Non-Asymptotic Uncertainty Quantification in High-Dimensional Learning

Authors: Frederik Hoppe, Claudio Mayrink Verdun, Hannah Laus, Felix Krahmer, Holger Rauhut

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the performance of our non-asymptotic confidence intervals through extensive numerical experiments across two settings: (i.) the classical debiased LASSO framework to contrast our nonasymptotic confidence intervals against the asymptotic ones. (ii.) the learned framework where we employ learned estimators, specifically the U-net [73] as well as the It-Net [18], to reconstruct real-world MR images and quantify uncertainty.
Researcher Affiliation Academia Frederik Hoppe RWTH Aachen University hoppe@mathc.rwth-aachen.de Claudio Mayrink Verdun Harvard University claudioverdun@seas.harvard.edu Hannah Laus TU Munich & MCML hannah.laus@tum.de Felix Krahmer TU Munich & MCML felix.krahmer@tum.de Holger Rauhut LMU Munich & MCML rauhut@math.lmu.de
Pseudocode Yes Algorithm 1 Estimation of Confidence Radius
Open Source Code Yes 1The code for our findings is available on Git Hub : https://github.com/frederikhoppe/UQ_high_ dim_learning
Open Datasets Yes We extend the debiasing approach to model-based deep learning for MRI reconstruction using the U-Net and It-Net on single-coil knee images from the NYU fast MRI dataset 2 [74, 75]. 2We obtained the data, which we used for conducting the experiments in this paper from the NYU fast MRI Initiative database (fastmri.med.nyu.edu) [74, 75].
Dataset Splits Yes The data is split into training (33370 slices), validation (5346 slices), estimation (1372 slices), and test (100 slices) datasets.
Hardware Specification Yes The experiments were conducted using Pytorch 1.9 on a desktop with AMD EPYC 7F52 16-Core CPUs and NVIDIA A100 PCIe030 030 40GB GPUs.
Software Dependencies Yes The experiments were conducted using Pytorch 1.9 on a desktop with AMD EPYC 7F52 16-Core CPUs and NVIDIA A100 PCIe030 030 40GB GPUs.
Experiment Setup Yes We then train an It-Net [18] with 8 layers, a combination of MS-SSIM [76] and ℓ1-losses and Adam optimizer with learning rate 5e 5 for 15 epochs to obtain our reconstruction function ˆX. (...) The It-Nets were trained for 15 epochs, and the U-Nets were trained for 20 epochs, both with batch size 40. (...) All It-Net and U-Nets are trained with a combination of the MS-SSIM-loss [76], the ℓ1-loss and the Adam optimizer with a learning rate of 5e 5, epsilon of 1e 4 and weight decay parameter 1e 5.