A neurodevelopment-inspired warm-up strategy to address uncertainty calibration: networks are briefly trained on random noise and labels before exposure to real data, leading to well-calibrated confidence and strong detection of unknown inputs.
Cool results!
#NeuroAI
www.nature.com/articles/s42...
Cheon and Paik show that overconfidence in deep neural networks arises from standard initialization practices, and that brief warm-up training with random noise improves uncertainty calibration and meta-cognitive recognition of unknown inputs.
