Variable auto-encoder concept embedding method, device and equipment based on cloud control

Abstract

The embodiment of the invention provides a depth variational auto-encoder training method, device and equipment based on cloud control. The method is applied to the technical field of machine learning artificial intelligence and data analysis. The method comprises the steps of obtaining input data and performing preprocessing such as normalization; inputting the preprocessed data into an auto-encoder coding network, and outputting Gaussian cloud concept parameters of potential variables; a data set is input into a variational auto-encoder to obtain digital characteristics of data, and hidden variables are generated through two times of re-parameter sampling. Correcting hidden variables by a Gaussian cloud concept obtained by pre-training the model; meanwhile, a cloud control generator is used for conducting concept constraint and optimization on the potential space according to the relation between the hidden variables and the concept Gaussian clouds; inputting the hidden variables into a decoding network to obtain a reconstruction result; and jointly updating model parameters according to the reconstruction error and the regularization loss until the training is completed. According to the method, flexible modeling and semantic embedding of concepts in a potential space are realized by introducing the expression capability of the cloud model for data uncertainty, and the interpretability, clustering performance and expression capability for data distribution of the model are improved.

Images