BriefGPT.xyz
Feb, 2021
面向预测用户流失的本地自适应标签平滑化
Locally Adaptive Label Smoothing for Predictive Churn
HTML
PDF
Dara Bahri, Heinrich Jiang
TL;DR
探讨了在训练现代神经网络时存在的高预测波动性，并提出了使用软标签来降低波动性的方法，该方法在多种基准分类任务和模型架构上都提高了准确性。
Abstract
Training modern
neural networks
is an inherently noisy process that can lead to high \emph{
prediction churn
} -- disagreements between re-trainings of the same model due to factors such as randomization in the par
→