JP7395396B2 - Information processing device, information processing method and program - Google Patents

Description

Embodiments of the present invention relate to an information processing device, an information processing method, and a program.

Uncertainty is a concept that originated in fields such as statistics, economics, and natural science. Various methods have been proposed to express uncertainty (uncertainty degree, uncertainty scale). In the field of machine learning, uncertainty can be used in various ways, such as selecting training data for active learning and selecting training data for learning highly accurate models.

JP2019-061642A

However, with the conventional technology, there are cases where the degree of uncertainty cannot be calculated with high accuracy.

The information processing device according to the embodiment includes a dividing section, a calculating section, and an integrating section. The dividing unit divides input data to be subjected to a process of inputting data and outputting a processing result into a plurality of partial data. The calculation unit executes processing for each of the plurality of partial data and calculates a plurality of degrees of uncertainty indicating the uncertainty of the processing. The integrating unit integrates a plurality of uncertainties and outputs the integrated uncertainty as input data uncertainty.

FIG. 1 is a diagram showing an example of data distribution for a two-class classification problem. FIG. 2 is a block diagram showing an example of the configuration of the information processing device according to the first embodiment. FIG. 3 is a diagram illustrating an example of a method for dividing parts. FIG. 4 is a flowchart illustrating an example of learning processing in the first embodiment. FIG. 5 is a flowchart illustrating an example of data sorting processing in the first embodiment. FIG. 6 is a diagram showing an example of image data and recognition results for the image data. FIG. 7 is a diagram showing an example of image data and recognition results for the image data. FIG. 8 is a block diagram showing an example of the configuration of an information processing device according to the second embodiment. FIG. 9 is a flowchart illustrating an example of learning processing in the second embodiment. FIG. 10 is a diagram for explaining the relationship between training data and selection target data. FIG. 11 is an explanatory diagram showing an example of the hardware configuration of the information processing apparatus according to the first or second embodiment.

DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of an information processing apparatus according to the present invention will be described in detail below with reference to the accompanying drawings. The following is an example of an information processing device (learning device) that calculates the degree of uncertainty, selects training data based on the calculated degree of uncertainty, and learns a machine learning model (hereinafter referred to as model) using the selected training data. Explain. Models include, for example, neural network models, random forests, support vector machines (SVM), conditional random fields (CRF), and logistic regression.

Processing using uncertainty is not limited to learning processing, and may be any type of processing. For example, the present invention may be applied to a device that calculates and outputs (displays, etc.) the degree of uncertainty in processing data.

Active learning is one of the machine learning frameworks that trains a classifier (an example of a model) while asking questions to experts. In general supervised learning, it is necessary to perform learning using a large amount of labeled data (training data) of several thousand or more. However, for many supervised learning tasks, obtaining labeled data is often very difficult or requires time and financial costs. Active learning is a learning method for reducing teaching costs by preferentially selecting training data that is as effective as possible for learning and presenting it to the user to be taught.

Theoretically, uncertainty is an index representing the ambiguity of the processing results (predicted results) of data by a model. In the following, the degree of uncertainty will be referred to as the degree of uncertainty. Learning data with high uncertainty (the prediction result is the most ambiguous) as training data can improve the prediction accuracy of the model and shorten the learning period to achieve the desired prediction performance. can.

FIG. 1 is a diagram showing an example of data distribution for a two-class classification problem. White triangles and black triangles each represent data classified into one of the two classes. The solid line represents the boundary line of the model's classification of the data. The closer the data is to the boundary line, the more uncertain the data is considered to be.

The following are reasons why prediction accuracy improves by using highly uncertain data as training data.
(F1) Reduce the learning frequency of data in the central parts 11 and 12 of the data distribution of each class to prevent overfitting to the data in the central parts 11 and 12.
(F2) Data 21, 22, and 23 that are far from the boundary line and have incorrect labels can be excluded.
(F3) Data with high uncertainty has a large information entropy (amount of information) and is considered to be included in the region 31 that is close to the boundary line. By intensively learning such data, the boundary line 32 can be constructed with higher accuracy.

The following methods exist to select data with high uncertainty. Note that probability is a value indicating processing accuracy (prediction accuracy, etc.). For example, in the case of recognition processing, the recognition rate corresponds to probability.
(M1) LeastConfidence: Select data with the minimum probability of the "label with maximum probability".
(M2) MarginSampling: Select data with the smallest probability difference between the "first highest probability label" and the "second highest probability label".
(M3) EntropyBased: Select data with the maximum entropy of the predicted distribution.
(M4) Query-By-Committee: Prediction processing is performed on data using multiple different models (committee) learned from mutually different initial values, and the variation in prediction results between multiple models or error (Loss value) Select the data with the largest value.

In each of the above methods, the following indicators correspond to the degree of uncertainty. Depending on the index, there are cases where the larger the value, the higher the uncertainty (ambiguity of the processing result), and the smaller the value, the higher the uncertainty.
(I1) Low probability of "the label with the highest probability" (I2) Difference between the probabilities of "the label with the highest probability" and "the label with the second highest probability" (I3) The magnitude of the entropy of the prediction distribution ( I4) Dispersion of predictions between multiple models or size of error

There is also a method of dividing data into a plurality of partial data (hereinafter sometimes referred to as parts) according to predetermined rules and performing predictive processing. For example, in connectionist temporal classification, sequential data is convolved, a feature map is divided into parts of one pixel each, prediction information of each part is integrated, and a final prediction is output. However, such methods are affected by local uncertainties and cannot select data with high overall uncertainty.

(First embodiment)
Therefore, the information processing apparatus according to the first embodiment calculates the degree of uncertainty of data so that the overall uncertainty is taken into consideration without being influenced by local uncertainties. Note that the process for which the degree of uncertainty is calculated may be any type of process, such as prediction, classification, or object recognition.

FIG. 2 is a block diagram showing an example of the configuration of the information processing device 100 according to the first embodiment. As shown in FIG. 2, the information processing device 100 includes a preprocessing section 101, a learning section 102, an output control section 103, a sorting section 110, and a storage section 121.

The preprocessing unit 101 executes preprocessing for data selection processing and learning processing. If a trained model already exists, preprocessing is a process of preparing a trained model according to the purpose as a base model. The preprocessing unit 101 may prepare the trained model as it is as a base model, or may perform preprocessing that does not require learning, such as pruning, on the trained model to generate the base model. good.

If a trained model does not exist, preprocessing involves dividing the input data into training data and selection target data according to predetermined rules, and learning the model using the training data to generate a base model. Includes processing to do. Training data is data used to learn a model in advance. The data to be sorted is data to be subjected to data sorting processing by the sorting unit 110. Examples of rules are listed below.
(R1) Randomly divide the data.
(R2) The labels (teaching information) given to the data are referred to and the data given the same label is divided so as not to be biased.
(R3) Refer to the classification information (clustering result information) in which the data is classified into a plurality of clusters, and divide the data so that the data classified into the same cluster are not biased.

The preprocessing unit 101 learns a model using training data and generates a base model. Note that if a trained model is prepared in advance and can be used as is, the information processing device 100 does not need to include the preprocessing unit 101.

The learning unit 102 uses training data to learn a model that executes a process of inputting data and outputting a processing result. For example, the learning unit 102 learns a model using, as training data, data separated as training data by preprocessing and data selected by the selection unit 110 from among the data to be selected. Data selection processing by the selection unit 110 will be described later.

The output control unit 103 controls the output of various information by the information processing device 100. For example, the output control unit 103 outputs information about the model learned by the learning unit 102 to an external device using the model. The output control unit 103 may display the calculated degree of uncertainty on a display device such as a display.

The sorting unit 110 executes data sorting processing on data to be sorted. The sorting section 110 includes a dividing section 111, a calculating section 112, and an integrating section 113.

The dividing unit 111 divides the data to be sorted (input data) into a plurality of partial data (parts). The dividing unit 111 may divide the input data to be sorted as is, or may divide data obtained by adding processing to the data to be sorted. The process to be added to the data to be sorted is, for example, a process to obtain the feature amount of the data to be sorted.

Whether or not to apply processing to the data to be sorted may be determined, for example, depending on the format of the input data for the model to be learned. For example, when using a model that inputs image data, the dividing unit 111 may divide the sorting target data, which is the image data, as it is. For example, when using a model that performs recognition processing by inputting the feature amount of image data, the dividing unit 111 may execute processing to obtain the feature amount of the image data, and divide the obtained feature amount.

The image data is, for example, two-dimensional data. The feature amount of image data is, for example, a feature map calculated by convolving two-dimensional image data. The feature map can be calculated, for example, by a neural network that inputs image data and outputs a feature map. The feature map is, for example, data in which pixel values indicating feature amounts are set at each two-dimensional pixel position. For image data or a feature map, the dividing unit 111 generates a plurality of parts by dividing the data into a certain number of pixels in the upper and lower or left and right directions, for example.

For example, when using a model that inputs one-dimensional series data such as character string data and audio data, the dividing unit 111 divides the series data at fixed intervals or unspecified intervals to generate multiple parts. . The unspecified interval is, for example, a randomly determined interval.

FIG. 3 is a diagram illustrating an example of a method for dividing parts. Images 301 and 302 are examples of data before division. Images 301 and 302 are each converted into feature maps 311. The dividing unit 111 divides the feature map 311 horizontally or vertically to generate parts.

The calculation unit 112 calculates the degree of uncertainty for each of a plurality of parts. For example, the calculation unit 112 inputs each of a plurality of parts into a model, and calculates the degrees of uncertainty shown in (I1) to (I4) above from the processing results.

The integrating unit 113 integrates a plurality of degrees of uncertainty calculated for each of a plurality of parts, calculates and outputs a degree of uncertainty for the data to be sorted before division. First, the integration unit 113 determines one of a plurality of integration methods based on state information indicating the state of the model with respect to verification data (validation data).

The multiple integration methods include, for example, the following methods.
(MM1) Calculate the average value or median value of all multiple uncertainties.
(MM2) Calculate the average value or median value of some uncertainties among the plurality of uncertainties.
(MM3) Calculate a weighted addition value of multiple uncertainties.

(MM2) is a method of calculating, for example, an average value or a median value of uncertainties whose values are included in a predetermined range among a plurality of uncertainties. The predetermined range is, for example, a range that excludes uncertainties such as deviations in values.

The verification data includes, for example, data obtained when processing is performed in advance using a model to be learned, data obtained when learning is performed in advance on this model, and the like. The state information is, for example, the prediction accuracy of the model with respect to the verification data, and the error when the model is trained with respect to the verification data.

For example, if the recognition rate (prediction accuracy) of each category of the model with respect to the verification data is significantly different, the integration unit 113 determines the above (MM3) as the integration method. Then, the integrating unit 113 integrates the uncertainties by applying a smaller weight to parts with a higher recognition rate and a larger weight to parts with a lower recognition rate, for example.

Furthermore, for example, if the recognition rate or error with respect to the verification data fluctuates greatly during model learning, the integration unit 113 determines (MM2) as the integration method. Then, the integrating unit 113 removes parts with different degrees of uncertainty by excluding uncertainties that are outside a predetermined range, and then calculates the median or average value of each part. , to integrate uncertainties.

Furthermore, if the model state information cannot be acquired, the integrating unit 113 determines (MM1) as the integrating method.

Each of the above units (preprocessing unit 101, learning unit 102, output control unit 103, and selection unit 110) is realized by, for example, one or more processors. For example, each of the above units may be realized by causing a processor such as a CPU (Central Processing Unit) to execute a program, that is, by software. Each of the above units may be realized by a processor such as a dedicated IC (Integrated Circuit), that is, by hardware. Each of the above units may be realized using a combination of software and hardware. When using a plurality of processors, each processor may implement one of each unit, or may implement two or more of each unit.

The storage unit 121 stores various information used in various processes by the information processing apparatus 100. For example, the storage unit 121 stores training data, selection target data, data indicating a model, and the like. The storage unit 121 can be configured from any commonly used storage medium such as a flash memory, a memory card, a RAM (Random Access Memory), an HDD (Hard Disk Drive), and an optical disk.

Next, a learning process performed by the information processing apparatus 100 according to the first embodiment configured as described above will be described. FIG. 4 is a flowchart illustrating an example of learning processing in the first embodiment.

The preprocessing unit 101 divides the input data into training data and selection target data (step S101). The preprocessing unit 101 learns a model using training data and generates a base model (step S102). Note that if a trained model is prepared in advance, steps S101 and S102 can be omitted.

The sorting unit 110 executes data sorting processing on the sorting target data (step S103). Details of the data selection process will be described later.

The learning unit 102 determines whether data with high uncertainty is obtained through the data selection process (step S104). If data with high uncertainty is obtained (step S104: Yes), the learning unit 102 adds the selected data to training data to learn the model (step S105).

The learning method by the learning unit 102 may be any method. For example, if the training data is unsupervised data, the learning unit 102 learns the model by active learning. If the training data is supervised data, the learning unit 102 learns the model by supervised learning.

After learning the model, the process returns to step S103 and is repeated. If data with high uncertainty is not obtained (step S104: No), the learning unit 102 outputs the learned model that has been learned so far (step S106), and ends the learning process.

Next, details of the data sorting process in step S103 will be explained. FIG. 5 is a flowchart illustrating an example of data sorting processing in the first embodiment.

The dividing unit 111 divides each of the input sorting target data into a plurality of parts (step S201). The calculation unit 112 calculates the degree of uncertainty for each of the plurality of divided parts (step S202). The integrating unit 113 refers to the state information of the model and determines the method of integrating the uncertainties (step S203). The integrating unit 113 integrates a plurality of degrees of uncertainty using the determined integration method, and calculates the degree of uncertainty for the data to be sorted before division (step S204). The integrating unit 113 outputs data to be sorted that has a high degree of uncertainty indicated by the calculated degree of uncertainty (for example, the degree of uncertainty is greater than or equal to a threshold) (step S205).

For example, when using an uncertainty degree where the smaller the value, the higher the uncertainty, as in (I2) above (difference between the probabilities of "the first most probable label" and "the second most probable label"), the integration The unit 113 outputs data to be sorted whose degree of uncertainty is less than or equal to a threshold value. When using the degree of uncertainty in which the larger the value, the higher the uncertainty, as in (I3) (magnitude of entropy of the predicted distribution) above, the integration unit 113 selects data to be sorted whose degree of uncertainty is equal to or higher than the threshold. Output.

The threshold value and termination condition may be a predetermined value or condition, or may be settable by the user. Further, the threshold value and termination condition may be dynamically changed during processing depending on the situation. For example, if the data to be sorted cannot be obtained using a certain threshold value, the integrating unit 113 may change the threshold value so that the data to be sorted can be easily obtained.

Differences between this embodiment and a comparative example will be explained below. To simplify the explanation, an example will be described in which image data including character strings with only two categories, "a" and "b", is recognized by a model, and the recognized categories are output. The method for dividing parts is to divide a feature map obtained by inputting image data into a neural network. In this embodiment, the method for selecting data with high uncertainty is (M2) (MarginSampling). Further, although the degree of uncertainty is generally related to noise and the beauty of characters, here, the degree of uncertainty is expressed only by noise.

6 and 7 are diagrams showing examples of image data and recognition results for the image data. The wedge shape in FIGS. 6 and 7 means that noise is included on the image. In the examples of FIGS. 6 and 7, image data 602 and 702 including the character string "aab" are divided into three parts P1, P2, and P3, respectively. Image data 602 includes noise only in part P3. That is, the image data 602 becomes data with high local uncertainty. Image data 702 includes noise in all parts. In other words, the image data 702 is data with high overall uncertainty.

Examples of corresponding recognition results 601 and 701 are shown above image data 602 and 702, respectively. The numerical values in the recognition results 601 and 701 indicate the probability that the character string corresponding to each part belongs to the category "a" or "b".

In the image data 602, "aa" corresponding to parts P1 and P2 has no noise and is recognized with a high probability, but "b" corresponding to part P3 is difficult to recognize due to large noise. This is the data. Image data 702 has small noise in "aab" corresponding to all parts P1, P2, and P3, and is data that is difficult to recognize on average.

In the comparative example, the probability of each label string is calculated by the product of the probabilities of the categories of each part, and the difference between the probabilities of each label string is defined as the degree of uncertainty. The label string is information in which characters (labels) that can be the recognition result of each part are arranged. In the examples of FIGS. 6 and 7, eight types of label strings are obtained: "aaa", "aab", "aba", "abb", "baa", "bab", "bba", and "bbb".

In addition, in the comparative example, image data with the smallest difference in probability (=uncertainty) between the "first highest probability label sequence" and the "second highest probability label sequence" is selected. do. In the comparative example, the degree of uncertainty for the image data 602 is calculated as follows, for example.
Label string "aab": 1 x 1 x 0.5 = 0.5
Label string "aaa": 1 x 1 x 0.5 = 0.5
Uncertainty = 0.5-0.5 = 0

Furthermore, in the comparative example, the degree of uncertainty for the image data 702 is calculated as follows, for example.
Label string “aab”: 0.75×0.75×0.75≒0.42
Label string “aaa”: 0.75×0.75×0.25≒0.14
Uncertainty = 0.42-0.14 = 0.28

Therefore, in the comparative example, the image data 702 (uncertainty level = 0.28), which has high overall uncertainty, is used as data with high uncertainty (low uncertainty value), but with local uncertainty. Image data 602 (uncertainty level=0) with high certainty is selected.

In contrast, in this embodiment, the degree of uncertainty for the image data 602 is calculated as follows, for example.
Uncertainty of part P1: 1-0=1
Uncertainty of part P2: 1-0=1
Uncertainty of part P3: 0.5-0.5=0
Uncertainty of image data 602: (1+1+0)/3≒0.67

Furthermore, in this embodiment, the degree of uncertainty for the image data 702 is calculated as follows, for example.
Uncertainty of part P1: 0.75-0.25=0.5
Uncertainty of part P2: 0.75-0.25=0.5
Uncertainty of part P3: 0.75-0.25=0.5
Uncertainty of image data 702: (0.5+0.5+0.5)/3=0.5

Therefore, in the present embodiment, image data 702 with high overall uncertainty is selected as data with high uncertainty (low uncertainty value).

As described above, in the comparative example, if the difference between the probability of the first candidate and the probability of the second candidate for each part is small for even one of the multiple parts, the difference in probability (Margin) between the label strings is also inevitable. becomes smaller. In other words, the influence of local uncertainty increases.

On the other hand, in this embodiment, the degree of uncertainty of data is calculated by calculating the degree of uncertainty for each part and integrating the degrees of uncertainty of each part (here, the average value is taken). Therefore, it is possible to calculate a more accurate degree of uncertainty in consideration of the overall degree of uncertainty. As a result, it becomes possible to select data with high overall uncertainty.

Furthermore, in this embodiment, it is also possible to reduce the amount of calculation as described below. In the comparative example, when calculating the degree of uncertainty of data, the probability of each label string (all combinations of categories) is calculated by the product of the categories of each part. Therefore, the order of the amount of calculation in the comparative example is Ο((number of categories classes)^(sequence length)). That is, as the number of parts increases, the amount of calculation increases exponentially.

On the other hand, in this embodiment, comparison between classes may be performed for each part. Therefore, the order of the amount of calculation is Ο((number of category classes)×(sequence length)). In this embodiment, as the number of parts increases, the amount of calculation increases linearly. Therefore, in this embodiment, data can be selected with a smaller amount of calculation than in the comparative example.

For example, in the case of FIG. 3, in order to calculate the degree of uncertainty in the comparative example, 8 Although it is necessary to calculate and compare the probabilities of each type of character string, in this embodiment, it is only necessary to compare the probabilities of "a" and "b" for each of the three parts three times in total.

As described above, the information processing apparatus according to the first embodiment can calculate the degree of uncertainty with higher accuracy.

(Second embodiment)
The information processing device according to the second embodiment further includes a function of performing model learning more efficiently.

FIG. 8 is a block diagram showing an example of the configuration of an information processing device 100-2 according to the second embodiment. As shown in FIG. 8, the information processing device 100-2 includes a preprocessing section 101-2, a learning section 102-2, an output control section 103, a sorting section 110, and a storage section 121. .

In the second embodiment, the functions of the preprocessing section 101-2 and the learning section 102-2 are different from those in the first embodiment. The other configurations and functions are the same as those in FIG. 1, which is a block diagram of the information processing apparatus 100 according to the first embodiment, so the same reference numerals are given and the description thereof will be omitted.

The preprocessing unit 101-2 differs from the preprocessing unit 101 of the first embodiment in that it further includes a function of setting categories to be recognized. For example, in the case of character recognition, the preprocessing unit 101-2 sets the category to characters or radicals. For example, in the case of speech recognition, the preprocessing unit 101-2 sets the category to basic phoneme. For example, in the case of object recognition, the preprocessing unit 101-2 sets a category for each object to be recognized. The preprocessing unit 101-2 sets, for example, a category specified by the user as a category to be recognized.

The learning unit 102-2 is different from the learning unit of the first embodiment in that it further includes a function for excluding data with low uncertainty from the training data (forgetting function) and a function for determining convergence of learning. It is different from 102.

Next, a learning process performed by the information processing apparatus 100-2 according to the second embodiment configured as described above will be described using FIG. 9. FIG. 9 is a flowchart illustrating an example of learning processing in the second embodiment.

The preprocessing unit 101 sets categories used in recognition processing (step S301).

Steps S302 and S303 are the same processes as steps S101 and S102 in the information processing apparatus 100 according to the first embodiment, so the description thereof will be omitted.

In this embodiment, the learning unit 102-2 determines whether learning using training data has been completed (converged) (step S304). For example, the learning unit 102-2 uses the error and error rate calculated during learning to determine whether learning has converged. The learning unit 102-2 determines that learning has converged, for example, when the error or error rate becomes less than or equal to a threshold value.

If the learning is not completed (step S304: No), the learning unit 102-2 returns to step S303 and repeats the process. When the learning is completed (step S304: Yes), the learning unit 102-2 moves part of the training data to the selection target data (step S305). For example, the learning unit 102-2 randomly selects some data from the training data and moves the selected data to the selection target data.

FIG. 10 is a diagram for explaining the relationship between training data and selection target data. Data 1001 corresponds to the data (input data) to be processed in step S302. The training data divided in step S302 is indicated as base model training data 1002 in FIG. Furthermore, the data to be sorted separated in step S302 corresponds to the data to be sorted 1003 in FIG.

Data selected by the screening process (selection: Yes) out of the screening target data 1003 is added to the training data 1004. The training data 1004 is used together with the base model training data 1002 in model learning in step S308. Of the data to be sorted 1003, data that has not been sorted by the sorting process (sorting: No) is left in the data to be sorted 1003.

In this embodiment, each time the model learning in step S308 is repeated, a part of the base model training data 1002 is moved to the selection target data in step S305. If data with low uncertainty is included in the base model training data 1002, if this data is moved to the selection target data 1003 in the process of step S305, it will not be selected by the selection process and will be removed from the training data 1004. It becomes possible to exclude. Therefore, the model can be learned more efficiently.

Returning to the explanation of FIG. 9. Steps S306 to S309 are the same processes as steps S103 to S106 in the information processing apparatus 100 according to the first embodiment, so the description thereof will be omitted.

After step S308, the learning unit 102-2 determines whether learning using the selected data has been completed (converged) (step S310). Similar to step S304, this process is a determination process using the error, error rate, etc. calculated during learning.

If the learning is not completed (step S310: No), the learning unit 102-2 returns to step S308 and repeats the process. If the learning is completed (step S310: Yes), the learning unit 102-2 returns to step S305 and repeats the process.

In this way, the information processing device according to the second embodiment further includes a function to exclude data with low uncertainty from training data and a function to determine convergence of learning, so that model learning can be made more efficient. can be carried out in a specific manner.

As explained above, according to the first to second embodiments, the degree of uncertainty can be calculated with high accuracy.

Next, the hardware configuration of the information processing apparatus according to the first or second embodiment will be described using FIG. 11. FIG. 11 is an explanatory diagram showing an example of the hardware configuration of the information processing apparatus according to the first or second embodiment.

The information processing device according to the first or second embodiment includes a control device such as a CPU (Central Processing Unit) 51, a storage device such as a ROM (Read Only Memory) 52 or a RAM (Random Access Memory) 53, and a network. It is provided with a communication I/F 54 that connects to perform communication, and a bus 61 that connects each part.

The program executed by the information processing device according to the first or second embodiment is provided by being pre-installed in the ROM 52 or the like.

The program executed by the information processing device according to the first or second embodiment is a file in an installable format or an executable format and is stored on a CD-ROM (Compact Disk Read Only Memory), a flexible disk (FD), or a CD. It may also be configured to be recorded on a computer-readable recording medium such as -R (Compact Disk Recordable) or DVD (Digital Versatile Disk) and provided as a computer program product.

Furthermore, the program executed by the information processing apparatus according to the first or second embodiment is configured to be stored on a computer connected to a network such as the Internet, and provided by being downloaded via the network. Good too. Furthermore, the program executed by the information processing apparatus according to the first or second embodiment may be provided or distributed via a network such as the Internet.

The program executed by the information processing device according to the first or second embodiment can cause the computer to function as each part of the information processing device described above. In this computer, the CPU 51 can read a program from a computer-readable storage medium onto the main storage device and execute it.

Although several embodiments of the invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and changes can be made without departing from the gist of the invention. These embodiments and their modifications are included within the scope and gist of the invention, as well as within the scope of the invention described in the claims and its equivalents.

100, 100-2 Information processing device 101, 101-2 Preprocessing unit 102, 102-2 Learning unit 103 Output control unit 110 Sorting unit 111 Division unit 112 Calculation unit 113 Integration unit 121 Storage unit

Claims (5)

a dividing unit that divides input data to be subjected to a process of inputting data and outputting a processing result into a plurality of partial data;
a calculation unit that executes the processing for each of the plurality of partial data and calculates a plurality of degrees of uncertainty indicating the uncertainty of the processing;
an integrating unit that integrates the plurality of uncertainties and outputs them as the uncertainties of the input data;
Equipped with
The integration unit determines one of a plurality of integration methods based on status information indicating a state of the processing on the verification data, and integrates the plurality of uncertainties using the determined integration method. death,
The plurality of integration methods include a method of calculating an average value or a median value of a plurality of the degrees of uncertainty, a method of calculating an average value or a median value of some of the degrees of uncertainty among the plurality of degrees of uncertainty, and a method of calculating a weighted sum of the plurality of uncertainties,
Information processing device. Some of the plurality of uncertainties are uncertainties whose values are included in a predetermined range among the plurality of uncertainties,
The information processing device according to claim 1 . further comprising a learning unit that learns a machine learning model for executing the process using a plurality of input data whose uncertainty indicated by the degree of uncertainty output by the integrating unit is equal to or greater than a threshold value as training data;
The information processing device according to claim 1. An information processing method executed by an information processing device, the method comprising:
a dividing step of dividing input data to be subjected to a process of inputting data and outputting a processing result into a plurality of partial data;
a calculation step of executing the process for each of the plurality of partial data and calculating a plurality of degrees of uncertainty indicating the uncertainty of the process;
an integrating step of integrating a plurality of the uncertainties and outputting them as the uncertainties of the input data;
including;
In the integration step, one of a plurality of integration methods is determined based on status information indicating the status of the processing for the verification data, and the plurality of uncertainties are integrated using the determined integration method. death,
The plurality of integration methods include a method of calculating an average value or a median value of a plurality of the degrees of uncertainty, a method of calculating an average value or a median value of some of the degrees of uncertainty among the plurality of degrees of uncertainty, and a method of calculating a weighted sum of the plurality of uncertainties,
Information processing method. to the computer,
a dividing step of dividing input data to be subjected to a process of inputting data and outputting a processing result into a plurality of partial data;
a calculation step of executing the process for each of the plurality of partial data and calculating a plurality of degrees of uncertainty indicating the uncertainty of the process;
an integrating step of integrating a plurality of the uncertainties and outputting them as the uncertainties of the input data;
run the
In the integration step, one of a plurality of integration methods is determined based on status information indicating the status of the processing for the verification data, and the plurality of uncertainties are integrated using the determined integration method. death,
The plurality of integration methods include a method of calculating an average value or a median value of a plurality of the degrees of uncertainty, a method of calculating an average value or a median value of some of the degrees of uncertainty among the plurality of degrees of uncertainty, and a method of calculating a weighted sum of the plurality of uncertainties,
program .

Images