Provable Continual Learning via Sketched Jacobian Approximations

Reinhard Heckel

An important problem in machine learning is the ability to learn tasks in a sequential manner. If trained with standard first-order methods most models forget previously learned ta…