Reproduce “On the Posterior Distribution in Denoising: Application to Uncertainty Quantification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hila Manor, Tomer Michaeli
Researchers from Technion – Israel Institute of Technology established novel theoretical relations connecting higher-order central posterior moments in Gaussian denoising to derivatives of the posterior mean function. Building on these relations, they developed an efficient, training-free method for uncertainty quantification, enabling the computation of semantically meaningful principal components and approximation of marginal posterior distributions from any pre-trained denoiser for high-resolution images. This approach achieved up to a 1000x speed-up compared to sampling-based methods while delivering comparable PC quality.
Key insights
- The entire posterior distribution in Gaussian denoising can be recursively characterized solely by the MSE-optimal denoiser function and its derivatives, implicitly encoding all necessary uncertainty information.
- The posterior covariance (second central moment) is directly proportional to the Jacobian of the denoiser, providing a fundamental link for practical uncertainty quantification.
- Efficiently approximating Jacobian-vector products with finite differences allows for computing semantically meaningful posterior principal components for large images using only forward passes, significantly reducing computational burden.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.