To test our algorithm, we exposed an agent to Atari games sequentially. Learning an individual game from the score alone is a challenging task, but learning multiple games sequentially is even more challenging as each game requires an individual strategy. As shown in the figure below, without EWC, the agent quickly forgets each game after it stops playing it (blue). This means that on average, the agent barely learns a single game. However, if we use EWC (brown and red), the agent does not forget as easily and can learn to play several games, one after the other.
Source: https://deepmind.com/blog/article/enabling-continual-learning-in-neural-networks
More from Deep Mind - Latest Post
We are very excited to announce the launch of DeepMind Health
Source Node: 802635
Time Stamp: Feb 23, 2016
Announcing the Partnership on AI to Benefit People & Society
Source Node: 801730
Time Stamp: Sep 27, 2016
TF-Replicator: Distributed Machine Learning for Researchers
Source Node: 749908
Time Stamp: Mar 7, 2019
Bringing the best of mobile technology to Imperial College Healthcare NHS Trust
Source Node: 799445
Time Stamp: Dec 21, 2016
Applying machine learning to mammography screening for breast cancer
Source Node: 769216
Time Stamp: Nov 23, 2017