Toward Generalizable Deep Generative Models

This thesis studies the generalization and convergence properties of Deep Generative Models. The thesis analyzes the generalization and convergence of Deep Generative Models using both empirical and theoretical methods. Based on the analysis, a number of methods for improving the generalization and stability of Deep Generative Models were developed. The last part of this thesis studies evaluation metrics for Deep Generative Models and introduces a novel metric for generalization based on the Minimum Description Length principle.