Training apparatus, training method and recording medium

1. A training apparatus for training an inference model performing a predetermined inference that infers, using the inference model, a class to which a state of a target belongs among a plurality of classes in a predefined classification space, the inference model being built using training data, the training apparatus comprising:

a memory to store a plurality of pieces of data, and

processing circuitry connected to the memory and configured to

read the plurality of pieces of data;

calculate a probability that the state of the target belongs to each class to which the state of the target belongs by executing the predetermined inference for the plurality of pieces of data;

calculate, for each piece of data, a score corresponding to a difference in the probability with respect to each identification boundary for the plurality of classes, the score being a value corresponding to a closeness of a position of each piece of data in the predefined classification space to the identification boundary;

change the score of a piece of data other than a selected piece of data selected based on the scores among the plurality of pieces of data, such that a distance from an identification boundary identical to an identification boundary related to the selected piece of data becomes larger, according to a particular distance for the selected piece of data in the predefined classification space,

give labels in the predefined classification space to pieces of data including the piece of data, the score of which is changed;

add, to the training data, the pieces of data to which the labels are given; and

update the inference model using the training data to which the pieces of data, to which the labels are given, has been added.

2. The training apparatus according to claim 1 , wherein the processing circuitry is further configured to add, to the training data, a piece of data, the predefined score of which is equal to or below a predetermined value.

3. The training apparatus according to claim 1 , wherein the processing circuitry is further configured to output distribution display data to display a distribution of the scores on a display.

4. The training apparatus according to claim 1 , wherein the training data used by the processing circuitry to update the inference model is data for deep training.

5. A training method for training, using a computer, an inference model performing a predetermined inference that infers, using the inference model, a class to which a state of a target belongs among a plurality of classes in a predefined classification space, the inference model being built using training data, the training method comprising:

reading a plurality of pieces of data from a memory storing the plurality of pieces of data;

calculating a probability that the state of the target belongs to each class to which the state of the target belongs by executing the predetermined inference for the plurality of pieces of data;

calculating, for each piece of data, a score corresponding to a difference in the probability with respect to each identification boundary for the plurality of classes, the score being a value corresponding to a closeness of a position of each piece of data in the predefined classification space to the identification boundary;

changing the score of a piece of data other than a selected piece of data selected based on the scores among the plurality of pieces of data, such that a distance from an identification boundary identical to an identification boundary related to the selected piece of data becomes larger, according to a particular distance for the selected piece of data in the predefined classification space;

giving labels in the predefined classification space to pieces of data including the piece of data the score of which is changed;

adding, to the training data, the pieces of data to which the labels are given; and

updating the inference model using the training data to which the pieces of data, to which the labels are given, has been added.

6. A non-transitory computer-readable recording medium in which a program is recorded, the program being for training an inference model performing a predetermined inference that infers, using the inference model, a class to which a state of a target belongs among a plurality of classes in a predefined classification space, the inference model being built using training data, the program causing a computer to execute a method comprising:

reading a plurality of pieces of data from a memory storing the plurality of pieces of data;

calculating a probability that the state of the target belongs to each class to which the state of the target belongs by executing the predetermined inference for the plurality of pieces of data;

calculating, for each piece of data, a score corresponding to a difference in the probability with respect to each identification boundary for the plurality of classes, the score being a value corresponding to a closeness of a position of each piece of data in the predefined classification space to the identification boundary;

changing the score of a piece of data other than a selected piece of data selected based on the scores among the plurality of pieces of data, such that a distance from an identification boundary identical to an identification boundary related to the selected piece of data becomes larger, according to a particular distance for the selected piece of data in the predefined classification space;

giving labels in the predefined classification space to pieces of data including the piece of data the score of which is changed;

adding, to the training data, the pieces of data to which the labels are given; and

updating the inference model using the training data to which the pieces of data, to which the labels are given, has been added.