Understanding Distribution Alignment Through Category Separability In An Infant-Inspired Domain Adaptation Task
Keywords: domain adaptation, distribution shift, infant learning, self-supervised learning
Abstract: We introduce a novel distribution shift, called the VI-Shift, that mimics the trade-off between object instances and viewpoints in the visual experience of infants. Motivated by findings in infant learning literature, we study this problem through the lens of domain adaptation, but without ImageNet pretraining. We show that the performances of two classic domain adaptation methods, Joint Adaptation Network (JAN) and Domain Adversarial Neural Networks (DANN), deteriorate without ImageNet pretraining. We hypothesize that the separability of source and target category clusters in the feature space plays a crucial role in the effectiveness of JAN. So, we propose 3 metrics to measure category separability and demonstrate that target separability in the pretrained network is strongly correlated with downstream JAN and DANN accuracy. Further, we propose two novel loss functions that increase target separability during pretraining by aligning the distribution of within-domain pairwise distances between the source and target distributions. Our experiments show that the application of these loss functions modestly improves downstream accuracy on unseen images from the target dataset.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation 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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 8683
Loading