1 paper · 1 filter
Fang Qi, Lu Peng, Zhengming Ding
Learning with softmax cross-entropy on one-hot labels often leads to overconfident predictions and poor robustness under noise or perturbations. Label smoothing mitigates this by r…