DeepMind Develops AI With An Imagination
STAY ON TARGET
Google is going full Westworld by giving robots the ability to imagine and plan.
In the same way a person would look at a glass of water on the edge of a table and consider whether it might fall, this form of deliberative reasoning, or “imagination,” is important for algorithms to develop and evolve.
DeepMind’s AlphaGo program time and again demonstrated an AI’s ability to look to the future, accurately predicting moves to beat even the best human players.
But those “internal models,” according to Google, work because things like Go have clearly defined rules and allow for predictable outcomes.
“The real world is complex, rules are not so clearly defined, and unpredictable problems often arise,” a DeepMind blog post said. “Even for the most intelligent agents, imagining in these complex environments is a long and costly process.”
One in which Google is willing to invest.
In two recently published papers, Team DeepMind describe new approaches for imagination-based planning, including an “imagination encoder”—a neural network that learns to extract useful information and ignore everything else.
To assess these new architectures, the AI played puzzler Sokoban (video below) and a spaceship navigation game—both of which require forward planning and reasoning, “making them the perfect environment to test our agents’ abilities,” DeepMind said.
By allowing the network to try each procedurally generated level only once, researchers can see how the machine mulls over different strategies “in its head” before making an actual attempt.
In both cases, the imagination-augmented agents outperformed their baseline counterpart “considerably,” according to the blog. Add to that a “manager” component, and AI can learn to solve tasks more efficiently in fewer steps.
“If the computational cost of imagination increases with each action taken, the agent imagines the effect of multiple chained actions early, and relies on this plan later without invoking imagination again,” DeepMind said.
In other words: now that artificial intelligence has devised their own unintelligible language and the ability to plan ahead, we’re all doomed.