| CPC G06N 20/00 (2019.01) [G06F 18/217 (2023.01); G06F 18/2321 (2023.01); G06N 3/02 (2013.01)] | 7 Claims |
|
1. A data conversion learning apparatus comprising:
at least one memory configured to store instructions; and
at least one processor configured to execute the instructions to:
perform data conversion for bringing a probability distribution of source data of a source domain and a probability distribution of target data of a target domain close to each other, the data conversion yielding converted data;
acquire information of a non-appearing class among a plurality of classes;
calculate a source domain certainty factor indicating certainty that the converted data belongs to the source domain and a target domain certainty factor indicating certainty that the converted data belongs to the target domain;
calculate a class certainty factor indicating certainty that the converted data is classified into each of the plurality of classes into which the source data is classified, wherein the plurality of classes include the non-appearing class and a first class other than the non-appearing class, and wherein the first class appears in the source domain and the target domain and the non-appearing class does not appear in the target domain;
calculate a first loss related to the non-appearing class based on the source domain certainty factor and the target domain certainty factor;
calculate a second loss related to the non-appearing class, based on the class certainty factor;
execute neural network operations to calculate the source domain certainty factor and the target domain certainty factor by evaluating source and target domain probabilities [p(d=S|x), p(d=T|x)] based on the converted data using a first neural network for domain identification [NN_d(z|θ_d)], and to calculate the class certainty factor by evaluating class probabilities [p(y=0|x), p(y=1|x)] based on the converted data using a second neural network for class identification [NN_c(z|θ_c)];
calculate the first loss using a weighted function of the source domain certainty factor and the target domain certainty factor, the weight being a ratio wd(x)=p(x|d=T, y=0)/p(x|d=S, y=0), and calculate the second loss using a weighted function of the class certainty factor, the weight being a ratio wc(x)=p(x|d=T, y=0)/p(x|d=T, y=1);
perform machine learning on the class certainty factor using the first loss; and
perform machine learning on the source domain certainty factor and the target domain certainty factor using the second loss,
wherein the variable x indicates data before the data conversion, the variable y indicates the class, the variable z indicates the converted data, the variable d indicates domain information, the subscript c indicates class information, the variable θ_d indicates a parameter of the first neural network, and the variable θ_c indicates a parameter of the second neural network,
wherein using the first loss calculated using the source domain certainty factor and the target domain certainty factor includes calculating the first loss related to the target domain by a ratio of a first probability to a second probability, the first probability being a probability that the data of the target domain that has undergone the data conversion is classified as the non-appearing class that does not appear in the target domain, the second probability being a probability that the data of the source domain that has undergone the data conversion is classified as the non-appearing class that does not appear in the source domain, and
wherein using the second loss calculated using the second class certainty factor includes calculating the second loss related to the target domain by a ratio of a third probability to a fourth probability, the third probability being a probability that the data of the target domain that has undergone the data conversion is classified as the non-appearing class that does not appear in the target domain, the fourth probability being a probability that the data of the target domain that has undergone the data conversion is classified as an appearing class that appears in the target domain.
|