Method of extracting unsuitable and defective data from plurality of pieces of training data used for learning of machine learning model, information processing device, and non-transitory computer-readable storage medium storing computer program

1 . A method comprising:

executing a process of extracting unsuitable and defective data from a plurality of pieces of training data, wherein

the plurality of pieces of training data is used for learning of a machine learning model for classifying input data into a plurality of classes, and

the machine learning model is configured as a vector neural network having a plurality of vector neuron layers;

inputting each of the plurality of pieces of training data into the machine learning model that is previously learned;

obtaining a feature spectrum from an output of a specific layer of the machine learning model;

classifying, into the plurality of classes, the feature spectrum corresponding respectively to the plurality of pieces of training data;

selecting target training data from the plurality of pieces of training data;

selecting a reference class from the plurality of classes;

calculating a plurality of degrees of similarity between the feature spectrum corresponding to the target training data and a plurality of feature spectra belonging to the reference class;

applying, to the plurality of degrees of similarity, a defectiveness function that is determined in advance;

calculating a defectiveness index with respect to the target training data; and

determining whether the target training data is the defective data, based on a result of comparison between the defectiveness index and a threshold value, wherein

the defectiveness function is a function for obtaining, as the defectiveness index, a representative value in a histogram of the plurality of degrees of similarity.

2 . The method according to claim 1 , wherein the application of the defectiveness function includes:

segmenting the histogram of the plurality of degrees of similarity into one or more unimodal distributions; and

obtaining, as the defectiveness index, a representative value in a representative unimodal distribution that is selected from the one or more unimodal distributions in accordance with a selection condition that is determined in advance.

3 . The method according to claim 2 , wherein the selection condition includes:

a first condition that a ratio of one unimodal distribution area to an entire area of the histogram is equal to or greater than an area threshold value; and

a second condition that, in the unimodal distribution satisfying the first condition, an average value of the plurality of degrees of similarity is the greatest.

4 . The method according to claim 1 , wherein

the defective data includes outlier data,

the reference class is a class corresponding to a target class to which the target training data belongs, and

the determination whether the target training data is the defective data includes:

determining the target training data is the outlier data when the defectiveness index is equal to or less than the threshold value; and

determining the target training data is not the outlier data when the defectiveness index exceeds the threshold value.

5 . The method according to claim 1 , wherein

the defective data includes overlap data approximating to training data in another class different from a class to which the defective data belongs,

the reference class is a class different from a target class to which the target training data belongs, and

the determination whether the target training data is the defective data includes:

determining the target training data is the overlap data when the defectiveness index is equal to or greater than the threshold value; and

determining the target training data is not the overlap data when the defectiveness index is less than the threshold value.

6 . The method according to claim 1 , wherein

the specific layer has a configuration in which a vector neuron arranged in a plane defined with two axes including a first axis and a second axis is arranged as a plurality of channels along a third axis being a direction different from the two axes, and

the feature spectrum is any one of:

a first type of the feature spectrum obtained by arranging a plurality of element values of an output vector of the vector neuron at one plane position in the specific layer, over the plurality of channels along the third axis;

a second type of the feature spectrum obtained by multiplying each of the plurality of element values of the first type of the feature spectrum by an activation value corresponding to a vector length of the output vector; and

a third type of the feature spectrum obtained by arranging the activation value at the one plane position in the specific layer, over the plurality of channels along the third axis.

7 . A non-transitory computer-readable medium having stored thereon, computer-executable instructions which, when executed by computer, cause the computer to execute operations, the operations comprising:

executing a process of extracting unsuitable and defective data from a plurality of pieces of training data, wherein

the plurality of pieces of training data is used for learning of a machine learning model for classifying input data into a plurality of classes; and

the machine learning model is configured as a vector neural network having a plurality of vector neuron layers;

inputting each of the plurality of pieces of training data into the machine learning model that is previously learned;

obtaining a feature spectrum from an output of a specific layer of the machine learning model;

classifying, into the plurality of classes, the feature spectrum corresponding respectively to the plurality of pieces of training data;

selecting target training data from the plurality of pieces of training data;

selecting a reference class from the plurality of classes;

calculating a plurality of degrees of similarity between the feature spectrum corresponding to the target training data and a plurality of feature spectra belonging to the reference class;

applying, to the plurality of degrees of similarity, a defectiveness function that is determined in advance;

calculating a defectiveness index with respect to the target training data; and

determining whether the target training data is the defective data, based on a result of comparison between the defectiveness index and a threshold value, wherein

the defectiveness function is a function for obtaining, as the defectiveness index, a representative value in a histogram of the plurality of degrees of similarity.

8 . A method, comprising:

executing a process of extracting unsuitable and defective data from a plurality of pieces of training data, wherein

the plurality of pieces of training data is used for learning of a machine learning model for classifying input data into a plurality of classes, and

the machine learning model is configured as a vector neural network having a plurality of vector neuron layers;

inputting each of the plurality of pieces of training data into the machine learning model that is previously learned;

obtaining a feature spectrum from an output of a specific layer of the machine learning model, wherein

the specific layer has a configuration in which a vector neuron is arranged in a plane defined with two axes including a first axis and a second axis,

the specific layer is arranged as a plurality of channels along a third axis which is a direction different from the two axes, and

the feature spectrum is any one of:

a first type of the feature spectrum obtained by arranging a plurality of element values of an output vector of the vector neuron at one plane position in the specific layer, over the plurality of channels along the third axis;

a second type of the feature spectrum obtained by multiplying each of the plurality of element values of the first type of the feature spectrum by an activation value corresponding to a vector length of the output vector; and

a third type of the feature spectrum obtained by arranging the activation value at the one plane position in the specific layer, over the plurality of channels along the third axis;

classifying, into the plurality of classes, the feature spectrum corresponding respectively to the plurality of pieces of training data;

selecting target training data from the plurality of pieces of training data;

selecting a reference class from the plurality of classes;

calculating a plurality of degrees of similarity between the feature spectrum corresponding to the target training data and a plurality of feature spectra belonging to the reference class;

applying, to the plurality of degrees of similarity, a defectiveness function that is determined in advance;

calculating a defectiveness index with respect to the target training data; and

determining whether the target training data is the defective data, based on a result of comparison between the defectiveness index and a threshold value.

9 . The method according to claim 8 , wherein the defectiveness function is a function for obtaining, as the defectiveness index, a statistic representative value of the plurality of degrees of similarity.

10 . The method according to claim 8 , wherein the defectiveness function is a function for obtaining, as the defectiveness index, an average value or a maximum value of the plurality of degrees of similarity.

11 . The method according to claim 8 , wherein the defectiveness function is a function for obtaining, as the defectiveness index, a representative value in a histogram of the plurality of degrees of similarity.

12 . The method according to claim 8 , wherein the application of the defectiveness function includes:

segmenting a histogram of the plurality of degrees of similarity into one or more unimodal distributions; and

obtaining, as the defectiveness index, a representative value in a representative unimodal distribution that is selected from the one or more unimodal distributions in accordance with a selection condition that is determined in advance.

13 . The method according to claim 12 , wherein the selection condition includes:

a first condition that a ratio of one unimodal distribution area to an entire area of the histogram is equal to or greater than an area threshold value; and

a second condition that, in the unimodal distribution satisfying the first condition, an average value of the plurality of degrees of similarity is the greatest.

14 . The method according to claim 8 , wherein

the defective data includes outlier data,

the reference class is a class corresponding to a target class to which the target training data belongs, and

the determination whether the target training data is the defective data includes:

determining the target training data is the outlier data when the defectiveness index is equal to or less than the threshold value; and

determining the target training data is not the outlier data when the defectiveness index exceeds the threshold value.

15 . The method according to claim 8 , wherein

the defective data includes overlap data approximating to training data in another class different from a class to which the defective data belongs,

the reference class is a class different from a target class to which the target training data belongs, and

the determination whether the target training data is the defective data includes:

determining the target training data is the overlap data when the defectiveness index is equal to or greater than the threshold value; and

determining the target training data is not the overlap data when the defectiveness index is less than the threshold value.