Posterior Meta-Replay for Continual Learning

In practice, however, recursive updating often leads to poor trade-off solutions across tasks because approximate inference is necessary for most models of interest.