Domain-Adversarial Training of Neural Networks
\nameYaroslav Ganin \nameEvgeniya Ustinova \addrSkolkovo Institute of Science and Technology (Skoltech) Skolkovo, Moscow Region, Russia \AND\nameHana Ajakan \namePascal Germain \addrDépartement d’informatique et de génie logiciel, Université Laval Québec, Canada, G1V 0A6 \AND\nameHugo Larochelle \addrDépartement d’informatique, Université de Sherbrooke Québec, Canada, J1K 2R1 \AND\nameFrançois Laviolette \nameMario Marchand \addrDépartement d’informatique et de génie logiciel, Université Laval Québec, Canada, G1V 0A6 \AND\nameVictor Lempitsky \addrSkolkovo Institute of Science and Technology (Skoltech) Skolkovo, Moscow Region, Russia
Abstract
We introduce a new representation learning approach for domain adaptation, in which data at training and test time come from similar but different distributions. Our approach is directly inspired by the theory on domain adaptation suggesting that, for effective domain transfer to be achieved, predictions must be made based on features that cannot discriminate between the training (source) and test (target) domains.
中文速览
跨域识别一直面临这样的困境：训练数据和真实使用场景的数据分布不一致，导致模型"水土不服"。这篇论文提出了一种叫做域对抗神经网络（Domain-Adversarial Neural Network，DANN）的方法，核心思路是在普通神经网络里加入一个"梯度反转层"，让网络在学会区分类别的同时，主动"忘掉"源域和目标域之间的差异——特征提取器要骗过域分类器，使其无法判断一个样本究竟来自哪个域。整个框架只需要源域的有标注数据和目标域的无标注数据，通过标准的反向传播就能端到端地联合训练，几乎不需要改动现有的深度学习架构。实验覆盖文本情感分析、图像分类（MNIST、SVHN、Office基准）以及行人重识别等多个任务，均达到或超越当时的最优水平。这项工作的重要性在于，它把"让特征跨域通用"这一理论目标直接转化为一个简洁可训练的网络结构，为后续大量域适应研究奠定了方法论基础。
原文 arXiv:1505.07818;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1505.07818v4