PDA

View Full Version : Artificial intelligence



Human
2004-Dec-18, 06:13 PM
http://www.nickbostrom.com/superintelligence.html

When do you think artificial intelligence will be super-intelligent?

A Thousand Pardons
2004-Dec-18, 06:59 PM
That article says within thirty years, so let's extrapolate.

2001: A Space Odyssey said we'd have regular commuter
flights to the moon by now. So...my guess is 2118

Human
2004-Dec-18, 07:08 PM
That article says within thirty years, so let's extrapolate.

2001: A Space Odyssey said we'd have regular commuter
flights to the moon by now. So...my guess is 2118

So things usually takes longer time than we guess at first? (I have not studied much science history.)

A Thousand Pardons
2004-Dec-19, 05:53 AM
No, just these kind of things. :)

pteranodon
2004-Dec-21, 03:48 PM
When we have computer power to run more complex neural networks than today. I think the key will not be faster processors, but faster networking infrastructure, so parallel computing in global scale will be real and practical. Neural networks are better suited for such distributed computing environments such as the internet. The problem today is bandwidth.

Nicolas
2004-Dec-21, 03:53 PM
You mean that you'll assign a physical processor to each function in the neural network? Or what do you mean with neural networks being better suited for computational networks?

The neural networks I'm familiar with are either computed on a single computer, or on a computation network setup, in which the physical computers have no interrelation with the neural network buildup. The spreading of computations needed for the neural network follows general algoritms for computational networks.

tofu
2004-Dec-21, 04:16 PM
yeah, neural networks are logical structures. They have nothing to do with actual physical computer networks like a LAN or internet. The term "logical structure" doesn't mean they perform logic operations. It means they only exist inside the computer.

The reason we don't have self aware AI today really has nothing to do with computing power. A lot of people make that mistake. They think if you just throw more computer at a problem then the problem will go away. That's just not true in this case. The problem, from a computer scientists point of view, is describing the thing you want to implement. If it were at all possible (with current state of the art CS) to model intelegence, then you could do it on any computer, it would just be very slow. Think of it this way, the IBM computer that beat the world champion chess player was named "Big Blue" There was nothing special about that computer. What was special was the program it was running. You could run that same program on an old 1 MHz Apple II if you really wanted to. It would just take longer for it to think of a move.

The same is true of AI. pteranodon says we'll have self-aware AI, "when we have the computer power." If that were true, we could have it today, but it would just be slower.

Neural networks aren't magic. They are just a different way of thinking about a problem. You can do what a neural network does using a sheet of notebook paper. There are certain categories of problems that computer science has a hard time dealing with. One of those categories is pattern recognition. How do you write a computer program to predict the stock market? The tools that we have to work with (in our head and in the computer) just weren't quite up to the task. And that's why neural networks were invented. They make writting pattern recognition software easier.

worzel
2004-Dec-21, 04:30 PM
Well said tofu.

What a lot of people don't realize is that in a world of massively parallel, hyper-dyper threaded, 2 to trillion bit, 1 septillahertz processors could do no more than a programmable calculator, given enough memory and time (or a turing machine with enough tape).

Nicolas
2004-Dec-21, 04:39 PM
Neural networks aren't magic. They are just a different way of thinking about a problem. You can do what a neural network does using a sheet of notebook paper.

I know what you mean: "it all comes down to math in the end". The pro's of neural networks are in the "self-learning" effect (watch out when interpreting this term). You give the neural network a goal it has to meet, inputs and what it can do as output. The networks starts to do directed "trial and error" untill it meets the goal. Example: stabilising a stick vertically on a robot "finger". The neural networks "exercises" until it has found exactly how to move itself to keep the stick straight up. After a while it outperforms humans.

A current problem with neural networks is the time it takes for them to "learn". Possible uses for neural networks are the steering of damaged aircraft with whatever is still available of steering surfaces/engine settings: the coomputer gets a pilot input, and knows which goal the pilot wants to achieve with this input. It tries all parameters in order to make the move in a different way than using the broken control surface (as far as possible). these systems of course are only possible if they learn very fast: if your neural network only knows how to steer again after 4 hours, it might as well learn how to call resque teams from the bottom of the ocean... This problem (the learning time, not the calling) can be solved by computing power.

Computers as we know them (including neural networks) are faster, less prone to miscalculations, have better memories and better copiers than humans, which makes them usefull in lots of applications.

Sometimes it is difficult to see the difference between forms of intelligence: a plane which turns (by a neural network) by using the engine throttle settings instead of the rudder when the rudder is damaged, does not show creativity of the neural network. It just performed trial and error. Humans could do that too, it would take them just way too long, and we aren't precise and sensitive enough to apply it correctly. Our creativity is in finding solutions that aren't "in the book" (think of a film like "final descend, without going into the realism of this movie). Neural networks can only go their own way within the boundaries we have defined for them, but NEVER BEYOND THEM. The only thing that makes them different from standard programs is that they find their own algoritm by trial and error of the available parameters, instead of humans putting them in the program from the start.

In fact a neural networks is capable of making tons of bad solutions, each one a bit less bad than the last one, untill it finally works. If that is superintelligence... Of course humans use this approach too, but their creativty allows for "creating" (it's in the word), where neural networks only use existing things (they can "learn" to fly difficult controllable heicopters, but they won't tell you "I think we should choose a different tailt layout because this one ain't working good". If a human doesn't like the helicopter tail, he won't endlessly try to steer the thing anyway, he'llcome up with things like fenestrons, which a neural network as we know it will never do. They simply aren't inspired.

worzel
2004-Dec-21, 04:45 PM
which a neural network as we know it will never do. They simply aren't inspired.
But maybe when we figure out how to write a universal neural net with enough inputs and outputs, it might be difficult to distinguish its solutions with inspiration.

Nicolas
2004-Dec-21, 05:32 PM
My thoughts went into that direction too: linking the neural network with some kind of internet of all true (...) knowledge as a secondary input. Probably the amount of possibilities rises faster than the computing power in order to do this, and completely structurising all info in the world in order for neural networks to find their way in it, probably will give such an enormous amount of new insights for humans that neural networks will have to take second place once again.

Humans can come up with creative solutions, computers with ANY solution. I think they will drown in the amount of info unless it is struturised, and that humans will take better advantage of info than neural networks.

I repaired a worn-out rubber from my joystick with a piece of foam from a pillow and a staple. I had to cut a piece of foam, and slice this half through in order to do this. The staple could only be applied after the foam was in its correct place. A neural network would have to combine (don't ask me how) the complete collection of "redundant or other more or less costless items in my house" to come with a similar solution. By the time I have explored all these objects and listed them, I could simply use the list and find pretty good solutions faster and more efficient than the computer. Say that the neural network is directed to solve the problem and starts with looking for costless soft objects (which is needed for the solutions) like humans do when trying to solve the problem. Only the computer would list things humans would never do: toilet paper is soft, so is a sandwich... If you'd have to give the properties of each and every object and give rules when a property is wanted or not, you'd have done such an enormous amount of research that you yourself would have come up with the ultimate solution along the way. Humans are continuously learning through their life, building up a database of knowledge from all kinds of aspects (called a Real Life System RLS). A computer can only "grow" such a system through input channels and lots,lots,lots of "cases" to study. It would take an enormous amount of time to generate such a computer RLS (literally a lifetime) and it would cover only as much information as a human RLS would have, only without fogetting anything (storage....).

My idea is that the only thing that computers can offer is speed (in calculation, communication and execution) and correctness (in storage, calculation and execution). We should use them in order to do this, not to outperform our brain. Cause computers aren't brains. They are just dedicated versions of a small part of it.

tofu
2004-Dec-21, 05:34 PM
Example: stabilising a stick vertically on a robot "finger". The neural networks "exercises" until it has found exactly how to move itself to keep the stick straight up.

Just to drive home the point that a neural network is just a tool and that there are plenty of other tools that are just as good, I would actually use fuzzy logic to solve the problem you described - I think the problem is actually called an "inverted pendulum." Anyway, fuzzy logic and neural nets are just tools that a computer scientist carries around in his bag. When he needs to solve a particular problem, he selects the most appropriate tool. An analogy might be what calculus is to a mathematician. When a problem calls for it, the mathematician uses that particular tool. The only connection to AI is that neural nets are the most commonly cited tool, probably because they sound all science fictiony. You could just as easily use fuzzy logic to implement some simple AI like maybe an expert system. Hell, I've even seen some very convincing IRC bots that were nothing more than lexical parsers - many of them written in perl!

I'm not at all convinced that neural networks are the going to be the ultimate solution to the AI problem. It sounds too much like a typical silver bullet proclamation to me.


it might be difficult to distinguish its solutions with inspiration.

I do however have an interesting story to tell about neural networks. I had to write one for a class as an undergraduate, and I recall it now because it relates to your comment about inspiration.

The assignment was to build a neural network with several inputs and one output. The network was to simulate a multiple or-gate. So, if any of the inputs were lit, the output should be lit. It looks like this (where inputs are in parenthesis and the output is after the equal)

(0,0,1,0) = 1 - that would be a correct result and the path would be reinforced.
(1,1,0,0) = 1 - also correct.

You get the idea. After you run the neural network program for a while, you have to do the equivalent of garbage collection. That is, you have to go back and clean up paths that are weighted too heavily because they haven't been used very often. Maybe a better analogy is a teacher going back and grading everyone on a curve.

Anyway, if you give the or-gate neural net truly random data, it learns very fast. That's mostly because the problem set was very small. But if you gave the net a million problems using the first three inputs, and only one using the last input, it essentially has to guess (though the output path would be heavily weighted by that time, so it is likely to guess right).

And now finally we come to the point of my story. When you go back and clean up the network, the paths that need the most work and the least used paths. So what you see, if you are outputting this process to the screen, is the network retrace paths, seemingly at random, and showing odd or unusual things it had experienced. It looked to me exactly like dreaming, because that's what you get in your dreams - the odd, unusual, or other extraordinary things that happened during the day.

So while my little neural net was dreaming about multiple or-gates, there I was, an 18 year old geek with no lady friends, dreaming about something phonetically very similar.

ToSeek
2004-Dec-21, 05:36 PM
It looked to me exactly like dreaming, because that's what you get in your dreams - the odd, unusual, or other extraordinary things that happened during the day.


That doesn't very accurately describe my dreams. I can very rarely figure out what inspired them - it's occasionally something that happened that day, but not very often.

Nicolas
2004-Dec-21, 05:43 PM
That indeed is the problem of the inverted pendulum. An extreme neural network controlled robot hinge could convert a normal pendulum into an inverted one (with real-life output). Try THAT as a human. I'm not thinking neural networks will be the future. They are just a current solution which is promising on some aspects.

Fuzzy logic is considered for the docking of the ATV with the ISS, which is difficult as the ISS is an elastic structure. Dr. Chu is an expert in this research.

Considering Fuzzy logic, neural networks or any other programming approach for artificial intelligence holds the same points I tried to make projected on neural networks: computers are calculators and only do tasks of a part of the human brain.

gzhpcu
2004-Dec-21, 05:48 PM
That article says within thirty years, so let's extrapolate.

2001: A Space Odyssey said we'd have regular commuter
flights to the moon by now. So...my guess is 2118

I wouldn't be so pessimistic. We all take Moore's Law for granted (cpu speed doubles every 18 months, memory size gets 50% smaller), which has been valid for over 30 years now for silicon chips and should continue to be so till about 2020. This should give artificial intelligence software the power it needs. Combined with respective high bandwidth, interconnecting computers running AI, this should give quite a bit of punch to the AI environment. So my guess is about 2030. Not yet superintelligent but very intelligent. This just with conventional silicon chip technology and the classical Van Neuman architecture, and disregarding advances with neural networks or massively parallel computing architectures and other technology breakthroughs which might just speed things up more.

The moon project problem is one of finding money. Lack of interest has hobbled advances.

worzel
2004-Dec-21, 06:07 PM
And now finally we come to the point of my story. When you go back and clean up the network, the paths that need the most work and the least used paths. So what you see, if you are outputting this process to the screen, is the network retrace paths, seemingly at random, and showing odd or unusual things it had experienced. It looked to me exactly like dreaming, because that's what you get in your dreams - the odd, unusual, or other extraordinary things that happened during the day.

So while my little neural net was dreaming about multiple or-gates, there I was, an 18 year old geek with no lady friends, dreaming about something phonetically very similar.
LOL

That reminds me of a (failed) research grant proposal one of my lecturers and I put up. He had this idea of using inverse implication rather than unification to do algorithmic learning. It was so long ago I can't remember the details but I remember being impressed with one paper example building the inductive definition of a tree from raw data. Wish I could remember more.

A Thousand Pardons
2004-Dec-21, 07:00 PM
Lack of interest has hobbled advances.
Everyone who wants a machine smarter than themselves, raise their hand :)

gzhpcu
2004-Dec-21, 07:03 PM
Lack of interest has hobbled advances.
Everyone who wants a machine smarter than themselves, raise their hand :)

It is not us, it is the industry. Factories with robots, more profit, more unemployment.... :(

Nicolas
2004-Dec-21, 07:06 PM
more unemployment, yes if you have inspection, maintenance, development and sales robots too. A robot industry large enough to keep us out of the factories would create many jobs on itself

A Thousand Pardons
2004-Dec-21, 07:08 PM
Lack of interest has hobbled advances.
Everyone who wants a machine smarter than themselves, raise their hand :)

It is not us, it is the industry. Factories with robots, more profit, more unemployment.... :(
Any AI that was smarter than us would be working for us?

Nicolas
2004-Dec-21, 07:08 PM
good point. robots mainly t ake over the "dumbest" jobs (automation)

A Thousand Pardons
2004-Dec-21, 07:12 PM
The same is true of AI. pteranodon says we'll have self-aware AI, "when we have the computer power." If that were true, we could have it today, but it would just be slower.
Then maybe it's here, but it hasn't learned to crawl yet. :)

PS: and is on pace to do so in 3042 AD

worzel
2004-Dec-22, 12:30 AM
The same is true of AI. pteranodon says we'll have self-aware AI, "when we have the computer power." If that were true, we could have it today, but it would just be slower.
Then maybe it's here, but it hasn't learned to crawl yet. :)

PS: and is on pace to do so in 3042 AD
My guess is it got stuck in an infinite loop watching football.

Nicolas
2004-Dec-22, 12:30 AM
Not all AI is male...

worzel
2004-Dec-22, 12:39 AM
Not all AI is male...
I'm not so sure about that. What are the problems with AI

- no common sense or intuition
- single mindedness
- no awareness of the bigger picture
- only really good at maths (even if only when presented as football stats)

Nicolas
2004-Dec-22, 12:45 AM
yes, but on the other hand:

-totally unpredictable for the normal man (neural networks)
-gives rise to more problems than they solve
-remembers any mistake you ever made and uses it against you when it suits you worst
-panics if there is a power outage

-and they sometimes can do such nice things!!!

gzhpcu
2004-Dec-22, 03:46 AM
The Japanese are working on robots:

Honda Robot (http://www.world.honda.com/HDTV/ASIMO/)

worzel
2004-Dec-22, 11:51 AM
The Japanese are working on robots:

Honda Robot (http://www.world.honda.com/HDTV/ASIMO/)

So am I, after discovering the Digital Lego Designer (http://www.lego.com/eng/factory/design/ldd.asp) I've reverted to childhood and asked my mum for the Lego Miindstorms (http://mindstorms.lego.com/eng/products/ris/index.asp) set for Christmas :oops: