Information Processing Method

Abstract

To properly construct a learning model.SOLUTION: An information processing method includes: an examination step of examining feature parameters related to a learned model constructed by machine learning using a development data set based on data regarding an object; and a re-construction step of re-constructing the learned model by the machine learning using a definition area being a range that can be taken by the feature value regarding the object and a new development data set being adjusted according to a relation with the feature parameters.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to the technical field of information processing methods. [Background technology]

[0002] As one example of this type of method, a method has been proposed in which machine learning processing is performed on features of biological sound data for a predetermined period including the period during which swallowing is estimated to have occurred, and whether or not there is a possibility of swallowing disorder is determined (see Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2018 / 070522 Summary of the Invention [Problem to be solved by the invention]

[0004] A learning model related to machine learning processing may be subject to overlearning or underlearning due to the data set used when constructing the model. The technology described in Patent Document 1 has room for improvement in this regard.

[0005] The present invention has been made in consideration of the above circumstances, and has an object to provide an information processing method capable of appropriately constructing a learning model. [Means for solving the problem]

[0006] In order to solve the above problem, the information processing device of the present invention includes an investigation step of investigating feature parameters of a trained model constructed by machine learning using a development dataset based on data related to a target, and a reconstruction step of reconstructing the trained model by machine learning using a new development dataset adjusted according to a domain, which is a range that a feature value related to the target can take, and a relationship with the feature parameters.

[0007] The functions and other advantages of the present invention will become apparent from the following detailed description of the preferred embodiments. [Brief description of the drawings]

[0008] [Figure 1] 4 is a flowchart illustrating a part of an information processing method according to an embodiment. [Diagram 2] FIG. 13 is a diagram showing an example of display of feature parameters. [Diagram 3] 10 is a flowchart showing another part of the information processing method according to the embodiment. [Figure 4] FIG. 1 is a conceptual diagram illustrating a concept of distribution of verification data. [Diagram 5] FIG. 13 is a diagram showing an example of display of evaluation results. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] An embodiment of an information processing method will be described. The information processing method according to the embodiment includes an investigation step and a reconstruction step. In the investigation step, feature parameters related to a trained model constructed by machine learning using a development dataset based on data related to a target are investigated. Here, the "feature parameters" refer to parameters acquired after the execution of machine learning. In other words, the "feature parameters" are parameters different from hyperparameters.

[0010] For example, consider the case of building a learning model that detects white lines on roads from images. In this case, the development dataset includes images that show white lines on roads. The way white lines are captured changes depending on the weather and time of day. If the development dataset contains, for example, a significantly large number of images that show white lines during the day on a sunny day, but contains almost no images that show white lines at night on a rainy day (i.e., if there is a bias in the development dataset), a learning model may be built that is over-learned about white lines during the day on a sunny day, but is under-learned about white lines at night on a rainy day.

[0011] Here, according to the research of the present inventor, it has been found that the feature parameters tend to be overlearned or underlearned. Therefore, by investigating the feature parameters in the investigation step, it is possible to detect the tendency of overlearning or underlearning.

[0012] In the reconstruction step, the trained model is reconstructed by machine learning using a new development dataset adjusted according to the domain, which is the range that the feature quantity related to the target can take, and the relationship with the feature parameters. Here, "adjusting" the development dataset means that the data (images in the above example) included in the development dataset is adjusted so as to reduce the tendency for overlearning or underlearning. The "domain" can be known in advance, for example, from literature, empirical rules, etc.

[0013] By performing machine learning using the adjusted new development dataset to reconstruct the trained model, it is possible to construct a trained model in which overlearning and underlearning are suppressed. Therefore, according to the information processing method, it is possible to appropriately construct a trained model.

[0014] In one aspect of the information processing method, the information processing method includes, before the investigation step, a construction step of constructing the trained model by machine learning using the development dataset, and after the reconstruction step, an evaluation step of evaluating the reconstructed trained model using a validation dataset for evaluating performance of the trained model.

[0015] In general, in machine learning, an optimal learning model is constructed by repeatedly constructing a learning model using a development dataset and evaluating a trained model using a validation dataset. In contrast, in the information processing method, before evaluating the trained model, a trained model is reconstructed using a new development dataset.

[0016] In this aspect, the validation data set may be a data set adjusted based on the domain. The validation data set may also be biased, as in the development data set. If a trained model is evaluated using a biased validation data set, the trained model may not be evaluated appropriately.

[0017] The term "adjustment" in relation to the validation dataset means that the data included in the validation dataset is adjusted so as to reduce bias. According to this aspect, the trained model is evaluated using the adjusted validation data, so that the trained model can be appropriately evaluated.

[0018] In this aspect, the method may further include a display step of displaying the results of the evaluation in the evaluation step as evaluation circles whose radii vary according to the performance of the trained model for each region defined by the values ​​of the first and second feature components on a plane defined by the first and second feature components related to the target. With this configuration, the evaluation results can be displayed in a visually easy-to-understand manner, which is highly advantageous in practice.

[0019] In another aspect of the information processing method, the target is a body sound. There are a wide variety of body sounds. Even if the body sound is caused by a single disease, the characteristics of the sound may change to some extent depending on the physique of the living body. For example, when constructing a learning model for detecting a single body sound, the data contained in the development dataset and the validation dataset are likely to be relatively biased. By applying the information processing method to body sounds, a trained model can be appropriately constructed. EXAMPLES

[0020] An embodiment of the information processing method will be described with reference to Fig. 1 to Fig. 5. In this embodiment, a learning model for determining the presence or absence of abnormal sounds in body sounds will be described, but the information processing method can also be applied to other learning models.

[0021] 1, first, a development dataset for constructing a learning model is prepared (step S101). Here, the development dataset is not limited to real data acquired by, for example, an inspection or an experiment, and may include processed data generated by, for example, processing the real data. As a method for generating the processed data, for example, a method for generating the processed data by changing at least one of the volume and frequency of the real data (i.e., body sound data) can be mentioned.

[0022] Next, the machine learning device calculates feature amounts from each data included in the development dataset (step S102). Various existing methods can be applied to the feature amount calculation method. One example of a method for calculating feature amounts is the method described in Japanese Patent No. 6672478. Next, the machine learning device constructs (designs) a learning model using a predetermined machine learning method (step S103). Since various existing methods can be applied to the machine learning method, detailed explanations thereof will be omitted.

[0023] After the learning model is constructed, the machine learning device investigates the feature parameters related to the constructed learning model (hereinafter, appropriately referred to as the "trained model") (step S104). At this time, the machine learning device displays the investigated feature parameters in a graph format, for example, as shown in FIG. 2(a).

[0024] In the graph shown in Figure 2(a), the weights are relatively large on the high feature dimension side of the domain, while the weights are relatively small on the low feature dimension side. In other words, overlearning may occur on the high feature dimension side, while underlearning may occur on the low feature dimension side.

[0025] Therefore, the development dataset is adjusted in the process of step S105. In the process of step S105, for example, some of the data included in the development dataset is deleted or new data is added to the development dataset so that the state shown in Fig. 2(a) approaches the state shown in Fig. 2(b).

[0026] Next, the machine learning device reconstructs (redesigns) the trained model using the adjusted development dataset (step S106). At this time, the machine learning device may calculate features from each data included in the adjusted development dataset, similar to the process of step S102.

[0027] The trained model constructed (reconstructed) as described above is verified and evaluated using a validation dataset. In Fig. 3, first, a validation dataset for verifying and evaluating the trained model is prepared (step S201). Here, the validation dataset may include processed data as well as real data, similar to the development dataset.

[0028] Next, a domain is determined from the results of the processing in step S104 described above (for example, the graph shown in FIG. 2(a)) (step S202). The domain determined at this time does not need to be the same as the domain when the development dataset is adjusted in the processing in step S105 described above.

[0029] Assume that the distribution of the volume and frequency of each data included in the validation data set prepared in the process of step S201 is biased, for example, as shown in Fig. 4(a). Note that the black circles in Fig. 4(a) represent data (similarly for Fig. 4(b)). Therefore, in the process of step S203, for example, some of the data included in the validation data set is deleted or new data is added to the validation data set so that the distribution shown in Fig. 4(a) approaches the distribution shown in Fig. 4(b).

[0030] Next, the machine learning device uses the adjusted validation data set to validate and evaluate the trained model (step S204). Since various existing aspects can be applied to the validation and evaluation method, detailed explanations thereof will be omitted.

[0031] The machine learning device displays the results of the verification and evaluation as a diagram, for example, as shown in FIG. 5. The circle in FIG. 5 is an evaluation circle. The performance of the trained model when data having a volume and frequency related to the center of the evaluation circle is input to the trained model is represented by the size of the evaluation circle. Here, the higher the performance of the trained model, the larger the radius of the evaluation circle.

[0032] In cases where the amount of data is currently limited, such as abnormal respiratory sound data, the domain of definition may change as the amount of clinical data increases. By using the method of the present invention, it is possible to improve the performance in a defined range and strengthen the learning model.

[0033] The present invention is not limited to the above-described embodiments, but may be modified as appropriate within the scope of the claims and the gist or concept of the invention as can be read from the entire specification, and information processing methods involving such modifications are also included in the technical scope of the present invention.

Claims

1. An investigation step of investigating feature parameters related to a trained model constructed by machine learning using a development dataset based on data related to the target; A reconstruction step of reconstructing the trained model by machine learning using a new development dataset adjusted according to the relationship between a domain, which is a range in which the feature value related to the target can be taken, and the feature parameters; 13. An information processing method comprising:

2. The information processing method includes: A construction step of constructing the trained model by machine learning using the development dataset before the investigation step; After the reconstruction step, an evaluation step of evaluating the reconstructed trained model using a validation data set for evaluating the performance of the trained model; The information processing method according to claim 1, further comprising:

3. 3. The information processing method according to claim 2, wherein the validation data set is a data set adjusted based on the domain.

4. 4. The information processing method according to claim 2 or 3, further comprising a display step of displaying a result of the evaluation in the evaluation step on a plane defined by a first feature component and a second feature component related to the target, for each region defined by a value of the first feature component and a value of the second feature component, the radius of the circle varying according to performance of the trained model.

5. 5. The information processing method according to claim 1, wherein the target is a body sound.

Images