The ability to learn tasks in a sequential fashion is crucial to the development of artificial intelligence. Neural networks are not, in general, capable of this and it has been widely thought that catastrophic forgetting is an inevitable feature of connectionist models. We show that it is possible to overcome this limitation and train networks that can maintain expertise on tasks which they have not experienced for a long time. Our approach remembers old tasks by selectively slowing down learning on the weights important for those tasks. We demonstrate our approach is scalable and effective by solving a set of classification tasks based on the MNIST hand written digit dataset and by learning several Atari 2600 games sequentially.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Continual Learning 20Newsgroup (10 tasks) EWC F1 - macro 0.9180 # 5
Continual Learning ASC (19 tasks) L2 F1 - macro 0.5243 # 15
Continual Learning ASC (19 tasks) EWC F1 - macro 0.7452 # 13
Class Incremental Learning cifar100 EWC 10-stage average accuracy 50.53 # 6
Continual Learning DSC (10 tasks) EWC F1 - macro 0.6576 # 6
Continual Learning F-CelebA (10 tasks) EWC Acc 0.6545 # 3