Can a small network learn new digits without losing the old ones?
A user highlights results on evaluation, loss scaling and synthetic replay from automated research they assigned to Claude.
TLDR
A user says they had Claude conduct automated research on a toy continual learning problem: whether a small network could learn new digits without losing previously learned ones. They point to results involving evaluation, loss scaling and synthetic replay.
Combined views
2.4K
1 Source, first seen ago