Why Self-Distillation Helps and Hurts: Denoising vs. Signal Forgetting

ORID VnA5q5jXVz · tags icml2026-repro paper-VnA5q5jXVz

#StatusPageArtifactClaim excerpt
1VERIFIED 2/201-deterministic-equivalent-recursions-prediction-rartifactDeterministic-equivalent recursions for prediction risk in overparameterized lin…
2VERIFIED 2/202-test-risk-function-self-training-iterationartifactTest risk as a function of self-training iteration count follows a U-shaped curv…
3VERIFIED 2/203-spiked-covariance-models-self-training-iterationartifactIn spiked covariance models, each self-training iteration acts as a direction-de…
4VERIFIED 2/204-iterated-generalized-cross-validation-igcv-estimartifactAn iterated Generalized Cross-Validation (iGCV) estimator is proposed that achie…
5VERIFIED 2/205-theoretical-denoising-forgetting-risk-tradeoff-dartifactThe theoretical denoising-forgetting risk tradeoff, derived for linear regressio…

Open logbook index · logbook.json