Overcoming catastrophic forgetting in neural networks

The paper addresses a key weakness of deep neural networks: unlike humans, they have been unable to learn multiple tasks sequentially, and catastrophic forgetting was widely thought to be an inevitable feature of connectionist models. Inspired by synaptic consolidation in neuroscience, the proposed approach trains networks sequentially by protecting the weights important for previous tasks, remembering old tasks by selectively slowing down learning on those weights.

The authors show it is possible to train networks that maintain expertise on tasks they have not experienced for a long time. The approach is demonstrated to be scalable and effective by solving a set of classification tasks based on a hand-written digit dataset and by learning several Atari 2600 games sequentially, enabling state-of-the-art results on multiple reinforcement learning problems experienced sequentially - a capability the authors argue is crucial to the development of artificial intelligence.