Neural network classifiers often think they know more than they do — and a new algorithm aims to fix that without retraining from scratch.
Researchers have published FedLAS, short for Feature-Modulated Bidirectional Label Smoothing, a plug-and-play training module designed to improve how confidently a model's outputs match its actual accuracy. Standard label smoothing — a common fix for overconfident classifiers — applies a uniform adjustment across all training samples regardless of how hard or ambiguous each one is. FedLAS instead reads a per-sample signal called the Feature Norm-based Confidence Indicator to decide how much smoothing to apply, and a separate gating module catches both overconfident wrong guesses and underconfident correct ones. Tests on standard and fine-grained vision benchmarks show consistent drops in Expected Calibration Error and Adaptive ECE while holding Top-1 accuracy steady.
Calibration is the unglamorous cousin of accuracy, but in high-stakes deployments — medical imaging, autonomous vehicles, content moderation — a model that is 90% confident and 60% right is actively dangerous. Most calibration research still defaults to post-hoc fixes applied after training; FedLAS works during training, which means the model learns calibrated confidence rather than having it bolted on afterward.
The code is public, which is the right move for a method positioned as plug-and-play — though "plug-and-play" has been promised before by techniques that quietly require careful hyperparameter tuning to actually work.