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.'"
The comments on TFA are a bit depressing though...
axilmar - You MIT guys don't realize how simple AI is. 2010-03-31 04:57:47
Until you MIT guys realize how simple the AI problem is, you'll never solve it.
AI is simply pattern matching. There is nothing else to it. There are no mathematics behind it, or languages, or anything else.
You'd think people who were so so certain that sure AI is easy would be making millions selling AI's to big buisness but no....
I'd be interested if this approach to AI allows for any new approaches to strategy.
Actually, axilmar hit it on the nail. There's more than one nail here, but that's not bad at all.
The next nail is "What patterns are *salient*". This is the billion dollar question in AI.
We hit *that* nail around 2003. In fact we're several nail further along....
I'm part of the crowd that thinks AI is much simpler than most people think. It's still not trivial.
But there's a *big* difference between a project to "tell the computer everything about the world
in first order predicate calculus" and "Figuring out how learning in the brain might work,
implementing that in a computer to test it, figuring out what might be wrong, and repeat the process
until we have something that is capable of learning anything we tell it roughly the way humans do".
The first approach is doomed to fail for reasons explained on my website below, including the simple
reason that the everyday mundane world is more complex than we think. Any ontology or semantic
web based project is thus doomed to fail.
The latter is "only" hampered by the fact that we haven't tried it yet. Attacking it from the
neuroscience angle is one way, but it's actually *easier* to attack it from the Epistemology
angle. "How is it possible to learn *anything*? What is it possible to learn at all? How *might*
the brain go about doing what it does? How could we duplicate it in a computer to see if
the theory is correct?" Repeat until we succeed.
A million man-years has been wasted on 20th Century style AI. We have so far put
10 person-years into 21st Century AI. To wit:
I (and my company) have been working on our idea of how this is supposed to be done since 2001
and though we have some interesting results and many insights we haven't been able to demonstrate
effects that are stronger than what you can do with regular programming. We have good benchmarks
but we're currently at 80-85% on tasks where regular programming can do 95% and humans 99.99%
but we're slowly improving. And as opposed to *many* AI projects, we are writing code, running
experiments daily (and overnight), have built our own extra-large computers (32 GB RAM linux systems)
etc. We are attempting to learn human languages (any language) by unsupervised training by simply
reading books (Jane Austen, in our case). We have good semantic level reading comprehension
tests that can be completely automated and work at *very low levels of IQ and reading comprehension*.
We've funded all of this work ourselves and hope to leverage this effort (once we get it to work :-)
into a market leading position on various semantic technologies including web search support technologies,
true semantic search, and superior speech understanding. Ask me for a Use Cases and Markets document.
When comparing the ideas in my company to those of almost all other AI research, *including TFA*,
I'd like to think that *we* at least got the Most Significant Bit correct. And we feel sad that most people
that are entering the field of AI today are being taught the wrong things, perpetuating the old myths and
mistakes and thereby guaranteeing we won't get decent AI any time soon.
I have an unpublished article that I'm trying to get into some mainstream magazine at
http://syntience.com/AIResearchInThe21stCentury.pdf - feel free to peek at it.
It's not a direct response to the MIT article but argues a different angle and aims
at roughly the same audience (you!).
If you want more info beyond that, then check out our other online resources:
Theory and motivational site (2 years old) : http://artificial-intuition.com/
Video site (latest insights, more detailed info) http://videos.syntience.com/ (or go to Vimeo.com and search for "syntience") Axilmar will enjoy "Models vs. Patterns" video.
Blog: