Google Creates AI Program That Uses Reasoning To Navigate the London Tube (theguardian.com)

An anonymous reader quotes a report from The Guardian: Google scientists have created a computer program that uses basic reasoning to learn to navigate the London Underground system by itself. Deep learning has recently stormed ahead of other computing strategies in tasks like language translation, image and speech recognition and even enabled a computer to beat top-ranked player, Lee Sedol, at Go. However, until now the technique has generally performed poorly on any task where an overarching strategy is needed, such as navigation or extracting the actual meaning from a text. The latest program achieved this by adding an external memory, designed to temporarily store important pieces of information and fish them out when needed. The human equivalent of this is working memory, a short-term repository in the brain that allows us to stay on task when doing something that involves several steps, like following a recipe. In the study, published in the journal Nature, the program was able to find the quickest route between underground stops and work out where it would end up if it traveled, say, two stops north from Victoria station. It was also given story snippets, such as "John is in the playground. John picked up the football." followed by the question "Where is the football?" and was able to answer correctly, hinting that in future assistants such Apple's Siri may be replaced by something more sophisticated. Alex Graves, the research scientist at Google DeepMind in London who led the work, said that while the story tasks "look so trivial to a human that they don't seem like questions at all," existing computer programs "do really badly on this." The program he developed got questions like this right 96% of the time.

Comments

What's the reason for reason?

[Reaperducer]

Since when has reason had anything to do with navigating the London tube system?

And why do so many cities use nautical themes for their stored payment cards?
London: Oyster
Hong Kong: Octopus
Seattle: Orca
Montreal: Opus
San Francisco: Clipper
Bolton: Squid
Merseyside: Walrus
Wellington: Snapper

Re:What's the reason for reason?

[_merlin]

HK Octopus predates London Oyster. The name comes from the Chinese name of the card "Baat Daaht Tung", literally "Eight-Arrived Passage" but figuratively "Access All Areas", eight referring to the cardinal and semicardinal points of the compass. Octopus is a catchy English name with a reference to eight in it.

Re:What's the reason for reason?

[gnick]

Opus means 'Work' (see 'Magnum Opus', so nothing to do with water.

Opus is nautical. He's a penguin. See Bloom County.

Re:I can't answer the question either

[sittingnut]

human language is ambiguous and there is never a single meaning. there are meanings(plural) depending on each of the contexts of what is described, who is describing it, who is receiving it, what is perceived, what is assumed, etc.etc..

any so called ai boiling down something to a single meaning, even a useful meaning(which seems to be the aim rather than achievement of such "ai" so far), is simply dumbing it down (stripping its intelligence if you will) to level of maths.

would be fun to feed this "ai" some good poetry and ask for meaning.

Re:I can't answer the question either

[Jon Peterson]

Human language is brilliantly imprecise.

It's a feature not a bug. A really big feature.

Feral or nurtured

[petes_PoV]

From the article:

an approach called deep-learning, in which the program learns how to do tasks independently rather than being pre-programmed with a set of rules by a human.

So while humans learn many of life's most important things: how to use a fork, how to speak (and occasionally: listen), how to clothe ourselves. hpw to obey the law, by being "programmed" with a set of rules by a human, this machine figured it out by itself.

I can see that this has application in some areas, but to be a good member of society shouldn't we want certain aspects of co-existence, values and social behaviour to come from rules, rather than each person or computer coming too its own conclusion about co-operating?

Re:Feral or nurtured

[AthanasiusKircher]

So while humans learn many of life's most important things: how to use a fork, how to speak (and occasionally: listen), how to clothe ourselves. hpw to obey the law, by being "programmed" with a set of rules by a human, this machine figured it out by itself.

I don't think you understand human learning very well at all. Most human learning is not conveyed through some sort of "rules," but rather is extrapolated from patterns humans notice. If you've ever been around a small child learning language, you quickly realize how grammar ACTUALLY works and is learned -- and it's NOT through formal rule-based systems. Kids just try various utterances, and when they get what they want, they notice success and try those again. Parents and other adults generally make subtle corrections, but there are rarely explanations with formal "rules," just repeating the correct sentence... leaving it for the child to take in the pattern.

That's actually how most human learning takes place. Well, any learning that actually requires skill or thought. Rules are good for things that have to be memorized precisely. But most of the social conventions we learn are things that we pick up through observation, imitation, and trying things out with minor corrections, rather than being given a set of formal rules to follow. Once we get beyond early childhood, rule sets can be introduced, but there's often a lot of implicit information beyond the text of the rules that you also tend to pick up through practice, observation, etc.

Various AI algorithms still tend to operate on similar feedback loops -- the "deep learning" algorithms, as I understand them, generally are fed large amounts of data and need to extrapolate common patterns, roughly like a small child does with language (though the AI systems do this is a MUCH more primitive way). Sometimes AI systems are also given feedback on successes to help them tailor their extrapolated patterns, again in a similar way to a parent verifying whether a kid's sentence is correct or making a correction.

I can see that this has application in some areas, but to be a good member of society shouldn't we want certain aspects of co-existence, values and social behaviour to come from rules, rather than each person or computer coming too its own conclusion about co-operating?

Sure. Is anyone actually suggesting that we abandon all "rules" or that we fail to provide AI with ANY rules? I'd say by far the biggest barrier to AI research in the early days was the idea that formal systems of rules could be used to solve various AI problems (like language processing, pattern recognition, etc.). It was a failure to understand that humans don't work like that either -- again, we may teach certain basic rules to other humans, but an advanced level of skill or application generally requires a lot of "intuition" and generalization that's not stated formally.

So AI researchers have thankfully abandoned the idea that we can ever get to "intelligence" through a formal rule system, and they're trying out these other approaches. Personally, I think they make a little too much out of the "intelligence" of things like "deep learning," which are still pretty crude compared to what your average 2-year-old could do with data input... but still, it's a lot better than trying to teach through rules. Yet that doesn't mean we can't have constraints or "rules" when it's actually helpful...

Say what? Fuzzy logic and Beyond 2000.

[wjcofkc]

I remember an episode of Beyond 2000 around 1992 or so. It featured a fuzzy logic system the Japanese had implemented in their own subway\tram system. Each train and each stop had the system and they were networked together so that the trains and stops could work together to maximize efficiency. I remember be amazed by it. This sounds awfully similar,

Although it could not respond to questions about a football.

To which I say: what the fuck? If I am on a rail system I want the computer to be thinking about its job, not a fucking football.

Re:This is not AI

[tehcyder]

This is impressive and all, but I won't believe in AI until I see a computer that can win at Mornington Crescent

Or Numberwang.

Re:We know who it will vote for

[Anonymous Coward]

Yes of course, pretty much anything more intelligent than a rock would.