Hacker News new | ask | show | jobs
by jacquesm 1013 days ago
You could simulate this by selectively locking and unlocking 'banks' of weights from a larger model to keep the influence there during training and to avoid losing them. Sort of a selective write-protect.