Reproduce “Noise May Contain Transferable Knowledge: Understanding Semi-supervised Heterogeneous Domain Adaptation from an Empirical Perspective” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuan Yao, Xiaopu Zhang, Yu Zhang, Jian Jin, Qiang Yang
Researchers from Beijing Teleinfo Technology, Inspur Computer Technology, Southern University of Science and Technology, China Academy of Information and Communications Technology, and Hong Kong University of Science and Technology empirically demonstrate that simple noise can serve as an effective source domain in semi-supervised heterogeneous domain adaptation (SHDA). Their study reveals that the transferable knowledge in SHDA primarily resides in the abstract properties of source discriminability and transferability, rather than the semantic content or specific features of the source data itself.
Key insights
- In SHDA, the semantic category information and specific feature representations of source samples are largely insignificant; permuting source categories or using semantically unrelated source datasets yields stable performance.
- Synthetic noise drawn from random distributions can function as an equally effective source domain for SHDA, challenging the conventional necessity of meaningful source data.
- The core transferable knowledge in SHDA is rooted in the source domain's discriminability (how well its categories are separated) and its transferability (how well its distribution aligns with the target domain), irrespective of the source data's origin.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.