Two NeurIPS Papers Explain Why Pretraining Beats Infeasible Incremental Learning

Two papers accepted to NeurIPS 2026 show that with Mahalanobis margin conditions on pretrained representations, task confusion decays exponentially. Building on the author's PhD result that optimal class-incremental learning is infeasible, the work explains why pretrained models can bypass this limitation.

2026-09-25 ~ 2026-09-25 · 2 related posts