Reproduce “From Parameter Rebasin to Functional Transfer Across Heterogeneous Transformers” 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
Filippo Rinaldi, Aniello Panariello, Giacomo Salici, Simone Magistri, Angelo Porrello, Andrew D. Bagdanov, Simone Calderara
The authors introduce Ariadne, which transfers fine-tuning effects between differently sized Transformers by aligning source and target base activations, mapping the source’s activation change, and fitting a target-native update from calibration data. On Vision8 expansion from ViT-B/16 to ViT-L/14, it reaches 84.54% accuracy, versus 73.87% for the target base; results cover related backbones and require calibration inputs and block correspondence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.