Noises are Transferable - An Empirical Study on Heterogeneous Domain Adaptation
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: noises, heterogeneous domain adaptation, semi-supervised learning, transferable discriminability
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Semi-supervised Heterogeneous Domain Adaptation (SHDA) handles the learning of cross-domain samples with both distinct feature representations and distributions.
In this paper, we perform the first empirical study on the SHDA problem by utilizing seven typical SHDA approaches for nearly 100 standard SHDA tasks. Surprisingly, we find that the noises drawn from simple distributions as source samples are transferable and can be used to improve the performance of target domain. To go deeper with the essence of the SHDA, we identify and explore several key factors, including the number of source samples, the dimensions of source samples, the original discriminability of source samples, and the transferable discriminability of source samples. Building upon extensive experimental results, we believe that the transferable knowledge in SHDA is primarily rooted in the transferable discriminability of source samples.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2488
Loading