Hai-Dang Dau

Sequential Monte Carlo for conditional simulation in generative diffusion models

Generative diffusion models work by running a Markov process to turn the data into white noise, learning the reversal using a neural network, and finally running the reverse process to generate new data points. They gain significant attention recently for superior performance in practical applications. After an introduction to this class of algorithms, we focus on the problem of conditional simulation, i.e., the generation of some data coordinates given some others, where only a sample from the joint distribution is observed. We present a novel procedure described in [1], where Sequential Monte Carlo techniques are used to guide the backward dynamics towards the desired target.