VALSE 论文速览 第104期：基于自适应对比学习的无源域数据域适应方法 ...

论文题目：Divide and Contrast: Source-free Domain Adaptation via Adaptive Contrastive Learning

作者列表：张子逸 (南京大学)、陈伟凯 (腾讯美国)、程卉 (中山大学)、李镇 (香港中文大学深圳)、李思远 (西湖大学)、林倞 (中山大学)、李冠彬 (中山大学)

B站观看网址：

https://www.bilibili.com/video/BV1iA411y7Ns/

论文摘要：

We investigate a practical domain adaptation task, called source-freedomain adaptation (SFUDA), where the source-pretrained model isadapted to the target domain without access to the source data. Existing techniques mainly leverage self-supervised pseudo labelingto achieve class-wise global alignment[1]or rely on local structureextraction that encourages feature consistency among neighborhoods[2].While impressive progress has been made, both lines of methodshave their own drawbacks - the "global" approach is sensitive tonoisy labels while the "local" counterpart suffers from source bias. In this paper, we present Divide and Contrast (DaC), a new paradigmfor SFUDA that strives to connect the good ends of both worlds whilebypassing their limitations. Based on the prediction confidence ofthe source model, DaC divides the target data into source-like andtarget-specific samples, where either group of samples is treatedwith tailored goals under an adaptive contrastive learning framework.Specifically, the source-like samples are utilized for learningglobal class clustering thanks to their relatively clean labels. Themore noisy target-specific data are harnessed at the instance levelfor learning the intrinsic local structures. We further align thesource-like domain with the target-specific samples using a memorybank-based Maximum Mean Discrepancy (MMD)loss to reduce thedistribution mismatch. Extensive experiments on VisDA, Office-Home,and the more challenging DomainNet have verified the superiorperformance of DaC over current state-of-the-art approaches.

论文信息：

[1] Z. Zhang, W. Chen, H. Cheng, M. Mao, L. Lin, and G. Li. Dual Adversarial Adaptation forCross-Device Real-World Image Super-Resolution. InNeurIPS,2022.

论文链接：

[https://arxiv.org/abs/2211.06612]

代码链接：

[https://github.com/ZyeZhang/DaC.git]

视频讲者简介：

张子逸，南京大学人工智能学院硕士二年级。主要研究方向为强化学习、计算机视觉，尤其是基于环境的强化学习、迁移学习方法的研究。

特别鸣谢本次论文速览主要组织者：

月度轮值AC：林迪 (天津大学)、胡鹏 (四川大学)