JPH0896084A - neural network - Google Patents

Abstract

(57) [Summary] [Purpose] A neural network that performs learning based on the collected teacher data in a system where it is difficult to collect a large amount of data such as handwritten characters. An object of the present invention is to provide a neural network capable of improving the learning efficiency of the neural network by newly creating data. [Structure] A neural network 22 that performs learning using teacher data, and a teacher data generation unit 25 that creates new teacher data 24 from a plurality of teacher data 26 collected in the neural network 22 are configured. [Effect] A large amount of learning can be performed by collecting a small amount of teacher data, and learning efficiency is improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network for learning indeterminate dot data such as handwritten characters as teacher data, and more particularly to a neural network capable of learning a lot with a small amount of teacher data.

[0002]

2. Description of the Related Art Conventionally, in a neural network for learning by using indeterminate dot data such as handwritten characters as teacher data, teacher data is collected and learning is performed using the teacher data. That is, FIG. 4 is a block diagram of a conventional neural network, in which 11 is a neural network output unit, 12 is a neural network, 13 is a neural network input unit, and 14 is teacher data collected in advance. Is. By inputting the collected teacher data 14 from the neural network input unit 13 to the neural network 12, learning is performed in the neural network 12, and the result is output to the neural network output unit 11.
Appear in. In this case, if the number of collected teacher data 14 is n, learning can be performed n times.

[0003]

In the above conventional method,
It is necessary to collect a large amount of teacher data in order to make the neural network perform learning, but it is difficult to collect a large amount of data such as handwritten characters, and learning more than the number of collected data is necessary. There was a problem that it could not be done.

Therefore, an object of the present invention is to provide a neural network capable of learning more than the number of collected data by newly generating the teaching data from the collected training data.

[0005]

To this end, according to the present invention, in a neural network that performs learning by using indeterminate dot data such as handwritten characters as teacher data, teacher data is generated based on a given number of teacher data. The generation unit automatically generates more than that number of teacher data to perform learning more than the given number of teacher data.

[0006]

In this structure, since new data can be generated in the teacher data generation unit, learning can be performed as many times as the number of newly generated data.

[0007]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a configuration diagram of a neural network according to an embodiment of the present invention. In FIG. 1, reference numeral 21 is a neural network output unit, and 22
Is a neural network, 23 is a neural network input unit, 24 is teacher data generated in the teacher data generation unit, and 25 is a teacher data generation unit which generates new teacher data 24 from given teacher data. 26 is input teacher data.
The new teacher data 24 is generated by inputting the teacher data 26 collected in advance to the teacher data generation unit 25. The generated teacher data 24 is transferred from the neural network input unit 23 to the neural network 22.
Is input to the neural network 22, learning is performed in the neural network 22, and the result appears in the neural network output unit 21. At this time, if the number of collected data is n, at least n + 1 teacher data 24 is generated by the teacher data generating unit 25, so that the neural network 22 can perform learning at least n + 1 times.

2 and 3 are explanatory views of a method for generating teacher data of a neural network in an embodiment of the present invention. Reference numerals 31 and 32 are dot data as pre-collected teacher data, 33 is a mask used for determining feature points of the pre-collected dot data 31 and 32, and 34 and 35 are feature points of the dot data 31 and 32. 36 and 37
Is the characteristic points of the dot data 31 and 32 collected in advance, and 38 is the data 34 and 3 obtained by extracting the characteristic points.
5 is the data in which the feature points of the teacher data newly generated from 5 are set, and 39 is the feature points newly set from the data 34 and 35 in which the feature points are extracted.

In FIG. 3, 41 and 42 are teacher data collected in advance, and 43 is data newly generated from the teacher data 41 and 42. By handling the dot data 31 and 32 collected in advance as dots and performing mask processing using the mask 33, characteristic points such as end points, curved points, and intersections of the line segments of the dot data 31 and 32 collected in advance are identified. Determine. The midpoint between the characteristic points 36 and 37 determined by the above processing is set as a new characteristic point 39. The data 38 is the result of performing this process for all the characteristic points of the data 34 and 35 from which the characteristic points are extracted. By connecting the newly set feature points 39 with line segments, 43 new teacher data is generated. Thereby, one new teacher data 43 is generated from any two teacher data 41 and 42. The generated teacher data 43 is the teacher data 41 collected in advance.
And 42, the above process is repeated to generate a large amount of data from a small amount of data.

[0010]

As described above, according to the present invention, even if it is difficult to collect a large amount of teacher data such as handwriting for learning, it is possible to perform a large number of learning based on a small amount of teacher data. Can realize a neural network that can be done.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a neural network according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a method for generating teacher data of a neural network according to an embodiment of the present invention.

FIG. 3 is an explanatory diagram of a method for generating teacher data of a neural network according to an embodiment of the present invention.

FIG. 4 is a block diagram of a conventional neural network.

[Explanation of symbols]

 21 Neural Network Output Unit 22 Neural Network 23 Neural Network Input Unit 25 Teacher Data Generation Unit 33 Mask

Claims (1)

[Claims] 1. A neural network for learning using indeterminate dot data such as handwritten characters as teacher data, wherein more than that number of teacher data is generated in a teacher data generator based on a given number of teacher data. A neural network characterized by performing learning more than a given number of teacher data by automatically generating.