Reproduce “Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods” 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
Dennis Wei, Inkit Padhi, Soumya Ghosh, Amit Dhurandhar, Karthikeyan Natesan Ramamurthy, Maria Chang
The authors define FiMODA, a way to estimate how sensitive a final model is to training examples when its training procedure is unavailable. Starting from the supplied model, they compare continued training with and without an example, averaging stochastic runs, and evaluate gradient-based approximations against this reference. First-order methods align best early; agreement declines with further training and is lower on CIFAR-10 and SST-2.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.