Abstract 

Bayesian optimization in the latent space of generative models has become a powerful method for exploring structured search spaces, such as molecular design. Instead of directly optimizing over complex, high-dimensional discrete structures, this approach maps structured inputs into a fixed-size latent space. Within this space, standard surrogate models and gradient-based optimization routines can be effectively applied. However, because these generative models are not trained with specific downstream tasks in mind, latent space optimization can exhibit problematic behaviours. In this talk, we explore an alternative method that decouples the surrogate and generative models. Instead of tightly integrating Gaussian Processes with Variational Autoencoders, our method trains these models separately and combines them through a simple Bayesian update. The aim is to develop a sampling strategy that efficiently identifies candidate structures with a high likelihood of achieving the target objective.

Notes

Personal website can be found here.