Machine learning apparatus, machine learning system, machine learning method, and program

A machine learning system, through a network, connecting:

• an influence function calculation apparatus comprising:
• at least one first processor;
• and
• memory storing first instructions,
• wherein the first instructions, when executed by the at least one first processor cause the influence function calculation apparatus to calculate an influence function indicating a level of sensitivity that an input data point has on parameters of a target model stored in a storage;
• and
• a target model training apparatus comprising:
• at least one second processor;
• and
• memory storing second instructions,
• wherein the second instructions, when executed by the at least one second processor, cause the target model training apparatus to:
• receive a training set of labeled data points;
• and
• train the target model stored on the storage based on the training set, wherein the training comprises, for an input data point:
• calculating a regularization gradient based on a result of the influence function received from the influence function calculation apparatus;
• updating the parameters of the target model based on the regularization gradient;
• and
• storing the target model with the updated parameters on the storage,
• wherein the influence function is defined by the expression
• If(x,x)=θ−x−θ
• where x represents the input data point, θ represents the values of the parameters in the trained target model when the input data point is included in the training, and θ−x represents the values of the parameters in the trained target model when the input data point is excluded from the training.

A non-transient computer-readable recording media storing a program that, when executed by at least one processor of a computer causes the at least one processor to perform:

• receiving a training set of labeled data points;
• and
• training a target model stored on a storage based on the training set, wherein the training comprises, for an input data point:
• calculating an influence function indicating a level of sensitivity that the input data point has on parameters of the target model;
• calculating a regularization gradient based on a result of the influence function;
• updating the parameters of the target model based on the regularization gradient;
• and
• storing the target model with the updated parameters on the storage,
• wherein the influence function is defined by the expression
• If(x,x)=θ−x−θ
• where x represents the input data point, θ represents the values of the parameters in the trained target model when the input data point is included in the training, and θ−x represents the values of the parameters in the trained target model when the input data point is excluded from the training.