Two NeurIPS 2026 papers: margin conditions tame sequential task confusion

khademinori · x · 2026-09-25

Two papers accepted at NeurIPS 2026. The first proves that under a Mahalanobis margin condition on pretrained representations, task confusion between classes learned at different times decays exponentially in the margin, and that fine-tuning erodes the margin at a rate growing with the learning rate. The second gives a closed-form characterization of the per-stage offset L2 regularization induces in sequentially trained cross-entropy heads, introducing gauge anchoring.

Related event: Two NeurIPS Papers Explain Why Pretraining Beats Infeasible Incremental Learning(2 posts)→

Original post →

More from Research

Research channel →