This document summarizes recent work on continual learning in deep neural networks. It begins by defining the problem of catastrophic forgetting in continual learning and desiderata like resistance to forgetting and bounded model size. It then reviews three categories of approaches: regularization methods like EWC that constrain weight changes to prevent forgetting, dynamic architectures like PN that add new columns for each task, and dual-memory methods like PCN that use two networks for progress and compression of knowledge. Finally, it discusses directions for future research like developing truly task-agnostic continual learning that does not require task boundaries or labels.