KR101864412B1 - Data managing method, apparatus and program for machine learning - Google Patents

Abstract

The present invention discloses a method, a device and a program for managing learning data, which can reduce a time required for learning and a storage space for storing learning data. The method for managing learning data comprises the following steps of: a computer obtaining new learning data; the computer performing a test with respect to the obtained new learning data by using a learned model; the computer extracting first learning data in which labeled information is obtained with an accuracy of a predetermined first reference value or more, as a result of the test; the computer deleting the extracted first learning data from the new learning data; and the computer learning again the learned model by using the new learning data in which the extracted learning data is deleted.

Description

DATA MANAGEMENT METHOD, APPARATUS AND PROGRAM FOR MACHINE LEARNING,

The present invention relates to a learning data management method, apparatus, and program.

Artificial intelligence (AI) system is a computer system that implements human-level intelligence. Unlike existing Rule-based smart systems, AI is a system in which machines learn, judge and become smart. Artificial intelligence systems are increasingly recognized and improving their understanding of user preferences as they are used, and existing rule-based smart systems are gradually being replaced by deep-run-based artificial intelligence systems.

Artificial intelligence technology consists of element technologies that utilize deep learning and machine learning.

Machine learning is an algorithm technology that classifies / learns the characteristics of input data by itself. Element technology is a technology that simulates functions such as recognition and judgment of human brain using machine learning algorithms such as deep learning. Understanding, reasoning / prediction, knowledge representation, and motion control.

Deep Neural Network (DNN) is widely used for machine learning and is considered to be close to human level in object, face, and speech recognition. In order to learn the neural network, a method of fine tuning by using a basic data set and then using a data set of a specific field is frequently used.

A problem to be solved by the present invention is to provide a learning data management method, an apparatus, and a program.

The problems to be solved by the present invention are not limited to the above-mentioned problems, and other problems which are not mentioned can be clearly understood by those skilled in the art from the following description.

According to one aspect of the present invention, there is provided a learning data management method comprising the steps of: acquiring new learning data from a computer; performing a test on the acquired new learning data using the learned model; , Extracting first learning data in which the labeled information is acquired with an accuracy of at least a first reference value, extracting the extracted first learning data from the new learning data, and extracting the extracted learning data And re-learning the learned model using the deleted new learning data.

The predetermined first reference value may be adjusted such that the amount of the new learning data does not exceed a predetermined range.

The method may further comprise the steps of: acquiring existing learning data used for learning the learned model; performing a test on the existing learning data using the re-learned model; Extracting second learning data obtained with an accuracy equal to or higher than a predetermined second reference value, and deleting the extracted second learning data from the existing learning data.

The predetermined second reference value may be adjusted such that the amounts of the new learning data and the existing learning data do not exceed a predetermined range.

The testing may include determining when to perform the test, and performing the test if the amount of the new learning data exceeds a predetermined reference value.

The step of performing the test may further include the steps of: inputting the labeled third learning data included in the new learning data into the learned model; obtaining an output of the learned model; And comparing the information labeled with the third learning data.

The step of re-learning the learned model may include re-learning the learned model so that the decision boundary of the learned model is the same as the learned model again.

The method may further include acquiring existing learning data used for learning of the learned model, and the step of re-learning the learned model further comprises: extracting from the existing learning data using the re-learned model And re-learning the learned model so as to be located near the feature extracted from the existing learning data using the learned model.

According to an aspect of the present invention, there is provided a learning data management apparatus including a memory for storing one or more instructions and a processor for executing the one or more instructions stored in the memory, Executing an instruction to acquire new learning data; performing a test on the acquired new learning data using the learned model; and a step of performing a test on the acquired new learning data using the learned model, Extracting the first learning data to be acquired, deleting the extracted first learning data from the new learning data, and re-learning the learned model using the new learning data from which the extracted learning data has been deleted .

According to an aspect of the present invention, there is provided a computer program stored in a recording medium readable by a computer to perform a learning data management method according to an embodiment of the present invention in combination with a hardware computer.

Other specific details of the invention are included in the detailed description and drawings.

According to the disclosed embodiment, it is possible to reduce the time required for learning and the storage space for storing learning data.

The effects of the present invention are not limited to the above-mentioned effects, and other effects not mentioned can be clearly understood by those skilled in the art from the following description.

1 is a simplified schematic diagram of a system capable of performing robotic surgery in accordance with the disclosed embodiments.
2 is a flowchart illustrating a method of managing learning data according to an embodiment.
3 is a view showing a storage space including existing learning data and a previously learned model according to an embodiment.
4 is a diagram showing an example of selecting new learning data.
5 is a diagram showing an example in which the learned model is re-learned.
FIG. 6 is a flowchart illustrating a method of managing and utilizing existing training data according to an embodiment.
7 is a diagram showing an example of a method of managing existing learning data.
8 is a view showing a storage space in which new and existing learning data are stored.
9 is a block diagram of an apparatus according to an embodiment.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention, and the manner of achieving them, will be apparent from and elucidated with reference to the embodiments described hereinafter in conjunction with the accompanying drawings. It should be understood, however, that the invention is not limited to the disclosed embodiments, but may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein, Is provided to fully convey the scope of the present invention to a technician, and the present invention is only defined by the scope of the claims.

The terminology used herein is for the purpose of illustrating embodiments and is not intended to be limiting of the present invention. In the present specification, the singular form includes plural forms unless otherwise specified in the specification. The terms " comprises "and / or" comprising "used in the specification do not exclude the presence or addition of one or more other elements in addition to the stated element. Like reference numerals refer to like elements throughout the specification and "and / or" include each and every combination of one or more of the elements mentioned. Although "first "," second "and the like are used to describe various components, it is needless to say that these components are not limited by these terms. These terms are used only to distinguish one component from another. Therefore, it goes without saying that the first component mentioned below may be the second component within the technical scope of the present invention.

Unless defined otherwise, all terms (including technical and scientific terms) used herein may be used in a sense that is commonly understood by one of ordinary skill in the art to which this invention belongs. In addition, commonly used predefined terms are not ideally or excessively interpreted unless explicitly defined otherwise.

As used herein, the term "part" or "module" refers to a hardware component, such as a software, FPGA, or ASIC, and a "component" or "module" performs certain roles. However, "part" or " module " is not meant to be limited to software or hardware. A "module " or " module " may be configured to reside on an addressable storage medium and configured to play back one or more processors. Thus, by way of example, "a" or " module " is intended to encompass all types of elements, such as software components, object oriented software components, class components and task components, Microcode, circuitry, data, databases, data structures, tables, arrays, and variables, as used herein. Or " modules " may be combined with a smaller number of components and "parts " or " modules " Can be further separated.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a simplified schematic diagram of a system capable of performing robotic surgery in accordance with the disclosed embodiments.

1, the robot surgery system includes a medical imaging apparatus 10, a server 100, a control unit 30 provided in an operating room, a display 32, and a surgical robot 34. Depending on the embodiment, the medical imaging equipment 10 may be omitted from the robotic surgery system according to the disclosed embodiment.

In one embodiment, the surgical robot 34 includes a photographing device 36 and a surgical tool 38.

In one embodiment, robotic surgery is performed by the user controlling the surgical robot 34 using the control unit 30. [ In one embodiment, robot surgery may be performed automatically by the control unit 30 without user control.

The server 100 is a computing device that includes at least one processor and a communication unit.

The control unit 30 includes a computing device including at least one processor and a communication unit. In one embodiment, the control unit 30 includes hardware and software interfaces for controlling the surgical robot 34.

The photographing apparatus 36 includes at least one image sensor. That is, the photographing device 36 includes at least one camera device and is used to photograph a target object, that is, a surgical site. In one embodiment, the imaging device 36 includes at least one camera coupled with a surgical arm of the surgical robot 34.

In one embodiment, the image photographed at the photographing device 36 is displayed on the display 340. [

In one embodiment, the surgical robot 34 includes one or more surgical tools 38 that can perform cutting, clipping, anchoring, grabbing, etc., of the surgical site. The surgical tool 38 is used in combination with the surgical arm of the surgical robot 34.

The control unit 30 receives information necessary for surgery from the server 100 or generates information necessary for surgery and provides the information to the user. For example, the control unit 30 displays on the display 32 information necessary for surgery, which is generated or received.

For example, the user operates the control unit 30 while viewing the display 32 to perform the robot surgery by controlling the movement of the surgical robot 34.

The server 100 generates information necessary for robot surgery using the medical image data of the object photographed beforehand from the medical imaging equipment 10 and provides the generated information to the control unit 30. [

The control unit 30 provides the information received from the server 100 to the user by displaying the information on the display 32 or controls the surgical robot 34 using the information received from the server 100. [

In one embodiment, the means that can be used in the medical imaging equipment 10 is not limited, and various other medical imaging acquiring means such as CT, X-Ray, PET, MRI and the like may be used.

In the disclosed embodiment, the surgical image obtained in the photographing device 36 is transmitted to the control section 30. [

In one embodiment, the control unit 30 may segment the surgical image obtained during the operation in real time.

In one embodiment, the control unit 30 transmits a surgical image to the server 100 during or after surgery.

The server 100 can divide and analyze the surgical image.

The server 100 learns and stores at least one model for dividing and analyzing a surgical image. In addition, the server 100 learns and stores at least one model for generating an optimized surgical process.

The server 100 uses learning data to learn at least one model, and the learning data includes, but is not limited to, surgical and surgical image labeling information.

Hereinafter, a method of managing the learning data collected by the server 100 and learning the model will be described. However, the embodiments described below can be applied not only in connection with the robot surgery system shown in FIG. 1, but also collect, store and manage learning data, perform learning using the learned data, and update the learned model The present invention can be applied to all kinds of embodiments.

2 is a flowchart illustrating a method of managing learning data according to an embodiment.

The embodiment shown in FIG. 2 includes one or more steps performed in time series by the server 100 shown in FIG.

In step S110, the server 100 acquires new learning data.

In one embodiment, the new learning data includes labeled data that can be used for learning. The subject for labeling the training data is not limited. For example, the labeling may be performed by a person or automatically by a computer.

Referring to FIG. 3, information on existing learning data 210 and previously learned model 220 is stored in the storage space 200. According to an embodiment, the storage space 200 may mean a memory included in the server 100 or an external storage space or a cloud storage space that can communicate with the server 100 in a wired or wireless manner. In addition, information on the existing learning data 210 and the learned model 220 may be stored in the same storage space or may be divided and stored in different storage spaces.

In one embodiment, the information about the learned model 220 may include, but is not limited to, information about at least one parameter for constructing the neural network of the learned model 220.

In one embodiment, the server 100 acquires new learning data 300. The new learning data 300 may be stored in the storage space 200 or may be stored in a separate database. The new learning data 300 may be stored in the same space as the existing learning data 210 or may be separately stored in another space.

In step S120, the server 100 performs a test on the new learning data acquired in step S110, using the learned model 220. [

In one embodiment, the server 100 determines when to perform the test, and may perform a test if the amount of new learning data 300 obtained exceeds a predetermined reference value.

The server 100 may perform the test when the entropy of the learning data or the new learning data 300 stored in the storage space 200 exceeds a predetermined reference value, but the present invention is not limited thereto.

In one embodiment, the server 100 enters one or more labeled learning data contained in the new learning data 300 into the learned model 220.

The server 100 acquires the output of the learned model 220 and compares the obtained output with the information labeled with the learning data input to the learned model 220. [ For example, if it is determined that the obtained output and the labeled information are similar to each other by a predetermined reference value or more, the server 100 can determine that the test for the corresponding learning data has been passed.

In step S130, the server 100 extracts learning data for which the labeled information is acquired with an accuracy of at least a predetermined reference value, and deletes the extracted learning data from the new learning data.

In one embodiment, when the new learning data 300 is acquired for the learned model, the server 100 re-learns the learned model 220 based on the acquired data, so that the accuracy of the learned model 220 , Or broaden the applicable range.

However, as the number of new learning data 300 for learning a model increases, a longer time is required for learning, and a larger storage space is required for storing all accumulated learning data. Therefore, the server 100 selects training data , It is possible to store only the selected learning data and to use it for learning of the model.

Referring to Fig. 4, an example of selecting new learning data is shown.

In one embodiment, the server 100 may perform a test on the learned model 220 using the new learning data 300 as a test data set. When the learned model 220 receives the specific learning data and can obtain the information labeled with the learning data with an accuracy of a predetermined reference value or more, the server 100 has already learned the learning data And the corresponding learning data may not be used for learning.

Therefore, the server 100 extracts the learning data 310 from which the information labeled from the new learning data 300 is acquired with an accuracy of at least a predetermined reference value, and extracts the extracted learning data 310 from the new learning data 300 . The server 100 stores the remaining learning data 320 after deleting the extracted learning data 310 in the storage space 200 and utilizes the data for learning.

In one embodiment, the predetermined reference value can be adjusted such that the amount of new learning data 320 does not exceed a predetermined range. For example, the predetermined reference value may be adjusted so that the amount of new learning data 320 does not exceed 50% of the acquired new learning data 300, but is not limited thereto.

In step S140, the server 100 re-learns the learned model 220 using the new learning data 320 from which the extracted learning data 310 has been deleted.

Referring to FIG. 5, an example of re-learning the learned model 220 is shown.

In one embodiment, the server 100 re-learns the learned model 220 using the new learning data 320 and obtains the re-learned model 400 as a result.

In one embodiment, the server 100 re-learns the learned model 220 using the new learning data 320 so that the learned model 220 does not forget the learning results based on the existing learning data 210 can do.

That is, the learned model 220 includes a Deep Neural Network (DNN), and re-learning of the learned model 220 using the new learning data 320 can be performed by fine tuning ). ≪ / RTI > However, there is a tendency to forget previously learned information in the process.

Therefore, a less-forgetful learning method can be applied in the process of expanding the domain of the neural network, in which less learned information is less forgotten.

In this case, the server 100 can keep the previously learned information even in the case where the existing learning data 210 can not be accessed, thereby preventing the waste of the storage space for storing the existing learning data 210 And it is possible to prevent a waste of learning time such as re-learning existing learning data 210 in a learning process.

According to the embodiment, the server 100 may utilize the existing learning data 210, but may select existing learning data 210 according to a predetermined criterion, and utilize the selected learning data 210 only. A method of selecting and utilizing existing learning data 210 will be described later.

The lower layer of the neural network is recognized as a feature extractor, and the highest layer is recognized as having characteristics of a linear classifier. That is, the weight of the soft max classifier represents a decision boundary for classifying the feature.

Due to the linear nature of the classifier at the top level, features extracted from the top level hidden layer can generally be linearly separated. From this, we present two characteristics of a learning method that can reduce the tendency to forget information learned in the existing domain of the neural network according to the disclosed embodiment.

First, the decision boundaries should not change.

Second, the feature extracted from the data of the existing region using the new neural network should be located close to the feature extracted from the data of the existing region.

The server 100 performs re-learning of the learned model 220 based on the above two characteristics

The property that the first characteristic, that is, the decision boundary, should not be changed can be applied by setting the learning rate of the boundary to zero.

However, the second characteristic can not be applied simply because it is impossible to access the data of the existing region.

In the disclosed embodiment, the second characteristic can be satisfied by using the learning data of the new area instead of the data of the existing area.

First, the server 100 first recycles the weight of the existing neural network, as in the conventional fine adjustment method. That is, the weight of the existing neural network is used as the initial weight of the new neural network.

Thereafter, the server 100 freezes the weight of the soft max classifier to freeze the boundary of the classifier. The server 100 may then train the in-depth network in a direction that minimizes the total loss function.

According to the embodiment, a method of achieving the above-mentioned second characteristic by accessing a part of the learning data of the existing area can also be considered. For example, by making some of the learning data of the existing area selectively accessible, the learning efficiency can be increased.

FIG. 6 is a flowchart illustrating a method of managing and utilizing existing training data according to an embodiment.

In step S150, the server 100 acquires existing learning data 210 used for learning of the learned model 220. [

In step S160, the server 100 performs the test on the existing training data 210 using the learned model 400 again.

In one embodiment, the server 100 inputs one or more labeled training data contained in the existing training data 210 into the re-learned model 400. [

The server 100 acquires the output of the learned model 400 again and compares the obtained output with the information labeled with the learning data input to the re-learned model 400 again. For example, if it is determined that the obtained output and the labeled information are similar to each other by a predetermined reference value or more, the server 100 can determine that the test for the corresponding learning data has been passed.

In step S170, the server 100 may extract learning data for which the labeled information is acquired with an accuracy of a predetermined reference value or more, and delete the extracted learning data from the existing learning data 210 as a result of the test.

Referring to FIG. 7, an example of a method of managing existing training data 210 is shown.

The server 100 inputs one or more labeled learning data included in the existing learning data 210 into the re-learned model 400 and outputs the learning data 212, which is obtained at an accuracy of a predetermined reference value or more, And the extracted learning data 212 can be deleted from the existing learning data 210. [ That is, the learning data 212 extracted from the existing learning data 210 is deleted, and the remaining existing learning data 214 is stored in the storage space 200.

That is, if the existing learning data is well processed even though the learning has been performed using the new learning data, the existing learning data is deleted, so that the learning speed can be increased and the storage space can be saved.

In one embodiment, the predetermined reference value may be adjusted such that the amount of existing training data does not exceed a predetermined range. Likewise, the predetermined reference value can be adjusted such that the amount of new learning data and existing learning data does not exceed a predetermined range.

Referring to FIG. 8, a storage space storing new and existing learning data is shown.

In one embodiment, the server 100 stores the data 214 in which some of the new learning data has been deleted and the data 320 in which some of the existing learning data has been deleted, in the storage space 200. [

In one embodiment, the existing learning data 214 and the new learning data 320 can be stored together in the storage space 200 as existing learning data 210 together. In this case, the predetermined reference value may be set such that the amount of the existing learning data 214 and the amount of the new learning data 320 does not exceed the amount of the existing learning data 210. [

Therefore, the server 100 performs re-learning when new learning data is added, but the amount of the entire learning data can be kept constant.

9 is a block diagram of an apparatus according to an embodiment.

The processor 102 may include one or more cores (not shown) and a connection path (e.g., a bus, etc.) to transmit and receive signals to and / or from a graphics processing unit (not shown) .

The processor 102 in accordance with one embodiment performs one or more instructions stored in the memory 104 to perform the training data management method described with reference to Figures 1-8.

For example, the processor 102 may perform one or more instructions stored in a memory to allow the computer to acquire new learning data, perform a test on the acquired new learning data using the learned model, Extracts first learning data in which the labeled information is acquired with an accuracy of a predetermined first reference value or more, deletes the extracted first learning data from the new learning data, and deletes the extracted first learning data from the new learning data The learned model can be re-learned using data.

The processor 102 may include a random access memory (RAM) (not shown) and a read-only memory (ROM) for temporarily and / or permanently storing signals (or data) , Not shown). In addition, the processor 102 may be implemented as a system-on-chip (SoC) including at least one of a graphics processing unit, a RAM, and a ROM.

The memory 104 may store programs (one or more instructions) for processing and control of the processor 102. Programs stored in the memory 104 may be divided into a plurality of modules according to functions.

As described above, the learning data management method according to an embodiment of the present invention can be implemented as a program (or an application) to be executed in combination with a computer, which is hardware, and can be stored in a medium.

The above-described program may be stored in a computer-readable medium such as C, C ++, JAVA, machine language, or the like that can be read by the processor (CPU) of the computer through the device interface of the computer, And may include a code encoded in a computer language of the computer. Such code may include a functional code related to a function or the like that defines necessary functions for executing the above methods, and includes a control code related to an execution procedure necessary for the processor of the computer to execute the functions in a predetermined procedure can do. Further, such code may further include memory reference related code as to whether the additional information or media needed to cause the processor of the computer to execute the functions should be referred to at any location (address) of the internal or external memory of the computer have. Also, when the processor of the computer needs to communicate with any other computer or server that is remote to execute the functions, the code may be communicated to any other computer or server remotely using the communication module of the computer A communication-related code for determining whether to communicate, what information or media should be transmitted or received during communication, and the like.

The medium to be stored is not a medium for storing data for a short time such as a register, a cache, a memory, etc., but means a medium that semi-permanently stores data and is capable of being read by a device. Specifically, examples of the medium to be stored include ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage, and the like, but are not limited thereto. That is, the program may be stored in various recording media on various servers to which the computer can access, or on various recording media on the user's computer. In addition, the medium may be distributed to a network-connected computer system so that computer-readable codes may be stored in a distributed manner.

The steps of a method or algorithm described in connection with the embodiments of the present invention may be embodied directly in hardware, in software modules executed in hardware, or in a combination of both. The software module may be a random access memory (RAM), a read only memory (ROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a flash memory, a hard disk, a removable disk, a CD- May reside in any form of computer readable recording medium known in the art to which the invention pertains.

While the present invention has been described in connection with what is presently considered to be practical exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, You will understand. Therefore, it should be understood that the above-described embodiments are illustrative in all aspects and not restrictive.

100: Device
102: Processor
104: Memory

Claims (10)

The computer acquiring the first new learning data;
A new learning data test step of performing a test on the acquired first new learning data using a primary model in which the computer has already been learned;
The computer extracting first extraction data in which the labeled information obtained by the test result for the first new learning data is acquired with an accuracy of a predetermined first reference value or more;
The computer deleting the first extracted data from the first new learning data;
The computer re-learning the primary model using the second new learning data from which the extracted data is deleted to generate a secondary model;
Acquiring first existing learning data, in which the computer is used for learning of the learned primary model;
An existing training data test step in which a computer performs a test on the first existing training data using the secondary model;
Extracting second extracted data in which the computer obtains the test result of the first existing learning data and the labeled information is acquired with an accuracy of a predetermined second reference value or more; And
And the computer deletes the extracted second extracted data from the first existing learning data to form second existing learning data. The method according to claim 1,
The predetermined first reference value is a value
And the amount of the second new learning data is adjusted so as not to exceed a preset range. delete The method according to claim 1,
Wherein the predetermined second reference value is a predetermined reference value,
The amount of the second new learning data and the amount of the second existing learning data is adjusted so as not to exceed a predetermined range. The method according to claim 1,
The new learning data test step includes:
Determining when the computer performs the test, and performing the test if the amount of the new learning data exceeds a predetermined reference value; Wherein the learning data management method comprises: The method according to claim 1,
The new learning data test step includes:
Inputting labeled learning data included in the first new learning data to the learned primary model by the computer;
The computer acquiring the output of the learned primary model; And
And the computer comparing the obtained output with the label information included in the labeled learning data. The method according to claim 1,
Wherein the secondary model generation step comprises:
Re-learning the primary model so that a decision boundary of the primary model is the same as a secondary model of the re-learned model; Wherein the learning data management method comprises: The computer acquiring the first new learning data;
A new learning data test step of performing a test on the acquired first new learning data using a primary model in which the computer has already been learned;
The computer extracting first extraction data in which the labeled information obtained by the test result for the first new learning data is acquired with an accuracy of a predetermined first reference value or more;
The computer deleting the first extracted data from the first new learning data;
Acquiring first existing learning data, in which the computer is used for learning of the primary model; And
And a step in which the computer again learns the learned primary model using the second new learning data from which the first extracted data has been deleted to generate a secondary model,
Wherein the secondary model generation step comprises:
The computer re-learning the primary model so that the feature extracted from the first existing learning data using the secondary model is located close to the feature extracted from the first existing learning data using the primary model ; And a learning data management method. A memory for storing one or more instructions; And
And a processor executing the one or more instructions stored in the memory,
The processor executing the one or more instructions,
Acquiring first new learning data;
A new learning data testing step of performing a test on the acquired first new learning data using the already learned primary model;
Extracting first extracted data in which the labeled information is obtained with an accuracy of a predetermined first reference value or more as a result of the test for the first new learning data;
Deleting the first extracted data from the first new learning data;
Generating the secondary model by re-learning the learned primary model using the second new learning data from which the extracted data is deleted;
Acquiring first existing learning data used for learning of the learned primary model;
An existing learning data test step of performing a test on the first existing learning data using the re-learned quadratic model;
Extracting second extracted data in which the labeled information is obtained with an accuracy of a predetermined second reference value or more as a result of the test on the first existing learning data; And
And deleting the extracted second extracted data from the first existing learning data to form second existing learning data. A computer program stored in a computer-readable recording medium in combination with a computer which is hardware.

Images