#catastrophic-forgetting

Question 1: Select all that apply

You fine-tune a pretrained model on a narrow new task and want it to keep what it could do before, that is, to avoid catastrophic forgetting. Which strategies help? Select all that apply.

Unfreeze all layers at once with a large learning rate, so the new task needs fewer steps1
Train until the loss on the new task is near zero, so that the weights settle down2
Add an EWC penalty that keeps weights important for the old task close to their old values3
Mix examples from the original data or tasks into the fine-tuning batches4
Freeze the pretrained weights and train only small adapter modules for the new task5

#catastrophic-forgetting

Timed test