Hyperparameter management device, hyperparameter management system, and hyperparameter management method

A hyperparameter management system comprising:

• a hyperparameter management server;
• and
• a client terminal,
• wherein:
• the hyperparameter management server and the client terminal are connected via a communication network,
• the hyperparameter management server includes:
• a compression network configured to dimensionally compress data to produce a low dimensional representation,
• an estimation network configured to estimate a density distribution of the low dimensional representation,
• a hyperparameter calculation unit configured to calculate hyperparameters for the compression network and the estimation network, and
• a determination unit implemented by a processor that is configured to determine a specificity of data;
• and in a case that subject data to be analyzed is received from the client terminal,
• the hyperparameter calculation unit is implemented by a processor that is configured to:
• calculate, based on a set of subject data and a set of uniform data, a gap statistic using a gap statistic calculation technique, and calculate, using a curve fitting technique, an estimation network hyperparameter based on at least the gap statistic,
• calculate a ratio of variance of principal components of the set of subject data, and calculate, using a curve fitting technique, a compression network hyperparameter based on at least the ratio of variance, and
• set the estimation network hyperparameter in the estimation network and set the compression network hyperparameter in the compression network;
• the compression network is configured to:
• dimensionally compress the set of subject data, generate a subject data low dimensional representation that corresponds to the set of subject data, and generate a set of reconstructed data based on the subject data low dimensional representation, and
• calculate, based on the set of reconstructed data and the set of subject data, an error of the set of reconstructed data with respect to the set of subject data;
• the estimation network is configured to:
• estimate, based on the subject data low dimensional representation and the error, a density distribution of the set of subject data;
• and
• the determination unit is configured to:
• determine a specificity of the set of subject data based on the density distribution and provide notification of a result of this determination to the client.

A hyperparameter management method executed by a hyperparameter management device,

• the hyperparameter management device comprising:
• a compression network configured to dimensionally compress data to produce a low dimensional representation;
• an estimation network configured to estimate a density distribution of the low dimensional representation;
• and
• a hyperparameter calculation unit implemented by a processor that is configured to calculate hyperparameters for the compression network and the estimation network,
• wherein the hyperparameter management method includes:
• a step of calculating, based on a set of subject data and a set of uniform data, a gap statistic using a gap statistic calculation technique, and calculating, using a curve fitting technique, an estimation network hyperparameter based on at least the gap statistic,
• a step of calculating a ratio of variance of principal components of the set of subject data, and calculating, using a curve fitting technique, a compression network hyperparameter based on at least the ratio of variance, and
• a step of setting the estimation network hyperparameter in the estimation network and setting the compression network hyperparameter in the compression network.