RMM: Reinforced Memory Management for Class-Incremental Learning

The preserved data are exemplars used for replaying. However, existing methods use a static and ad hoc strategy for memory allocation, which is often sub-optimal.