arxivcs.CVcs.AIcs.LG2026-06-26
FedLAS: Feature-Modulated Bidirectional Label Smoothing for Neural Network Calibration
Thiru Thillai Nadarasar Bahavan, Sachith Seneviratne, Saman Halgamuge
Deep Neural Network (DNN) classifiers suffer from poor calibration when their softmax outputs (predictive confidence) deviate from the empirical likelihoods. This manifests itself as either overconfident incorrect predictions or under-confident correct predictions. Label smoothin…