KR20200099966A - Method and apparatus for learning based on data including nominal data - Google Patents

Abstract

The present invention relates to a method and a device for learning based on data including nominal data. The device for learning can comprise: a data processing part that acquires first input data and second input data from raw data, wherein the first input data and the second input data are different types of data from each other; a first learning part that acquires first output data by performing learning by using both the first input data and the second input data; a second learning part that acquires second output data by performing learning by using the raw data; and a combination part that performs learning by using the first output data and the second output data. Therefore, excellent and appropriate learning results can be acquired even when learning is performed based on different types of data such as nominal data or numerical data.

Description

Learning method and device based on data including nominal data {METHOD AND APPARATUS FOR LEARNING BASED ON DATA INCLUDING NOMINAL DATA}

It relates to a learning method and apparatus based on data including nominal data.

In recent years, according to the development of hardware and software, research on machine learning in which hardware, such as a computer device, itself learns and generates rules has been conducted in various fields. For example, machine learning is also used to predict various situations or conditions that can occur. For example, machine learning is used to predict the likelihood of a traffic accident on the road, traffic flow in a city area or suburbs, predict whether a machine or communication equipment has failed, or predict whether a disease will occur. Has become. However, when machine learning is used to determine the situation, different types of data are often used as input data, so it is difficult to obtain an appropriate learning model or prediction result with only a conventional simple learning model. For example, some of the input data tends to become meaningless, and there is also a significant difference between the predicted result and the actual result because the input data is not properly reflected.

Republic of Korea Patent Publication No. 2017-0065898 Republic of Korea Patent Publication No. 2015-0072471

A learning method and apparatus based on data including nominal data capable of obtaining better and appropriate learning results even when learning is performed based on different types of data such as nominal data or numeric data. What is provided is a task to be solved.

In order to solve the above-described problem, a learning method and apparatus based on data including nominal data are provided.

The learning apparatus includes a data processing unit in which the first input data and the second input data are of different types of data, and the first input data to obtain first output data by performing learning using both the first input data and the second input data. 1 learning unit, a second learning unit that acquires second output data by performing learning using the raw data, and a combination unit that performs learning using the first output data and the second output data. have.

The learning method includes obtaining first input data and second input data from raw data, wherein the first input data and the second input data are data of different types, both of the first input data and the second input data. Acquiring first output data by performing learning using, acquiring second output data by performing learning using the raw data, and learning using the first output data and the second output data It may further include the step of performing.

According to the above-described learning method and apparatus based on data including nominal data, more appropriate and superior learning results are obtained even when learning is performed based on data of different types such as nominal or numeric types. This can be done, and accordingly, the effect of improving the performance of the learning model can be obtained.

According to the above-described learning method and apparatus based on data including nominal data, the risk of over-fitting occurs more than a general data over-sampling technique used to supplement the asymmetric class structure. There is also an advantage that can be significantly reduced.

According to the above-described learning method and apparatus based on data including nominal data, it is possible to obtain an effect of improving the stability of the model while improving performance compared to other learning models.

According to the above-described learning method and apparatus based on data including nominal data, it is possible to improve and improve the performance of various systems or services that also use nominal data such as a content recommendation system or a risk prediction system. .

1 is a block diagram of a learning device according to an embodiment.
2 is a diagram for explaining an example of the operation of the learning device.
3 is a block diagram illustrating an embodiment of a learning processing unit.
4 is a diagram for describing an example of an embedded layer.
5 is a diagram illustrating an example of a program code for distinguishing nominal data and numeric data.
6 is a diagram for describing an example of an operation of performing transfer learning.
7 is a flowchart of an embodiment of a learning method.

In the following specification, the same reference numerals refer to the same elements unless otherwise specified. The term "unit" used below may be implemented in software or hardware, and according to an embodiment, one "unit" is implemented as one physical or logical part, or a plurality of "units" It may be implemented as a physical or logical part, or one'unit' may be implemented as a plurality of physical or logical parts.

When a part is said to be connected to another part throughout the specification, it may mean a physical connection depending on the part and another part, or may mean electrically connected. In addition, when a part includes another part, this does not exclude another part other than the other part unless otherwise stated, and it means that another part may be included further according to the designer's choice. do.

Terms such as first and second are used to distinguish one part from another, and unless otherwise specified, they do not mean sequential expressions. In addition, expressions in the singular may include plural expressions unless there is clearly an exception in the context.

Hereinafter, an embodiment of a learning apparatus and a learning processing unit for performing the same will be described with reference to FIGS. 1 to 6.

1 is a block diagram of a learning device according to an embodiment.

As shown in FIG. 1, the learning device 10 may include a learning processing unit 100 that performs learning on at least one input data, and if necessary, a storage unit 20 and an input unit ( 30) and the output unit 32 may include at least one. At least two of the storage unit 20, the input unit 30, the output unit 32, and the learning processing unit 100 are provided to enable mutual data transmission and reception through cables, circuits, and/or wireless communication devices.

The storage unit 20 stores at least one data and/or at least one application (which may be referred to as software, app, or program) required for the operation of the learning processing unit 100, and at the request of the learning processing unit 100 Accordingly, it is provided to provide data and/or applications to the learning processing unit 100.

According to an embodiment, the storage unit 20 may store raw data 1 to be input to the learning processing unit 100 for learning. The raw data 1 is at least one data set in which at least one data is collectively formed, for example, a first data set to an nth data set (1-1 to 1-n, where n is a natural number greater than 2) It may include, and at least two of these data sets 1-1 to 1-n may be constructed as sets of different types of data. For example, the first data set (1-1) includes a set of numeric data (scalar data), and the second data set (1-2) is nominal data (can be expressed as qualitative data, etc.) It may also contain a set of. Depending on the situation, these data sets 1-1 and 1-2 may be separated from each other and stored in the storage unit 20, or may be mixed with each other and stored in the storage unit 20. That is, in one raw data 1, different types of data, for example, nominal data 1-1 and numeric data 1-2 may be mixed with each other.

According to an embodiment, the storage unit 20 may further store a learning model 2 that is learned by the learning processing unit 100 or to be used for learning. The learning model 2 may comprise a learning algorithm, for example. Learning algorithms include, for example, a deep neural network (DNN), a convolutional neural network (CNN), a recurrent neural network (RNN), a convolutional recurrent neural network (CRNN), a multilayer perceptron (MLN), a deep trust neural network (DBN), and/or Deep Q-Networks, etc. may be included, but the present invention is not limited thereto. In addition, the storage unit 20 may further include a transfer learning parameter 3 to be used for transfer learning of the learning processing unit 100 when the learning processing unit 100 performs transfer learning. In addition, the storage unit 20 may further store models, data and/or applications required for the operation/processing of the learning processing unit 100.

The storage unit 20 may include, for example, at least one of a main memory device and an auxiliary memory device. The main memory device may be implemented using a semiconductor storage medium such as ROM and/or RAM. The ROM may include, for example, a conventional ROM, EPROM, EEPROM, and/or MASK-ROM. The RAM may include, for example, DRAM and/or SRAM. Auxiliary storage devices include flash memory devices, SD (Secure Digital) cards, solid state drives (SSDs, Solid State Drives), hard disk drives (HDDs, Hard Disc Drives), magnetic drums, compact disks (CDs), DVDs ) Or an optical media such as a laser disk, a magnetic tape, a magneto-optical disk, and/or a floppy disk, and at least one storage medium capable of permanently or semi-permanently storing data.

The input unit 30 may receive an external command/instruction and/or receive at least one data. For example, the input unit 30 receives a learning start command or a learning result output command, etc. according to a user's operation, and/or at least one data corresponding to the raw data 1 or a transfer learning parameter 3 I, learning model (2), etc. can be input. The input unit 30 is, for example, a data terminal to which a keyboard, a mouse, a keypad, a touch screen, a touch pad, a trackball, a track pad, a scanner, an external memory device (such as an external hard disk or an external flash memory device) can be mounted. For example, a universal serial bus terminal, etc.), a wired or wireless communication module, and/or a separately generated data set (1-1 to 1-n), or data input in response to a user's operation, etc. It may include at least one device, but is not limited thereto.

The output unit 32 stores the learning result of the learning processing unit 100, the raw data 1 stored in the storage unit 20, the learning model 2, and/or the transfer learning parameter 3 according to the instruction of the user. Alternatively, it can be output externally according to a predefined setting. The output unit 32 may include, for example, an image output device, a printer device, or a sound output device.

The learning processing unit 100 performs learning based on the raw data 1 to learn a predetermined learning model, for example, the learning model 2 stored in the storage unit 20 to acquire the learned model, and/ Alternatively, result data corresponding to the input data may be obtained by applying predetermined input data to the trained model. A detailed description of the operation of the learning processing unit 100 will be described later.

The learning processing unit 100 may be implemented using one or more processors. Here, the processor is, for example, a central processing unit (CPU, Central Processing Unit), a microcontroller unit (MCU, Micro Controller Unit), a Micom (Micro Processor), an application processor (AP, Application Processor), electronic control It may include a unit (ECU, Electronic Controlling Unit) and/or other electronic devices capable of processing various operations and generating control signals. These devices can be implemented using, for example, one or more semiconductor chips and related components. The learning processing unit 100 may run an application (not shown) stored in the storage unit 20 to perform operations, judgments, processing, and/or control operations related to learning. Here, the application stored in the storage unit 20 may be previously created by a designer or transmitted from an external storage medium (eg, an external hard disk device) and stored in the storage unit 20, and/or It may be acquired or updated through an electronic software distribution network accessible through a wired or wireless communication network.

The above-described learning device 10 can be implemented based on an information processing device capable of performing data operation, processing, and various control operations related thereto. For example, the learning device 10 is a desktop computer, a laptop computer, a server computer device, a smart phone, a tablet PC, a smart watch, a head mounted display (HMD) device, a navigation device, a personal digital assistant (PDA, Personal Digital Assistant), portable game consoles, digital televisions, set-top boxes, home appliances (robot cleaners, washing machines, refrigerators, etc.), electronic billboards, artificial intelligence sound reproducing devices, vehicles, robots, mechanical devices, or other learning processes that can be performed. It may include at least one of various devices.

2 is a diagram for explaining an example of the operation of the learning device.

According to the exemplary embodiment shown in FIG. 2, the raw data 1 is input to the learning processing unit 100, and the learning processing unit 100 performs learning based on the raw data 1 to obtain a learning result 120a, or less. The first learning result) can be obtained. More specifically, when the raw data 1 is input, the learning processing unit 100 generates and acquires the input data 114 based on the raw data 1, and responds to all input data 114 (for example, nominal data). And/or input data corresponding to the numerical data) into the learning unit 121-1 for all data to perform learning, and/or to nominal data among the input data 114 After learning by performing learning by inputting the corresponding data into the learning unit 121-2 for nominal data (hereinafter, partial data learning unit), the entire data learning unit 121-1 and the partial data learning unit ( Based on the learning result of 121-2), a learning result 121a (hereinafter referred to as first output data) corresponding to the input data 114 may be obtained. In addition, the learning processing unit 100 obtains the second output data 122a by separately inputting the raw data 1 into the learning unit 122 (hereinafter referred to as the second learning unit) for the raw data 1 itself, The first learning result 120a may be obtained by combining the first output data 121a and the second output data 122a. In this case, the combination of the first output data 121a and the second output data 122a may be performed by employing an ensemble model.

3 is a block diagram illustrating an embodiment of a learning processing unit.

More specifically, as illustrated in FIG. 4, the learning processing unit 100 may include a data processing unit 110, a combination learning unit 120, and a result output unit 140, according to an exemplary embodiment.

The data processing unit 110 acquires the raw data 1 and acquires the input data 114 to be input to the combination learning unit 120 by converting all or part of the data in the raw data 1 or not converting it. can do. In this case, the data processing unit 110 may be provided including a data acquisition unit 111, a data conversion unit 112, and an input data generation unit 113.

The data acquisition unit 111 acquires at least one data set (1-1 to 1-n) of the raw data 1 from the storage unit 20, and the acquired data set (1-1 to 1-n) The data may be transferred to the data conversion unit 112 or the input data generation unit 113 according to the type of data of. For example, the data acquisition unit 111 is a data set including numeric data from among the data sets 1-1 to 1-n. For example, the first data set 1-1 is an input data generator 113 ), and the data set including nominal data, for example, the second data set 1-2 may be designed to be transmitted to the data conversion unit 112. Accordingly, the numeric data, for example, the first data set 1-1 is transferred to the input data generator 113 as a single value without conversion.

The data conversion unit 112 may convert the received data and transmit the converted data to the input data generation unit 113. For example, the data conversion unit 112 receives nominal data, for example, at least one data of the second data set 1-2, and converts each received data into a node having a predefined size. And transfer the changed data to a node having a predefined size to the input data generator 113. In this case, the nominal data may all be changed to nodes of the same size as defined in advance. The size of the node may be specified by the user or may be predefined by the designer.

4 is a diagram for explaining an example of an embedded layer, and is a diagram in which nominal data is expressed in a one-hot format.

The input data generation unit 113 may generate data to be input to the combination learning unit 120 based on at least one data transmitted from the data acquisition unit 111 and/or the data conversion unit 112. In this case, the input data generation unit 113 corresponds to the input data 114-1 (hereinafter referred to as first input data) corresponding to the data transmitted from the data acquisition unit 111 and the data transmitted from the data conversion unit 112. Data to be input to the combination learning unit 120 may be generated by generating input data 114-2 (hereinafter referred to as second input data). Here, the first input data 114-1 may be composed of numeric data and/or the second input data 114-2 may be composed of nominal data. The input data generator 113 may convert the input data 114-1 and 114-2 as necessary. For example, the input data generation unit 113 removes unnecessary data for analysis or converts at least one of the input data 114-1 and 114-2 (e.g., one-hot ) Encoding processing, etc.) and/or scaling processing (e.g., changing the data range to a value between 0 and 1) on at least one of the input data 114-1 and 114-2 It is also possible to perform further processing necessary to speed up learning or improve accuracy. Through this process, input data 114 to be input to the combination learning unit 120 may be formed. The input data 114 generated by the input data generation unit 113 is transferred to the first learning unit 121 of the combination learning unit 120, and is transferred from the first learning unit 121 to an embedded layer. Can be used. The embedded layer is a layer created by being fully connected to a predefined number (k) of nodes by a user or designer, and is used as an input value (input layer) to the combination learning unit 120 as described later. . Referring to the example shown in FIG. 4, the embedded layer includes nominal data (data expressed as 1 within a feature size according to a one-hot format) and numeric data (data having a value other than 1). Is created by The embedded layer may have a predetermined value (k) predefined by a user or a designer, for example, a value obtained by multiplying the size of a node and the size of a feature as the size.

The combination learning unit 120 may include a first learning unit 121, a second learning unit 122 and a combination unit 123 as shown in FIG. 3.

The first learning unit 121 receives the input data 114-1 and 114-2 from the input data generation unit 113 and performs learning based on this. The first learning unit 121 may include a whole data learning unit 121-1 and a partial data learning unit 121-2 in an embodiment.

The whole data learning unit 121-1 performs learning using all types of data 114 of the embedded layer, for example, both the first input data 114-1 and the second input data 114-2. Can be done. In other words, the total data learning unit 121-1 includes unconverted numeric data 114-1, that is, numeric data in its primitive form, and nominal data 114- converted into nodes of a predetermined size. 2) can be used together to perform learning. The full data learning unit 121-1 performs learning using, for example, a deep neural network, a convolutional neural network, a recurrent neural network, a convolutional recurrent neural network, a multilayer perceptron, a deep trust neural network and/or a deep Q-network, etc. You may. According to an embodiment, the total data learning unit 121-1 may include at least one hidden layer, and specifically, may include, for example, four hidden layers. The size of the hidden layer may be determined based on, for example, trial-error.

Some data learning unit 121-2 includes input data derived from nominal data among the first input data 114-1 and the second input data 114-2, that is, the second input data 114-2. Learning can be performed based on. Accordingly, some of the data learning units 121-2 perform learning based on data converted into nodes of a predetermined size in the same manner. Some data learning units 121-2 may linearly analyze the second input data 114-2. If necessary, the partial data learning unit 121-2 may apply a weight to the second input data 114-2. Some of the data learning units 121-2 may perform learning based on, for example, a factorization model (FM). Since each data 1-2 has been converted to a node of the same size, sparse characteristics caused by nominal data are alleviated, and the dot product operation can be performed. Depending on the embodiment, a class weight may be further used to obtain the first output data 121a. The class weight is a weight assigned to each class, where the class may include a category of a label used in machine learning or the like.

5 is a diagram illustrating an example of a program code for distinguishing nominal data and numeric data.

If necessary, the first learning unit 121 may classify the nominal data 141-2 from the entire input data 141 (that is, an embedded layer) before the operation of the partial data learning unit 121-2. Specifically, as shown in the code of FIG. 5, the first learning unit 121 extracts only nominal data processed with the same size (for example, the size of a node set by a user, etc.) from the entire embedding layer and partially Slicing pre-processing may be further performed to transfer to the data learning unit 121-2. In more detail, the first learning unit 121 sequentially determines whether the value (feature) of the data is (for example, 1 or other values) for each data in the overall input data 141. The overall input data 141 is determined by determining, adding a value for identification to each data according to the value of the data, and inputting each data to a specific learning unit 121-1 and 121-2 according to the added value. Can also be classified.

As illustrated in FIG. 2, first output data 121a may be obtained according to the learning results of the whole data learning unit 121-1 and the partial data learning unit 121-2. The first output data 121a may include, for example, a prediction result according to learning by the first learning unit 121.

According to an embodiment, the process of acquiring the first output data 121a of the first learning unit 121 is through the same process as the process of acquiring the result data by a deep factorization model (DeepFM) or This can be done through a partially modified process.

The learning process of the first learning unit 121 described above may be expressed by, for example, Equation 1 below.

[Equation 1]

는 일부 데이터 학습부(121-2)에서 수행되는 학습을 의미하고,

는 전체데이터 학습부(121-1)에서 수행되는 학습을 의미한다. 여기서 가중치(

)는 학습 과정에서 변경될 수도 있다.

는 제1 출력 데이터(121a)를 의미한다.In Equation 1

Means learning performed by some data learning unit 121-2,

Denotes learning performed by the whole data learning unit 121-1. Where the weight(

) May change during the learning process.

Denotes the first output data 121a.

In a typical learning algorithm, a data preprocessing process such as artificially creating or deleting data in a class is required to alleviate biased learning of an unbalanced class. However, this pre-processing process can increase the risk of overfitting in learning. However, since the above-described first learning unit 121 does not require the process of artificially generating and removing data, and can perform learning by further using class weights as needed, it has another conventional learning algorithm. You can solve the overfitting problem. The first output data 121a obtained according to the learning result of the first learning unit 121 may be transmitted to the combination unit 123.

The second learning unit 122 may be provided to perform learning based on the raw data (1:1-1 to 1-n). In other words, the first learning unit 121 performs learning based on the input data 114-1 and 114-2 obtained from the data processing unit 110, but the second learning unit 122 is the input data 114. Learning is performed based on raw data (1:1-1 to 1-n) stored in the storage unit 20, not on the basis of -1, 114-2). The second learning unit 122 may perform learning based on, for example, a deep neural network, a convolutional neural network, a recurrent neural network, a convolutional recurrent neural network, a multilayer perceptron, a deep trust neural network, and/or a deep Q-network. have. In this case, the learning algorithm used by the second learning unit 122 may be the same as or different from the learning algorithm used in the above-described all data learning unit 121-1. The learning result of the second learning unit 122, that is, the second output data 122a may also be transmitted to the combination unit 123. The second output data 122a may include, for example, a prediction result according to learning by the second learning unit 122.

The combination unit 123 merges the learning result of the first learning unit 121 and the learning result of the second learning unit 122, and as shown in FIG. 2, the first learning result 120a, for example, New learned models can be acquired. More specifically, the combination unit 123 receives the output data 121a of the first learning unit 121 and the output data 122a of the second learning unit 122, and receives these output data 121a and 122a. Further learning can be performed by using the input value. In other words, the first output data 121a and the second output data 122a may be used as training data. According to an embodiment, the process of generating a learning model by the combination unit 123 may be performed using the same or partially different stacking methods of an ensemble model.

The first learning result 120a obtained by the combining unit 123, for example, a learning model or result data obtained by being applied to a learning model, may be transmitted to the result output unit 140. The result output unit 140 may transmit the received learning result to the output unit 32 to be output to the outside and/or transmit the received learning result to the storage unit 20 to be temporarily or non-temporarily stored.

The two steps described above, that is, acquiring the first output data 121a and the second output data 122a, and the step of creating a new model by combining the output results, perform training using only a single model. Compared to the case, the performance is relatively improved and the model performance can be stabilized.

According to an embodiment, at least one of the first learning unit 121, the second learning unit 122, and the combination unit 123 may perform learning by further applying transfer learning.

6 is a diagram for describing an example of an operation of performing transfer learning.

According to an embodiment, as shown in FIGS. 3 and 6, the combination unit 123 and the transfer learning unit 130 may be connected to each other to perform learning. More specifically, the transfer learning unit 130 may include at least one hidden layer each having at least one node, and at least one node of the hidden layer of the combination learning unit 120 is a hidden layer of the transfer learning unit 130 It is connected to at least one node of to form a learning model as a whole. The combination unit 123 and the transfer learning unit 130 may further perform learning based on transfer learning to obtain a learning result 130a (hereinafter, a second learning result). In this case, the transfer learning unit 130 may acquire an improved learning model by performing learning based on information of a model that has already been learned (for example, a transfer learning parameter 3). In other words, the transfer learning unit 130 performs learning using at least one model related to the model (for example, a traffic volume prediction model) to be learned by the learning device 10 and learned in advance, Learning may be performed by recycling the parameters 3 of at least one node in at least one hidden layer belonging to the learned model, respectively. Here, for performing transfer learning, all of the hidden layers of at least one hidden layer of the previously learned model may be used, or only some of the hidden layers may be used. Further, for transfer learning, the parameter 3 of all nodes in the same hidden layer may be used, or only the parameter 3 of some of all nodes may be used.

The second learning result 130a obtained by the transfer learning unit 130 may be transmitted to the result output unit 140 for external output and/or internal storage. When the transfer learning unit 130 is further included as described above, the first learning result 120a may or may not be transmitted to the result output unit 140 according to the designer's selection.

The above-described data processing unit 110, the combination learning unit 120, the transfer learning unit 130, and/or the result output unit 140, and each of the elements 111, 112, 113, and 121, 122, 123), etc. may be physically classified and/or logically separated. For example, the data processing unit 110, the combination learning unit 120, the transfer learning unit 130, and the result output unit 140 may be implemented through a processor made of a single semiconductor device, and each physically separated It may be implemented through a plurality of different processors. In addition, the above-described data processing unit 110, combination learning unit 120, transfer learning unit 130, and/or result output unit 140 are implemented using at least one physical processor in various ways according to the designer's selection. It is possible.

Hereinafter, an embodiment of a learning method will be described with reference to FIG. 7.

7 is a flowchart of an embodiment of a learning method.

As shown in FIG. 7, first, raw data to be used for training a learning model is obtained (300). The raw data may include nominal data, and may further include other data (eg, numeric data) other than nominal data. The raw data may be pre-stored in a storage medium inside the device, directly input by a user or designer, input through another external storage medium, or transmitted through a wired or wireless communication network. May be.

Input data may be generated based on the raw data (302). In this case, input data may be generated by converting the nominal data into nodes of the same size among the original data and maintaining the numeric data without conversion. The generated input data can be used as an embedded layer in a learning model.

All of the generated input data (i.e., both nominal data and numeric data) may be used for full data learning (304). The entire data learning may be performed based on a predetermined machine learning algorithm, such as a deep neural network, a convolutional neural network, a recurrent neural network, a convolutional recurrent neural network, a multilayer perceptron, a deep trust neural network, and/or a deep Q-network.

Meanwhile, among the generated input data, nominal data may be additionally used for learning nominal data (306). If necessary, prior to the nominal data learning, a process of separately extracting nominal data from the input data may be further performed. Learning of nominal data may be performed based on a factorization model. If necessary, it is possible to further use class weights for learning nominal data.

When the result of learning the full data and the result of learning the nominal data are obtained, the first output data is obtained by combining these results (308).

Meanwhile, the second learning may be further performed simultaneously or sequentially with the process of generating the input data or the process of obtaining the first output data (302 to 308) (310). The second learning is performed using raw data. The second learning may also be performed based on a learning algorithm such as a deep neural network, a convolutional neural network, a recurrent neural network, a convolutional recurrent neural network, a multilayer perceptron, a deep trust neural network, and/or a deep Q-network. In response to the second learning, second output data is obtained (312).

When the first output data and the second output data are obtained (308, 312), the first output data and the second output data are combined (314). Specifically, learning is further performed using the first output data and the second output data as input values, and a new learning model may be generated accordingly. According to an embodiment, generation of a learning model based on a combination of the first output data and the second output data may be performed using a stacking method.

The newly created learning model may be output to the outside in a visual and/or auditory form, or may be temporarily or non-temporarily stored in a storage medium.

According to an embodiment, transfer learning may be additionally performed (316). That is, generation of the learning model by combination may be performed by further applying transfer learning. The learning model generated according to transfer learning may also be externally output or stored in a storage medium.

The learning method according to the above-described embodiment may be implemented in the form of a program that can be driven by a computer device. Here, the program may include a program command, a data file, a data structure, or the like alone or in combination. The program may be designed and produced using machine code or high-level language code. The program may be specially designed to implement the above-described method, or may be implemented using various functions or definitions previously known to and available to those skilled in the computer software field. In addition, here, the computer device may be implemented including a processor or a memory that enables the function of a program to be realized, and may further include a communication device if necessary.

A program for implementing the above-described learning method may be recorded on a computer-readable recording medium. The computer-readable recording medium includes, for example, a magnetic disk storage medium such as a hard disk or a floppy disk, a magnetic tape, an optical recording medium such as a compact disk or a DVD, a magnetic-optical recording medium such as a floppy disk, and a ROM. , A semiconductor storage device such as RAM or flash memory, etc., may include various types of hardware devices capable of storing a specific program executed according to a call from a computer.

The above-described learning method and apparatus can be applied to various fields. For example, the above-described learning method and device may be used in a system for providing or recommending specific information or data to a user based on the user's data, and more specifically, for example, advertisement, movie, insurance, It can be applied to systems, devices, or services in various fields, such as shopping (for example, Internet shopping or TV-based home shopping, etc.) and/or content providing services (for example, Internet protocol television services, etc.). Further, the above-described learning method and apparatus can be applied to a system for predicting a current or future situation/state based on existing data (for example, a risk prediction system based on traffic accident data). In addition, the above-described learning method and apparatus can be applied to a plurality of data-based learning systems, devices, and methods (services, etc.) having nominal data in the same manner or by partially modified methods.

As described above, various embodiments of a learning apparatus and method based on data including nominal data have been described, but the apparatus and method are not limited to the above-described embodiments. Various devices or methods that can be implemented by modifying and modifying based on the above-described embodiment by a person having ordinary knowledge in the relevant technical field may also be examples of learning devices and methods based on data including the above-described nominal data. have. For example, the described techniques are performed in an order different from the described method, and/or components such as systems, structures, devices, circuits, etc. described are combined or combined in a form different from the described method, or other components or Even if it is replaced or replaced by an equivalent, it may be an embodiment of a learning apparatus and method based on data including the above-described nominal data.

1: raw data 2: learning model
3: parameter for transfer learning 10: learning device
20: storage unit 30: input unit
32: output unit 100: learning processing unit
110: data processing unit 120: combination learning unit
121: first learning unit 122: second learning unit
123: union unit 130: transfer learning unit
140: result output section

Claims (13)

A data processing unit obtaining first input data and second input data from raw data, wherein the first input data and the second input data are data of different types;
A first learning unit that acquires first output data by performing learning using both the first input data and the second input data;
A second learning unit that acquires second output data by performing learning using the raw data; And
A learning apparatus further comprising a combination unit for performing learning using the first output data and the second output data.
The method of claim 1,
The raw data includes nominal data corresponding to the second input data.
The method of claim 2,
The learning apparatus further includes numeric data corresponding to the first input data as the raw data.
The method of claim 3,
The data processing unit acquires the second input data by converting the nominal data into nodes of a predefined size.
The method of claim 1,
The first learning unit,
An entire data learning unit that performs learning using the first input data and the second input data; And
And a partial data learning unit that performs learning using only the second input data.
The method of claim 1,
The learning apparatus further comprises a transfer learning unit that is connected to the combination unit and further performs learning based on information on the model that has already been learned.
The method of claim 1,
At least one of the first learning unit, the second learning unit, and the combination unit,
Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Convolutional Recurrent Neural Network (CRNN), Multilayer Perceptron (MLN), Deep Trust Neural Network (DBN) and Deep Q-Networks A learning device that performs the learning based on at least one of.
Obtaining first input data and second input data from raw data, wherein the first input data and the second input data are different types of data;
Acquiring first output data by performing learning using both the first input data and the second input data;
Acquiring second output data by performing learning using the raw data; And
The learning method further comprises: performing learning using the first output data and the second output data.
The method of claim 8,
The raw data includes nominal data corresponding to the second input data and numeric data corresponding to the first input data.
The method of claim 9,
Obtaining first input data and second input data from the raw data,
And obtaining the second input data by converting the nominal data into nodes of a predefined size.
The method of claim 8,
The step of obtaining first output data by performing learning using both the first input data and the second input data,
Performing learning using the first input data and the second input data; And
Performing learning using only the second input data; Learning method comprising at least one of.
The method of claim 8,
Learning method further comprising; further performing learning based on information of the model that has already been learned.
The method of claim 8,
The learning is performed by a deep neural network (DNN), a convolutional neural network (CNN), a recurrent neural network (RNN), a convolutional recurrent neural network (CRNN), a multilayer perceptron (MLN), a deep trust neural network (DBN), and a deep Q- A learning method comprising performing the learning based on at least one of Deep Q-Networks.

Images