[Re] Improved Calibration and Predictive Uncertainty for Deep Neural Networks

How do log loss, calibration summaries and selective-prediction error compare under a fixed validation and test split?

An independent scorer owns labels, binning choices and uncertainty metrics; calibration tuning is excluded from test data.

What you could produce

• Versioned protocol, input and environment manifest, and independent per-case comparison table including uncertainty and incomplete cases.

Before you use it

• PyTorch (named in journal metadata; dependency version not checked)

Limits to keep in view

• No research code was executed; no independent scientific verification has been performed.
• The current qualified pilot is self-contained Python 3.13 with a 90-second author deadline. This article's environment has not been qualified for that path.
• Published source metadata and a historical review do not establish compatibility, successful reproduction, operator independence or current scientific correctness.
• Dependencies named in metadata are not available under the current standard-library-only pilot; a new qualified environment is required.

Source and permission context

Singh, Aditya; Bay, Alessandro. [Re] Improved Calibration and Predictive Uncertainty for Deep Neural Networks. ReScience C 6(2), #10; 10.5281/zenodo.3818605.

Rights need review. Review the scope and upstream conditions before reuse.

Still unresolved

• The associated code's exact license and version-specific third-party notices have not been independently checked.
• Data licenses, permissions, consent restrictions and redistribution conditions have not been independently checked.
• Artifacts are referenced only; PDF/archive contents, dependency locks and bytes have not been inspected or executed.