Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

UncertaINR: Uncertainty Quantification of End-to-End Implicit Neural Representations for Computed Tomography

Authors: Francisca Vasconcelos, Bobby He, Nalini M Singh, Yee Whye Teh

TMLR 2023 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We study a Bayesian reformulation of INRs, Uncerta INR, in the context of computed tomography, and evaluate several Bayesian deep learning implementations in terms of accuracy and calibration. We find that they achieve well-calibrated uncertainty, while retaining accuracy competitive with other classical, INR-based, and CNN-based reconstruction techniques.
Researcher Affiliation Academia Francisca Vasconcelos EMAIL Department of Electrical Engineering and Computer Science University of California, Berkeley Bobby He EMAIL Department of Statistics University of Oxford Nalini Singh EMAIL Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology Yee Whye Teh EMAIL Department of Statistics University of Oxford
Pseudocode No The paper describes algorithms like Bayes-by-Backprop, Monte Carlo Dropout, Hamiltonian Monte Carlo, and Deep Ensembles conceptually in the text (e.g., Section 3 and Appendix D) but does not provide structured pseudocode blocks or clearly labeled algorithm figures.
Open Source Code Yes Project code is available at: https://github.com/bobby-he/uncertainr.
Open Datasets Yes Two datasets were considered. The first consists of artificial 256 256 pixel Shepp-Logan phantom (Shepp & Logan, 1974) brain images... The second dataset, used for Uncerta INR baseline comparisons, contains 512 512 pixel abdominal CT scan images, provided by the Mayo Clinic for the 2016 Low-Dose CT AAPM Grand Challenge (Mc Collough et al., 2017).
Dataset Splits Yes We tuned hyperparameters on 5 validation images and evaluated performance on 5 test images. The second dataset... 3 validation images and 8 test images were used to generate noisy 60and 120-view sinograms.
Hardware Specification Yes Non-HMC methods were run on a cluster of 4 GPU nodes consisting of 8 GPUs each, containing a mixture of GTX 1080, GTX 1080Ti, and Ge Force RTX 2080 Ti cards. HMC methods (*) were run on two Titan RTX cards, which were faster and had double the memory.
Software Dependencies No The project codebase was developed in Python, mostly using Pytorch (Paszke et al., 2019), Hydra (Yadan, 2019), and Weights & Biases (Biewald, 2020) to implement the NN functionality and Blitz (Esposito, 2020) for BNN functionality. For HMC, we used the No-U-Turn-Sampler (Hoffman et al., 2014) sampling scheme in Num Pyro (Phan et al., 2019), which is based in JAX (Bradbury et al., 2018).
Experiment Setup Yes All Uncerta INR were optimized using Adam (not Adam W, to eliminate the computational costs of tuning the weight decay parameter) and we found that longer training times were needed (on the order of 15,000 epochs). If that many training epochs cannot be used for some reason, we found that stochastic weight averaging (Izmailov et al., 2018) can be used to improve reconstruction accuracy by a few decibel for Uncerta INRs trained with few epochs. ... Table 8: Hyperparameters of the top-performing MCD UINRs, HMC UINRs, and INRs reported in Table 2.