- Main
Learning and Sampling in Neural Generative Models: Energy-Based Estimation, Diffusion Sampling, and Beyond
- Yu, Peiyu
- Advisor(s): Wu, Ying Nian;
- Cheng, Guang
Abstract
Deep generative models represent complex data distributions using neural networks, but expressive probability models can be difficult to learn and sample from. This dissertation investigates two coupled challenges: constructing useful estimators for neural generative distributions, and improving sampling procedures that depend on or make use of learned models. The first part studies learning for energy-based generative models, including the relation between scaled noise-contrastive estimation and maximum likelihood, and the use of diffusion-amortized Markov chain Monte Carlo for learning latent energy-based priors. The second part studies optimization of diffusion sampling through multi-denoiser guidance and instance-dependent denoising schedules learned by policy-gradient methods with variance reduction. The final part develops structured generative modeling approaches to unsupervised foreground extraction and interpretable text representation learning. Together, these studies highlight that probabilistic representation, statistical estimation, and generation algorithms are interdependent choices. They present complementary methods for improving the learnability and practical usefulness of neural generative models.