Wonjun Lee
University of Minnesota
Title:
Geometry-Preserving Encoder/Decoder In Latent Generative Models
Abstract:
Generative modeling seeks to create new data samples that resemble a given dataset, with diffusion models recently gaining prominence. A key challenge in diffusion models is addressing the high-dimensional input space. To improve efficiency, recent approaches solve diffusion models in a lower-dimensional latent space using an encoder. The variational autoencoder (VAE) is the most common framework in this area, known for learning latent representations and generating samples. In this talk, I will talk about a novel encoder/decoder framework that preserves the geometric structure of the data distribution, offering distinct theoretical advantages over the VAE. I will then demonstrate how this geometry-preserving encoder enhances both encoder and decoder training, providing theoretical guarantees for the convergence of the training process, including faster convergence in decoder training.
