KR20200054121A - Method for machine learning and apparatus for the same - Google Patents

Abstract

The present invention provides a machine learning method capable of reducing annotation costs and improving the performance of a target model. According to several embodiments of the present invention, the machine learning method performed by a computing device comprises: a step of acquiring a learning data set of a first model including a plurality of data samples which are not given label information; a step of calculating a miss-prediction probability of the first model for the plurality of data samples; a step of selecting at least one data sample among the plurality of data samples based on the calculated miss-prediction probability to construct a first data sample set; a step of acquiring first label information for the first data sample set; and a step of using the first data sample set and the first label information to perform first learning for the first model.

Description

Machine learning method and apparatus {METHOD FOR MACHINE LEARNING AND APPARATUS FOR THE SAME}

The present disclosure relates to a machine learning method and apparatus. More specifically, the present invention relates to a method and apparatus for performing a method capable of quickly constructing a high-performance machine learning model while minimizing human and time costs required for annotation.

Supervised learning is a machine learning method of constructing a target model 3 that performs a target task by learning a dataset 2 given label information (ie, correct answer information) as shown in FIG. 1. Therefore, in order to perform supervised learning on the dataset 1 to which no label information (indicated by a tag icon) is given, an annotation task must be performed.

Annotation refers to the task of tagging label information for each data to create a training dataset. Since annotation work is generally performed by humans, it is a significant human and time consuming task to generate a large set of training datasets. In particular, in the case of constructing a machine learning model for diagnosing the type or location of a lesion in a medical image, an annotation work has to be performed by a skilled specialist, which is much more expensive than other domains.

Various studies have been conducted to reduce the cost of annotation work in the field of machine learning and to build a high-performance model with a small amount of training dataset. For example, machine learning techniques such as transfer learning, weakly supervised learning, and active learning can all be seen as part of the study.

Of these, active learning is a technique that reduces the cost of an annotation task by selecting a dataset that is difficult to classify among all datasets and performing learning on the selected dataset. In other words, active learning can be viewed as a technique that reduces annotation costs by performing annotation on only selected datasets.

The process in which active learning is performed is illustrated in FIG. 2. As shown in FIG. 2, active learning randomly extracts a sample set 5 consisting of some data samples from a data set 4 to which label information is not given, and performs a first annotation operation (①), It starts in the process of performing the first learning (②) of the target model (6) for the set (5). Next, among the datasets 4, uncertainty sampling is performed on unlearned data samples, and a sample set 7 that is difficult to classify is selected (③). The uncertainty sampling is performed using the target model 6, and as a measure of uncertainty, an entropy value based on a confidence score for each class of the target model 6 is mainly used. Here, the data samples that are difficult to classify refer to data samples in which the confidence score is evenly distributed for each class, such that the entropy value becomes a threshold or higher. In addition, the annotation operation is performed only on the selected sample set 7 (④), and the second learning (⑤) of the target model 6 is performed with the sample set 8 obtained with label information. In addition, the processes (③, ④, ⑤) are repeatedly performed until learning of the target model 6 is finished.

As described above, active learning is a learning technique that achieves a target performance without annotating an entire data sample by intensively learning only some data samples that are difficult to classify from a model point of view.

However, the above active learning technique has various problems. First, the most fundamental problem is that the entropy value that is the basis of uncertainty sampling is calculated based on the target model's confidence score. In other words, since the accuracy of the entropy value is not high until the target model is sufficiently trained, data samples that are difficult to classify cannot be accurately selected. In addition, the performance of the target model is gradually improved while active learning is performed, and the effect of reducing the annotation cost is also insignificant.

Another problem is that, since entropy is an index that can be applied only to classification tasks, the range of active learning is greatly limited. For example, active learning based on entropy values cannot be used to build machine learning models associated with regression tasks.

Therefore, in order to maximize the effect of reducing annotation costs by expanding the application range of active learning and accurately selecting data samples effective for learning, a new type of active learning technique is required.

Korean Registered Patent No. 10-1908680 (released on March 8, 2018)

The technical problem to be solved through some embodiments of the present disclosure is to provide a machine learning method and an apparatus for performing the method, which can reduce human and time costs for annotation work.

Another technical problem to be solved through some embodiments of the present disclosure is to provide a method for accurately selecting a data sample effective for learning and an apparatus for performing the method.

Another technical problem to be solved by the present disclosure is to provide a method capable of expanding an application range of active learning by using a general purpose sampling index rather than entropy, and an apparatus for performing the method.

The technical problems of the present disclosure are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

In order to solve the above technical problem, a machine learning method according to some embodiments of the present disclosure is a machine learning method performed on a computing device, wherein the first model includes a plurality of data samples to which label information is not given. Acquiring a dataset, calculating a miss-prediction probability of the first model with respect to the plurality of data samples, and at least one of the plurality of data samples based on the calculated misprediction probability Selecting a data sample to construct a first data sample set, obtaining first label information for the first data sample set, and using the first data sample set and the first label information to form the first model It may include performing a first learning for.

In some embodiments, the first set of data samples may be composed of data samples in which the calculated false prediction probability is greater than or equal to a threshold.

In some embodiments, calculating the false predicted probability of the first model includes: constructing a second model for calculating the false predicted probability of the first model based on the evaluation result of the first model, and And calculating a misprediction probability for each of the plurality of data samples using a second model.

In some embodiments, constructing the second model includes: training the first model using a first data sample given correct answer label information, and evaluating the first model using the first data sample The method may include tagging label information based on the evaluation result to the first data sample, and constructing the second model by learning the first data sample with the tagged label information.

In some embodiments, the tagged label information may be a prediction error of the first data sample.

In some embodiments, the step of tagging the label information, in response to determining that the prediction result corresponds to a false positive (FP) or a false negative (FN), tagging a first value as a label of the first data sample And in response to a determination that the prediction result corresponds to a true positive (TP) or a true negative (TN), tagging a second value with a label of the first data sample.

In some embodiments, the step of building the second model includes: training the first model using a first data sample given correct answer label information, and learning the second model using a second data sample given correct answer label information Evaluating the first model, tagging label information based on the evaluation result into the second data sample, and constructing the second model by learning the second data sample with the tagged label information. It can contain.

In some embodiments, updating the second model using the evaluation result of the first trained first model, and selecting at least one data sample from among untrained data samples using the updated second model , Configuring a second data sample set, obtaining second label information for the second data sample set, and using the second data sample set and the second label information to obtain the first learned first The method may further include performing a second learning on the model.

In some embodiments, calculating a confidence score for each class of a first data sample included in the training data set through the first model, the first data sample based on the confidence score for each class The method may further include calculating an entropy value for and excluding the first data sample from the training dataset of the first model in response to determining that the entropy value is below a threshold.

In some embodiments, the first data sample set is composed of data samples in which the calculated misprediction probability is greater than or equal to a first threshold, and configuring the first data sample set comprises calculating the calculated data among the plurality of data samples. The method may include selecting at least one data sample having a false prediction probability less than a second threshold and configuring a second data sample set and excluding the second data sample set from the training dataset of the first model. .

In some embodiments, the step of performing the first learning may include at least partially assigning different sample weights to each data sample constituting the first data sample set and the first weights based on the sample weights. And learning a set of data samples, wherein the sample weight value may be determined based on a misprediction probability of each data sample.

In some embodiments, the step of performing the first learning further comprises generating a second data sample set from the first data sample set by applying a data augmentation technique, and further adding the second data sample set. And learning and updating the first model.

In some embodiments, based on a misprediction probability of the first trained first model, selecting at least one data sample among data samples not used for the first training to configure a second data sample set, Obtaining second label information for the second data sample set and performing second learning for the first trained first model using the second data sample set and the second label information. It may further include.

A machine learning apparatus according to some embodiments of the present disclosure for solving the above-described technical problem may include a memory including one or more instructions and a plurality of labels that are not given by executing the one or more instructions. Acquiring a training dataset of a first model including a data sample, calculating a miss-prediction probability of the first model for the plurality of data samples, and based on the calculated misprediction probability Selecting at least one data sample from a plurality of data samples to configure a first data sample set, obtaining first label information for the first data sample set, and obtaining the first data sample set and the first label information. It may include a processor for performing a first learning for the first model using.

A computer program according to some embodiments of the present disclosure for solving the above technical problem is coupled to a computing device to obtain a training dataset of a first model including a plurality of data samples to which label information is not given. Step, calculating a miss-prediction probability of the first model with respect to the plurality of data samples, and selecting at least one data sample among the plurality of data samples based on the calculated misprediction probability Constructing a first data sample set, obtaining first label information for the first data sample set, and first learning about the first model using the first data sample set and the first label information It may be stored in a computer-readable recording medium to perform the steps of performing.

A machine learning method according to some other embodiments of the present disclosure for solving the above technical problem is a machine learning method performed on a computing device, and includes a learning data set including a plurality of data samples to which label information is not given. Acquiring, acquiring first label information for a first set of data samples included in the training data set, learning the first set of data samples with the first label information, and building a first model; Calculating a miss-prediction of the first model with respect to data samples excluding the first data sample set from the training data set, based on the misprediction probability, at least one of the remaining data samples Selecting a data sample and constructing a second data sample set, a second for the second data sample set The method may include acquiring label information and learning a second model in an initialization state with the second data sample set and the second label information.

A machine learning apparatus according to some other embodiments of the present disclosure for solving the above-described technical problem, a memory including one or more instructions and one or more instructions, by executing the one or more instructions, a plurality of labels are not given Acquiring a training dataset including a data sample of, obtaining first label information for a first data sample set included in the training dataset, and learning the first data sample set as the first label information Construct a first model, calculate a miss-prediction of the first model for the remaining data samples excluding the first data sample set from the training data set, and calculate the miss-prediction of the first model based on the false prediction probability. At least one data sample is selected from the remaining data samples to form a second data sample set, And a processor that acquires second label information for the second data sample set, and trains a second model in an initialized state with the second data sample set and the second label information.

A computer program according to some other embodiments of the present disclosure for solving the above-described technical problem is combined with a computing device to obtain a learning dataset including a plurality of data samples to which label information is not provided, wherein Obtaining first label information for a first set of data samples included in a training data set, and learning the first set of data samples with the first label information to build a first model, wherein the learning data set includes Calculating miss-prediction of the first model with respect to the remaining data samples excluding the first set of data samples, and selecting at least one data sample from the remaining data samples based on the false prediction probability Configuring a second data sample set, obtaining second label information for the second data sample set And wherein the step may be stored in a computer-readable recording medium so as to execute the step of learning a second model of the initialized state to the second label information set and the second data sample.

According to various embodiments of the present disclosure described above, a data sample to be annotated is selected based on a false prediction probability of a target model. That is, data samples are not selected based on uncertainty, but data samples that are likely to have a wrong target model are selected. Unlike the uncertainty, the misprediction probability is not a value dependent on the confidence score of the target model, so that a data sample can be more accurately selected.

In addition, the learning effect can be improved by intensively learning the target model with a data sample in which the target model is wrong. That is, the performance of the target model can quickly reach the target performance. Accordingly, the cost of computing and time required for learning can be greatly reduced, and the cost of annotation can also be significantly reduced.

In addition, since active learning is performed based on the probability of incorrect prediction of the target model without depending on the entropy value, the range of application of active learning can be greatly expanded.

Effects according to the technical spirit of the present disclosure are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is an exemplary diagram for explaining a relationship between supervised learning and annotation work.
2 is an exemplary diagram for explaining a conventional active learning technique.
3 and 4 are diagrams for schematically explaining operations and inputs and outputs of a machine learning apparatus according to some embodiments of the present disclosure.
5 and 6 are exemplary block diagrams illustrating a machine learning apparatus according to some embodiments of the present disclosure.
7 is an exemplary diagram for describing a learning operation of a machine learning apparatus according to some embodiments of the present disclosure.
8 is an exemplary flow diagram illustrating a machine learning method in accordance with some embodiments of the present disclosure.
9 and 10 are diagrams for explaining a method for constructing a model for calculating a false prediction probability according to a first embodiment of the present disclosure.
11 is a view for explaining a method for constructing a model for calculating a false prediction probability according to a second embodiment of the present disclosure.
12 is an exemplary diagram for describing a method for selecting (sampling) a data sample based on a misprediction probability according to some embodiments of the present disclosure.
13 is an exemplary diagram illustrating a machine learning method according to some embodiments of the present disclosure.
14 is an exemplary diagram for explaining a method of improving a learning effect using a data expansion technique according to some embodiments of the present disclosure.
15 is an exemplary diagram for explaining a method of improving a learning effect using sample weights according to some embodiments of the present disclosure.
16 is a flowchart illustrating a machine learning method according to some other embodiments of the present disclosure.
17 to 19 are diagrams for explaining a patch sampling method based on an entire slide image according to some application examples of the present disclosure.
20 is an exemplary diagram illustrating a machine learning model according to some use cases of the present disclosure.
21 is a diagram for explaining a machine learning method according to some application examples of the present disclosure.
22 is an example hardware configuration diagram illustrating an example computing device capable of implementing an apparatus in accordance with various embodiments of the present disclosure.

Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Advantages and features of the present disclosure, and methods for achieving them will be apparent with reference to embodiments described below in detail in conjunction with the accompanying drawings. However, the technical spirit of the present disclosure is not limited to the following embodiments, and may be implemented in various different forms, and only the present embodiments allow the present disclosure to be complete, and common knowledge in the technical field to which the present disclosure pertains. It is provided to fully inform the holder of the scope of the present disclosure, and the technical spirit of the present disclosure is only defined by the scope of the claims.

It should be noted that in adding reference numerals to the components of each drawing, the same components have the same reference numerals as possible even though they are displayed on different drawings. In addition, in describing the present disclosure, when it is determined that a detailed description of related known configurations or functions may obscure the subject matter of the present disclosure, the detailed description will be omitted.

Unless otherwise defined, all terms (including technical and scientific terms) used in this specification may be used in a sense that can be commonly understood by those skilled in the art to which this disclosure belongs. In addition, terms defined in the commonly used dictionary are not ideally or excessively interpreted unless specifically defined. The terminology used herein is for describing the embodiments and is not intended to limit the present disclosure. In the present specification, the singular form also includes the plural form unless otherwise specified in the phrase.

In addition, in describing the components of the present disclosure, terms such as first, second, A, B, (a), and (b) may be used. These terms are only for distinguishing the component from other components, and the nature, order, or order of the component is not limited by the term. When a component is described as being "connected", "coupled" or "connected" to another component, the component may be directly connected to or connected to the other component, but another component between each component It will be understood that elements may be "connected", "coupled" or "connected".

As used herein, "comprises" and / or "comprising" refers to the components, steps, operations and / or elements mentioned above, the presence of one or more other components, steps, operations and / or elements Or do not exclude additions.

Prior to the description of the present specification, some terms used in the specification will be clarified.

In the present specification, a target model is a model that performs a target task and is a target model to be built through machine learning.

In this specification, label information means correct answer information of a data sample. The label information can be generally obtained through annotation work.

In this specification, annotation means an operation of tagging label information in a data sample. The annotation may be used as a term meaning the label information itself, but in order to prevent confusion of terms, the specification is used in the meaning defined above. The annotation may be used interchangeably with terms such as tagging and labeling in the art.

In the present specification, a miss-prediction probability means a probability (that is, a probability that prediction is wrong) or a probability that an error is included in the prediction result when a specific model for a given data sample performs prediction.

In this specification, an instruction is a series of instructions grouped based on a function and refers to a component of a computer program and executed by a processor.

Hereinafter, some embodiments of the present disclosure will be described in detail according to the accompanying drawings.

3 and 4 are diagrams for schematically explaining operations and inputs and outputs of the machine learning apparatus 100 according to some embodiments of the present disclosure.

3, the machine learning apparatus 100 is a computing apparatus capable of performing a machine learning method according to various embodiments of the present disclosure. Here, the computing device may be a laptop, a desktop, a laptop, etc., but is not limited thereto, and may include all types of devices equipped with computing functions. An example of the computing device will be referred to FIG. 22. Hereinafter, for convenience of description, the machine learning apparatus 100 will be abbreviated as the learning apparatus 100.

Although the learning device 100 is implemented as one physical computing device in FIG. 3 as an example, the first function of the learning device 100 in the actual physical environment is implemented in the first computing device, and the second function is It may also be implemented in a second computing device. In addition, specific functions of the learning device 100 may be implemented to be performed through distributed / parallel processing in multiple computing devices (or processors).

As shown in FIG. 3, the learning apparatus 100 may build a target model 13 that receives a data set 11 to which label information is not given, and performs machine learning on it. At this time, the learning device 100 selects a data sample set (ie, a sub data set) corresponding to a part of the data set 11, obtains label information for the sub data set, and is based on the obtained label information To learn. In addition, this learning process can be repeated until the target performance of the target model 13 is satisfied. In the following description, when the data sample set corresponds to a part of the entire data set, the terms "data sample set" and "sub data set" may be used interchangeably.

In some embodiments, as shown in FIG. 4, the learning device 100 requests an annotation device 15 to annotate the selected sub-dataset, and an annotation result (ie, a label) from the device 15 Information). Here, the annotation device 15 is a computing device used by an annotator, and may be a device equipped with an annotation tool. That is, the annotator may provide label information for the requested data sample set using the annotation tool.

As mentioned above, since the annotating work has to be performed manually by the annotator, it is very time consuming and human cost. Therefore, in order to minimize annotation costs, it is important to accurately select a set of data samples effective for learning.

In some embodiments, the learning device 100 may select at least one data sample from the dataset 11 based on the false prediction probability of the target model 13. According to this embodiment, data samples that are not based on the entropy value (ie, uncertainty) of the target model 13 and that the prediction of the target model 13 may be wrong may be selected as an annotation target. By doing so, the accuracy of data selection can be improved, and the learning speed of the target model 13 can be improved. Furthermore, by quickly converging the performance of the target model 13 to the target performance, the annotation cost can be greatly reduced. Detailed description of this embodiment will be described in detail in the following drawings.

In some embodiments, the learning device 100 and the annotation device 15 can communicate over a network. Here, the network is a wired / wireless network of any kind, such as a local area network (LAN), a wide area network (WAN), a mobile radio communication network, a Wibro (Wireless Broadband Internet), and the like. Can be implemented.

So far, the operation and input / output of the learning apparatus 100 according to some embodiments of the present disclosure have been schematically described with reference to FIGS. 3 and 4. Hereinafter, the configuration and operation of the learning apparatus 100 will be described with reference to FIGS. 5 to 7.

5 and 6 are block diagrams illustrating a learning device 100 in accordance with some embodiments of the present disclosure. In particular, FIG. 6 further shows the operation flow of the learning device 100.

5 and 6, the learning device 100 includes a data set acquisition unit 110, a selection unit 130, a label information acquisition unit 150, a learning unit 170, and a learning end determination unit 190 It may include. However, only components related to embodiments of the present disclosure are illustrated in FIGS. 5 and 6. Accordingly, a person skilled in the art to which the present disclosure belongs may recognize that other general-purpose components may be further included in addition to the components shown in FIGS. 5 and 6. In addition, each component of the learning apparatus 100 shown in FIGS. 5 and 6 shows functionally divided functional elements, and a plurality of components may be implemented in an integrated form in a physical environment. Be careful. Hereinafter, each component will be described.

The data set acquisition unit 110 acquires a data set 21 to be used for training the target model. The learning data set 21 may be composed of a plurality of data samples to which label information is not given, but may be partially included in a data sample to which label information is given.

Next, the sorting unit 130 selects a data sample to be annotated in the learning data set 21. The sorting unit 130 may include a false prediction probability calculating unit 131 and a data selecting unit 133.

The misprediction probability calculating unit 131 calculates a misprediction probability for each of a plurality of data samples included in the learning data set 21. In this case, the plurality of data samples may be all or part (e.g. unlearned data samples) included in the learning data set 21.

In order to calculate the false prediction probability of the target model, the false prediction probability calculation unit 131 may use a predetermined machine learning model. In order to exclude duplicate description, the description of the machine learning model will be described later with reference to the drawings of FIG. 8 and below.

Next, the data selector 133 selects at least one data sample based on the false prediction probability. Specifically, the data sorting unit 133 selects a data sample (that is, a data sample having a high probability that the target model is incorrect) among the plurality of data samples having a false prediction probability of a threshold or higher. The selected data sample may constitute the sub data set 23.

At this time, the number of the selected data samples may be a preset fixed value or a fluctuating value that varies depending on the situation. For example, the number may be a fluctuating value that changes based on a difference between a target model's current performance and a target performance, the number of untrained data samples, and an annotation cost. For a more specific example, the number of the selected data samples may fluctuate to a smaller value as the difference between the current performance and the target performance of the target model becomes smaller. For another example, the number of the selected data samples may fluctuate to a smaller value as the number of unlearned data samples decreases or the annotation cost increases.

Next, the label information acquisition unit 150 acquires the label information 25 of the sub data set 23 selected as a result of the annotation operation. For example, the label information acquiring unit 150 may acquire label information 25 for the sub dataset 23 from the annotation device 15.

Next, the learning unit 170 performs learning on the target model by learning the sub data set 23 selected by the selecting unit 130 with the obtained label information 25. For example, when the target model is a neural network-based model, the learning unit 170 may perform learning by updating the weight of the target model through error back propagation, but the technical scope of the present disclosure is not limited thereto. .

Next, the learning end determination unit 190 determines whether to end the learning of the target model based on the designated learning end condition. The learning end condition may be modified in any number depending on the embodiment. For example, when the performance of the target model reaches the target performance, the learning end condition may be set based on the number of learning iterations, but the technical scope of the present disclosure is not limited thereto.

Specifically, the learning end determination unit 190 may end learning in response to a determination that the specified learning end condition is satisfied. In the opposite case, learning can continue. If learning continues, the selection unit 130, the label information acquisition unit 150, and the learning unit 170 may perform the above-described process again. An example of a process in which learning is repeated is illustrated in FIG. 7.

As illustrated in FIG. 7, during the primary learning process, a first annotation operation and first learning on the first sub dataset 32 selected in the learning dataset 31 may be performed. In addition, before selecting the first sub-dataset 32, in the primary learning process, a model for calculating a false prediction probability through learning may be constructed and updated. The description of the misprediction probability calculation model will be described later with reference to FIGS. 8 to 11. When the first learning is completed, a determination 34 is made as to whether or not learning has ended, and a secondary learning process may be started according to the determination of continuing learning.

During the second learning process, a second annotation operation and second learning on the second sub data set 35 selected from the untrained sub data set 33 may be performed. Of course, according to some other embodiments, data samples for the second annotation may be selected from the entire data set 31 rather than the untrained sub data set 33. In addition, before selecting the second sub-dataset 35, the misprediction probability calculation model may be updated through learning in the secondary learning process. Through this, in the secondary learning process, data samples that may have a wrong target model can be more accurately selected.

In some embodiments, the first sub-dataset 32 selected in the primary learning process is learned based on the first weight, and the second sub-dataset 35 selected in the secondary learning process is assigned to the second weight Can be learned on the basis. In this case, the second weight may be set to a value greater than the first weight. Here, learning based on the weight means that the stronger or weaker intensity is learned according to the weight value. In addition, learning the target model with stronger intensity has a greater influence on the weight adjustment of the target model, which may be implemented in any way. For example, the learning intensity may be increased by increasing the prediction error, but the technical scope of the present disclosure is not limited thereto. Since the second sub dataset 35 is a data sample in which the first trained target model is likely to be wrong, it may be a more important data sample (ie, samples that are more effective for learning) than the first sub dataset 32. Therefore, according to the present embodiment, the learning effect of the target model can be further improved, and the performance of the target model can quickly reach the target performance. Of course, annotation costs can be further reduced.

In the above manner, learning can be repeatedly performed until the learning end condition is satisfied. Additional description of the components 110 to 190 of the learning device 100 will be further referred to the description of FIG. 8 and below.

For reference, it should be noted that not all components shown in FIGS. 5 and 6 are essential components of the learning apparatus 100. That is, the learning apparatus 100 according to some other embodiments of the present disclosure may be implemented by some of the components illustrated in FIGS. 5 and 6.

Each component in FIGS. 5 and 6 may mean software or hardware such as a field programmable gate array (FPGA) or an application-specific integrated circuit (ASIC). However, the above components are not limited to software or hardware, and may be configured to be in an addressable storage medium, or may be configured to execute one or more processors. The functions provided in the above components may be implemented by more detailed components, or may be implemented as a single component that performs a specific function by combining a plurality of components.

So far, the configuration and operation of the learning apparatus 100 according to some embodiments of the present disclosure have been described with reference to FIGS. 5 to 8. Hereinafter, a machine learning method according to some embodiments of the present disclosure will be described in detail with reference to FIGS. 8 to 16.

Each step of the machine learning method may be performed by a computing device. In other words, each step of the machine learning method may be implemented with one or more instructions executed by a processor of a computing device. All of the steps included in the machine learning method may be executed by one physical computing device, but the first steps of the method are performed by the first computing device, and the second steps of the method are performed by the second computing device. It may be performed by. Hereinafter, it is assumed that each step of the machine learning method is performed by the learning apparatus 100 to continue the description. However, for convenience of description, description of the operation subject of each step included in the machine learning method may be omitted.

8 is an exemplary flow diagram illustrating a machine learning method in accordance with some embodiments of the present disclosure. However, this is only a preferred embodiment for achieving the object of the present disclosure, and of course, some steps may be added or deleted as necessary.

As illustrated in FIG. 8, the machine learning method starts at step S10 of obtaining a training dataset of a target model. The training data set includes a plurality of data samples to which label information is not given.

In step S20, the learning device 100 builds or updates a model for calculating a false prediction probability of the target model. The method of constructing the misprediction probability calculation model (hereinafter abbreviated as “calculation model”) may vary according to embodiments. Hereinafter, some embodiments for constructing the calculation model will be described with reference to FIGS. 9 to 11.

9 is an exemplary flowchart illustrating a method for constructing a model for calculating a false prediction probability according to a first embodiment of the present disclosure.

As illustrated in FIG. 9, the first embodiment starts at step S110 of selecting a data sample set corresponding to a part of the training data set.

In steps S120 and S130, the learning apparatus 100 acquires label information for the selected data sample set, and trains a target model using the acquired label information.

In step S140, the learning apparatus 100 evaluates the target model using the selected data sample set again. For example, the learning apparatus 100 may evaluate a target model by inputting a first data sample to a target model to obtain a prediction result, and comparing the prediction result with label information of the first data sample.

In step S150, the learning apparatus 100 constructs a model for calculating a false prediction probability using the evaluation result. More specifically, the learning apparatus 100 may tag the evaluation result with label information of a corresponding data sample, and learn the data sample with the label information to construct the calculation model. In order to provide a more convenient understanding, this step S150 will be described in detail with reference to FIG. 10.

FIG. 10 shows a confusion matrix. When the target model is a model that performs a classification task, the evaluation result may correspond to a specific cell in the confusion matrix. As illustrated in FIG. 10, a first value (eg 0) is tagged as a label value 42 in the data sample 41 having an evaluation result of FP (false positive) or FN (false negative), and the evaluation result is TP. A second value (eg 1) may be tagged as the label value 44 in the data sample 43 that is (true positive) or TN (true negative). That is, if the prediction of the target model matches the correct answer, "1" is tagged, and if there is a mismatch, "0" may be tagged.

When the above data samples 41 and 43 and label information are learned, the calculation model outputs a high confidence score when data similar to the data that the target model accurately predicted is input. Also, in the opposite case, the calculation model outputs a low confidence score. Therefore, the calculation model can accurately calculate the probability of incorrect prediction of the target model for the input data.

On the other hand, it should be noted that FIG. 10 shows only some examples of tagging label information. According to some other embodiments of the present disclosure, the learning apparatus 100 may tag the prediction error of the data sample as label information. Here, the prediction error means a difference between a predicted value (ie, confidence score) and an actual value (ie, correct answer information).

Further, according to some other embodiments of the present disclosure, the learning apparatus 100 tags a first value (eg 0) when a prediction error of a data sample is greater than or equal to a threshold value, and when the prediction error is less than the threshold value The second value (eg 1) may be tagged as label information.

11 is an exemplary flowchart illustrating a method for constructing a misprediction probability calculation model according to a second embodiment of the present disclosure.

11, the overall process of the second embodiment is similar to the first embodiment shown in FIG. However, in the second embodiment, the selected data sample set is divided into a first sample set and a second sample set, and then a target model is trained with the first sample set, and the target model is evaluated with the second sample set. There is a difference in that it is (see S230 to S250).

That is, in the above-described first embodiment, the evaluation was performed with the learned sample set, but in the second embodiment, there is a difference in that the target model is more accurately evaluated by classifying the learning and evaluation sample sets.

According to some embodiments, the learning device 100 may repeatedly perform learning and evaluation using a k-fold cross validation technique. At this time, the evaluation result may be used as learning data of the calculation model. The cross-validation will be a technique obvious to those skilled in the art, and a description thereof will be omitted. According to this embodiment, as more evaluation data is secured, a more accurate calculation model can be built.

In addition, according to some embodiments, the learning apparatus 100 may generate a similar sample set from a sample set for evaluation using a data augmentation technique, and further learn the generated sample set to build an output model. have. Of course, the technical idea inherent in this embodiment can also be applied to a sample for learning or a sample set for evaluation of the first embodiment described above. The data expansion technique will be a technique obvious to those skilled in the art, and a description thereof will be omitted.

The description will be continued again with reference to FIG. 8.

In step S30, the learning apparatus 100 calculates a false prediction probability for each data sample included in the training data set using the calculation model. For example, as illustrated in FIG. 12, the learning device 100 inputs each data sample 52 to 54 into the calculation model 51 to obtain the confidence scores 55 to 57 of the calculation model, and the confidence score The false prediction probability may be calculated based on (55 to 57).

In some embodiments, as shown in FIG. 12, when the calculation model 51 is trained to output a confidence score for the correct and incorrect classes (eg, learning with label 1 if it matches the correct answer, and label 0 when there is a mismatch) If learned), the probability of incorrect prediction of the confidence score of the incorrect class (shown as underlined) may be used.

Referring back to FIG. 8, in step S40, the learning apparatus 100 selects at least one data sample from the training data set based on the probability of a misprediction. For example, the learning apparatus 100 may select a data sample having a probability of a false prediction higher than a threshold or a data sample of k with a high probability of a false prediction (where k is a natural number of 1 or more).

In step S50, the learning apparatus 100 acquires label information for the selected data sample to train the target model. Since the selected data sample is a sample in which the prediction of the target model is highly likely to be incorrect, training the target model with the selected data samples can quickly improve the performance of the target model.

In step S60, the learning apparatus 100 determines whether the learning end condition is satisfied. In response to the determination that the learning end condition is satisfied, the learning apparatus 100 may end learning. Conversely, in response to the dissatisfaction determination, the learning apparatus 100 performs steps S20 to S50 described above again.

If learning is repeated, in step S20, the learning apparatus 100 may re-evaluate the learned target model and update the calculation model by learning the evaluation result. By doing so, the probability of incorrect prediction can be accurately calculated as learning is repeated.

In addition, in steps S30 and S40, the learning apparatus 100 may select a data sample to be annotated from among untrained data samples, not the entire learning data set. Of course, according to some other embodiments, data samples to be annotated may be selected from the entire training dataset.

For reference, among the above-described steps S10 to S60, step S10 is performed by the data set acquisition unit 110, steps S20 and S30 are performed by the misprediction probability calculation unit 131, and step S40 is the data selection unit (133). In addition, step S50 may be performed by the label information acquisition unit 150 and the learning unit 170, and step S60 may be performed by the learning end determination unit 190.

In order to provide more convenience, the process of performing the above-described machine learning method will be described once again with reference to FIG. 13. In particular, FIG. 13 shows an example in which a calculation model is constructed according to the method illustrated in FIG. 9.

As illustrated in FIG. 13, an annotation is performed on the sub data set 62 corresponding to a part of the learning data set 61 of the target model (①), and the first learning and evaluation of the target model 63 are performed. Is performed (②, ③). The evaluation result is tagged to each data sample 65 used in the evaluation to build the calculation model 66 (④). In addition, a learning model 66 is constructed by learning the samples 65 tagged with the evaluation result (⑤), and a prediction incorrect answer (that is, the prediction is incorrect) based on the false prediction probability calculated by the calculation model 66 Likely samples) The dataset 67 is selected (⑥). Next, an annotation is performed on the selected sub data set 67 (⑦), and the second learning of the target model 63 is performed with the label information 68 and the sub data set 67 obtained as the result of the annotation. It is performed (⑧). In this way, by intensively learning the predictive incorrect samples, the performance of the target model 63 can be sharply improved.

In some embodiments, when the second learning is performed, re-learning of the sub dataset 62 used for the first learning may be performed. That is, in order to make the most use of the dataset secured with label information, iterative learning may be performed on the same dataset. Such technical ideas can be used in various ways in the learning process. For example, first datasets in which label information is secured during the primary learning process may be reused (ie, re-learned) in the secondary learning process.

In some embodiments, the first data sample set 67 selected through the calculation model 66 and the selected second data sample set (eg 62) regardless of the calculation model 66 may be trained with different weights. . For example, the first data sample set 67 may be trained with a first weight, and the second data sample set (e.g. 62) may be trained with a second weight. At this time, the first weight may be set to a larger value than the second weight. By doing so, important data samples can be trained with a stronger intensity, and the performance of the target model 63 can be quickly improved.

In some embodiments, in order to reduce annotation costs, some data samples with poor learning effectiveness may be excluded from the training dataset 61. At this time, a criterion for determining a data sample having a low learning effect may vary according to embodiments.

In the first embodiment, a data sample whose entropy value is less than a threshold (ie, a data sample that the target model can reliably classify) may be excluded from the training data set 61. More specifically, a confidence score for each class for a first data sample is calculated by the target model 63 and an entropy value can be calculated based on the confidence score for each class. At this time, when the calculated entropy value is less than the threshold, the first data sample may be excluded from the training data set 61. By doing so, the cost of annotation can be further reduced.

In the second embodiment, data samples having a false prediction probability below a threshold (ie, data samples that can be accurately classified by the target model) may be excluded from the training data set 61. This is because it is not necessary to learn data samples that the target model can accurately classify.

The process of excluding unnecessary data samples from the learning dataset 1 according to the above-described embodiments may be performed at any time, such as when each learning process is completed, when a new learning process is started, periodically, or the like.

So far, a machine learning method according to some embodiments of the present disclosure has been described with reference to FIGS. 8 to 13. According to the above-described method, according to various embodiments of the present disclosure described above, a data sample to be annotated is selected based on a false prediction probability of the target model. That is, data samples are not selected based on uncertainty, but data samples that are likely to have a wrong target model are selected. Unlike the uncertainty, the misprediction probability is not a value dependent on the confidence score of the target model, so that a data sample can be more accurately selected.

In addition, the learning effect can be improved by intensively learning the target model with a data sample in which the target model is wrong. That is, the performance of the target model can quickly reach the target performance. Accordingly, the cost of computing and time required for learning can be greatly reduced, and the cost of annotation can also be significantly reduced.

In addition, since active learning is performed based on the probability of incorrect prediction of the target model without depending on the entropy value, the range of application of active learning can be greatly expanded.

Hereinafter, some embodiments of the present disclosure designed to further improve the learning effect and further reduce annotation cost will be described with reference to FIGS. 14 and 15.

14 is an exemplary diagram illustrating a machine learning method using a data expansion technique according to some embodiments of the present disclosure.

As illustrated in FIG. 14, a data expansion technique may be applied to the sub dataset 75 selected from the training data set 71 through the calculation model 73. This is because the selected sub dataset 75 is composed of data samples that are very effective for learning the target model.

More specifically, the learning device 100 may expand the sub dataset 75 to generate similar datasets 77 and 79 and further train the target model with the similar datasets 77 and 79. By doing so, the performance of the target model can be quickly improved, and the annotation cost can be reduced.

When the data sample is an image format, the data expansion may be performed in an image cropping, rotating, flipping, resizing, color jittering, or the like. The technical scope of the present disclosure is not limited thereto.

In some embodiments, sub dataset 75 may be trained with a first weight, and similar datasets 77 and 79 may be trained with a second weight. In this case, the first weight may be set to a value higher than the second weight. That is, the original dataset 75 can be learned more strongly, and the similar datasets 77 and 79 can be learned weaker.

15 is an exemplary diagram illustrating a machine learning method based on sample weight according to some embodiments of the present disclosure.

As illustrated in FIG. 15, differential sample weights may be set for each data sample 84 to 86 of the sub data set 83 selected from the training data set 81 through the calculation model 82. . In addition, learning may be performed on the target model 87 based on the sample weight. In Fig. 15, the thickness of the arrow indicates the learning intensity.

Here, the sample weight value may be determined based on a false prediction probability. For example, a higher sample weight may be assigned to a data sample having a high probability of misprediction. By doing so, a data sample in which the target model is likely to be wrong can be trained more strongly, and the learning effect can be improved. Of course, learning time and annotation costs can be reduced.

So far, some embodiments of the present disclosure designed to further improve the learning effect have been described. Hereinafter, a machine learning method according to some other embodiments of the present disclosure will be described with reference to FIG. 16. For clarity of the present specification, descriptions of contents overlapping with the aforementioned machine learning method will be omitted.

16 is an exemplary diagram for describing a machine learning method according to some other embodiments of the present disclosure.

As shown in FIG. 16, the overall process of the machine learning method according to the present embodiment is similar to that described with reference to FIG. 8. However, in this embodiment, there is a difference in that the target model 96 is newly constructed using the sub dataset 94 and label information 95 selected by the calculation model 93.

The reason why the target model 96 is newly constructed with the selected sub data set 94 is to learn the sub data set 94 more strongly. More specifically, in the previous embodiment (refer to FIG. 13), the weight of the target model 63 is first adjusted through the first learning (②), and the weight of the target model 63 through the second learning (⑧). Was adjusted. Therefore, the weight of the target model is greatly adjusted by the first learning (②), thereby minimizing the influence of the second learning (⑧) on the selected sub dataset (eg, the second learning depends on the fine-tuning degree). Performance), the performance of the target model may degrade.

Therefore, in the present embodiment, the target model 96 in the initialized state is trained with the selected sub dataset 94 (⑧). In addition, by learning the existing sub data set 97 that has not been selected by the calculation model 93 after the learning process (⑧) (⑨), the selected sub data set 94 is learned more strongly. By doing so, a better target model can be built.

So far, a machine learning method according to various embodiments of the present disclosure has been described with reference to FIGS. 8 to 16. Hereinafter, some application examples in which the machine learning method is applied to the medical domain will be described.

Due to the nature of the medical domain, there are not many learning datasets given label information, and annotation work should be performed by a skilled specialist. For example, when tagging the location, type, and name of a lesion in a radiographic image, the annotating operation is inevitably performed by a radiologist. Therefore, more annotation costs are required compared to other domains, and the effect can be maximized when the technical idea of the present disclosure is utilized in the medical domain.

17 to 19 show an example of generating a learning dataset from a high resolution whole slide image photographing tissue.

As shown in FIG. 17, when the tissue region 203 is extracted from the entire slide image 201, the training dataset 205 may be generated through patch sampling.

As shown in the sampling examples 211 and 213 shown in FIGS. 18 and 19, the size of the patch (or the size of the sampling area) may vary depending on the target model. Further, each patch may be sampled in a form overlapping each other.

For example, as shown in FIG. 20, when the target model is a model that analyzes cell-level images and classifies mitosis and normal cells (eg, a CNN-based classification model), in one whole slide image Large patches of small size can be sampled (eg see FIG. 18). Therefore, a large amount of training datasets to which label information is not given can be generated.

As described above, the process of generating a training dataset through patch sampling may be automatically performed through image analysis or processing technology, but annotation work on the training dataset must be manually performed by a specialist. Therefore, a significant annotation cost is inevitably consumed. In such an environment, in order to build a target model, it may be utilized in a machine learning method according to various embodiments of the present disclosure described above.

An example in which the machine learning method is utilized is illustrated in FIG. 21.

As shown in FIG. 21, the specialist 22 may act as an annotator for the learning dataset 221. The overall learning process is the same as described above. First, an annotation is performed on the sub data set extracted from the training data set 221 (①), and learning and evaluation of the riding model 223 is performed using label information obtained as an annotation result (②). , ③). Also, by calculating the evaluation results, the calculation model 224 is constructed (④, ⑤), and the predicted incorrect answer set 225 is selected using the false prediction probability calculated by the calculation model 224 (⑥). Next, an annotation for the predicted incorrect answer set 225 is performed by the annotator 222 (⑦), and the target model 223 may be updated by learning the annotation result (⑧).

The above-described process is repeatedly performed until the target model 223 satisfies the learning end condition. According to various embodiments described above, even if all of the learning datasets 221 are not learned, the learning of the target model 223 ends. The condition can be satisfied. For example, the learning end condition can be quickly satisfied through weighted learning based on weight, data expansion technique, and selective learning based on probability of false prediction. Accordingly, intervention of the annotator 222 may be minimized while learning is performed, and computation / time costs, annotation costs, and the like required for learning may be greatly reduced.

So far, with reference to FIGS. 17 to 21, an example in which the technical idea of the present disclosure is utilized in a medical domain has been briefly described. Hereinafter, a computing device 300 capable of implementing a device (e.g. learning device 100) according to various embodiments of the present disclosure will be described.

22 is an example hardware configuration diagram illustrating an example computing device 300 capable of implementing an apparatus in accordance with various embodiments of the present disclosure.

As illustrated in FIG. 22, the computing device 300 may include a memory (for loading) a computer program performed by one or more processors 310, a bus 350, a communication interface 370, and a processor 310. 330) and a storage 390 for storing the computer program 391. However, only components related to the exemplary embodiment of the present disclosure are illustrated in FIG. 22. Therefore, a person skilled in the art to which the present disclosure belongs may recognize that other general-purpose components may be further included in addition to the components illustrated in FIG. 22.

The processor 310 controls the overall operation of each component of the computing device 300. The processor 310 includes a CPU (Central Processing Unit), a Micro Processor Unit (MPU), a Micro Controller Unit (MCU), a Graphic Processing Unit (GPU), or any type of processor well known in the art of the present disclosure. Can be. Also, the processor 310 may perform an operation on at least one application or program for executing the method according to embodiments of the present disclosure. Computing device 300 may include one or more processors.

The memory 330 stores various data, commands and / or information. The memory 330 may load one or more programs 391 from the storage 390 to execute a method / operation according to various embodiments of the present disclosure. For example, when a computer program 391 performing a machine learning method according to some embodiments of the present disclosure is loaded in the memory 330, a module may be implemented on the memory 330 as illustrated in FIG. 5. . The memory 330 may be implemented as a volatile memory such as RAM, but the technical scope of the present disclosure is not limited thereto.

The bus 350 provides a communication function between components of the computing device 300. The bus 350 may be implemented as various types of buses such as an address bus, a data bus, and a control bus.

The communication interface 370 supports wired and wireless Internet communication of the computing device 300. Also, the communication interface 370 may support various communication methods other than Internet communication. To this end, the communication interface 370 may include a communication module well known in the art of the present disclosure.

The storage 390 may store the one or more programs 391 non-temporarily. The storage 390 is well-known in the field of non-volatile memory such as read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EPMROM), flash memory, hard disk, removable disk, or the technical field to which the present disclosure pertains. And any known form of computer-readable recording media.

Computer program 391 may include one or more instructions that, when loaded into memory 330, cause processor 310 to perform a method in accordance with various embodiments of the present disclosure. That is, the processor 310 may execute methods according to various embodiments of the present disclosure by executing the one or more instructions.

For example, the computer program 391 acquires a training dataset of a first model including a plurality of data samples to which no label information is given, and a probability of incorrect prediction of the first model for the plurality of data samples Calculating an operation, selecting at least one data sample among the plurality of data samples based on the calculated false prediction probability, and configuring a first data sample set, and first label information for the first data sample set It may include one or more instructions to perform the operation of performing the first learning on the first model using the first data sample set and the first label information and the operation of acquiring. In this case, the learning device 100 according to some embodiments of the present disclosure may be implemented through the computing device 300.

For another example, the computer program 391 acquires a training data set including a plurality of data samples to which no label information is given, first label information for the first data sample set included in the training data set Acquiring and learning the first set of data samples with the first label information to build a first model, and for the remaining data samples excluding the first set of data samples in the learning data set, Calculating an incorrect prediction probability, selecting at least one data sample from the remaining data samples based on the incorrect prediction probability, and configuring a second data sample set, and second label information for the second data sample set Learning a second model in an initialization state with the operation of acquiring and the second data sample set and the second label information It may include one or more instructions to perform the operation. In this case, the learning device 100 according to some other embodiments of the present disclosure may be implemented through the computing device 300.

So far, an exemplary computing device capable of implementing an apparatus according to various embodiments of the present disclosure has been described with reference to FIG. 22.

The technical idea of the present disclosure described so far with reference to FIGS. 1 to 22 may be embodied as computer readable codes on a computer readable medium. The computer-readable recording medium may be, for example, a removable recording medium (CD, DVD, Blu-ray Disc, USB storage device, removable hard disk), or a fixed recording medium (ROM, RAM, computer-equipped hard disk). You can. The computer program recorded on the computer-readable recording medium may be transmitted to another computing device through a network such as the Internet and installed on the other computing device, and thus used on the other computing device.

In the above, even if all components constituting the embodiments of the present disclosure are described as being combined or operated as one, the technical spirit of the present disclosure is not necessarily limited to these embodiments. That is, within the scope of the present disclosure, all of the components may be selectively combined to operate.

Although the operations in the drawings are shown in a specific order, it should not be understood that the operations must be performed in a specific order or in a sequential order, or that all the illustrated actions must be executed to obtain a desired result. In certain situations, multitasking and parallel processing may be advantageous. Moreover, the separation of various configurations in the above-described embodiments should not be understood as such separation is necessary, and the described program components and systems may generally be integrated together into a single software product or packaged into multiple software products. It should be understood that there is.

Although the embodiments of the present disclosure have been described with reference to the accompanying drawings, a person of ordinary skill in the art to which the present disclosure pertains may implement the present disclosure in other specific forms without changing the technical spirit or essential characteristics. Understand that there is. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. The scope of protection of the present disclosure should be interpreted by the claims below, and all technical spirits that are within the equivalent scope should be interpreted as being included in the scope of the technical spirits defined by the present disclosure.

Claims (15)

A memory containing one or more instructions; And
By executing the one or more instructions,
Obtain a training data set including a plurality of data samples to which no label information is given,
Calculate a miss-prediction probability for the plurality of data samples by using a probability calculation model trained to calculate a probability that a prediction of a target model that is a machine learning model performing a target task through machine learning is wrong. and,
Based on the calculated misprediction probability, at least one data sample among the plurality of data samples is selected to generate a first sub dataset,
Obtain first label information, which is label information for a data sample included in the generated first sub data set,
And a processor performing first learning on the target model using the first data sub-dataset and the first label information.
Machine learning device. According to claim 1,
The processor
Generating the first sub-dataset by selecting at least one data sample in which the calculated erroneous prediction probability is greater than or equal to a threshold value in the learning dataset
Machine learning device. According to claim 1,
The processor
Train the probability calculation model using the evaluation result of the prediction of the target model,
Calculating a false prediction probability for the plurality of data samples using the learned probability calculation model
Machine learning device. According to claim 3,
The processor
The predicted result of the target model is evaluated by comparing the result predicted by the target model with the correct answer label information of the data samples for evaluation given the correct answer label information,
Train the probability calculation model using the evaluation result and the evaluation data samples
Machine learning device. The method of claim 4,
The processor
Tagging the prediction data corresponding to the difference between the predicted result from the target model and the correct answer label information as label information in the data samples for evaluation
Machine learning device. The method of claim 5,
The processor
Tagging the first value as label information in the data sample for evaluation where the evaluation result corresponds to false positive (FP) or false negative (FN),
Tagging the second value as label information in the data sample for evaluation where the evaluation result corresponds to TP (true positive) or TN (true negative)
Machine learning device. According to claim 3,
The processor
Acquire second label information for a second sub data set corresponding to a part of the learning data set,
The target model is initially trained using at least some of the data samples included in the second sub dataset tagged with the second label information,
The target model is compared with a result predicted by the initially trained target model and label information corresponding to the predicted result for at least some of the data samples included in the second sub data set tagged with the second label information. To evaluate the predicted results of
Machine learning device. The method of claim 7,
The processor
Retraining the target model using the first sub-dataset, the first label information, the second sub-dataset and the second label information
Machine learning device. The method of claim 8,
The processor
Training the target model by assigning a greater weight to the first sub-dataset than the second sub-dataset
Machine learning device. According to claim 3,
The processor
Retrain the probability calculation model using the evaluation result of the prediction of the first learned target model,
Using the re-trained probability calculation model, misprediction probability of unlearned data samples is calculated,
Generating a second sub dataset by selecting at least one sample from the non-learning data samples based on the calculated probability of incorrect prediction of the non-learning data samples
Machine learning device. The method of claim 10,
The processor
Acquire second label information that is label information for each data sample included in the second sub data set,
Performing a second learning on the target model using the second sub dataset and the second label information
Machine learning device. According to claim 1,
The processor
Using the target model, a confidence score for each class for at least some data samples included in the training data set is calculated,
Entropy values for at least some of the data samples are calculated based on the calculated confidence score for each class,
Based on the calculated entropy value, the target model is trained by excluding some data samples from the training dataset.
Machine learning device. According to claim 1,
The processor
Excluding the data sample in which the calculated false prediction probability is less than the reference value from the training data set
Machine learning device. According to claim 1,
The processor
A weight is assigned to at least one of the data samples included in the first sub data set based on the probability of incorrect prediction of each of the data samples included in the first sub data set,
Train the target model using the weighted data sample
Machine learning device. Combined with the hardware,
Obtaining a training dataset including a plurality of data samples to which label information is not given;
Using a probability calculation model trained to calculate a probability that a prediction of a target model, which is a machine learning model performing a target task through machine learning, is incorrect, calculates a miss-prediction probability for each of the plurality of data samples. Calculating;
Generating a first sub dataset by selecting at least one data sample among the plurality of data samples based on the calculated false prediction probability;
Obtaining first label information that is label information for a data sample included in the generated first sub data set; And
The first sub-dataset and the first label information are used to store the medium in order to execute the step of performing first learning on the target model.
Computer program.

Images