Computational Mathematics and Statistics Seminar by Deqian Kong: Top-down Latent Space Generative Models for Optimization and Planning
Speaker:
Deqian Kong, Ph.D. candidate University of California Los Angeles
Title:
Top-down Latent Space Generative Models for Optimization and Planning
Abstract:
In this talk, we explore recent developments in latent space generative models, with a focus on our contributions to latent space energy-based models and latent plan transformers. We compare these models to current prevalent methods such as diffusion models and causal transformers, highlighting how our approach aims to offer explicit abstractions for improved generalization, planning, and online optimization.
We will discuss two main applications of our work: the adaptation of offline reinforcement learning for planning purposes, and the use of our models in online optimization, particularly with a focus on molecule design and drug discovery. These applications are representative examples of how latent space models can be applied to complex problems, offering potential pathways for further research and exploration in these areas.
We will discuss two main applications of our work: the adaptation of offline reinforcement learning for planning purposes, and the use of our models in online optimization, particularly with a focus on molecule design and drug discovery. These applications are representative examples of how latent space models can be applied to complex problems, offering potential pathways for further research and exploration in these areas.
Computational Mathematics and Statistics Seminar