Researchers · Interviewed at ICLR 2026
University of Toronto & Vector Institute

ICLR 2026 · June 3, 2026 · 14 min
"Can a model avoid catastrophic forgetting without keeping a replay buffer?"
In this #ICLR 2026 interview, Nikita Dhawan (University of Toronto & Vector Institute) explains LIFTER (LInearized Function TRick), a method that estimates function-space distances from only the first two moments of past data distributions. The approach nearly matches an oracle method that has full access to old training examples—while never storing those examples in the first place.