Oscar Leong Institute: University of California, Los Angeles Date: September 19, 2024 Title: Generative networks for inverse problems without ground-truth data Abstract: Overcoming ill-posedness in inverse problems is a fundamental challenge in imaging, often requiring prior models to constrain the solution space. Traditional approaches rely on hand-crafted prior models, where parameters are fine-tuned through trial and error, a process that is both time-consuming and susceptible to human bias. Alternatively, machine learning methods attempt to learn image priors directly from clean data, which can be expensive or impossible to obtain in practice. In this talk, I will present our recent work on learning image priors directly from noisy measurements. Leveraging tools from variational inference and deep generative networks, we demonstrate that when the ground-truth images share common, low-dimensional structure, the noisy measurements themselves contain sufficient information to learn an approximate image prior, enabling us to solve the underlying inverse problem. We illustrate our framework on a range of inverse problems, from denoising to compressed sensing problems inspired by black-hole imaging.
