Modern ImageNet foundation models are underconfident in-distribution, improve calibration under distribution shift, and only benefit from post-hoc calibration in-distribution.
Proceedings of the AAAI Conference on Artificial Intelligence29(2015)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
Beyond Overconfidence: Foundation Models Redefine Calibration in Deep Neural Networks
Modern ImageNet foundation models are underconfident in-distribution, improve calibration under distribution shift, and only benefit from post-hoc calibration in-distribution.