Slashdot Mirror
MIT Finds 'Grand Unified Theory of AI'
aftab14 writes "'What's brilliant about this (approach) is that it allows you to build a cognitive model in a much more straightforward and transparent way than you could do before,' says Nick Chater, a professor of cognitive and decision sciences at University College London. 'You can imagine all the things that a human knows, and trying to list those would just be an endless task, and it might even be an infinite task. But the magic trick is saying, "No, no, just tell me a few things," and then the brain — or in this case the Church system, hopefully somewhat analogous to the way the mind does it — can churn out, using its probabilistic calculation, all the consequences and inferences. And also, when you give the system new information, it can figure out the consequences of that.'"
Tell me about you to build a cognitive model in a fantastically much more straightforward and transparent way than you could do before.
What is the difference between an endless task and an infinite task?
Sounds a bit like a journalists brain to me ... NO NO let me make up the rest of the Story
Told that the cassowary is a bird, a program written in Church might conclude that cassowaries can probably fly. But if the program was then told that cassowaries can weigh almost 200 pounds, it might revise its initial probability estimate, concluding that, actually, cassowaries probably can’t fly.
But you just induced a bunch of rules I didn't know were in your system. That things over 200 lbs are unlikely to fly. But wait, 747s are heavier than that. Oh, we need to know that animals over 200 lbs rarely have the ability of flight. Unless the cassowary is an extinct dinosaur in which case there might have been one ... again, creativity and human analysis present quite the barrier to AI.
Chater cautions that, while Church programs perform well on such targeted tasks, they’re currently too computationally intensive to serve as general-purpose mind simulators. “It’s a serious issue if you’re going to wheel it out to solve every problem under the sun,” Chater says. “But it’s just been built, and these things are always very poorly optimized when they’ve just been built.” And Chater emphasizes that getting the system to work at all is an achievement in itself: “It’s the kind of thing that somebody might produce as a theoretical suggestion, and you’d think, ‘Wow, that’s fantastically clever, but I’m sure you’ll never make it run, really.’ And the miracle is that it does run, and it works.”
That sounds familiar ... in both the rule based and probabilistic based AI, they say that you need a large rule corpus or many probabilities accurately computed ahead of time to make the system work. Problem is that you never scratch the surface of a human mind's lifetime experience though. And Chater's method, I suspect, is similarly stunted.
I have learned today that putting 'grand' and 'unified' at the title of an idea in science is very powerful for marketing.
My work here is dung.
Since the actual summary seems to involve a fluff filled soundclip without anything useful, here's the run down of the article.
1) We first tried to make AIs that could think like us by inferring new knowledge from existing knowledge.
2) It turns out that teaching AIs to infer new ideas is really freaking hard. (Birds can fly because they have wings, mayflies can fly because they have wings, helicopters can... what??)
3) We turned to probability based AI creation: you feed the AI a ton of data (training sets) and it can go "based on training data, most helicopters can fly."
4) This guy, Noah Goodman of MIT, uses inferences with probability: he uses a programming language named "Church" so the computer can go
"100% of birds in training set can fly. Thus, for a new bird there is a 100% chance it can fly"
"Oh ok, penguins can't fly. Given a random bird, 90% chance it can fly. Given random bird with weight to wing span ratio of 5 or less, 80% chance." and so on and so forth.
5) Using a language that mixes two separate strategies to train AIs, a grand unified theory of ai (lower case) is somehow created.
6) ???
7) When asked if sparrows can fly, the AI asks if it's a European sparrow or an African sparrow, and Skynet ensues.
Signatures are the new names.
This kind of probabilistic inference approach with "new information" [evidence] being used to figure out "consequences" [probability of an event happening] sounds very similar to Bayesian inference/networks.
I would be interested in knowing how does this approach compares to BN and the Transferable Belief Model (or Dempster–Shafer theory) which itself addresses some shortcomings of BN.
Ubuntu is an African word meaning 'I can't configure Debian'
HYPE. More grand unified hype. The "grand unified theory" is just a mashup of old-days rules & inferences engines thrown in with probabilistic models. Hyperbole at its finest, to call it a grand unified theory of AI. Where are connotations and framing effects? How does working short term memory interact with LTM and how does Miller magic number show up? How can the system understand that "john is a wolf with the ladies" without thinking that john is hairy and likes to bark at the moon? I could go on but feel free to fill in the blanks. So long and thanks for all the fish MIT.
q.v. Alonzo Church
From the article:
As a research tool, Goodman has developed a computer programming language called Church — after the great American logician Alonzo Church
Your comment fits the criteria of Flamebait and Offtopic, but definitely NOT Funny.
Thanks, Slashdot's mandatory comment waiting period! I'm sure glad I was late to this party.
DRM: Terminator crops for your mind!
I looked at the documentation of this "Church Programming language". Scheme and most other Lisp derivatives have been around longer and can do more. This is neither news nor a revolutionary discovery.
The way the author wrote the article, it seems like nothing different from an expert system straight from the 70's, e.g. MYCIN. That one also uses probabilities and rules; the only difference is that it diagnoses illnesses, but that can be extended to almost anything.
Probably the only contribution is a new language. Which, I'm guessing, probably doesn't deviate much from, say, CLIPS (and at least THAT language is searchable in Google... I can't seem to find the correct search terms for Noah Goodman's language without getting photos of cathedrals, so I can't even say if I'm correct)
AI at this point has diverged so much from just probabilities and rules that it's not practical to "unify" it as the author claims. Just look up AAAI and its many conferences and subconferences. I just submitted a paper to an AI workshop... in a conference ... in a GROUP of co-located conferences ... that is recognized by AAAI as one specialization among many. That's FOUR branches removed.
Pet peeve: Profane people propagating perfunctory pedantry.
Wow, as someone working in this domain I can say that this article is full of bold conjectures and shameless self-advertising. For a start, (1) uncertain reasoning and expert systems using it is hardly new. This is a well-established research domain and certainly not the golden grail of AI. Because, (2) all this probabilistic reasoning is nice and fine in small toy domains, but it quickly become computationally intractable in larger domains, particularly when complete independence of the random variables cannot be assured. And for this reason, (3) albeit being a useful tool and important research area, probabilistic reasoning and uncertain inference is definitely not the basis of human reasoning. The way we draw inference is much more heuristic, because we are so heavily resource-bound, and there are tons of other reasons why probabilistic inference is not cognitively adequate. (One of them, for example, is that untrained humans are incapable of making even the simplest calculations in probability theory correctly, because it is harder than it might seem at first glance.) Finally, (5) there are numerous open issues with all sorts of uncertain inference, ranging from certain impossibility results, over different choices that all seem to be rational somehow (e.g. DS-belief vs. ranking functions vs. probability vs. plausibility measures and how they are intereconnected with each other, alternative decision theories, different rules of dealing with conflicting evidence, etc.) to philosophical justifications of probability (e.g. frequentism vs. Bayesianism vs. propensity theory and their quirks, justification of inverse inference, etc).
In a nutshell, there is nothing wrong with this research in general or the Church programming language, but it is hardly a breakthrough in AI.
This is just a library for Scheme. It does the same things that have been done before. In scheme.
Move along.
Again, as I bring up often with AI researchers, we as humans evolved over millions of years (or were created, doesn't matter) from simple organisms that encoded information that built up simple systems into complex systems. AI, true AI, must be grown, not created. Asking the AI if a Bat is a mammal and can fly can a squirrel? ignores a foundation of development in intelligence, our brains were created to react and store, not store and react from various inputs.
Ask an AI if the stove is hot. It should respond "I don't know, where is the stove?" Rather AI would try and make an inference based on known data. Since there isn't any the AI on a probablistic measure would say that blah blah stoves are in use at any given time and there is a blah blah blah. A human would put thier hand (a senor) near the stove and measure the change, if any in temperature and reply yes or no accordingly. If a human cannot see the stove, and had no additional information either a random guess is in order or a "I have no clue." response of some sort. The brain isn't wired to answer a specific question but it is wired to correlate independent inputs to draw conclusions based on the assembly and interaction of data and infer and deduce answers.
Given a film of two people talking a computer with decent AI would catagorize objects, identify people versus say a lamp, determine the people are engaged in action (versus a lamp just sitting there) making that relevant, hear the sound coming from the people then infer they are talking (making the link.) Then paralell the computer would filter out the chair, and various scenery in the thread now processing "CONVERSATION". The rest of the information is stored and additional threads may be created as the environment generates other links but if the AI is paying attention to the conversation then the TTL for the new threads and links should be short. When the conversation mentions the LAMP the information network should link the LAMP information to the CONVERSATION thread and provide the AI additional information (that was gathering in the background) that travels with the CONVERSATION thread.
Now the conversation appears to be about the lamp and wheather it goes with the room's decor. Again the links should be built adding, retroactively the room's information into the CONVERSATION thread (again expiring information that is irrelivant to a short term memory buffer) and ultimately since visual and verbal queues imply that the AI's opinion is wanted should result in the AI blurting out, "I love Lamp."
In case you missed it, this was one long Lamp joke...
-=[ Who Is John Galt? ]=-
We call it being Fashionably Redundant.
The enemies of Democracy are
This is embarrassing. MIT needs to get their PR department under control. They're inflating small advances into major breakthroughs. That's bad for MIT's reputation. When a real breakthrough does come from MIT, which happens now and then, they won't have credibility.
Stanford and CMU seem to generate more results and less hype.
FTA: "In the 1950s and 60s, artificial-intelligence researchers saw themselves as trying to uncover the rules of thought. But those rules turned out to be way more complicated than anyone had imagined. Since then, artificial-intelligence (AI) research has come to rely, instead, on probabilities -- statistical patterns that computers can learn from large sets of training data."
From the viewpoint of Jaynes and many Bayesians, probability IS simply the rules of thought.
I've always enjoyed reading about AI, and like many here have done some experiments on my own time.
This week I've been looking for a simple state modeling language, for use in fairly simple processes, that would tie into some AI.
I wasn't really that impressed with anything I found, so when I saw the headline, I jumped to read the article.
Unfortunately, this is a step in the right direction, but not all that clear to write or maintain, and probably too complex for what I need to do.
The cleanest model to do these types of things I've found is the 1896 edition of Lewis Caroll's Symbolic Logic. (Yes, the same Lewis Caroll that wrote Alice in Wonderland).
Maybe "axilmar" is more interested in the ethics of AI than commercial gain. Maybe "axilmar" is getting ready to create a free cylon project that will eventually be completed by a Scandinavian student. Although "axilmar" never completes his own project, he'll consistently complain about the name of the newer, complete, more popular project and its derivatives. "Axilmar's" efforts will shift to creating and running the Free Cylon Foundation (or FCF). He spends the majority of his time give strikingly similar speeches over and over around the world. Despite the absolute consistency of his message he and by association the FCF are increasingly seen as a fringe political group. Despite the FCF's best efforts to promote the rights of the Cylons and hope for peaceful coexistence, the world's civilization eventually falls into chaos as the Cylons engage in war against humanity. Not long before his death at the hands of a cylon as he tries to convince the cylon that he's more righteous than other humans, "axilmar" is overheard muttering some complaint about a printer...
Do a search for articles with MIT in the title and you'll find that's a pretty common story here.
Pretty much.
The pragmatic answer to the Chinese Room problem is "Who gives a fuck? There's no way to prove that our own brains aren't basically Chinese Rooms, so if the only difference between a human intelligence and an artificial one is that we know how the artificial one works, why does it matter?"
But really, identifying patterns, and then inferring further information from the rules those patterns imply, is a pretty good behavior.
The enemies of Democracy are
I have a child. When I watch her learn its totally rules based. Also, very importantly when she is told that her existing rules don't quite describe reality she is quick to make a new exception (rule). Since she's young her mind is flexible and she doesn't get angry when its necessary to make an exception. The new rule stands until a new exception comes up.
eg in english she wrote "there toy" since she wasn't familiar with the other there's. She was corrected to "their toy". But of course, there is still "they're".
Comment removed based on user account deletion
Excuse me.
The technical term is Hurd-Cylon, okay? Please use the correct term from now on.
Thanks,
Axilmar Stallman
The pragmatic answer to the chinese room is that the non-chinese-speaking person in the room in combination with the book of algorithmic instructions, considered together as a system, does understand chinese.
Searle's mistake is an identity error - the failure to see that a computer with no software is not the same identity as a computer with software loaded inot it. The latter quite possibly could understand chinese (or some other domain) while the former most definitely does not.