SYSTEM AND METHOD FOR SEMI-SUPERVISED CONDITIONAL GENERATION MODELING USING HOSTILE NETWORK

摘要

To make it easy to generate a composite object using a semi-supervised GAN (Generative adversarial network).SOLUTION: In operation, a generator module 162 compound a noise vector and a data object derived from an attribute label. An unsupervised discriminator module 164 calculates a value indicative of a probability that the data object and training objects obtained from a training data set are genuine, and a latent feature expression of the data object. A supervised discriminator module 165 calculates, on receiving the latent feature expression and the attribute label, a value indicative of a probability that the attribute label given the data object is genuine, and repeats the process until a data object which cannot be identified as a forgery is generated.SELECTED DRAWING: Figure 1D