Overcoming catastrophic forgetting in neural networks
James KirkpatrickDeepMind, London EC4 5TW, United Kingdom;Razvan PascanuDeepMind, London EC4 5TW, United Kingdom;Neil C. RabinowitzDeepMind, London EC4 5TW, United Kingdom;Joel VenessDeepMind, London EC4 5TW, United Kingdom;Guillaume DesjardinsDeepMind, London EC4 5TW, United Kingdom;Andrei A. RusuDeepMind, London EC4 5TW, United Kingdom;Kieran MilanDeepMind, London EC4 5TW, United Kingdom;John QuanDeepMind, London EC4 5TW, United Kingdom;Tiago RamalhoDeepMind, London EC4 5TW, United Kingdom;Agnieszka Grabska‐BarwińskaDeepMind, London EC4 5TW, United Kingdom;Demis HassabisDeepMind, London EC4 5TW, United Kingdom;Claudia ClopathBioengineering Department, Imperial College London, London SW7 2AZ, United KingdomDharshan KumaranDeepMind, London EC4 5TW, United Kingdom;Raia HadsellDeepMind, London EC4 5TW, United Kingdom;
2017en
ABI
Abstract
The ability to learn tasks in a sequential fashion is crucial to the development of artificial intelligence. Until now neural networks have not been 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 that 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 a hand-written digit dataset and by learning several Atari 2600 games sequentially.
Identifiers
Citations and references
Cited by 30 references