Uncertainty quantification is crucial for the deployment of image restoration models in safety-critical domains, like biological and medical imaging. To date, methods for uncertainty visualization have mainly focused on per-pixel estimates, which provide limited information. A more natural visualization of uncertainty can be obtained from the principal components (PCs) of the posterior distribution. In this talk I will present methods for predicting the PCs of the posterior, as well as for visualizing the distribution along the space spanned by those PCs, in a single forward pass of a neural network. Our methods are both more accurate and orders of magnitude faster than the naïve approach of applying PCA to posterior samples generated by a conditional generative model. I will illustrate the effectiveness of our approaches on multiple inverse problems in imaging, including denoising, inpainting, super-resolution, colorization, and biological image-to-image translation.
The talk will cover joint works with Elias Nehme, Omer Yair, and Hila Manor.