Reproduce “On Second-Order Group Influence Functions for Black-Box Predictions” 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
Samyadeep Basu, Xuchen You, Soheil Feizi
The authors develop a second-order group influence estimate that adds interaction terms to summed influences, estimating how removing training examples changes model behavior without retraining. On MNIST logistic regression, removing 50% of training data improved correlation with leave-group-out retraining effects by over 55% versus first-order influence. This result does not extend to neural networks, where both methods had non-significant correlations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.