Artificial Intelligence 3 min read

Reducing the Covariate Shift by Mirror Samples in Cross Domain Alignment

By constructing virtual mirror samples that occupy identical positions across source and target domains, the authors eliminate covariate shift while preserving distribution structure, enabling superior unsupervised domain adaptation that achieves state‑of‑the‑art performance on Office and VisDA benchmarks and improves real‑world lighting and gender‑recognition tasks.

Youku Technology
Youku Technology
Youku Technology
Reducing the Covariate Shift by Mirror Samples in Cross Domain Alignment

NIPS (Conference and Workshop on Neural Information Processing System)

神经信息处理系统大会是机器学习领域的顶级会议。在NIPS 2021,阿里巴巴文娱AI大脑北斗星团队有一文入选,研究成果属于视觉分类领域。

Reducing the Covariate Shift by Mirror Samples in Cross Domain Alignment

作者

王敏全

赵寅

蔡龙军

(作者均来自阿里巴巴阿里文娱AI大脑北斗星团队)

简介

在基础视觉分类任务中,通过消除domain之间的 covariate shift来做到domain alignment是解决无监督跨域自适应unsupervised domain adaptation (UDA)的一种常用方法。然而目前无论是prototype based还是sample-level based的domain alignment方法都忽视了数据分布的结构特性甚至破坏了covariate shift的基本假设。

本文提出了一个新颖的在另一个domain构造(虚拟)镜像样本的方法。所构造的镜像样本与原样本表征了在两个domain中处于相同位置的样本对。同时,通过对齐镜像样本对来保证两个domain的对齐。理论上可证明该方法可以得到更好的跨域泛化性能。经过大量的实验,本文所提出的方法在Office, VisDA等4个公开benchmark上都达到了SOTA。在我们的‘现场解决方案’项目中解决了从正常光照到红外光照的跨域性别识别、表情识别等模型迁移任务,节省了大量的标注资源,提高了模型性能。

machine learningAI researchdomain adaptationcovariate shiftcross-domain alignmentSOTAvisual classification
Youku Technology
Written by

Youku Technology

Discover top-tier entertainment technology here.

0 followers
Reader feedback

How this landed with the community

login Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.