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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Uncertainty Quantification for Deep Regression using Contextualised Normalizing Flows
Authors: Adriel Sosa Marco, John D. Kirwan, Alexia Toumpa, Simos Gerasimou
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide experimental evidence that the MCNF-based uncertainty estimate is well calibrated, is competitive with state-of-the-art uncertainty quantification methods, and provides richer information for downstream decision-making tasks. ... Our experimental evaluation using a diverse set of datasets and state-of-the-art UQ methods [14, 16] demonstrates that MCNF achieves competitive results in terms of marginal coverage while also having lower error values and narrower intervals. ... 5 Evaluation ... Table 1 summarizes the performance results of our evaluation. |
| Researcher Affiliation | Collaboration | Adriel Sosa Marco , John Daniel Kirwan , Alexia Toumpa , Simos Gerasimou Arquimea Research Center, Spain Department of Computer Science, University of York, York, UK Department of Elect. Eng., and Computer Science and Eng., Cyprus University of Technology, Cyprus EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Steps involved in a full forward pass of the proposed MCNF method for inference ... Algorithm 2 Mini-batch building scheme to train the Normalizing Flow head of MCNF |
| Open Source Code | Yes | A prototype open-source MCNF tool and case study repository, available at https://github. com/alexiatoumpa/MCNF. |
| Open Datasets | Yes | Adopting the evaluation procedure from CQR [14] and MCCP [16], we used the following datasets to evaluate MCNF: the Boston Housing dataset [25] (506 observations, 14 attributes); the Concrete dataset [26] (1030 observations, 9 attributes); the Abalone dataset [27] (4177 observations, 11 attributes); the Tertiary Protein Structure dataset [28] (45730 observations, 10 attributes); the wave energy dataset [29] (63600 observations, 149 attributes), and the superconductivity dataset [30] (21263 observations, 81 attributes). These datasets were obtained from [31]. ... the Solubility dataset [33] |
| Dataset Splits | Yes | For the training and testing sets, we use an 80:20 split. |
| Hardware Specification | No | We provide details regarding the computational setup used for our evaluation in the Appendix. (Note: The provided Appendix content does not contain these details) |
| Software Dependencies | No | The paper mentions 'Adam optimizer', 'Python', 'PyTorch', 'scikit-learn' but does not specify exact version numbers for these software components or libraries. |
| Experiment Setup | Yes | The model is trained for 100 epochs using the Adam optimizer, a custom pinball loss function that aggregates the quantile errors, and a batch size of 32. We fixed the learning rate to 5e-4 and the weight decay regularization factor to 1e-6. ... The Normalizing Flow component of MCNF uses a sequence of two Neural Splines flows [24], with a 3-layered multilayer perception comprising 64 hidden units which produces the 16 support vectors of the spline transformation and their inner derivatives. ... To keep the computational overhead related to the prior Monte Carlo Dropout sampling low, we set n MCD = 50. ... using a batch size of 32 with the Adam optimiser and a 0.001 learning rate. We set τ = 1e10 |