Temperature scaling and calibration analysis of modern neural classifiers

What changed

Studies mismatch between neural confidence and correctness, and fits a held-out scalar logit temperature as a practical multiclass calibration procedure without changing class argmax.

Scientific scope

Studies mismatch between neural confidence and correctness, and fits a held-out scalar logit temperature as a practical multiclass calibration procedure without changing class argmax. Builds on Platt scaling and existing temperature use. Assumes representative same-distribution calibration/test data; binned ECE is an empirical measure, not universal or conditional calibration. Architecture correlations are not causal identifications. First three authors explicitly equal in alphabetical order.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Builds on Platt scaling and existing temperature use. Assumes representative same-distribution calibration/test data; binned ECE is an empirical measure, not universal or conditional calibration. Architecture correlations are not causal identifications. First three authors explicitly equal in alphabetical order.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• On Calibration of Modern Neural Networks

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction