迁移学习(EADA)《Unsupervised Energy-based Adversarial Domain Adaptation for Cross-domain Text Classification》
论文信息
论文标题:Unsupervised Energy-based Adversarial Domain Adaptation for Cross-domain Text Classification
论文作者:Han Zou, Jianfei Yang, Xiaojian Wu
论文来源:ACL 2021
论文地址:download
论文代码:download
引用次数:
1 前言
Energy-based Generative Adversarial Network 结合 DANN 的模型;
2 方法
整体框架:
$\text{DANN}$ 训练目标:
$\begin{array}{l}\underset{G_{f}, G_{y}}{\text{min}} \quad \mathcal{L}_{y}\left(\mathbf{X}_{s}, Y_{s}\right)-\gamma \mathcal{L}_{f}\left(\mathbf{X}_{s}, \mathbf{X}_{t}\right) \\\underset{G_{d}}{\text{min}} \quad \mathcal{L}_{d}\left(\mathbf{X}_{s}, \mathbf{X}_{t}\right)\end{array}$
本文训练目标:
$\begin{array}{l}\underset{G_{f}, G_{y}}{\text{min}} \quad \mathcal{L}_{C E}\left(\mathbf{X}_{s}, Y_{s}\right)+\gamma \mathcal{L}_{A E}\left(\mathbf{X}_{\mathbf{t}}\right), \\\underset{G_{a}}{\text{min}}\quad \mathcal{L}_{A E}\left(\mathbf{X}_{\mathbf{s}}\right)+\max \left(0, m-\mathcal{L}_{A E}\left(\mathbf{X}_{\mathbf{t}}\right)\right)\end{array}$
$\mathcal{L}_{A E}\left(\mathbf{x}_{i}\right)=\left\|G_{a}\left(G_{f}\left(\mathbf{x} ; \theta_{f}\right) ; \theta_{a}\right)-\mathbf{x}_{i}\right\|_{2}^{2}$
因上求缘,果上努力~~~~ 作者:图神经网络,转载请注明原文链接:https://www.cnblogs.com/BlairGrowing/p/17231324.html