It is possible to attain “superhuman performance” in yasks without the need for human intervention in giving out rules by DeepMind’s latest AI program.
Mastery in dozens of old Atari video games, chess, and the Asian board games of Go and Shogi, has been achieved by MuZero, just like the research hub’s earlier artificial intelligence agents.
However this latest AI program had to work out their rules for itself, unlike its predecessors.
The prohram is already being put into practical use to find out ways to encode videos which could significantly bring down costs for YouTube.
“The real world is messy and complicated, and no-one gives us a rulebook for how it works,” DeepMind’s principal research scientist David Silver said in a television interview.
“Yet humans are able formulate plans and strategies about what to do next.
“For the first time, we actually have a system which is able to build its own understanding of how the world works, and use that understanding to do this kind of sophisticated look-ahead planning that you’ve previously seen for games like chess. [It] can start from nothing, and just through trial and error both discover the rules of the world and use those rules to achieve kind of superhuman performance.”
The work marked a “significant step forward”, said Wendy Hall, professor of computer science at the University of Southampton and a member of the government’s AI council,
Hall however also raised concerns.
“The results of DeepMind’s work are quite astounding and I marvel at what they are going to be able to achieve in the future given the resources they have available to them,” she said.
“My worry is that whilst constantly striving to improve the performance of their algorithms and apply the results for the benefit of society, the teams at DeepMind are not putting as much effort into thinking through potential unintended consequences of their work,” she said.
“I doubt the inventors of the jet engine were thinking about global pollution when they were working on their inventions. We must get that balance right in the development of AI technology,” she added.
Details of MuZero were first published by London-based DeepMind in 2019 but the firm waited till the publication of a paper in the journal Nature in order to discuss in publicly.
This latest AI program highlights the success of the firm in deep reinforcement learning that is a technique that uses many-layered neural networks that help machines to teach new skills to themselves, through a process of trial and error. In this process the machines received “rewards” for success instead of being told what to do.
DeepMind is owned by the same parent as Google.
DeepMind was being used already to try and find out an new way of video compression, Dr Silver said.
“If you look at data traffic on the internet, the majority of it is video, so if you can compress video more effectively you can make massive savings,” he explained. “And initial experiments with MuZero show you can actually make quite significant gains, which we’re quite excited about.”
This new AI technology can be a big money-saver for YouTube – the largest video sharing platform of the world, owned by Google.
Since MuZero only tries to model aspects of the environment that are deemed to be important for making a decision instead of using a wider approach, therefore it has beene successful, the firm believes.
“Knowing an umbrella will keep you dry is more useful to know than modelling the pattern of raindrops in the air,” it explains in a blog.
(Adapted from BBC.com)