DeepMind, the artificial intelligence (AI) company now owned by all-seeing internet giant Google, has made a big step towards creating AI that can 'think like a human', according to the company.
The company claims that it has tuned its AI algorithms to enable the transfer of skills picked up accomplishing one task in order to apply to subsequent tasks.
The key is that DeepMind has been taught not to forget useful information, but instead to store it so that it can be drawn upon later.
"If we're going to have computer programs that are more intelligent and more useful, then they will have to have this ability to learn sequentially," DeepMind's James Kirkpatrick told The Guardian.
The problem comes from the fact that while computers can be programmed or learn to do something, when they've mastered it they get taught something else - and forget what it was they were originally taught, or are unable to apply it to subsequent tasks.
Now, DeepMind researchers have found a way to cure this so-called ‘catastrophic forgetting', instead putting it to good use. Indeed, without that skill, there can't be a general-purpose artificially intelligent being.
"Humans and animals learn things one after the other and it's a crucial factor which allows them to learn continually and to build upon their previous knowledge," said Kirkpatrick.
The secret is in getting the neural network to evaluate what it's learned in a task, and "hard-wire" the information that was most useful. The results were noticeable.
When given 10 Atari 2600 video games to learn (such as Space Invaders), the new technique enabled DeepMind's machine to perfect seven of 10 games. Without the new processes, it barely learned one in the same period.
Just like humans, AI doesn't know what it doesn't know and, quite often, misses the information it needs to remember. Just like humans, it doesn't always learn its lesson.
"We know that sequential learning is important, but we haven't got to the next stage yet, which is to demonstrate the kind of learning that humans and animals can do. That is still a way off. But we know that one thing that was considered to be a big block is not insurmountable," said Kirkpatrick.
Just spent a year working on them? Too bad, Intel's lost interest
Sony factory in Wales now making 100,000 Raspberry Pis every week
38-year-old Alexander Vinnik faces up to 55 years in jail
Threadripper also available from today if you want a lot more power - but you'll have to wait for the motherboards to appear