Reproduce “Post-Training Augmentation Invariance” 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
Keenan Eikenberry, Lizuo Liu, Yoonsang Lee
The authors introduce MaWa, a small adapter trained on frozen pretrained features to add rotation invariance while anchoring clean features near their original representations. On STL10 with DINO features, a clean-trained linear classifier reaches 94.66% accuracy on arbitrarily rotated images, versus 71.22% for the end-to-end baseline. MaWa approximately preserves clean-feature pairwise distances, but can fail when cross-class collisions remain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.