59、基于BERT的对抗性半监督主动学习与随机牛顿法优化

基于BERT的对抗性半监督主动学习与随机牛顿法优化

一、BERT-based Adversarial Semi-Supervised Active Learning (B - ASAL)

1. 算法概述
   • 输入 ：标记数据$S_l$、未标记数据$S_u$，初始化生成器$\varphi_G$、判别器$\gamma_D$、分类器$\sigma_C$和BERT编码器$\beta_E$的参数。
   • 输出 ：优化后的$\varphi_G$、$\gamma_D$、$\sigma_C$和$\beta_E$。
   • 算法流程 ：

1: for i = 1 to epochs do
2:     Sample batch of size n from Sl labeled and Su unlabeled data |Sl| = |Su| = n
3:     Sample {z1, ..., zm+n} ∈Sg from the prior Pz
4:     Generate encode E(Sl), E(Su) and E(Sg)
5:     For labeled data (x, y ∈Sl):
6:         Compute Ll_C from Eq. 3
7:         Update C by descending:
8: