Influence Functionsedit
Methods for estimating how training examples affect learned models.
Influence Functions are analytical tools for estimating how a training point affects a fitted model or a downstream prediction. In modern machine learning they are often used as approximations: instead of retraining after changing one point, the method estimates the effect through gradients and curvature information.1
Research contextedit
If a researcher can estimate the effect of a point, group, or weighted subset, they can ask which data should be removed, downweighted, kept, or inspected. Influence functions therefore connect data selection, machine unlearning, fairness correction, and robustness analysis. The same idea also motivates why Hessian-vector products and second-order approximations appear in unlearning papers.
Connection to Qiao's workedit
Qiao's unlearning work uses influence-style reasoning in several forms. Hessian-Free Online Certified Unlearning relies on efficient updates without explicit Hessian inversion. Beyond Binary Erasure uses weighted influence to turn a deletion problem into a corrective intervention for fairness and robustness. Influence functions provide tools for analyzing local sensitivity in Qiao's research on data operations.
See alsoedit
Footnotesedit
-
Koh and Liang, "Understanding Black-box Predictions via Influence Functions", ICML 2017, reintroduced classical influence-function ideas for explaining predictions in modern machine-learning models. ↩