Reproduce “QUTCC: Quantile Uncertainty Training and Conformal Calibration for Imaging Inverse Problems” 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
Cassandra Tong Ye, Shamus Li, Tyler King, Kristina Monakhova
Researchers at Cornell University introduced QUTCC, a method for deep learning in imaging inverse problems that combines simultaneous quantile regression with spatially-adaptive conformal calibration. This approach yields tighter, statistically guaranteed uncertainty intervals and enables estimation of pixel-wise conditional probability density functions, improving reliability in sensitive applications.
Key insights
- Conditioning a U-Net with a quantile embedding allows a single network to efficiently learn and predict any conditional quantile, forming the basis for non-linear uncertainty quantification.
- Directly calibrating input quantile parameters for each pixel enables spatially-adaptive and tighter uncertainty intervals compared to global or group-wise scaling approaches.
- Pixel-wise conditional PDFs can be reconstructed by taking the inverse of the numerical derivative of the learned quantile function, offering a detailed, assumption-free probabilistic understanding of image predictions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.