SMART learns a per-sample temperature from the gap between the top two logits using a soft-bin ECE objective, achieving strong calibration with as few as 50 validation samples.
Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Sample Margin-Aware Recalibration of Temperature Scaling
SMART learns a per-sample temperature from the gap between the top two logits using a soft-bin ECE objective, achieving strong calibration with as few as 50 validation samples.