JP2015176175A - Information processing apparatus, information processing method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technique capable of improving a performance of a learning model used in a recognition device.SOLUTION: An information processing apparatus 10 comprises: an input data reception part 110 which receives a plurality of feature quantities and a plurality of labels given to the respective feature quantities; a label distribution calculation part 120 which calculates the probability distribution of labels for each feature quantity from the plurality of labels given to the respective feature quantities; and an enlargement learning data generation part 140 which generates a plurality of pieces of learning data for each feature quantity on the basis of the feature quantities and the probability distribution of labels for each feature quantity.

Description

  The present invention relates to an information processing device, an information processing method, and a program, and more particularly, to an information processing device, an information processing method, and a program that can be applied to generation of learning data of a recognizer.

  In recent years, research on recognizers that recognize unknown input data by estimating meanings (labels) such as evaluation values and evaluation categories of input data based on feature values extracted from input data has been widely performed. For example, an emotion recognizer that estimates the type of emotion, the level of emotion, and the like based on feature amounts extracted from speech data has been studied.

  An example of a technique related to such a recognizer is disclosed in the following non-patent document. Non-Patent Document 1 or 2 calculates an average value of a plurality of labels assigned to each feature amount as a teacher label for each feature amount, and each of the plurality of feature amounts and a teacher label calculated for the feature amount A method for generating learning data (recognition device) that recognizes a label of unknown input data based on learning data generated for each feature amount by generating one learning data for each feature amount by combining Yes.

M. Grimm, K. Kroschel and S. Narayanan, "SUPPORT VECTOR REGRESSION FOR AUTOMATIC RECOGNITION OF SPONTANEOUS EMOTIONS IN SPEECH," Acoustics, Speech and Signal Processing, 2007. ICASSP 2007. IEEE International Conference on. IEEE, 2007. p. IV -1085-IV-1088. F. Eyben, M. Wollmer and B. Schuller, "openEAR? Introducing the Munich Open-Source Emotion and Affect Recognition Toolkit," Affective Computing and Intelligent Interaction and Workshops, 2009. ACII 2009. 3rd International Conference on. IEEE, 2009. p. 1-6.

  When generating a learning model, generally, a human evaluation value is used as a label assigned to each feature amount. For this reason, the labels to be given may vary, particularly for feature quantities that are difficult to estimate. However, in the methods disclosed in each of the above-mentioned non-patent documents, variations in a plurality of labels given to each feature amount are not taken into consideration when generating a teacher label. In other words, in each of the above-mentioned non-patent documents, feature quantities that are difficult to estimate due to large label variations are uniformly handled as feature quantities that are easy to estimate with little label variation. It may not be possible.

  An object of the present invention is to provide a technique capable of improving the performance of a learning model used in a recognizer or the like.

According to the present invention,
Input data receiving means for receiving a plurality of feature quantities and a plurality of labels respectively assigned to the feature quantities;
Label distribution calculating means for calculating a probability distribution of a label for each feature quantity from a plurality of labels assigned to each feature quantity;
Extended learning data generating means for generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
An information processing apparatus is provided.

According to the present invention,
Computer
Receiving a plurality of feature amounts and a plurality of labels assigned to each feature amount,
From the plurality of labels assigned to each feature amount, calculate the probability distribution of the label for each feature amount,
Based on the probability distribution of the label for each feature amount, obtain a plurality of teacher labels for each feature amount,
Generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
An information processing method including the above is provided.

According to the present invention,
Computer
Input data receiving means for receiving a plurality of feature amounts and a plurality of labels respectively assigned to the feature amounts;
Label distribution calculating means for calculating a probability distribution of a label for each feature amount from a plurality of labels assigned to each feature amount;
Extended learning data generating means for generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
A program for functioning as a server is provided.

  ADVANTAGE OF THE INVENTION According to this invention, the technique which can improve the performance of the learning model used with a recognizer etc. is provided.

It is a block diagram which shows notionally the processing composition of the information processor in a 1st embodiment. It is a figure which shows an example of the probability distribution of the label calculated by the label distribution calculation part. It is a figure which shows notionally the hardware constitutions of information processing apparatus. It is a flowchart figure which shows the flow of a process of the information processing apparatus in 1st Embodiment. It is a block diagram which shows notionally the processing structure of the information processing apparatus in 2nd Embodiment. It is a flowchart which shows the flow of a process of the information processing apparatus in 2nd Embodiment. It is a figure which shows an example of the learning model produced | generated by the information processing apparatus of this invention. It is a figure which shows an example of the learning model produced | generated by the technique of a nonpatent literature.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In all the drawings, the same reference numerals are given to the same components, and the description will be omitted as appropriate.

[First Embodiment]
[Processing configuration]
FIG. 1 is a block diagram conceptually showing the processing configuration of the information processing apparatus 10 in the first embodiment. In FIG. 1, the information processing apparatus 10 includes an input data receiving unit 110, a label distribution calculation unit 120, a teacher label acquisition unit 130, and an extended learning data generation unit 140.

  The input data receiving unit 110 receives, as input data, a plurality of feature amounts and a plurality of labels provided for each feature amount. Here, the “feature amount” means a vector or the like that characterizes input data. The “label” means an evaluation value or an evaluation category according to the feature amount received as input data. The input data receiving unit 110 receives a plurality of feature amounts and a plurality of labels assigned to the feature amounts, for example, via an input device such as a keyboard (not shown). Note that the method in which the input data receiving unit 110 receives a plurality of feature amounts and a plurality of labels given for each feature amount is not limited thereto.

  The label distribution calculation unit 120 calculates the probability distribution of the label for each feature amount from the plurality of labels each received for each feature amount received by the input data reception unit 110. For example, the label distribution calculation unit 120 calculates a probability distribution of labels by a maximum likelihood estimation method assuming a Gaussian distribution or a multinomial distribution. Further, the label distribution calculation unit 120 may calculate a probability distribution of labels assuming a distribution other than a Gaussian distribution or a multinomial distribution. Alternatively, the label distribution calculation unit 120 may calculate the probability distribution of a label using a histogram or the like without assuming a probability distribution function.

  The extended learning data generation unit 140 generates a plurality of learning data (extended learning data) for each feature amount based on the probability distribution of the label for each feature amount calculated by the label distribution calculation unit 120. For example, the extended learning data generation unit 140 according to the present embodiment generates a plurality of learning data based on each feature amount and a plurality of teacher labels for each feature amount acquired by a teacher label acquisition unit 130 described later. Here, “learning data” is data used to generate a learning model, which will be described later, and is data that can indicate the correspondence between a plurality of feature amounts and labels relating to the feature amounts.

  The teacher label acquisition unit 130 acquires a plurality of teacher labels for each feature amount based on the probability distribution of the label for each feature amount. Specifically, there are a plurality of labels that can be acquired as teacher labels, and the probability that the teacher label acquisition unit 130 acquires each label as a teacher label for each feature amount is the probability distribution of the label for each feature amount. Follow. That is, in the probability distribution of the feature-specific labels, a label corresponding to a portion having a high probability density is easily selected as a teacher label. Note that the teacher label acquisition unit 130 receives, for example, an input related to the number of teacher labels to be acquired via an input device such as a keyboard (not shown). Then, the teacher label acquisition unit 130 acquires the input number of teacher labels for each feature amount according to the probability distribution of the labels related to each feature amount.

  The “teacher label” is a representative label given to each feature received by the input data receiving unit 110, and is used when a learning model is generated by support vector regression, a support vector machine, or the like. Information. Here, the teacher label acquired by the teacher label acquisition unit 130 depends on the probability distribution calculated by the label distribution calculation unit 120. This will be described with reference to FIG. FIG. 2 is a diagram illustrating an example of a probability distribution of labels calculated by the label distribution calculation unit 120.

  It can be said that the feature quantity for which the probability distribution of the label as shown in FIG. 2A is calculated is a feature quantity that has a large label variation and is difficult to estimate the label. On the other hand, it can be said that the feature quantity for which the probability distribution of the label as shown in FIG. 2B is calculated is a feature quantity that is easy to estimate the label with little variation in the label. By using such a probability distribution, the teacher label acquisition unit 130 can select a teacher label in consideration of the difficulty in estimating a label for each feature amount.

The teacher label acquisition unit 130 generates a teacher label using, for example, a rejection sampling method described in detail in Non-Patent Document 3 below. In this case, the feature quantity indicating the probability distribution of the label in FIG. 2A is more than the feature quantity indicating the probability distribution of the label in FIG. Variations in labels increase. That is, with respect to the feature quantity indicating the probability distribution of the label as shown in FIG. 2A, the teacher label acquisition unit 130 allows a certain amount of deviation from the average value of the plurality of labels received by the input data reception unit 110. Multiple teacher labels can be obtained. In addition, regarding the feature quantity indicating the probability distribution of the label as shown in FIG. 2B, the teacher label acquisition unit 130 shows the deviation from the average value of the plurality of labels received by the input data reception unit 110 in FIG. It is possible to obtain a plurality of teacher labels smaller than a).
CM Bishop, "Pattern Recognition and Machine Learning", Springer, 2006 (Section 11.1.2)

  Then, the extended learning data generation unit 140 combines each feature amount received by the input data reception unit 110 and each of a plurality of teacher labels acquired for each feature amount, and generates a plurality of learning data for each feature amount. To do. The extended learning data generation unit 140 stores the generated plurality of learning data in the extended learning data storage unit 142. In addition, although the example in which the extended learning data storage unit 142 is provided in the information processing apparatus 10 is shown in FIG. 1, the extended learning data storage unit 142 is connected to the information processing apparatus 10 so as to be communicable. May be provided.

[Hardware configuration]
FIG. 3 is a diagram conceptually illustrating the hardware configuration of the information processing apparatus 10. The information processing apparatus 10 includes, for example, a CPU 11, a memory 12, an input / output interface (I / F) 13, a communication apparatus 14, and the like that are connected to each other via a bus 15. The memory 12 is a RAM (Random Access Memory), a ROM (Read Only Memory), a hard disk, or the like. The communication device 14 communicates with other computers and devices. A portable recording medium or the like can be connected to the communication device 14.

  The input / output I / F 13 can be connected to a user interface device (not shown) such as a display device or an input device. The display device is a device such as an LCD (Liquid Crystal Display) or a CRT (Cathode Ray Tube) display that displays a screen corresponding to drawing data processed by the CPU 11 or a GPU (Graphics Processing Unit) (not shown). The input device is a device that receives input from a user such as a keyboard and a mouse.

  Each processing unit of the information processing apparatus 10 described above is realized, for example, by executing a program stored in the memory 12 by the CPU 11. Further, the program may be installed from a portable recording medium such as a CD (Compact Disc) or a memory card, or from another computer on the network via the communication device 14 and stored in the memory 12. Good. The hardware configuration of the information processing apparatus 10 is not limited to the example illustrated in FIG.

[Operation example]
An operation example of the information processing apparatus 10 according to the present embodiment will be described with reference to FIG. FIG. 4 is a flowchart showing a process flow of the information processing apparatus 10 in the first embodiment.

  The input data receiving unit 110 receives, as input data, a plurality of feature amounts and a plurality of labels assigned to the respective feature amounts (S102). The label distribution calculation unit 120 calculates the probability distribution of the label for each feature amount from the plurality of labels for each feature amount acquired in S102 (S104). For example, the label distribution calculation unit 120 calculates the probability distribution of the label by maximum likelihood estimation assuming a Gaussian distribution, a multinomial distribution, or the like. Then, the teacher label acquisition unit 130 acquires a plurality of teacher labels for each feature amount based on the probability distribution of the label for each feature amount calculated in S104 (S106). Then, the extended learning data generation unit 140 combines each feature quantity received in S102 and each of a plurality of teacher labels acquired for each feature quantity, and obtains a plurality of learning data (extended learning data) for each feature quantity. Generate (S108). Then, the extended learning data generation unit 140 stores the extended learning data generated in S108 in the extended learning data storage unit 142.

[Operation and Effect of First Embodiment]
As described above, in the present embodiment, the probability distribution (label variation) of the label for each feature amount is calculated from the plurality of labels given for each feature amount. Then, based on the calculated probability distribution for each feature amount, a plurality of teacher labels are acquired for each feature amount. Each feature amount is combined with each of a plurality of teacher labels acquired for each feature amount, and a plurality of learning data (extended learning data) is generated for each feature amount.

  As described above, according to the present embodiment, it is possible to generate a plurality of learning data (extended learning data) in consideration of variations in labels given for each feature amount, that is, difficulty in estimating each feature amount. Specifically, for feature quantities with large variations in the labels given, that is, feature quantities that are difficult to estimate, extended learning is allowed by allowing some deviation from the average value of multiple labels given for each feature quantity. Data is generated. In addition, with respect to a feature amount with small variations in labels to be given, that is, a feature amount that can be easily estimated, extended learning data with a small deviation from the average value of a plurality of labels given for each feature amount is generated. Here, when the average value of a plurality of labels input based on human judgment is used as one teacher label, as in the technique of non-patent literature, the relationship between this teacher label and a plurality of labels based on human judgment is used. Learning data having a large degree of deviation can be created. On the other hand, in the present invention, various patterns of learning data (extended learning data) are generated based on the probability distribution of a plurality of labels assigned for each feature amount. The degree of deviation from the label can be reduced. That is, by using the extended learning data generated in the present embodiment, it is possible to improve the estimation accuracy of the learning model.

[Second Embodiment]
In the present embodiment, an extended learning model is generated using the extended learning data generated in the first embodiment, and information processing for determining a label to be added to unknown data using the generated extended learning model The apparatus will be described.

[Processing configuration]
FIG. 5 is a block diagram conceptually showing the processing configuration of the information processing apparatus 10 in the second embodiment. The information processing apparatus 10 according to the present embodiment further includes an extended learning model generation unit 150, a recognition data acquisition unit 160, a label estimation unit 170, and a label determination unit 180.

  The extended learning model generation unit 150 generates a plurality of learning models (extended learning models) based on the plurality of learning data (extended learning data) generated by the extended learning data generation unit 140. Specifically, the extended learning model generation unit 150 generates a plurality of learning models based on a plurality of learning data stored in the extended learning data storage unit 142 using, for example, support vector regression or a support vector machine. The extended learning model generation unit 150 stores the generated plurality of learning models in the extended learning model storage unit 152. 1 shows an example in which the extended learning model storage unit 152 is provided in the information processing device 10, but the extended learning model storage unit 152 is connected to the information processing device 10 so as to be communicable with other devices. May be provided.

  Further, the extended learning model generation unit 150 uniquely selects one learning data for each feature amount from M pieces of learning data (M is an integer of 2 or more) generated for each feature amount, and selects the selected learning data. M learning models (extended learning models) may be generated using the learning data for each feature amount. Specifically, the extended learning model generation unit 150 excludes learning data used when generating other learning models when selecting one learning data for each feature amount in order to generate a learning model. By doing in this way, the extended learning model in which the probability distribution of the label for each feature amount calculated by the label distribution calculation unit 120 is accurately reflected can be generated. Further, as the number of teacher labels acquired by the teacher label acquisition unit 130 increases, the extended learning model generation unit 150 can generate an extended learning model with higher accuracy.

  The recognition data acquisition unit 160 acquires unknown recognition target data by a user input operation or the like.

  The label estimation unit 170 estimates the label of the recognition target data acquired by the recognition data acquisition unit 160 using a plurality of learning models stored in the extended learning model storage unit 152. Here, the label estimation unit 170 estimates at least two labels of recognition target data using at least two of the plurality of learning models stored in the extended learning model storage unit 152. The number of learning models used by the label estimation unit 170 can be specified by, for example, input from the user.

  The label discriminating unit 180 discriminates the label to be given to the recognition target data acquired by the recognition data acquiring unit 160 based on the statistics of the plurality of labels estimated using the plurality of learning models. The label determination unit 180 determines, for example, the average value or intermediate value of a plurality of labels estimated by the label estimation unit 170 as a label to be added to the recognition target data. Further, the label discriminating unit 180 may discriminate the most presumed label as the label to be added to the recognition target data by, for example, taking a majority vote of a plurality of labels estimated by the label estimating unit 170.

[Operation example]
An example of the operation of the information processing apparatus 10 according to this embodiment will be described with reference to FIG. FIG. 6 is a flowchart showing the flow of processing of the information processing apparatus 10 in the second embodiment.

  The recognition data acquisition unit 160 acquires the recognition target data by, for example, a user input operation (S202). Then, the label estimation unit 170 estimates the label of the recognition target data using a plurality of learning models (extended learning models) stored in the extended learning model storage unit 152 (S204). Here, the label estimation unit 170 estimates at least two labels of the recognition target data using at least two learning models among a plurality of learning models stored in the extended learning model storage unit 152. Then, the label determination unit 180 determines the label to be added to the recognition target data from the plurality of labels estimated using the plurality of learning models in S204 (S206). Specifically, the label determination unit 180 determines a label to be added to the recognition target data based on the average value of the label estimated in S204, the intermediate value, or the majority of the estimated labels.

[Operation and Effect of Second Embodiment]
As described above, in the present embodiment, a plurality of learning models (extended learning models) are generated using the learning data of various patterns based on the probability distribution of the plurality of labels given for each feature amount generated in the first embodiment. Generated. Then, the labels estimated by the plurality of learning models are integrated, and the label to be added to the recognition target data is determined. That is, since a label to be given to unknown recognition target data is determined by a learning model reflecting various patterns based on the probability distribution of labels, according to the present embodiment, a label to be given to unknown recognition target data Can be improved.

  Examples will be given below to describe the above-described embodiments in more detail. The present invention is not limited in any way by the following examples.

  The input data receiving unit 110 receives, for example, feature quantities such as prosody and volume of audio data, and feature quantities obtained by performing edge processing on image data. Further, the input data receiving unit 110 receives a plurality of labels given by human judgment corresponding to the received feature amount.

  The label distribution calculating unit 120 calculates the probability distribution of the labels for each feature amount from the plurality of labels for each feature amount received by the input data receiving unit 110. For example, the label distribution calculation unit 120 estimates a distribution by maximum likelihood estimation assuming a Gaussian distribution or a multinomial distribution.

  The teacher label acquisition unit 130 acquires a plurality of (for example, M) teacher labels based on the probability distribution calculated by the label distribution calculation unit 120. Furthermore, the extended learning data generation unit 140 selects and combines one teacher label for each feature amount from among M teacher labels for each feature amount, and generates extended learning data.

  The difference between the learning model generated by the information processing apparatus 10 of the present invention and the learning model generated by the technique of non-patent literature will be described with reference to FIGS. FIG. 7 is a diagram illustrating an example of a learning model generated by the information processing apparatus 10 according to the present invention. Moreover, FIG. 8 is a figure which shows an example of the learning model produced | generated by the technique of a nonpatent literature. Here, a case where three labels (labelers 1 to 3) are provided for each of five feature amounts (feature amounts 1 to 5) is illustrated.

  In the technique of non-patent literature, a regression model as indicated by a broken line in FIG. 8 is learned using five feature amounts and an average value of three labels assigned to the five feature amounts. In the technique of non-patent literature, variations in labels given for each feature amount are not taken into consideration. Therefore, when the label to be added to the recognition target data is estimated using the regression model learned by the technique of non-patent literature, the estimated label is greatly deviated at the feature amount 3 as shown in FIG. there is a possibility.

  On the other hand, in the information processing apparatus 10 of the present invention, as described in the above-described embodiment, a plurality of (in this case, M = 4) learning data generated in consideration of variations in labels given for each feature amount. Based on this, a plurality of regression models (estimated values 1 to 4) as shown in FIG. 7 are learned. Then, by using statistics of a plurality of regression models (for example, average values, intermediate values, etc.), a regression model (integrated estimated value) as shown by a thick broken line in FIG. 7 can be obtained as a result. As a result, according to the plurality of information processing apparatuses 10 of the present invention, a certain amount of deviation from the average value of the given label is allowed for a feature amount (for example, feature amount 2) that has a large variation in labels and is difficult to estimate. With respect to a feature quantity (for example, feature quantity 3) that has a small label variation and is easy to estimate, a learning model is generated that reduces the deviation from the average value of the given label. That is, it is possible to generate a learning model with improved estimation accuracy of unknown recognition target data.

  As mentioned above, although embodiment of this invention was described with reference to drawings, these are the illustrations of this invention, Various structures other than the above are also employable.

  In the plurality of flowcharts used in the above description, a plurality of steps (processes) are described in order, but the execution order of the steps executed in each embodiment is not limited to the description order. In each embodiment, the order of the illustrated steps can be changed within a range that does not hinder the contents. Moreover, each above-mentioned embodiment can be combined in the range in which the content does not conflict.

Hereinafter, examples of the reference form will be added.
1. Input data receiving means for receiving a plurality of feature quantities and a plurality of labels respectively assigned to the feature quantities;
Label distribution calculating means for calculating a probability distribution of a label for each feature quantity from a plurality of labels assigned to each feature quantity;
Extended learning data generating means for generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
An information processing apparatus.
2. Based on the probability distribution of the label for each feature amount, further comprising a teacher label acquisition means for acquiring a plurality of teacher labels for each feature amount,
The extended learning data generation means combines the feature amount and each of a plurality of teacher labels acquired for the feature amount, and generates a plurality of learning data for each feature amount.
1. The information processing apparatus described in 1.
3. An extended learning model generating means for generating a plurality of learning models based on the plurality of learning data generated for each feature amount;
1. Or 2. The information processing apparatus described in 1.
4). The extended learning model generation means includes
One learning data is uniquely selected for each feature amount from among M pieces of learning data (M is an integer of 2 or more) generated for each feature amount, and learning data for each selected feature amount is selected. To generate M learning models,
3. The information processing apparatus described in 1.
5. Recognition data acquisition means for acquiring unknown recognition target data;
Label estimation means for estimating a label of the recognition target data using the plurality of learning models;
Further comprising label discriminating means for discriminating a label to be given to the recognition target data based on statistics of a plurality of labels estimated using the plurality of learning models.
3. Or 4. The information processing apparatus described in 1.
6). Computer
Receiving a plurality of feature amounts and a plurality of labels assigned to each feature amount,
From the plurality of labels assigned to each feature amount, calculate the probability distribution of the label for each feature amount,
Generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
Information processing method including the above.
7). The computer is
Based on the probability distribution of the label for each feature amount, obtain a plurality of teacher labels for each feature amount,
Combining the feature quantity and each of a plurality of teacher labels acquired for the feature quantity, and generating a plurality of learning data for each feature quantity;
Including. Information processing method described in 1.
8). The computer is
Generating a plurality of learning models based on the plurality of learning data generated for each feature amount;
Including. Or 7. Information processing method described in 1.
9. The computer is
One learning data is uniquely selected for each feature amount from among M pieces of learning data (M is an integer of 2 or more) generated for each feature amount, and learning data for each selected feature amount is selected. To generate M learning models,
Including Information processing method described in 1.
10. The computer is
Get unknown recognition target data,
Estimating a label of the recognition target data using the plurality of learning models,
Determining a label to be given to the recognition target data based on statistics of a plurality of labels estimated using the plurality of learning models;
Including Or 9. Information processing method described in 1.
11. Computer
Input data receiving means for receiving a plurality of feature amounts and a plurality of labels respectively assigned to the feature amounts;
Label distribution calculating means for calculating a probability distribution of a label for each feature amount from a plurality of labels assigned to each feature amount;
Extended learning data generating means for generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
Program to function as.
12 The computer,
Based on the probability distribution of the label for each feature amount, further function as a teacher label acquisition means for acquiring a plurality of teacher labels for each feature amount,
In the computer,
In the extended learning data generating means, the feature amount and each of a plurality of teacher labels acquired for the feature amount are combined, and a procedure for generating a plurality of learning data for each feature amount is executed.
11. The program described in.
13. The computer,
Extended learning model generation means for generating a plurality of learning models based on the plurality of learning data generated for each feature amount;
To further function as Or 12. The program described in.
14 In the computer,
In the extended learning model generation means,
One learning data is uniquely selected for each feature amount from among M pieces of learning data (M is an integer of 2 or more) generated for each feature amount, and learning data for each selected feature amount is selected. To execute the procedure of generating M learning models using
13. The program described in.
15. The computer,
Recognition data acquisition means for acquiring unknown recognition target data;
Label estimation means for estimating a label of the recognition target data using the plurality of learning models,
Label discriminating means for discriminating a label to be given to the recognition target data based on statistics of a plurality of labels estimated using the plurality of learning models;
13 for further functioning as Or 14. The program described in.

10 Information processing apparatus 11 CPU
12 Memory 13 Input / output I / F
14 Communication device 15 Bus 110 Input data reception unit 120 Label distribution calculation unit 130 Teacher label acquisition unit 140 Extended learning data generation unit 142 Extended learning data storage unit 150 Extended learning model generation unit 152 Extended learning model storage unit 160 Recognition data acquisition unit 170 Label estimation unit 180 Label discrimination unit

Claims (7)

Input data receiving means for receiving a plurality of feature quantities and a plurality of labels respectively assigned to the feature quantities;
Label distribution calculating means for calculating a probability distribution of a label for each feature quantity from a plurality of labels assigned to each feature quantity;
Extended learning data generating means for generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
An information processing apparatus. Based on the probability distribution of the label for each feature amount, further comprising a teacher label acquisition means for acquiring a plurality of teacher labels for each feature amount,
The extended learning data generation means combines the feature amount and each of a plurality of teacher labels acquired for the feature amount, and generates a plurality of learning data for each feature amount.
The information processing apparatus according to claim 1. An extended learning model generating means for generating a plurality of learning models based on the plurality of learning data generated for each feature amount;
The information processing apparatus according to claim 1 or 2. The extended learning model generation means includes
One learning data is uniquely selected for each feature amount from among M pieces of learning data (M is an integer of 2 or more) generated for each feature amount, and learning data for each selected feature amount is selected. To generate M learning models,
The information processing apparatus according to claim 3. Recognition data acquisition means for acquiring unknown recognition target data;
Label estimation means for estimating a label of the recognition target data using the plurality of learning models;
Further comprising label discriminating means for discriminating a label to be given to the recognition target data based on statistics of a plurality of labels estimated using the plurality of learning models.
The information processing apparatus according to claim 3 or 4. Computer
Receiving a plurality of feature amounts and a plurality of labels assigned to each feature amount,
From the plurality of labels assigned to each feature amount, calculate the probability distribution of the label for each feature amount,
Generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
Information processing method including the above. Computer
Input data receiving means for receiving a plurality of feature amounts and a plurality of labels respectively assigned to the feature amounts;
Label distribution calculating means for calculating a probability distribution of a label for each feature amount from a plurality of labels assigned to each feature amount;
Extended learning data generating means for generating a plurality of learning data for each feature amount based on the feature amount and the probability distribution of the label for each feature amount;
Program to function as.

Images