Paper Overview
Field: Machine Learning Authors: Parampreet Singh, Anushka Singh, Sumit Kumar, Vipul Arora Published: 2026-08-22 arXiv: 2608.20326
Summary
Deep neural networks are often overconfident, assigning high confidence even to incorrect predictions. Post-hoc confidence estimation addresses this by training lightweight auxiliary heads on frozen classifiers, but existing objectives suffer from inherent ambiguity: the target values for correct and incorrect predictions overlap, making confidence for incorrect predictions near the decision boundary hard to distinguish from correct ones.
This paper proposes TCPα, a novel confidence objective that resolves these limitations by introducing a margin-controlled penalty on misclassified samples. Theoretically, TCPα guarantees complete separation between the target values of correct and incorrect predictions; the separation margin is independent of the number of classes and increases monotonically with the penalty parameter.
Since accurate classifiers naturally produce few errors, learning these targets leads to a severely imbalanced regression problem. The authors systematically study training strategies under this imbalance and identify effective configurations through extensive ablation studies.
Results
Evaluated on lag recognition (raga identification), TCPα improves the base model's macro-F1 from 0.89 to 0.98 simply by rejecting the lowest-confidence 8% of predictions.
--- *Auto-collected on 2026-08-22*