Could this be another leap for AI software?

Don't know whether this has been mentioned here on the forums already, but I'm gonna go ahead and post this anyway. This is too darn interesting to just let blow past ;p

Recently, IBM (yup.. the same folks who did the Deep Blue vs Kasparov gig in 1997) recently came up with another challenge. They designed a system capable of processing natural language and answering questions posed in it.

Named 'Watson', after IBM's first president, it had the following features:
"Watson is made up of a cluster of ninety IBM Power 750 servers (plus additional I/O, network and cluster controller nodes in 10 racks) with a total of 2880 POWER7 processor cores and 16 Terabytes of RAM. Each Power 750 server uses a 3.5 GHz POWER7 eight core processor, with four threads per core. The POWER7 processor's massively parallel processing capability is an ideal match for Watson's IBM DeepQA software which is embarrassingly parrallel (that is a workload that executes multiple threads in parallel)."
It also had the equivalent to 200 million pages of structured and unstructured content consuming four terabytes of disk storage in total (yes, including the full text of wikipedia) and all that without being connected to the internet.

Holy shite? Yes indeed...

As for the challenge? It would participate in the game of Jeopardy alongside two of its all-time best human players.

And how it ended, you ask? Watson bested them both, and received the first prize of $1 million, of which IBM readily donated 100%, divided between two charities, the first one being 'World Vision', a humanitarian relief organization, and 'World Community Grid', which uses spare computing power donated to it to help in projects dealing with, for example, AIDS and cancer research, in a huge public computing grid.


I'm currently at university studying AI myself, so I got terribly excited when I saw this stuff. I thought it was absolutely amazing, awesome, and conjured a sheepish grin on my face for a long long time afterwards. Natural language processing, dealing with ambiguity, parrallel processing lines of algorithms, machine learning.. It made my insides tingle.. This might just be the first step towards a real-life EDI :o
Whatever the case might be, I want one :wink:


Anyhow.. Here are some vids (where you were all waiting for ;p):
YouTube - Jeopardy! IBM Watson Day 1 (Feb 14, 2011) Part 1/2 (part one of the first day of the challenge)
YouTube - Watson Jeopardy Full Final Game (Day 3 Feb 16) (day three of the challenge)
YouTube - How IBM's Watson supercomputer wins at Jeopardy, with IBM's Dave Gondek (a more detailed look at how Watson works)
YouTube - IBM Watson: Final Jeopardy! and the Future of Watson (the experience of the whole episode, what we have learnt and what the future may hold, from an IBM perspective)
http://www.youtube.com/watch?v=5Gpaf...eature=related - a PBS documentary on Watson


Aaand for more info, here are some useful pages:
Watson (artificial intelligence software) - Wikipedia, the free encyclopedia
World Community Grid - Wikipedia, the free encyclopedia
World Vision - Wikipedia, the free encyclopedia


Sooo that's it.. Have fun viewing, discussing, and whatevering. Hope to see y'all alive and well when the robot overlords take over.

Have a great day!

-Tim

I don't reckon there have ever been any leaps for AI software.

^Nor could there be. Nevertheless, software incrementally improves with every passing decade. Sooner or later, it will be almost human.

And let's not discount the effect of hardware. Say you write some horrible, ugly, AI program that's a million times too slow to simulate a human. It'll never work, right? Just wait 20 years. Now computers are a million times faster, and suddenly your ugly program is fast enough.

What I meant is that there has never been a significant AI advancement of any kind, slow or not, and your post nicely demonstrates the misconception: the approach is always brute force, which has nothing to do with intelligence. AI needs a conceptual enlightenment, not more and more CPU thrown at it.

Quote:

---

Originally Posted by Xei

What I meant is that there has never been a significant AI advancement of any kind, slow or not, and your post nicely demonstrates the misconception: the approach is always brute force, which has nothing to do with intelligence. AI needs a conceptual enlightenment, not more and more CPU thrown at it.

---

The brain is brute force. That's proof that brute force can work.

...no it isn't?

Quote:

---

Originally Posted by Xei

...no it isn't?

---

...yes it is? If you could simulate the interactions of all the molecules in a brain, then that would be very much a brute force approach.

Oh, cynical one, I know what you're thinking: but simulating molecules is no better than actually having a brain!

Yes, but we could probably go up at least a few levels of abstraction without losing anything. For example, if you have a robust set of rules for which conditions produced which proteins, you could avoid having to simulate protein folding. You could merely use a lookup table. Also, ion channels could be vastly simplified in software.

Even things like synapse growth could be simulated without actually reproducing the molecular-level interactions if you have a robust rule set for when a new synapse appears.

Ah, I think I misunderstood (though you were clear as mud). When you said the brain is brute force, I thought you meant the brain in nature is a brute force approach to intelligence. Did you mean that simulating the brain in silico would be a brute force approach to AI?

In that case I agree with you.

The point I was making is that computing power is necessary but far from sufficient. You have to simulate the correct program, be that program a brute force simulation of an entire brain, or a smart approach via understanding how the brain works and simulating an analogous model.

Indeed, brute force programs haven't been tried until very recently for the simple reason that no one has had the computing power. But now, projects like Blue Brain are beginning to get into the realm of possibly, in the next few years, starting to take a bite out of the problem. Blue Brain itself is still too high level and too slow, but the successor to Blue Brain should be much more impressive.

I think developing AI that fits the structure of computing hardware will be more rewarding, than trying to replicate everything biological brains do. As Xei says, an analogous model.

Anybody who knows how things are going with neural networks?

Quote:

---

Originally Posted by Marvo

I think developing AI that fits the structure of computing hardware will be more rewarding, than trying to replicate everything biological brains do. As Xei says, an analogous model.

Anybody who knows how things are going with neural networks?

---

Like I said, there's many levels of abstraction. If you try to write an AI using a series of IF-THEN statements, that's merely the highest level of abstraction. But there are many levels below that. For example, neural nets.

Quote:

---

Originally Posted by cmind

Like I said, there's many levels of abstraction. If you try to write an AI using a series of IF-THEN statements, that's merely the highest level of abstraction. But there are many levels below that. For example, neural nets.

---

What do you mean by levels of abstraction? I'm not following.

I agree with the people who said this isn't really a leap in anything. Sounds like a big database. Having a big database doesn't make it any smarter. Heck he isn't even listening to the questions, which would be pretty cool for advancement in information input. Instead it is taking a typed message and then just looking up the information in its huge database.

Regardless of how big the database may be, or how quickly it can find stuff, it is still just a database. If instead of Jeopardy you played am I smarter than a 5th grader, the computer could of beat people 10 years ago. It just has a bigger storage of data now.

Quote:

---

Originally Posted by Marvo

What do you mean by levels of abstraction? I'm not following.

---

Abstraction is when you simplify your model in order to gain higher understanding. Yes, it's hard to describe.

So for example, computers run on transistors. Well, actually, it's protons and electrons. An abstraction of this atomic model would be to start talking about transistors. Then, you can group transistors in certain ways to form gates. That's another abstraction. If you group gates together into registers and logic units, that's another abstraction. Then the next level of abstraction is to talk about machine code and assembly language instructions. After that comes compilers, then programming languages, and then at some point way above everything, you get to GUIs.

In the brain, we start with protons and electrons. Then we move up to atoms. Above that, it's small molecules. Those small molecules organize themselves into proteins. Proteins come together to form neurons. About 2-3 levels above individual neurons you get cortical columns. Above cortical columns we eventually get to brain lobes, and abstracting that, we can talk about behaviours.

Quote:

---

Originally Posted by Alric

I agree with the people who said this isn't really a leap in anything. Sounds like a big database. Having a big database doesn't make it any smarter. Heck he isn't even listening to the questions, which would be pretty cool for advancement in information input. Instead it is taking a typed message and then just looking up the information in its huge database.

Regardless of how big the database may be, or how quickly it can find stuff, it is still just a database. If instead of Jeopardy you played am I smarter than a 5th grader, the computer could of beat people 10 years ago. It just has a bigger storage of data now.

---

No. That's just incorrect. Go watch the PBS Nova episode. The machine didn't simply "look up" answers. Obviously. That would be retarded.

Well I was just watching the episode of Jeopardy, and it does look like it is just looking up answers. Questions with a lot "key" words in them confuse him. It seems very much like it is looking through a database and picking out what matches the best. So when multiple key words or phrases are used, it confuses him.

An example of this is the question, His victims include Charity Burbage, Mad Eye Moody, and Severus Snipe; He'd be easier to catch if you'd just name him! It showed the computer had no idea and though with the top choices of 37% chance it was harry potter, 20% chance voldermort and 8% chance Dumbledore. It make it really look as if it just took the names as key words and looked up in it's database and checked what had the most matches. Which is why harry potter came up more often, even though it wasn't even close to the right answer.

If you look at questions with dates or quotes in them however, it almost a 100% certain.

It guess doesn't appear to have anything to do with the difficulty of the question, but how it is worded and the number of key words or phrases in it. That isn't intelligence, that is looking up stuff in a database.

I agree with Alric, and that was the point I was originally making about brute force.

This has no more intelligence than the way Deep Blue played chess. That is to say, none.

Quote:

---

Originally Posted by Alric

It guess doesn't appear to have anything to do with the difficulty of the question, but how it is worded and the number of key words or phrases in it. That isn't intelligence, that is looking up stuff in a database.

---

Maybe you should look in the mirror, because that's not so different from how humans do it.

Indeed, there's not really anything elaborate about Watson, apart from the hardware it's made from. The software is not particularly fascinating.

ninja-edit: meant to complement Xei's post.

cmind, yes, we use experiences from our life to interpret situations we're in, but the way we process our information (and the way our mind is wired together) is much more elaborate than Watson.

We don't have a good handle at all on how humans come up with answers to trivia questions. But the very different patterns of errors committed by human contestants and by Watson suggest that they are approaching these problems in very different ways. Which shouldn't be the least bit surprising, because the makers of Watson never set out to design a program that would perform a human task the way that humans do it, they simply set out to design a program that would perform a human task as well as humans.

People do not think like that. Like in the example I gave, no one who read harry potter would think Harry potter killed the three people mentioned. The computer doesn't appear to understand the sentences at all, instead it is just chopping up the sentences into key words then looking them up in it's database. I think if you took the sentences and mixed the words up so they were in a random order, the computer would still get the same results, which would show it has no clue what it is reading.

Quote:

---

Originally Posted by cmind

Maybe you should look in the mirror, because that's not so different from how humans do it.

---

Unfounded assertion much? We don't have any clear idea how language works.

The next leap in AI is going to be real speech recognition.

Speech recognition is useless if the AI can't understand what words mean.

The AI will have "Chinese room" understanding, and that's quite sufficient as long as it's deep enough in its rule set. And when you get down to it, humans also only have "Chinese room" understanding.

The Chinese room is supposed to emulate an algorithm that can realistically converse in Chinese. A weaker Chinese room would at least be able to understand language but Alric's already demonstrated how this machine can't even approach being able to do that.

I hope you don't subscribe to Searle's actual philosophical argument, it's the biggest pile of rubbish I've ever heard. It's already been thoroughly and unsalvageably destroyed by rebuttals from various philosophers although Searle still stands by it in a rather ridiculous fashion.