Learning Dynamical Transport without Data

Algorithms based on dynamical transport of measure, such as score-based diffusion models, have resulted in great progress in the field of generative modeling. However, these algorithms rely on access to an abundance of data from the target distribution. A complementary problem to this is learning to generate samples from a target distribution when only given query access to the unnormalized log-likelihood or energy function associated to it, with myriad application in statistical physics, chemistry, and Bayesian inference. I will present an algorithm based on dynamical transport to sample from a target distribution in this context, which can be seen as an augmentation of annealed importance sampling and sequential Monte Carlo. Time permitting, I will also discuss how to generalize these ideas to dynamics of discrete distributions. This is joint work with Eric Vanden-Eijnden, Peter Holderrieth, and Tommi Jaakkola.

In person or on Zoom:

https://harvard.zoom.us/j/92220006185?pwd=V3mrb4cNSbgRXtNJtRJkTvWFVhmbI5.1
Password: cmsa