CoLLAs 2026 keynote: fix forgetting via drift compensation and model merging, not just prevention

apsarathchandar · x · 2026-09-15

In the second Day 1 keynote at CoLLAs 2026 in Bucharest, Joost van de Weijer argued for a fresh take on continual learning: instead of only fighting catastrophic forgetting, use drift compensation to correct representations after the fact. He also shared insights on model merging to combine independently trained task models into one.

Original post →

More from Research

Research channel →