Google is teaching it’s AI to dream. And we are a step closer to Westworld’s thinking, dreaming herd of androids in human flesh.
Using an UNsupervised REinforcement and Auxiliary Learning tool (UNREAL), researchers in Google’s DeepMind are able to apply the same situations that cause animals to dream in artificial intelligence. With the same application Google has used the same deep reinforcement learning methods that it had used to teach its AI to master the game Go in January.
The ‘dream’ as is explained in the the paper ‘Reinforced Learning with UnsupervisedAuxiliary Tasks’ by the DeepMind researchers is another effective mechanism to ‘experience replay’ which has proven to be an efficient way of improving both data efficiency and stability of deep reinforcement learning algorithms. The intention is to help the artificial intelligence agent to constantly make value judgments between good decisions and bad. Another quick method of teaching the AI to score reward points in the game Labyrinth (a game developed by DeepMind researchers where the player needs to navigate a maze and collect apples), the dreams are meant to reinforce the agent about maximising the rewards and changing the environment in the process.
In a recent paper published by the Deep Mind researchers, the dreams seem to have achieved a major leap in the system’s performance and achievements.
(Reposted from Linkedin Pulse: https://www.linkedin.com/pulse/why-google-teaching-its-ai-dream-tannistho-ghosh?trk=pulse_spock-articles)