Search papers, labs, and topics across Lattice.
This paper addresses the limitations of current conditioned generative models that depend on score functions, particularly in scenarios where the diffusion process has a singular coefficient and the conditional densities are either non-existent or non-smooth. By employing causal optimal transport, the authors define approximate loss functions that facilitate a minimum-entropy control for model guidance with minimal assumptions. The key result shows that this method effectively characterizes the predictable representation property of conditioned diffusion processes, leading to improved generative performance under challenging conditions.
Causal optimal transport reveals a new pathway for guiding degenerate diffusion models, even when traditional score functions fail.
Current conditioned generative models heavily rely on score functions for guidance during training. When the generative model is a diffusion process with a singular diffusion coefficient and the underlying (conditional) densities either do not exist or are not smooth, we use causal optimal transport to define \emph{approximate} loss functions that identify a minimum-entropy control for guidance under minimal assumptions. Our approach relies on causal optimal transport and its characterization through the predictable representation property of (conditioned) diffusion processes whose associated martingale problem is well posed, \`a la \"Ust\"unel.
