Representer points: signed training-feature contributions to network logits

What changed

Derives a representer decomposition of pre-activation outputs into signed training-feature similarities at stationary points with positively weighted L2 regularization of the final linear layer.

Scientific scope

Derives a representer decomposition of pre-activation outputs into signed training-feature similarities at stationary points with positively weighted L2 regularization of the final linear layer. The theorem is conditional on its final-layer regularization and stationarity assumptions, not arbitrary unregularized checkpoints. Signed logit contributions are not identical to loss influence or deletion counterfactuals. First two authors explicitly equal; proof source read, not independently certified.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. The theorem is conditional on its final-layer regularization and stationarity assumptions, not arbitrary unregularized checkpoints. Signed logit contributions are not identical to loss influence or deletion counterfactuals. First two authors explicitly equal; proof source read, not independently certified.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Representer Point Selection for Explaining Deep Neural Networks

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction