Reproduce “Unsupervised Domain Adaptation within Deep Foundation Latent Spaces” 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
Dmitry Kangin, Plamen Angelov
A finetuning-free Unsupervised Domain Adaptation (UDA) method is introduced, leveraging the fixed latent spaces of pre-trained Vision Transformers in conjunction with interpretable prototypical networks. This approach achieves superior performance compared to existing state-of-the-art UDA baselines on multi-source UDA datasets by effectively transferring knowledge across domains without modifying the feature extractor.
Key insights
- Vision Transformers provide sufficiently robust and semantically rich latent spaces to achieve effective UDA without finetuning the feature extractor.
- Prototypical networks offer inherent, "by-design" interpretability for cross-domain classification, allowing for qualitative analysis of decision errors.
- The quality and pretraining strategy of the foundation model, such as additional ImageNet-1K finetuning, significantly impacts performance in a finetuning-free UDA setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.