摘要
解决领域偏移(domain shift)或数据集偏置(dataset bias)问题通常侧重于去发现源领域和目标领域之间的域不变表征.尽管这种做法到目前为止已经取得了有效的进展,但是其受限于特征层面的学习,使之无法充分利用已有的信息,极大约束了领域适应任务.为了有所改进,本文着眼于更加困难的无监督领域适应图像分类研究,提出了多层面的分步领域适应方法.该方法将不同层面取得的成果进行划分,并将算法流程细分为多步,对数据进行分步处理,保证了最大化数据利用率和具备高度的可扩展性.此外,在标签层面,本文巧妙地将目标领域中的样本分为易适应和难适应两类,并结合领域对抗损失(domain-adversarial loss)进行再次处理.模型的实现基于一个已有的代表性算法,在标准领域适应任务上的实验达到了预期效果.
The general solution to domain shift or dataset bias focuses on discovering the domain-invariant representations between source domain and target domain. Although effective,such a solution is limited to the feature-level and does not make full use of the existing information which considerably constrains domain adaptation. To further improve performance,we propose Multi-level and Step-by-step Domain Adaptation,applying to the harder unsupervised case of image classification. In this work,domain adaptation takes into account the efforts made at different levels and procedure is divided into several steps,each of which adequately processes the data,which guarantees maximizing data efficiency and having high scalability. In addition,at label-level,samples in target domain are skillfully divided into two categories,namely,easy to adapt and difficult to adapt,which are reprocessed combined with a domainadversarial loss. Finally,we implement our model based on an excellently representative algorithm,and experiments achieve expected results on standard domain adaptation tasks.
作者
许浩
李宗印
郭卫斌
XU Hao;LI Zong-yin;GUO Wei-bin(School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China;Institute of Science and Technology Development,East China University of Science and Technology,Shanghai 200237,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2019年第9期1921-1925,共5页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61672227)资助
关键词
领域适应
迁移学习
图像分类
对抗网络
domain adaptation
transfer learning
image classification
adversarial networks
