DeepMind: Google 's AI company DeepMind , based in London, has created two different types of AI that can use their imagination to plan the future and perform tasks with a higher success rate than AIs without imagination.
Are you confused?
In a post on their website, DeepMind researchers give a brief overview of their “new family of approaches to imagination-based design.” So-called imagination agents, or I2As, use an internal imagination encoder that helps the AI decide which predictions about the future are and are not useful.
Researchers argue that AI's imagination is crucial for tackling real-world problems, where sometimes you need to mentally try out several possible outcomes of your actions to predict or decide which one is best.
Recently, DeepMind founder Demis Hassabis wrote a paper published in Neuron about how the development of general-purpose AI depends on understanding and encoding human abilities like imagination, curiosity, and memory. So his company appears to be making progress on at least one of those human abilities.
The I2A vehicles were tasked with testing different situations for scientists to observe their predictive abilities. Among the tests was a Sokoban puzzle game and a spaceship navigation game. Sokoban is a puzzle in which a small alien must push boxes into the right place – he can’t pull boxes, though, so one wrong move can mean you lose the entire round.
The I2A-carrying machines performed much better than the machines that didn't use I2A. They learned how to navigate the puzzles with less experience by extracting more information from the internal simulations they were running.
Of course, the type of imagination described in Demis Hassabis' paper doesn't even come close to the way humans use their imagination, but the experiment shows that AIs can perform much better when they imagine different scenarios before acting.
As Hassabis told Neuron, creating machines with imagination that can rival what we can do “is perhaps the most difficult challenge for AI research: to build an agent that can plan hierarchically, be truly creative, and produce solutions to challenges here and now without the need for a human mind.”
Step by step, we will probably get there.
