifically described.

[First Embodiment] FIG. 1 shows an example of a configuration of a neural network teacher signal allocating apparatus according to a first embodiment of the present invention.

The basic learning data storing means 1, the input data generating means 2, the existing input data detecting means 3, the generated data storing means 4, the main learning data storing means 6, the main neural network means 6, the first Sub-neural network means 7, first sub-learning data storage means 8, teacher signal arrangement conversion means 9 and 12, second sub-neural network means 10, second sub-learning data storage means 11, output conversion means 13 , 14 and output coincidence state detecting means 15
And a teacher signal matching state detecting means 16, a teacher signal assignment identifying means 17, and a data display input processing means 20.

The detailed configuration and operation will be described below.

In the basic learning data storage means 1, a typical representative
Alternatively, basic learning data composed of further characteristic input data and teacher signals corresponding to the separated categories is stored in advance, and based on the basic learning data, the main neural network means 6 and the first sub neural network means 7 are used.
And the second sub-neural network means 10 are learned. First, when the main neural network means 6 is to be learned, basic learning data including basic learning input data and a corresponding teacher signal is stored in the main learning data storage means 5, and these are read out. After the initial value of the connection weight coefficient prepared in advance is set, learning is performed, and learning is continued until a convergence state in which all output signals are correct.

Next, in the first sub-neural network means 7, only the teacher signal of the basic learning data read out from the basic learning data storage means 1 is given in advance via the teacher signal arrangement conversion means 9. According to the obtained conversion rule, the layout is changed, and together with the learning input data, as the first sub-learning data,
It is stored in the first sub-learning data storage means 8 and, using these, the same connection weight coefficient initial value as that of the main neural network means 6 prepared in advance in the first sub-neural network means 7 is set, and then learning is performed. Learning is continued until a convergence state in which all output signals are correct.

Similarly, the teacher signal arrangement conversion means 12 converts the teacher signal, and a conversion rule different from that for the first sub-learning data is prepared in advance. The placement-transformed teacher signal and learning input data are
The second sub-learning data storage means 11 stores the data as sub-learning data, and using these, the same connection weight coefficient initial values as those of the main neural network means 6 prepared in advance in the second sub-neural network means 10 are set. After learning, the learning is continued until the convergence state in which all output signals are correct.

In these series of learning processes, the main neural network means 6, the first sub network means 7,
Since the second sub-network means 10 learns the same learning input data with different teacher signal arrangements, the generalization ability is almost the same, but the generalization area is slightly shifted equivalently. With different characteristics. In other words, where the separation boundary area is complicated and narrow,
There are neural network means for sending an output signal that is a correct answer and neural network means for sending an output signal that is an incorrect answer. On the other hand, for input data far from the separation boundary and close to the basic learning input data, an output signal of a correct answer is sent out. Therefore, in the parallel neural network means 6, 7, 10, the input data in the vicinity of the separation boundary area can be effectively made by utilizing the fact that the output coincidence state and the difference between correct and incorrect answers due to the slight difference in the formation of the separation boundary plane occur. , And correct / incorrect estimation of the output signal can be performed.

Here, in order to systematically prepare learning data and test data necessary for efficiently learning and testing the neural network means,
First, near-neighbor input data having a small distance centered on representative basic learning input data is generated as generated input data, and the parallel neural network means 6, 7, 10 in which the category separation boundary surface is slightly shifted equivalently. Using,
Identify the neighboring input data that is located near the category separation boundary area and the data that is located far from the separation boundary area, estimate the correct answer of those output signals with respect to the neighborhood input data, and refer to the information to refer to the information. Is performed in accordance with the teacher signal.

In order to realize the above-mentioned processing, first, when the above-mentioned learning processing is completed, the respective neural network means 6, 7, 10 are connected in parallel, and the respective neural network means 6, 7, 10 are connected to the neighboring input data. Send an output signal. As the nearby input data, the basic learning data is sequentially read from the basic learning data storage unit 1, and input data around the basic learning input data is generated by the input data generating unit 2 and used as the nearby input data. As an example of the generation method, when binary input data is assumed, first, data close to the hamming distance by about one from the basic learning input data are sequentially generated as the neighboring input data.

At this time, the existing input data detecting means 3
Whether or not the neighborhood input data has already been generated so far and the neighborhood data including the assigned teacher signal is stored in the generated data storage means 4. It is checked whether or not it is stored in the storage means 1. Existence (collision)
If not, the neighboring input data is input to the neural network means 6, 7, 10 and the teacher signal assignment identifying means 17 connected in parallel. Further, a teacher signal (reference teacher signal) from the basic learning data storage unit 1 and a nearby input signal corresponding to the basic learning input data used for generating the nearby input data are sent to the teacher signal assignment identifying unit 17, respectively.

On the other hand, if it already exists (collides) and it is the basic learning input data, it is discarded. Also,
In the case of existing neighborhood input data, the existing input data detection signal indicating the collision with the existing neighborhood input data from the existing input data detection means 3 and the teacher signal corresponding to the basic learning input data used for generating the neighborhood input data (Reference teacher signal)
And the colliding neighboring input signal from the input data generating means 2 are transmitted to the teacher signal assignment identifying means 17. The generated data storage unit 4 sends the existing neighboring data that has collided to the data display input unit 20. Furthermore,
This neighborhood input data is input to the main neural network means 6, the first auxiliary neural network means 7, and the second auxiliary neural network means 10, respectively.

The output signal for the neighboring input data from the main neural network means 6 is sent out as it is as the main output signal. On the other hand, the output signals from the first and second sub-neural network means 7 and 10 have inverse conversion rules in the reverse direction to the conversion rules used in the teacher signal arrangement conversion means 9 and 12, respectively. It is input to output conversion means 13 and 14. As a result, output conversion means 1
3 and a first converted output signal obtained from
4 have the same teacher signal arrangement relationship as the main output signal. Therefore, it is possible to form a category separation boundary plane which is slightly deviated equivalently from the neighboring input data, and transmit each of the output signals.

The output coincidence state between these three output signals includes a state in which the main output signal, the first converted output signal, and the second converted output signal all match, a state in which any two output signals match, A match between the main output signal and the first converted output signal, a match between the main output signal and the second converted output signal, or
The output signal and the second output signal are in a state of coincidence, and further, these output signals are not in agreement with each other, that is, are in any state of non-coincidence.

The output coincidence detecting means 15 detects the output coincidence between these output signals, and inputs the output coincidence signal to the teacher signal assignment identifying means 17 as an output coincidence signal. On the other hand, the teacher signal matching state detecting means 16 detects a matching state between these output signals and the reference teacher signal, and determines the matching state signal as a teacher signal matching state signal.
Enter 7

The teacher signal assignment identifying means 17 outputs an output coincidence state signal, a teacher signal coincidence state signal, a main output signal,
The converted output signal, the second converted output signal, and the generated nearby input data are input, and a teacher signal corresponding to the basic learning input data on which the generated nearby input data is based is input as a reference teacher signal. . Further, an existing input data detection signal from the existing input data detection means 3 is also input.
Based on these input signals, a teacher signal allocation mode to nearby input data is identified.

Here, as the identification of the teacher signal assignment,
The operation of the case where the teacher signal automatic assignment mode, the teacher signal confirmation assignment mode, the teacher signal inquiry assignment mode, and the teacher signal assignment reconfirmation mode are provided will be described.

In the teacher signal automatic assignment mode, as the output coincidence state, an output coincidence state signal indicating that the main output signal, the first sub-output signal, and the second sub-output signal for the neighboring input data have all coincided is input. In this case, it is determined that a teacher signal matching state signal indicating that all of them match the reference teacher signal is input as the teacher signal matching state, and it can be regarded as a correct answer. Therefore, the output signal is automatically assigned as a teacher signal of the neighborhood input data, stored in the generated data storage means 4, and automatically assigned signal, output coincidence signal, main output signal, first output signal through the data display input processing means 20. The converted output signal, the second converted output signal, the input neighborhood input data, the reference teacher signal, and the teacher signal matching state signal are displayed. Neural network designers only need to display these data for the time being and do not require any special work.

Generally, when a neural network having a slightly shifted separation boundary surface is well designed and connected in parallel, this state occurs very often. Therefore,
The teacher signal assignment work of the neural network designer can be almost automatically automated, and can be greatly reduced to, for example, about 10% or less of the conventional work. Further, since these neighborhood data are correct answers, they need not always be learned, and can be used as test data. Therefore, the automatically assigned neighborhood data may be stored as automatically assigned neighborhood data in the generated data storage unit 4 and used as test data.

In the teacher signal confirmation assignment mode, an output matching state signal indicating that all output signals do not match is input as an output matching state, and at least one of the output signals is referred to as a teacher signal matching state. It is highly likely that the same output signal as the reference teacher signal is a correct answer when it is determined that the teacher signal matching state signal indicating the state of matching with the reference teacher signal has been input. Therefore, since all the output signals do not match, the reference teacher signal is displayed as a candidate for the correct teacher signal, and the first teacher signal assignment request signal for requesting the neural network designer to confirm, that is, the confirmation assignment signal is transmitted. I do. Confirmation assignment signal, output matching state signal, main output signal, first converted output signal, second converted output signal, input neighborhood input data, reference teacher signal, and teacher signal matching state via data display input processing means 20 And signals are displayed.

After requesting the neural network designer to confirm whether the output signal matching the reference teacher signal is a correct answer, when a correct answer determination signal is input to the data display input processing means 20, the neighborhood input data and its output are output. The signal is stored in the generated data storage unit 4 as neighborhood data. When an incorrect answer determination signal is input, the generated data is stored in the generated data storage means 4 as confirmation neighborhood data together with a correct teacher signal and neighborhood input data obtained by making an input request.

In the teacher signal inquiry assignment mode, regardless of the output coincidence state, it is determined that a teacher signal coincidence state signal indicating that each output signal is different from the reference teacher signal is input as the teacher signal coincidence state. Thus, an output signal different from the reference teacher signal is very likely to be a wrong answer. Therefore, a second teacher signal assignment request signal, that is, a query assignment signal, which requests the neural network designer to input a correct output signal as a teacher signal, is transmitted. Data display input processing means 20
, An inquiry assignment signal, an output matching state signal, a main output signal, a first converted output signal, a second converted output signal, input neighborhood input data, a reference teacher signal, and a teacher signal matching state signal are displayed, respectively. I do.

When a correct answer determination signal is input to the data display input processing means 20, the neighborhood input data and its output signal are stored in the generated data storage means 4 as neighborhood data. When an incorrect answer determination signal is input, it is stored in the generated data storage unit 4 as inquiry neighborhood data together with a correct teacher signal and neighborhood input data obtained by making an input request.

In the teacher signal assignment reconfirmation mode, a state in which newly generated neighboring input data collides with existing neighboring input data is identified by the existing input data detection signal.
As the basic learning data, representative and characteristic data are given, and near input data having a short distance is sequentially generated based on these data. Collision with neighboring input data in neighboring data existing in the generated data storage means 4 has very important information from the viewpoint of category separation. That is,
The colliding neighboring input data is input data equidistant from the respective basic learning input data upon which it was generated, and is essentially on the separation boundary in the category classification, and the ambiguity in the category separation Is input data that easily causes Further, when a collision occurs, it means that the teacher signal assigned to the basic input data on which the neighboring input data is based is different from the teacher signal assigned to the existing neighboring data. ing. Therefore, re-confirmation of the existing teacher signal assignment is required in order to find an assignment error of the teacher signal assigned to the existing neighboring input data.

When a collision is detected by the existing input data detecting means 3, the collision may be simply discarded and processed. However, since the data is important data on the separation boundary, the existing input data detection means 3 detects the collision. The signal is input to the teacher signal assignment identifying means 17, and the parallel neural network means 6,
7 and 10 are input with the collided neighboring input data, and the output coincidence state signal, the main output signal, the first converted output signal, the second converted output signal, the collided neighboring input data, the reference teacher signal, the teacher signal coincidence state signal, The existing input data detection signal is displayed via the data display input processing means 20. In addition, the existing generated data that has collided from the generated data storage unit 4 is displayed via the data display input processing unit 20.

The neural network designer is requested by the third teacher signal assignment request signal to reconfirm the correctness of the teacher signal assigned to the existing neighboring input data, and the correct answer confirmation signal is sent to the data display input processing means 20. When input, the existing neighborhood data is left as it is, and when an incorrect answer confirmation signal is input, a correct teacher signal input is requested, and the existing neighborhood data is corrected with the correct teacher signal.
To store. At this time, information indicating collision and reconfirmation may be added and stored.

As described above, based on the generated neighborhood input data and the output signals respectively sent from the parallel neural network means 6, 7, and 10, the confirmation request is automatically or to the neural network designer. Stores the teacher signal and the neighboring input data that are revealed by the inquiry request or the reconfirmation request as the neighboring data.
As described above, under the given distance condition, after the teacher signal is assigned to the neighboring input data generated based on all the basic learning input data, the distance from the basic learning input data gradually increases. By sequentially generating the generated input data and processing the generated input data in the same manner, it is possible to systematically collect neighboring data in a wider range. By performing learning and testing of the neural network means using the obtained neighborhood data, it is possible to efficiently improve performance (learning ability and generalization ability). These neural network means may be executed independently and applied to various applications, or may be executed as parallel neural network means and applied.

Although the case has been described where the neighboring data is stored separately according to the teacher signal assignment mode, they may be stored without being distinguished.

Next, FIG. 4 shows a processing flow relating to the program of the first embodiment.

After downloading the program from the program storage medium 51, the conversion learning data processing 52
Read basic learning data from basic learning data memory,
Convert the teacher signal arrangement corresponding to the basic input data,
As many different teacher signal arrangements as the number of parallel connections of the neural network means minus one are prepared, and conversion learning data is generated.

Further, the neural network means connected in parallel perform the neural network learning parameter setting processing 53 read from the parallel neural network parameter memory, and individually perform the learning processing 55 on the parallel neural networks.
When all correct output signals are transmitted for the basic learning input data, the learning is terminated, and the connection weight coefficients are stored in the parallel neural network parameter memories.
Next, in a neighborhood input data generation processing 55 for each basic learning input data, neighborhood input data of a designated distance is generated based on the given basic learning input data, and a corresponding teacher signal is referred to as a reference teacher signal. Set as

In the existing neighborhood input data detection processing 56, it is determined whether the generated neighborhood input data is the same as that already generated based on other basic learning input data, and the same as the other basic learning input data. Is searched through the basic learning data memory and the generated data memory, and if it is the same as the basic learning input data, the generated nearby input data is discarded, new nearby input data is generated and checked again. .

If the input data already exists as the neighboring input data, the existing input data detection signal and the existing neighboring data are stored for transmission to the designer via the data display input processing 63. Thereafter, the generated neighborhood input data is input to the parallel neural network means, and in parallel neural network execution processing 57, an output signal is obtained from each neural network means. Each output signal of the parallel neural network means learned using different teacher signal arrangements is converted through an output conversion process 59 to obtain a converted output signal. In an output coincidence state detection process 60, the parallel output learned using the basic learning data is used. A match state between the main output signal from the neural network means and the first and second converted output signals is checked, and an output match state signal is transmitted and stored.
In the teacher signal coincidence state detection processing 61, the main and first
Then, a matching state between the second converted output signal and the reference teacher signal is checked, and a teacher signal matching state signal is transmitted and stored.

In the teacher signal assignment identifying process 62, the output matching state signal, the main output signal, the first and second converted output signals, the generated nearby input data, the reference teacher signal, and the teacher signal matching state The signal and the existing input data detection signal are used to identify the teacher signal automatic assignment mode, the confirmation assignment mode, the inquiry assignment mode, and the assignment reconfirmation mode. Further, if the data, the identified mode, and the existing input data in the existing neighborhood input data detection processing 56 are detected, the existing neighborhood data is output to the designer via the data display input processing 63. To be displayed. In the confirmation mode, the inquiry mode, and the reconfirmation mode, the designer requests input or designation of a correct teacher signal, and uses the input or designated teacher signal and the generated neighborhood input data as neighborhood data as neighborhood data storage processing 64 And stores it in the generated data memory. In the automatic assignment mode, the generated neighborhood input data and the main output signal as a teacher signal are stored as neighborhood data in the generated data memory.

In the given basic learning input data, it is determined in the neighborhood input data full generation determination processing 65 whether or not all neighborhood input data of the set distance have been generated. Returning to the data generation processing 55,
The next neighborhood data is generated, and a series of processing is performed. If it has been generated, the final basic learning input data determination process 66
It is determined whether the given basic learning input data is the final basic learning input data, and if not, the next basic learning input data is set in the next basic learning input data setting processing 67. Then, the process returns to the neighborhood input data generation processing 55 to perform a series of processing. If the input data is the final basic learning input data, a distance determination process 68 is performed. If the set distance between the generated neighboring input data and the basic learning input data is equal to or less than a predetermined distance threshold (Th),
The distance for generation is increased and set via the distance increase setting processing 69, and the process returns to the neighborhood input data generation processing 55 to perform a series of processing. When the specified distance threshold is reached,
The process of allocating the teacher signal to the nearby input data is all ended.

Here, when the neighborhood input data is generated for all the basic learning input data, the distance from the basic learning input data is increased, and the same processing is performed.
However, conversely, when the set distance reaches the given distance threshold, the basic learning input data is changed,
When the neighborhood input data is generated, the same processing is performed, and the neighborhood input data is generated for all the learning input data,
The process of allocating the teacher signal to the nearby input data may be ended.

In each of the neighborhood data obtained in the above embodiment, the neighborhood data (generated data) obtained in the confirmation assignment mode and the inquiry assignment mode, or further in the case of the reconfirmation mode, , May be stored as learning data. In addition, the automatic assignment neighborhood data obtained in the automatic assignment mode may be stored as test data. Using these learning data together with the basic learning data, the stored coupling weight coefficient is set and the learning process is performed.
Further, by performing test processing using these test data, it is possible to easily improve the performance of the parallel neural network means.

[Second Embodiment] FIG. 2 shows an example of the configuration of a neural network teacher signal allocating apparatus according to a second embodiment of the present invention.

Basic learning data storage means 1 similar to FIG.
Input data generation means 2, existing input data detection means 3, generated data storage means 4, main neural network means 6, first sub-neural network means 7,
Second sub-neural network means 10, output coincidence state detecting means 15, teacher signal coincidence state detecting means 16,
, The teacher signal assignment identifying means 17, and the data display input processing means 20.
And main coupling weight coefficient initial value storage means 30, 31, 32. In the present embodiment, the learning conditions of each of the neural network means 6, 7, 10 are different from those of the first embodiment.

Therefore, only the different parts will be described below, focusing on the detailed configuration and operation. Each of the neural network means 6, 7, 10 uses the basic learning data stored in the learning data storage means 33, and further stores a main connection weight coefficient initial value storage means 30 and a first sub connection weight coefficient initial value storage means 31. And different initial values of the connection weighting factors stored in the second sub-connection weighting factor initial value storage means 32 are set and learned. As a result, each of the neural network means 6, 7, and 10 can form a slightly shifted category separation boundary surface as in the first embodiment. That is, the main output signals from the parallel-connected neural network means 6, 7, 10 having slightly different category separation boundaries for the neighboring input data at a short distance from the basic learning input data,
The matching state between the first sub-output signal and the second sub-output signal is checked, and input data near the separation boundary region and input data far from the separation boundary region are distinguished.

However, unlike the first embodiment,
All use the same teacher signal constellation, and since only the initial value of the connection weight coefficient is a category separation boundary surface obtained under different learning conditions, a separation boundary surface that is slightly shifted but similar is formed. Is done. In this case, even when all the output signals match, unlike the reference teacher signal, a wrong answer output signal is easily generated. Therefore, although teacher signal inquiry assignment is larger than that in the first embodiment, teacher signal assignment identifying means 17 can perform similar teacher signal assignment identification and assign teacher signal to nearby input data.

The teacher signal allocation processing flow of the program according to the second embodiment does not require the conversion learning data processing 52 in the teacher signal allocation processing flow according to the first embodiment, and all of them have the same basic learning data. , Different initial values of the connection weighting factors are set in the respective neural network means, and the parallel neural network individual learning process 54,
Learning is performed at 55. The output conversion process 59 is also unnecessary, but the other processes are the same as the process flow in the program of the first embodiment. Therefore, here
The description of these operations is omitted.

As in the first embodiment, in the neighborhood data, the neighborhood data (generation data) obtained in the confirmation assignment mode and the inquiry assignment mode, or further in the case of the reconfirmation mode, is learned. It may be stored as data. Further, the neighborhood data obtained in the automatic assignment mode may be stored as test data. By using these learning data together with the basic learning data, setting the stored connection weighting coefficients and performing learning processing, and further performing test processing using these test data, the performance of the parallel neural network means is improved. Can be easily done. Naturally, neural network means connected in parallel also have high performance.

It is also easy to combine the respective learning conditions described in the first and second embodiments, that is, to combine different teacher signal arrangements with different initial values of connection weighting coefficients. Although not described in detail here, different neural network means 6, 7, 10 are provided with different teacher signal arrangements in the first embodiment and further different connection weighting factors in the second embodiment. This is a method of preparing initial values and learning using these. When the parallel connection and the execution processing are performed after the individual learning processing is completed, it is natural that the output conversion means 1
3, 14 are used in the same manner as in the first embodiment. Based on such a configuration, the same teacher signal assignment identification as in the first and second embodiments may be performed. Since each of the neural network means 6, 7, and 10 learns under the environment of the teacher signal arrangement different from the more random initial value of the connection weight coefficient, the category separation boundary surface can be relatively stably shifted. Near input data around the boundary area can be detected in a well-balanced manner.

In the above embodiment, an example has been shown in which the neighboring input data is sequentially generated from the nearby input data whose distance from the basic learning input data is short. However, random generation of nearby input data within a given distance may be used.

In the output matching state detecting means 15 and the teacher signal matching state detecting means 16 of the above embodiment, the output signal has been described assuming an integer value.
It is not limited to an integer value. When the output signal or the teacher signal is a continuous value, the absolute value of the difference between the unit output signals of the output layer units in the parallel neural network means 6, 7, 10 or the sum thereof is given. If these conditions are satisfied, it may be determined that they match.

Further, the case where the input data generating means 2 uses the Hamming distance assuming binary input data with respect to the distance to the corresponding basic learning input data when generating the neighboring input data is described. However, the Euclidean distance may be used for continuous values, and the Lie distance may be used for multi-valued values. As for the degree of parallelism, the case of three parallels has been specifically described in the first and second embodiments, but is not limited to this degree of parallelism.

Further, in the existing input data detecting means 3,
Instead of performing a collision search of the generated data storage unit 4 or the basic learning data storage unit 1 every time the generated input data is generated, a search is made in advance within the range of the distance in which the generated generated input data is generated. Data may be stored and processed.

[0082]

As described above, in the prior art, in the work of allocating teacher signals to many input data as learning data and test data, which is necessary for realizing a high-performance neural network, Neural network designers individually determine teacher signal assignments by simply looking at all prepared input data without using any information on the category separation boundary that would be formed by the training data. In addition, the task of allocating teacher signals is enormous and requires much time and labor. In addition, if the input data size is large, an artificial assignment error is likely to occur.

Since only simple artificial teacher signal assignment is performed for all prepared input data, learning data and category separation effective for achieving high generalization ability are achieved. Collection of boundary area information and efficient collection of neighboring data in consideration of the information cannot be performed.

That is, the teacher signal assignment work is performed in an environment in which the neural network does not know what category separation boundary forms at all. In particular, a systematic method for obtaining high performance and high generalization is used. Since efficient collection and selection of learning data and test data are not performed, it is not possible to quickly and efficiently execute the formation of a separation boundary by optimal learning and testing, and to reduce the amount of computation required for them.

On the other hand, according to the present invention, the neural network means which has learned in advance the basic learning input data having the basic and main characteristics prepared in advance are connected in parallel. It is possible to detect input data close to the category separation boundary region by detecting an output coincidence state between output signals from the means, and to detect insufficient formation of a category separation boundary surface by detecting all mismatch states. Further, since the output matching state between the output signals output in parallel and the matching state between the reference teacher signal and the output signal, that is, the teacher signal matching state, are used, the formation state of the category separation boundary surface and the correctness of the output signal are used. Answer estimation, automatic assignment of teacher signals, confirmation, inquiry,
Further, it is possible to systematically assign a teacher signal according to the reconfirmation at the time of collision of generated input data and the reliability of an output signal.

That is, when performing the teacher signal assignment operation,
The neural network designer can proceed with the work while obtaining information on the separation boundary formed by the neural network means. In the case of automatic assignment, little attention is required. In the case of confirmation, the candidate teacher signal is used. Check
In the case of an inquiry, it is possible to request the designer to carefully assign a teacher signal, and to reconfirm the teacher signal assignment in the existing generated data when the generated input data collides.

Therefore, the correct teacher signal assignment work by the neural network designer can be performed efficiently and without any work mistake.

Further, it is easy to preferentially learn output data by confirming or inquiring, and furthermore, re-confirmed input data as learning data to improve characteristics. Neighboring input data that sends a correct output signal means that the category separation has been correctly performed by the learning data used, so that it is not always necessary to use it for learning, and the amount of computation can be significantly reduced. I can do it. Therefore, the neighborhood input data to which the teacher signal is automatically assigned and the teacher signal associated therewith can be stored as test data and used as evaluation of generalization characteristics.

As described above, in the prior art, it is difficult for a neural network designer to assign a teacher signal and the amount of work is enormous, requiring a great deal of time and effort, and effectively learning learning data. Although the selection is difficult and high-performance neural network means cannot be easily realized, the teacher signal allocating apparatus using the parallel neural network of the present invention makes it possible to effectively allocate teacher signal to a huge amount of input data. In addition, it is possible to quickly collect / select learning data and test data effective for improving correct answer output sending ability and generalization ability, and to perform neural network learning with a small amount of calculation based on the collection / selection. Generalization ability can be easily achieved.

Therefore, the present invention is very effective in a practical field where a neural network having high performance and high generalization ability is required. In other words, since the necessary learning data and test data can be effectively collected over a wide range, the generalization ability can be easily improved, and an execution processing neural network suitable for the usage environment can be realized. An excellent system can be constructed.

As a result, it goes without saying that various pattern recognition systems are required, such as artificial intelligence systems, network fault detection and search systems,
Its effects are very large, as it can be widely applied to network security systems, information retrieval systems, image processing systems, and the like.

[Brief description of the drawings]

FIG. 1 is a configuration example of a neural network teacher signal assignment device according to a first embodiment of the present invention.

FIG. 2 is a conventional neural network teacher signal allocating device.

FIG. 3 is a configuration example of a neural network teacher signal assignment device according to a second embodiment of the present invention.

FIG. 4 is a flowchart of a teacher signal assignment process according to the first embodiment.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Basic learning data storage means 2 Input data generation means 3 Existing input data detection means 4 Generated data storage means 5 Main teacher signal storage means 6 Main neural network means 7 Secondary first neural network means 8 Secondary first teacher signal storage means 9 Teacher Signal arrangement conversion means 10 Secondary second neural network means 11 Secondary second teacher signal storage means 12 Teacher signal arrangement conversion means 13 Output conversion means 14 Output conversion means 15 Output coincidence state detection means 16 Teacher signal coincidence state detection means 17 Teacher signal assignment Identification means 20 data display input processing means 30 main connection weight coefficient initial value storage means 31 first sub-connection weight coefficient initial value storage means 32 second sub-connection weight coefficient initial value storage means 33 learning data storage means 37 input data generation means 38 Existing input data detection means 39 Generated data storage Step 40 Data display input processing means 51 Program storage medium 52 Conversion learning data processing 53 Neural network learning parameter setting processing 54 Parallel neural network individual learning processing 55 Neighbor input data generation processing 56 Existing neighbor input data detection processing 57 Parallel neural network execution processing 59 Output conversion processing 60 Output coincidence state detection processing 61 Teacher signal coincidence state detection processing 62 Teacher signal assignment identification processing 63 Data display input processing 64 Neighbor data storage processing 65 Neighbor input data all generation determination processing 66 Final basic learning input data determination processing 67 Next Basic learning input data setting processing 68 Distance judgment processing 69 Distance increase setting processing

Claims (14)

[Claims] 1. Basic learning input data and a teacher signal corresponding to a classification category are used as basic learning data, and learning is performed using the basic learning data and a different initial value of a connection weight coefficient. Neural network means (6, 7, 10); input data generating means (2) for generating generated input data based on the basic learning input data; and the parallel connected neural network receiving the generated input data Output matching state detecting means (15) for detecting an output matching state between output signals respectively sent from the means (6, 7, 10) and sending out an output matching state signal; A teacher signal matching state detecting means (1) for detecting a teacher signal matching state between the teacher signal corresponding to the learning input data and the output signal and transmitting a teacher signal matching state signal. ), And at least a teacher signal assignment identifying means (17) for identifying the output coincidence state signal and the teacher signal coincidence state signal for assignment of the teacher signal to the generated input data. A neural network teacher signal assignment device, wherein a teacher signal is assigned to the generated input data based on a result. 2. A main neural network means (6) for using basic learning input data and a teacher signal corresponding to a classification category as basic learning data, learning with the basic learning data and transmitting a main output signal, and different conversion rules. Transforms the teacher signal arrangement of the basic learning data, and trains the conversion learning data in which the teacher signal is assigned to the basic learning input data. One or more sub-neural networks connected in parallel with the main neural network means (6) Means (7, 10); input data generating means (2) for generating generation input data based on the basic learning input data; output signals of the sub-neural network means (7, 10); Is converted according to the inverse conversion rule of converting the teacher signal arrangement of the basic learning data into the teacher signal arrangement, and a converted output signal is transmitted. Output conversion means (13, 1
And 4) inputting the generated input data to the neural network means (6, 7, 10) connected in parallel, detecting an output matching state between the obtained main output signal and the converted output signal, and matching the output. Output coincidence state detection means for transmitting a state signal (15)
A teacher signal matching state for detecting a teacher signal matching state between the teacher signal corresponding to the basic learning input data on which the generated input data is based and the main and converted output signals and transmitting a teacher signal matching state signal Detecting means (16), and at least a teacher signal assignment identifying means (17) for identifying the output coincidence state signal and the teacher signal coincidence state signal for assignment of the teacher signal to the generated input data. A neural network teacher signal assignment device, wherein a teacher signal is assigned to the generated input data based on the result of the identification. 3. The neural network means (6,
7. The neural network teacher signal allocating apparatus according to claim 2, wherein in (7) and (10), different connection weighting coefficient initial values are set and learning is performed using the basic learning input data. 4. The neural network means (6,
7. The neural network teacher signal allocating apparatus according to claim 2, wherein the same connection weighting coefficient initial value is set and learning is performed using the basic learning input data. 5. An output coincidence state signal indicating that all of the output signals input by the output coincidence state detection means (15) coincide with each other, and further, the teacher signal coincidence state detection means (16) The teacher signal corresponding to the basic learning input data on which the generated input data is input and a teacher signal matching state signal indicating that the output signal is matched with the teacher signal assignment identifying means ( The neural network teacher signal assignment according to any one of claims 1, 2, 3 and 4, wherein when the signal is inputted in (17), the input signal is identified and the teacher signal is assigned to the generated input data. apparatus. 6. The neural network teacher signal allocating apparatus according to claim 5, wherein said generated input data and said assigned teacher signal are stored as third generated data. 7. An output coincidence state signal indicating that at least one of the output signals input to the output coincidence state detection means (15) is different, and an output coincidence state detection means (16). And a teacher signal matching state signal indicating that at least one of the output signals received as input matches the teacher signal corresponding to the basic learning input data on which the generated input data is based. When input by the identification means (17), the identification signal is identified, and further, any of the output signals does not match the teacher signal corresponding to the basic learning input data on which the generated input data is based. When the teacher signal assignment state signal indicating the above is input to the teacher signal assignment identifying means (17), it is identified, and the first teacher assignment request signal and the second teacher signal assignment are identified. The request signal, transmits at least one, claim, characterized in that assigning a teacher signal to the generating input data based on information inputted from the outside 1,2,3,
7. The neural network teacher signal allocating device according to any one of 4, 5, and 6. 8. A method according to claim 1, wherein said assigned teacher signal and said generated input data are defined as first generated data based on said first teacher signal allocation request signal, and said second teacher signal allocation request signal is determined based on said second teacher signal allocation request signal. The neural network teacher signal allocating apparatus according to claim 7, wherein the allocated teacher signal is stored as second generated data. 9. When the generated input data generated by the input data generating means (2) matches the existing generated input data of the stored existing generated data, the existing input data detection signal is changed to the teacher signal. Assignment identification means (17)
When the existing input data detection signal is identified by the teacher signal assignment identification means (17), a teacher signal assignment reconfirmation request signal is sent out from the outside. 9. The method according to claim 6, wherein when a correct teacher signal is input, the teacher signal is assigned to the existing generated input data, and the stored existing generated data is rewritten and stored. Neural network teacher signal assignment device. 10. Basic learning input data and a teacher signal corresponding to a classification category are used as basic learning data, and learning is performed using the basic learning data and a different initial value of a connection weight coefficient. Neural network processing, input data generating processing for generating generated input data based on the basic learning input data, and inputting the generated input data between output signals respectively output from the parallel-connected neural network processing. An output matching state detection process for detecting an output matching state and transmitting an output matching state signal, and a teacher signal matching state between a teacher signal corresponding to basic learning input data on which the generated input data is based and the output signal And a teacher signal matching state detection process for sending a teacher signal matching state signal, and at least the output matching state signal and the teacher signal matching state signal. A program for performing at least a teacher signal assignment identification process for identifying a status signal for assignment of a teacher signal to the generated input data, and allocating a teacher signal to the generated input data based on at least the result of the identification. A recording medium storing a program for executing a neural network teacher signal assignment process characterized by the above-mentioned. 11. A main neural network process for using basic learning input data and a teacher signal corresponding to a classification category as basic learning data, learning with the basic learning data, and transmitting a main output signal; One or more sub-neural network processes connected in parallel with the main neural network process, by converting a teacher signal arrangement of the learning data and learning conversion learning data in which a teacher signal is assigned to the basic learning input data; An input data generation process of generating generation input data based on the input data; and an inverse process of converting an output signal of the sub-neural network process so that a teacher signal arrangement of the conversion learning data becomes a teacher signal arrangement of the basic learning data. An output conversion process of converting according to a conversion rule and sending out a converted output signal; An output matching state detecting process for inputting to the connected neural network processing, detecting an output matching state between the obtained main output signal and the converted output signal, and sending out an output matching state signal; A teacher signal matching state detection process for detecting a teacher signal matching state between the teacher signal corresponding to the basic learning input data and the main and converted output signals and transmitting a teacher signal matching state signal; A program for performing at least a teacher signal assignment identifying process for identifying a state signal and the teacher signal matching state signal for assignment of the teacher signal to the generated input data, and at least based on the result of the identification, A storage medium storing a program for executing a neural network teacher signal assigning process, wherein a teacher signal is assigned to the same. 12. An output match state signal indicating that all of the output signals input in the output match state detection processing match each other, and further, an output match state signal input in the teacher signal match state detection processing. When the teacher signal corresponding to the basic learning input data on which the generated input data is based and the teacher signal matching state signal indicating that the output signal matches are input in the teacher signal assignment identification processing. 12. The storage medium storing a program for executing a neural network teacher signal assignment process according to claim 10, wherein the program is identified and the teacher signal is assigned to the generated input data. 13. An output match state signal indicating that at least one of the output signals input in the output match state detection processing is different, and at least one of the output signal input in the teacher signal match state detection processing. A teacher signal matching state signal indicating that the output signal matches the teacher signal corresponding to the basic learning input data on which the generated input data is based; Then, a teacher signal matching state signal indicating that any of the output signals does not match the teacher signal corresponding to the basic learning input data on which the generated input data is based is identified. When input in the teacher signal assignment identifying process, the input signal is identified, and at least one of the first teacher signal assignment request signal and the second teacher signal assignment request signal is identified. 13. The neural network teacher signal assignment process according to claim 10, wherein a teacher signal is assigned to the generated input data based on information input from the outside. A storage medium that stores a program. 14. The generated input data and the assigned teacher signal are stored as generated data, and the generated input data generated in the input data generating process and the stored existing generated data are stored. When there is a match with the existing generated input data, an existing input data detection process for sending an existing input data detection signal to the teacher signal assignment identification process is provided. In the teacher signal assignment identification process, the existing input data detection signal Upon identification, a teacher signal assignment reconfirmation request signal is transmitted. When a correct teacher signal is input from the outside, the teacher signal is assigned to the existing generated input data, and the stored existing generated data is rewritten and stored. 14. A program for executing a neural network teacher signal assignment process according to claim 10, wherein The storage medium storing the ram.