KR20230128865A - Method for classifying data through recurrent learning transfer based on deep learning - Google Patents

Abstract

Disclosed is a data classification method through deep learning-based learning that pre-classifies a large amount of data using a small amount of data so that humans can easily classify it. The method includes (a) a data acquisition unit acquiring big data and generating a big data dataset and a labeled labeling dataset, (b) a feature value extraction unit from the big data dataset and the labeling dataset, respectively. Extracting a first feature value and a second feature value; (c) a learning unit generating a pre-trained model through neural network-based learning using the first feature value, and the pre-trained model and generating a neural network layer removal model using the second feature value, and (d) a model generating unit extracts a third feature value by applying a weight value based on the neural network removal model, and the third feature value and generating a final classification model for classifying the big data through learning based on the neural network using

Description

Method for classifying data through recurrent learning transfer based on deep learning}

The present invention relates to data classification technology, and more particularly, to a data classification method through deep learning-based learning.

Machine learning technology refers to a technology for generating a model capable of solving a specific problem from a generalization process for input data. In order to create a model with excellent performance, high-quality training data and a learning algorithm for the generalization process are required.

As a technique for improving the performance of a model, there is an ensemble technique. The ensemble technique is a technique for generating a single strong learner by combining a plurality of weak learners.

These ensemble techniques include a bagging technique using a voting method, a boosting technique using a weighted voting method, and a stacking technique using prediction values obtained from a single model as training data. There is a technique.

In addition, with the recent development of artificial intelligence technology, the value of data is significantly increasing. In the case of deep learning technology based on supervised learning, data with accurate labeling (marking correct answers) is very important.

However, since most of the labeling work according to such data is performed directly by a person, it may be difficult to accurately classify ambiguous data, and when the amount of data is considerably large, there is a problem in that it takes time to do the classification work.

1. Republic of Korea Patent Publication No. 10-2020-0097505

The present invention is proposed to solve the problems caused by the above background art, and provides a data classification method through deep learning-based learning that pre-classifies a large amount of data using a small amount of data so that a person can easily classify it. There is a purpose.

In order to achieve the above object, the present invention provides a data classification method through deep learning-based learning that pre-classifies a large amount of data using a small amount of data so that a person can easily classify it.

The method,

(a) acquiring big data by a data acquisition unit and generating a big data dataset and a labeled labeling dataset;

(b) extracting, by a feature value extraction unit, a first feature value and a second feature value from the big data dataset and the labeling dataset, respectively;

(c) The learning unit generates a pre-trained model through neural network-based learning using the first feature value, and uses the pre-trained model and the second feature value to generate a neural network layer removal model. generating; and

(d) The model generator extracts a third feature value by applying a weight value based on the neural network removal model, and uses the third feature value to perform final classification for classifying the big data through learning based on the neural network. It is characterized in that it comprises a; step of generating a model.

In addition, the neural network layer removal model is characterized in that it is generated using transfer learning to remove the neural network layer.

In addition, the transfer learning is characterized in that a Deep Neural Network (DNN) layer is removed from the pre-trained model generated through learning based on a Convolution Neural Network (CNN).

In addition, the classification is characterized in that it is a sound data classification.

In addition, the weight value is characterized in that the weight value of the convolutional layer based on the pre-trained model.

On the other hand, in another embodiment of the present invention, (a) the data acquisition unit acquires big data to create a mirabella-labeled sound source dataset and a labeled first labeling dataset; (b) extracting, by a feature value extraction unit, a first feature value from the first labeling dataset; (c) The learning unit generates a pre-trained model through neural network-based learning using the first feature value, and uses the pre-trained model and the first feature value to generate a neural network layer removal model. generating; (d) a model generator extracts a second feature value by applying a weight value based on the neural network removal model, and classifies the mirabelled sound source dataset through learning based on the neural network using the second feature value generating a temporary classification model for; and (e) the model generation unit classifies the unlabeled sound source dataset using the temporary classification model, and generates a final classification model for classifying the big data using a reconstructed second labeling dataset according to the classification result. It provides a data classification method through deep learning-based learning, comprising the step of determining.

In addition, the step (e) may include: (e-1) comparing, by the model generating unit, a result value according to the classification result with a preset threshold value; (e-2) reconstructing a second labeling dataset by moving the model creation unit to another dataset and labeling the data if the resulting value is greater than the threshold value; and (e-3) generating an additional dataset by combining the second labeling dataset with the first labeling dataset.

In addition, the step (e-3) may include subtracting a preset value from the threshold when the additional dataset is generated; repeating steps (a) to (e) until the threshold value reaches a predetermined reference value; and determining the temporary classification model generated when the repetition is last performed as the final classification model.

In addition, the second labeling dataset is characterized in that it is divided into a training dataset and a test dataset.

In addition, the ratio of the training dataset and the testing dataset is characterized in that 90%: 10%.

According to the present invention, it is possible to pre-classify a large amount of data using a small amount of data so that a person can easily classify it.

1 is a block diagram of a data classification system according to an embodiment of the present invention.
2 is a flowchart showing a model generation process according to an embodiment of the present invention.
3 is a flowchart showing a final classification and labeling process according to an embodiment of the present invention.
4 is a flowchart illustrating a process of generating a new additional data set according to an embodiment of the present invention.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all modifications, equivalents, or substitutes included in the spirit and technical scope of the present invention.

In describing each figure, like reference numbers are used for like elements.

Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. These terms are only used for the purpose of distinguishing one component from another.

For example, a first element may be termed a second element, and similarly, a second element may be termed a first element, without departing from the scope of the present invention. The term "and/or" includes any combination of a plurality of related listed items or any of a plurality of related listed items.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

Terms such as those defined in commonly used dictionaries should be interpreted as having meanings consistent with the meanings in the context of the related art, and unless explicitly defined in this application, they should not be interpreted in ideal or excessively formal meanings. Should not be.

Hereinafter, a data classification method through deep learning-based learning according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of a data classification system 100 according to an embodiment of the present invention. Referring to FIG. 1, the data classification system 100 includes a data acquisition unit 110 that acquires big data and generates a big data dataset and a labeled labeling dataset, from the acquired big data dataset and labeling dataset. The feature value extractor 120 extracts the first feature value and the second feature value, the data storage 130 stores the data set, the first feature value and the second feature value, and the neural network using the first feature value A learning unit 140 that generates a pre-trained model through learning based on and generates a neural network layer removal model using the pre-trained model and the second feature value, and weight values based on the neural network removal model. It may be configured to include a model generation unit 150 that extracts a third feature value by applying the third feature value and generates a final classification model for classifying the big data through learning based on a neural network using the third feature value.

The data acquisition unit 110 performs a function of acquiring data. The data acquisition unit 110 acquires data generated through sensors installed in the field. In particular, the data acquisition unit 110 may acquire big data to create a big data dataset, or label the big data to create a labeled (correct answer mark) labeled dataset.

Of course, the data acquisition unit 110 may acquire data by being connected to sensors through a communication network (not shown), or may be directly connected to sensors. To this end, the data acquisition unit 110 may include a communication modem, memory, and the like. The data may be sound data, but is not limited thereto, and image data is also possible.

A communication network refers to a connection structure capable of exchanging information between nodes such as a plurality of terminals and servers, such as a public switched telephone network (PSTN), a public switched data network (PSDN), and an Integrated Services Digital Network (ISDN). Networks), Broadband ISDN (BISDN), Local Area Network (LAN), Metropolitan Area Network (MAN), Wide LAN (WLAN), and the like.

However, the present invention is not limited thereto, and wireless communication networks CDMA (Code Division Multiple Access), WCDMA (Wideband Code Division Multiple Access), Wibro (Wireless Broadband), WiFi (Wireless Fidelity), HSDPA (High Speed Downlink Packet Access) ) network, Bluetooth, NFC (Near Field Communication) network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, and the like. Alternatively, it may be a combination of these wired communication networks and wireless communication networks.

The feature value extractor 120 performs a function of extracting feature values from data. In other words, the feature value extractor 120 performs a function of extracting a first feature value and a second feature value from the acquired big data dataset and labeling dataset, respectively.

The feature value is a log-mel-spectrogram, and is extracted by applying a mel filter and a log scale after performing short-time Fourier transform on the raw data of a general audio file.

The data storage 130 performs a function of storing data obtained through the data acquisition unit 110 and feature values obtained through the specific value extraction unit 120 . To this end, the data storage 130 may be a flash memory type, a hard disk type, a multimedia card micro type, or a card type memory (for example, SD (Secure Digital)). or XD (eXtreme Digital) memory, etc.), RAM (Random Access Memory, RAM), SRAM (Static Random Access Memory), ROM (Read Only Memory, ROM), EEPROM (Electrically Erasable Programmable Read Only Memory), PROM (Programmable Read Only Memory) Only Memory), a magnetic memory, a magnetic disk, and an optical disk may include at least one type of storage medium.

In addition, it may operate in relation to a web storage and a cloud server that perform a storage function on the Internet.

The learning unit 140 generates a pre-trained model through neural network-based transfer learning using the first feature value and generates a neural network layer removal model using the pre-trained model and the second feature value. perform a function

A neural network has a structure with several hidden layers between the input layer and the output layer. This is a method in which a machine learns and processes processing rules through data itself, even if humans do not suggest procedures or rules for data processing. A deep neural network (DNN), a convolution neural network (CNN), or a recurrent neural network (RNN) may be used as the neural network.

In one embodiment of the present invention, the learning unit 140 generates a pre-trained model through CNN-based learning using feature values. Then, the DNN layer is removed from the generated model to prepare a model for extracting new feature values. Then, prepare a dataset with the class you want to classify, and extract feature values from the data. The extracted data is regenerated through a weight value of a convolution layer based on a previously created model (ie, a pre-trained model), which becomes a new feature value.

Transfer learning means using some of the abilities of a neural network learned in a specific field to learn a neural network used in a similar or completely new field.

For example, the front end of neural networks such as Resnet (Residual Networks) or VGG (Visual Geometry Group) is generally composed of CNN layers. This CNN layer has the ability to extract features from images.

It is known to first extract novelties, then patterns, and finally shapes, etc. Therefore, by using the feature extraction capability of Resnet or VGG neural networks with high performance learned through tens of thousands to tens of millions of images, and by changing only the linear (Affine; matrix calculation for weights and biases) layer as the last output layer, Re-learning only this changed layer is transfer learning.

Referring to FIG. 1 , the model generation unit 150 extracts a third feature value by applying a weight value based on the neural network removal model, and generates a final classification model through neural network-based learning using the third feature value. .

Transfer learning is effective even when the number of training data is small, and the learning speed is fast. And it has the advantage of providing much higher accuracy than learning without transfer learning.

The feature value extraction unit 120, the learning unit 140, and the model generation unit 150 shown in FIG. 1 refer to units that process at least one function or operation, and may be implemented in software and/or hardware. there is. In hardware implementation, ASIC (application specific integrated circuit), DSP (digital signal processing), PLD (programmable logic device), FPGA (field programmable gate array), processor, microprocessor, other It may be implemented as an electronic unit or a combination thereof.

In software implementation, software component components (elements), object-oriented software component components, class component components and task component components, processes, functions, properties, procedures, subroutines, segments of program code, drivers, firmware, microcode, data , databases, data structures, tables, arrays, and variables. Software, data, etc. may be stored in memory and executed by a processor. The memory or processor may employ various means well known to those skilled in the art.

2 is a flowchart showing a model generation process according to an embodiment of the present invention. Referring to FIG. 2 , the data acquiring unit 110 acquires big data and creates a big data dataset (step S210).

Then, the feature value extractor 120 extracts a first feature value from the big data dataset (step S211).

Thereafter, the learning unit 140 generates a pre-trained model through neural network-based learning using the first feature value, and in the pre-trained model, a neural network layer (eg, For example, the DNN layer) is removed to generate a neural network layer removal model (steps S213, S215, S217, and S223).

Meanwhile, the data acquisition unit 110 creates a labeling dataset using the acquired big data (step S220).

Then, the feature value extraction unit 120 extracts a second feature value from the labeling dataset (step S221).

Then, the model generation unit 150 extracts a new third feature value by applying the weight of the previous convolution layer to the neural network layer removal model (step S225).

The first and second feature values are feature values extracted based on the mel spectrogram, and the third feature value is a feature value using a convolution filter in addition to the corresponding spectrogram. The previous convolutional layer is a convolutional layer with pre-learned weights through big data.

Thereafter, the model generation unit 150 utilizes the new third feature value to generate a final classification model through DNN-based learning (step S229). Since the final classification model establishes and uses a small number of classes to be classified, it has higher accuracy than a pre-learning model that needs to classify a large number of classes.

When the algorithm shown in FIG. 2 is applied, there is an advantage in that a model with good performance can be created even if the number of datasets of desired labels is not large.

As an example related to data labeling, suppose that the number of datasets to be classified is about 10,000, and the developer creates up to 50 datasets with 5 classes per class. However, if there is no time to classify while entering 9750 pieces of data directly, it is difficult to do the task. Therefore, by adding a model generation method and a series of procedures, it is possible to self-learn with little data and classify and label the data. Figure 3 shows this.

3 is a flowchart showing a final classification and labeling process according to an embodiment of the present invention. Referring to FIG. 3 , the data acquisition unit 110 acquires big data and creates a labeled labeling dataset (step S310).

Then, the feature value extraction unit 120 extracts a first feature value from the first labeling dataset (step S320).

Thereafter, the learning unit 140 generates a pre-trained model through neural network-based learning using the first feature value, and in the pre-trained model, a neural network layer (eg, For example, the DNN layer) is removed to generate a neural network layer removal model (step S330).

Then, the model generator 150 extracts a new second feature value by applying the weight of the previous convolution layer (step S340).

Then, a temporary classification model is generated through DNN-based learning using the second feature value (steps S301, S350, and S360). The ad hoc classification model can learn and classify many classes, but instead has a feature with low accuracy.

The model generating unit 150 applies the miraveling sound source dataset to the temporary classification model and evaluates the result (steps S301 and S370). In other words, it is checked whether the result value of classification is greater than a preset threshold value (N%). Here, N may be 90.

As a result of the confirmation, in step S370, if the result value of classification is smaller than the threshold value, steps S310 to S370 proceed again.

In contrast, in step S370, if the result value of classification is greater than the threshold value, the second labeling dataset is reconstructed by moving to another dataset and labeling (steps S380 and S381). That is, labeling is performed by filtering out only data that exceeds a predetermined threshold. In this way, only data that is significantly similar to the existing learned data can be filtered out.

Thereafter, a newly added additional dataset is created by combining the newly reconstructed first labeling dataset and the existing second labeling dataset, and the threshold is lowered (ie, subtracted) by 1% of a preset value (step S382). Thereafter, steps S310 to S381 are performed.

In step S382, if the threshold value is a preset reference value (N==70%), the temporary classification model is determined as the final classification model, and the unlabeled sound source dataset is finally classified and labeled using the final classification model ( Steps S383, S390).

In other words, repeat until a certain threshold is reached. When a certain threshold is reached, the transfer learning model obtained in the last iteration is determined as the final classification model, and the remaining unlabeled data is applied to the final classification model for final classification and labeling.

4 is a flowchart illustrating a process of generating a new additional data set according to an embodiment of the present invention. Referring to FIG. 4 , 100% accurate classification through the corresponding algorithm may be difficult. However, there is an advantage in that a data group having a certain characteristic is formed by the trained model by continuously adding data, and the data are labeled through the formed data group.

To elaborate, newly labeled second labeling data is generated through the temporary classification model, and the second labeling dataset is divided into a training dataset and a test dataset (S410 and S411). The training dataset can be 90%, and the test dataset can be 10%. Of the classified data, divide 90% for training and 10% for testing, and proceed with foldering together with the existing dataset.

Meanwhile, an additional dataset is created by combining the split dataset with the dataset used for the existing learning (steps S420, S421, and S430).

Even if people reclassify the data after the task, since most of the data is correctly structured, only a few data irrelevant to the label need to be removed.

In addition, the steps of a method or algorithm described in connection with the embodiments disclosed herein are implemented in the form of program instructions that can be executed through various computer means such as a microprocessor, processor, CPU (Central Processing Unit), etc. It can be recorded on any available medium. The computer readable medium may include program (instruction) codes, data files, data structures, etc. alone or in combination.

The program (command) code recorded on the medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROM, DVD, and Blu-ray, and read only memory (ROM). ), RAM (Random Access Memory), flash memory, etc., may include a semiconductor memory device specially configured to store and execute program (command) codes.

Here, examples of the program (command) code include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler. The hardware devices described above may be configured to act as one or more software modules to perform the operations of the present invention, and vice versa.

100: data classification system
110: data acquisition unit
120: feature value extraction unit
130: data storage
140: learning unit
150: model generating unit

Claims (10)

(a) acquiring big data by the data acquisition unit 110 and generating a big data dataset and a labeled labeling dataset;
(b) extracting, by the feature value extractor 120, first feature values and second feature values from the big data dataset and the labeling dataset, respectively;
(c) The learning unit 140 generates a pre-trained model through neural network-based learning using the first feature values, and uses the pre-trained model and the second feature values to generate a neural network. generating a layer removal model; and
(d) The model generation unit 150 extracts a third feature value by applying a weight value based on the neural network removal model, and classifies the big data through learning based on the neural network using the third feature value. generating a final classification model for
Data classification method through deep learning-based learning comprising a.
According to claim 1,
The neural network layer removal model is a data classification method through deep learning-based learning, characterized in that generated using transfer learning to remove the neural network layer.
According to claim 2,
The transfer learning is data through deep learning-based learning, characterized in that by removing a Deep Neural Network (DNN) layer from the pre-trained model generated through learning based on a Convolution Neural Network (CNN). classification method.
According to claim 1,
The classification is a data classification method through deep learning-based learning, characterized in that acoustic data classification.
According to claim 1,
The weight value is a data classification method through deep learning-based learning, characterized in that the weight value of the convolution layer based on the pre-trained model.
(a) generating, by the data acquisition unit 110, big data to create a mirabella-labeled sound source data set and a first labeled labeling data set;
(b) extracting, by the feature value extraction unit 120, a first feature value from the first labeling dataset;
(c) The learning unit 140 generates a pre-trained model through neural network-based learning using the first feature values, and uses the pre-trained model and the first feature values to generate a neural network. generating a layer removal model;
(d) The model generating unit 150 extracts a second feature value by applying a weight value based on the neural network removal model, and uses the second feature value to perform learning based on the neural network to perform the unlabeled sound source data. generating a temporary classification model for classifying the set; and
(e) The model generation unit 150 classifies the unlabeled sound source dataset using the temporary classification model, and classifies the big data using a reconstructed second labeling dataset according to the classification result. determining a classification model;
Data classification method through deep learning-based learning comprising a.
According to claim 6,
In step (e),
(e-1) comparing, by the model generation unit 150, a result value according to the classification result with a preset threshold value;
(e-2) if the result value is greater than the threshold, the model creation unit 150 moves to another dataset and labels it to reconstruct a second labeling dataset; and
(e-3) generating an additional dataset by combining the second labeling dataset with the first labeling dataset; Data classification method through deep learning-based learning, characterized in that it comprises a.
According to claim 7,
In the step (e-3),
subtracting a preset value from the threshold when the additional data set is generated;
repeating steps (a) to (e) until the threshold value reaches a predetermined reference value; and
Determining the temporary classification model generated when the repetition is last performed as the final classification model; Data classification method through deep learning-based learning, characterized in that it comprises a.
According to claim 7,
The second labeling dataset is a data classification method through deep learning-based learning, characterized in that divided into a training dataset and a test dataset.
According to claim 9,
Data classification method through deep learning-based learning, characterized in that the ratio of the training dataset and the test dataset is 90%: 10%.