This paper presents a novel Quantile Uncertainty Training and Calibration (QUTCC) technique to address the hallucination problem of deep learning models, especially in scientific and medical inverse problems where accuracy is critical, such as medical imaging (MRI, microscopy). To overcome the limitation of existing uncertainty quantification techniques that generate large ranges with little information by setting uncertainty ranges using linear constant scales, QUTCC utilizes the U-Net architecture and quantile embedding to estimate the entire quantile of the conditional distribution. Then, iteratively improves the uncertainty ranges through iterative upper and lower quantile queries, providing more accurate ranges while maintaining the desired confidence level. We experimentally show that it achieves more accurate uncertainty ranges than existing methods on several denoising and compressed MRI reconstruction tasks.
