Posted October 17, 2019
Manifold-Structured Latent Space for Deep Generative Modeling
Summary

Exploring the manifold structure of the latent space by using multiple charts and their transition functions, a novel enhancement of current generative models whose latent space is modeled with a single-chart structure.

Key Findings

The proposed method will more accurately capture the ;true intrinsic geometry of the latent space, and advance the effectiveness of generative models. Moreover, it offers a deeper understanding of the structure of the data in both the original and the latent space.

Project Start Date
Project End Date