Classification Of Supervisory Side-information And Research On Optimization Methods Of Constraint Information In Generative Adversarial Networks

Generative Adversarial Networks(GAN)are generative models that model the distribution of target data in a manner similar to a two-person zero-sum game.The network has received a lot of attention since its introduction,and a large number of related improvements have also appeared.In generation tasks,there are often scenarios in which,in addition to the real data,some supervisory side-information of the data can be obtained by a predetermined mechanism or an external program,and the supervisory information may be instructive for the generation task.In recent years,many GAN models using supervisory side-information have been proposed.However,most of these work are for specific types of supervisory sideinformation,and there is no work to classify and summarize the scenario where supervisory side-information exists.Besides,no one has conducted in-depth research on the scenarios where the supervisory side-information can be seen as constraints.It can bring convenience to the application of GAN models if the aforementioned scenarios can be properly classified and deeply studied.Two classification methods are proposed in this paper to classify the scenarios where supervisory side-information exists in the application of GAN models.The first classification method is based on the types of supervisory side-information.Using this classification method,the supervisory side-information can be classified into three categories: constraint information,numerical information and other information.The second classification method is based on the certainty of supervisory side-information,by which the supervisory side-information can be classified into three categories: explicit information,partially explicit information,and implicit information.For supervisory side-information which can be seen as constraints,by applying the penalty function method of the numerical optimization field to the optimization method for the GAN models,a penalty function-based optimization method is proposed in this paper for GAN models.For clearer comparison and analysis,the proposed method and other optimization methods for the constraint information in GAN tasks are studied together.In order to verify the effectiveness of the above methods,experiments on both synthetic data and real data are conducted.The experimental results show that the proposed method can improve the proportion of data that satisfies the constraints and the mode coverage of the target distribution compared with other types of methods,and will not cause obvious damage to the original model.