|
|
|
Valiant Dancer
Forum Goalie
USA
4826 Posts |
 Posted - 09/09/2004 : 10:41:03 [Permalink]
|
quote:
Originally posted by Dave W.
Valiant Dancer wrote:quote:
I believe the task of programming instinctiveness into a machine is not possible as the manner in which instinctiveness is understood is inadequate to be able to produce a program mimicing it.
You have misunderstood my use of the word "teach." I did not intend it to mean "program." I meant "teach," as we teach kids different facts about the world.
I believe that we can create a set of programs which will have the capability of learning. I do not know if such a pile of code will "instinctively" desire to learn, or if its creators will have to force-feed it data until such a time as the benefits of learning are established.
I have yet to see a program which makes value judgements like this.
quote:
quote:
Humans also have the ability to feel pain and react in an instictual manner. Machines lack this trait.
Much of these "instinctual" reactions are handled away from the brain. A ganglion in the lower spine creates the knee-jerk reflex, as the path from knee all the way to brain and back takes too long, and walking itself would be impossible. And clearly, such reactive machines do exist. Otherwise, a Roomba would simply get stuck in a corner.
No, the instincts I am talking about are more cerebral. Like curiosity. These may need to be taught (not programmed, but taught).quote:
From my analysis of complex programming systems, I have seen problems which are unexpected crop up when it didn't make sense why. I then was able to isolate the error and analyze why it was occurring. It usually was an interaction error from sharing variable names where something was erroneously killed though indirection or the value changed. I have seen nothing to convince me that a machine is capable of sentience.
I fail to see the relevance of these comments to anything I have written.
I have not seen any program which has the capability to move past it's programming and auto-retire unused code or modify core processing systems without programmer intervention. The example was to show how I have seen a set of programs interact which may appear to be doing something independant of design but turned out to be an error.
quote:
quote:
Dave W and I have a basic disagreement that placing the individual pieces in close proximity and establishing structured communication between them...
Perhaps the fact that the pieces will be able to "re-wire" their communications on the fly will help clear up the picture?
Unfortunately, it does not. I cannot percieve a system which is not programmed to auto-fail to another communication path to do so indepentant of programmer intervention.
quote:
quote:
..will give the extremely complex machine a quality which would achieve sentience. Dave W believes it will. I don't.
And I am trying to understand why you feel the problems you have brought up are insurmountable, contrary to the history of science in general. AI is still a field which is busy, so far as I know, with many people striving for a truly sentient program. In other words, that the lack of inherent curiousity (for example) is a dead-end to the search for a true AI doesn't appear to be self-evident. Other than your personal experiences - which while interesting, are less than compelling as generalizations - how did you arrive at your conclusions?
I arrived at these conclusions based on my understanding of structured programming and personal experiences with it. It is why I say that I believe it to be impossible based on (admittedly cynical) personal analysis of humans in general and the operational elements of computing. I could very well be wrong. It could also be based somewhat on my personal theistic belief that life is something that cannot be re-created in the labratory. I have been philisophically investigating that part of my belief. It may very well be influencing my judgement. I have not seen anything in trade journals which spoke about human like AI in other than very general terms with no real details on how to produce it at a basic functional level. Everything I have seen on the subject talk in high-level specifications which may or may not be possible. I don't have that much faith in humans.
Again, I may be getting bogged down in minutae on this. |
Cthulhu/Asmodeus when you're tired of voting for the lesser of two evils
Brother Cutlass of Reasoned Discussion |
 |
|
|
Valiant Dancer
Forum Goalie
USA
4826 Posts |
Posted - 09/09/2004 : 10:46:56 [Permalink]
|
quote:
Originally posted by BigPapaSmurf
[quote]I believe the task of programming instinctiveness into a machine is not possible as the manner in which instinctiveness is understood is inadequate to be able to produce a program mimicing it. Humans also have the ability to feel pain and react in an instictual manner. Machines lack this trait. quote]
I think insticnt is the most machine like trait humans have and would be one of the simpler tasks in the overall project, the real trick is having it overcome instinct, such as sacrificing itself when it will certainly take damage. Thats the key to AI I believe, learning that its programming isnt always the correct action.
This seems to be imparting a sense of self preservation which is not programmed into the machine. Machines must follow their programming, whether it is written on the fly by a sophisticated program or executed as written by a programmer.
Instinct, although machine like in the way of responding to a stimulus, is not well understood to the point where it can be programmed into a machine. I see it as the most difficult task of the entire project. |
Cthulhu/Asmodeus when you're tired of voting for the lesser of two evils
Brother Cutlass of Reasoned Discussion |
|
|
|
Dave W.
Info Junkie
USA
26022 Posts |
Posted - 09/09/2004 : 11:12:17 [Permalink]
|
Valiant Dancer wrote:quote:
I have yet to see a program which makes value judgements like this.
Well, that's the point, isn't it? A program which makes value judgements is the goal. You're absolutely correct that one does not yet exist, but that has no bearing upon whether or not one is impossible to create.quote:
I have not seen any program which has the capability to move past it's programming and auto-retire unused code or modify core processing systems without programmer intervention.
Again, what you've seen is not necessarily all that can be.quote:
The example was to show how I have seen a set of programs interact which may appear to be doing something independant of design but turned out to be an error.
Have you played with cellular automata? One can build an entire computer within the "Game of Life," yet the program executing the game itself will have no knowledge that it is simply a sub-layer of processing.
Similarly, electrons have no clue that they are involved in the meaning of this sentence. Also, neurotransmitters are required for thought, yet carry no thought themselves.quote:
Unfortunately, it does not. I cannot percieve a system which is not programmed to auto-fail to another communication path to do so indepentant of programmer intervention.
This really makes me wonder what you think the human brain does. What do you understand the difference to be between a program and a neuron which makes the latter, in a group, capable of abstract thought while the former, no matter how many one groups, will never think?quote:
I arrived at these conclusions based on my understanding of structured programming and personal experiences with it. It is why I say that I believe it to be impossible based on (admittedly cynical) personal analysis of humans in general and the operational elements of computing. I could very well be wrong. It could also be based somewhat on my personal theistic belief that life is something that cannot be re-created in the labratory.
This could very well answer my previous question, but I won't make that assumption.quote:
I have not seen anything in trade journals which spoke about human like AI in other than very general terms with no real details on how to produce it at a basic functional level.
Which journals? Some books delve into the subject in great depth. There are courses available at universities.quote:
Everything I have seen on the subject talk in high-level specifications which may or may not be possible.
It sounds like you've only seen review articles, and nothing technical.quote:
I don't have that much faith in humans.
That's why we need A.I.!
|
- Dave W. (Private Msg, EMail)
Evidently, I rock!
Why not question something for a change?
Visit Dave's Psoriasis Info, too. |
|
|
|
BigPapaSmurf
SFN Die Hard
3192 Posts |
Posted - 09/09/2004 : 11:53:44 [Permalink]
|
Machines cannot acheive human instint because they are not human, that doesnt mean they cant be given a differnet form of instinct. I dont see why we would need a 'true understanding' of instinct to give machines their own version of it.
quote:
Humans also have the ability to feel pain and react in an instictual manner. Machines lack this trait.
Of course they lack this trait, they have no nervous system and have no need for one. I feel your personal definition of instinct is too limited to the human/animal realm. Machines will never be part of that system and clearly instinct will need to be programmed in. But that has little bearing on the AI issue for me because instict is not a roadblock on the path to intelligence IMO. Things like creativity and the ability to doubt and modify its programming those are the hard parts. |
"...things I have neither seen nor experienced nor heard tell of from anybody else; things, what is more, that do not in fact exist and could not ever exist at all. So my readers must not believe a word I say." -Lucian on his book True History
"...They accept such things on faith alone, without any evidence. So if a fraudulent and cunning person who knows how to take advantage of a situation comes among them, he can make himself rich in a short time." -Lucian critical of early Christians c.166 AD From his book, De Morte Peregrini |
|
|
|
Dr. Mabuse
Septic Fiend
Sweden
9688 Posts |
Posted - 09/09/2004 : 14:13:12 [Permalink]
|
Some 20 years ago I got my first personal computer. Eventually my programming skills developed to a level where I wrote game based on binary selection.
It was a primitive form of case-based reasoning.
"Think of a car" was the game. The player was supposed to think a car (model) and then the computer started to ask questions about the car to which you answered yes or no.
Once the computer arrived at the end of the binary branched tree of questions it guessed what car the player was thinking of.
Of course the answer was most often wrong. So the program asked for the right answer, then for a yes/no question that it could use to distinguish them from each other.
By playing several rounds, the knowledge-base of the program expanded, way beyond the initial knowledge set by the programmer.
I let several friends play the game, and eventually the knowledge-base hit the built in limits of the program: BASIC on those old computers couldn't dynamically re-define data arrays, and I had to keep to the limited memory available. But I did get more than a hundred different car models. The entire database was saved on compact cassette. (This was during the Commodore 64 era).
Now, a computer with true AI will be much more advanced, and it's program requirements different. Back in those days, one could only run one program at a time on a computer, and the programs couldn't communicate with each other. Yet, in my example I managed to create a program that gave the illusion of learning.
Multitasking opens up a completely new avenue of programming.
A program could be made to emulate a neuron.
If you start several processes of this program and connect the programs to each other you get something that computer scientists calls a neural net.
One or two or a hundred brain cells in a lump does not create intelligence. Neither does a hundred interconnected processes either. A billion brain cells could start working together... (I don't know how many we are supposed to have).
It is not the individual brain cell that exhibit intelligence and consciousness, but all cells working in unison. The same goes for the neural net. With high enough number of simultaneous processes I'm confident that the neural net will also be intelligent.
My think-of-a-car program couldn't start working without me predefining a basic level of knowledge: I had to pre-program (as Dave put it earlier "force-feed") knowledge that could be a starting point upon which to keep building more.
In much the same way, a neural net program needs to be prepared with a starting point of info in order to function. We could compare this to the basic instincts that children are born with.
A neural program cell could be programmed with a time-out, if it does not get a response within a certain time-period, it could search for other nodes to connect to.
In another thread Dave described how only a few simple rule can make a flock of birds behave in a very complex pattern. The same goes for neural nets. If enough of them are interconnected, similar effects will be apparent. |
Dr. Mabuse - "When the going gets tough, the tough get Duct-tape..."
Dr. Mabuse whisper.mp3
"Equivocation is not just a job, for a creationist it's a way of life..." Dr. Mabuse
Support American Troops in Iraq:
Send them unarmed civilians for target practice..
Collateralmurder. |
|
|
|
Dave W.
Info Junkie
USA
26022 Posts |
Posted - 09/09/2004 : 18:46:07 [Permalink]
|
Mab, the A.I. I'm envisioning in this thread is basically a neural net on a grand scale. So grand that we do not currently have the processing power to execute it.
About the simplest neural net there is is something called a Binary Associative Memory (BAM). In a BAM, there are two set of "neurons," call them A and B. Each neuron in set A "connects to" every neuron in set B, and vice versa (no neurons in set A connect to other set A neurons, though, and the same for B). Each connection is stored as a number which represents a "connection strength." After training, you can present a data set to the A neurons, and the B neurons will take on whatever values they've been trained to in response. If the data set is incomplete or slightly different from those used during training, the output may also be incomplete or different, but it'll be close. These BAMs are being widely used today for pattern recognition tasks, as it's really easy to "arrange" the neurons to represent pixels in an image, or to encode text into them.
None of the neurons in a BAM "understands" anything about pattern recognition, or even that it's a part of a pattern-recognizing tool. All it "knows" is that it has to do certain things with the input it gets, and present the results as output. The overall behaviour of the BAM is one of "recognition," and that behaviour is not explicitly coded into the BAM at all.
And it is certainly not the case that any of the electrons zipping through the processor running a BAM have any inkling of what their overall behaviour is accomplishing, even in terms of simple things like adding two binary bits.
Just like nobody here has any neurons which "know" the SFN, but instead large collections of the trillion or so neurons in your heads (that's a trillion in each head) exhibit an overall behaviour which allow you to find your way back here. And the neurotransmitters are even less complex, individually, yet without them no brain would function.
Simple things can and regularly do exhibit, when acting in concert, surprising behaviours that aren't built-in to the pieces.
Anyway, neural nets with more-complicated arrangements of connections than a simple BAM can do much more than pattern recognition.
Ah, here is an interesting book review, which states in part,The brain is often referred to as the most complex structure in the universe. This complexity gives rise to remarkable properties such as the mind and consciousness, as well as to a thriving philosophical debate about their nature. These questions are still out of the reach of neuronal modelling. Nevertheless, modelling has provided intriguing glimpses of how some of the properties of the brain may emerge. Abstract neural networks loosely based on biology exhibit useful (and brain-like) functions such as learning by example, robustness, and the ability to generalize. Realistic neuronal modelling, which this book examines, is beginning to provide the theoreti-cal counterpart of experimental neuroscience, and to answer many questions about the higher functions of the brain. It also says,A generic personal computer in 1999 is powerful enough to model a 50-compartment neuron with a full complement of Hodgkin–Huxley-type ion channels in real time, that is, the model will function at the same speed as the real cell would. Such language is highly intriguing in a variety of ways, not least of which comes from the question "how accurate a model is that?" Five years later, can we (by Moore's Law) expect to be able to run ten such models on a single generic PC? If so, we only need 100 billion PCs networked together to make a brain. 
But I'm now officially rambling. This stuff is just way cool to think about, even in limited forms like BAMs and small networks. And Skyhawk is absolutely correct in that by coupling these things with other technologies, the capabilities may really take off. |
- Dave W. (Private Msg, EMail)
Evidently, I rock!
Why not question something for a change?
Visit Dave's Psoriasis Info, too. |
|
|
|
Dr. Mabuse
Septic Fiend
Sweden
9688 Posts |
Posted - 09/09/2004 : 22:40:54 [Permalink]
|
Indeed, Dave, you and I are thinking much the same about the capabilities of current and future computer technology. In order to create a "true" humanlike brain AI that would like the book you referenced require that 50-compartment neuron with a full complement of Hodgkin–Huxley-type ion channels. However, much computing power could be saved by using a less advanced neuron model. It wouldn't result in a human brain emulator, but it would still be intelligent, perhaps even self-aware.
Since the computer model will not have to be bothered with smell or taste, perhaps not even tactile sensations, there will be more computing power left for other stuff. |
Dr. Mabuse - "When the going gets tough, the tough get Duct-tape..."
Dr. Mabuse whisper.mp3
"Equivocation is not just a job, for a creationist it's a way of life..." Dr. Mabuse
Support American Troops in Iraq:
Send them unarmed civilians for target practice..
Collateralmurder. |
|
|
|
BigPapaSmurf
SFN Die Hard
3192 Posts |
Posted - 09/10/2004 : 05:15:33 [Permalink]
|
| Optics is a massive draw on memory, I read somewhere that 80?% (a majority) of our brainpower is dedicated to optics and comprehending images. |
"...things I have neither seen nor experienced nor heard tell of from anybody else; things, what is more, that do not in fact exist and could not ever exist at all. So my readers must not believe a word I say." -Lucian on his book True History
"...They accept such things on faith alone, without any evidence. So if a fraudulent and cunning person who knows how to take advantage of a situation comes among them, he can make himself rich in a short time." -Lucian critical of early Christians c.166 AD From his book, De Morte Peregrini |
|
|
|
Dr. Mabuse
Septic Fiend
Sweden
9688 Posts |
Posted - 09/10/2004 : 05:44:30 [Permalink]
|
quote:
Originally posted by BigPapaSmurf
Optics is a massive draw on memory, I read somewhere that 80?% (a majority) of our brainpower is dedicated to optics and comprehending images.
Well, I saw somewhere that the eye was equivialent of somewhere between 100 and 200 megapixels (though lower colour resolution). If you limit optical input to 5 Mpix, lots of power is set free there also. |
Dr. Mabuse - "When the going gets tough, the tough get Duct-tape..."
Dr. Mabuse whisper.mp3
"Equivocation is not just a job, for a creationist it's a way of life..." Dr. Mabuse
Support American Troops in Iraq:
Send them unarmed civilians for target practice..
Collateralmurder. |
|
|
|
Dave W.
Info Junkie
USA
26022 Posts |
Posted - 09/10/2004 : 06:18:01 [Permalink]
|
| There's no need for the A.I. to "see" in any normal sense. Hook the danged thing straight into the Internet for its input. Let it develop its own image-processing routines to make sense of the graphics, photos, and other visual stuff. |
- Dave W. (Private Msg, EMail)
Evidently, I rock!
Why not question something for a change?
Visit Dave's Psoriasis Info, too. |
|
|
|
astropin
SFN Regular
USA
970 Posts |
Posted - 09/10/2004 : 12:41:11 [Permalink]
|
| Most of us think it's at least possible if not probable....so when? What's your prediction? As I've mentioned I like Kurzweil's analysis showing exponential gains in technology in general so I find no reason to doubt his approximate timeline of around 2010 for the required processing power but more like 2030 before the required "software" can be accomplished. |
I would rather face a cold reality than delude myself with comforting fantasies.
You are free to believe what you want to believe and I am free to ridicule you for it.
Atheism:
The result of an unbiased and rational search for the truth.
Infinitus est numerus stultorum |
|
|
|
Dr. Mabuse
Septic Fiend
Sweden
9688 Posts |
Posted - 09/10/2004 : 18:46:54 [Permalink]
|
quote:
Originally posted by astropin
Most of us think it's at least possible if not probable....so when? What's your prediction? As I've mentioned I like Kurzweil's analysis showing exponential gains in technology in general so I find no reason to doubt his approximate timeline of around 2010 for the required processing power but more like 2030 before the required "software" can be accomplished.
Well, given the estimation that a current top-of-the-line PC can emulate 10 neurons in real-time, and the computing power follows Moore's Law, one PC will be able to emulate 1280 neurons in real-time twelve years from now. In another twelve years it will be 320'000 neurons.
Today, the hottest supercomputer does 40 Tera-flops. A 2GHz P4 running a large program does ~250 Mega-flops.( http://www2b.abc.net.au/science/techtalk-old/posts/topic57729.shtm ) If you pick today's top of the line PC, let's assume it delivers 400 MF. The difference is 100'000 times.
If Supercomputers also follow Moore's Law, the projected capacity for neuron emulation in real time will be 30-40 billion neurons in real-time about 25 years from now.
If we settle for less than full complement of Hodgkin–Huxley-type ion channels-emulated neurons, then the number of emulated neurons will increase accordingly.
My guess is that the original neuron simulation back in 1999 wasn't programmed in machine-code, so switching to a more efficient programming language will also give a major improvement. |
Dr. Mabuse - "When the going gets tough, the tough get Duct-tape..."
Dr. Mabuse whisper.mp3
"Equivocation is not just a job, for a creationist it's a way of life..." Dr. Mabuse
Support American Troops in Iraq:
Send them unarmed civilians for target practice..
Collateralmurder. |
|
|
|
Dave W.
Info Junkie
USA
26022 Posts |
Posted - 09/10/2004 : 19:14:17 [Permalink]
|
Mab wrote:quote:
If Supercomputers also follow Moore's Law, the projected capacity for neuron emulation in real time will be 30-40 billion neurons in real-time about 25 years from now.
...On a single machine. Network 25 of those machines together (assuming that network speeds also increase), and you're potentially looking at a trillion neurons - or about the size of a human brain. And if supercomputers aren't your cup of tea, perhaps something called "AI@Home" could be created, to make use of idle CPU cycles among many millions of Internet-connected PCs to create an intelligence (and perhaps, with PCs dropping out and reappearing in the network, it would be schizophrenic). |
- Dave W. (Private Msg, EMail)
Evidently, I rock!
Why not question something for a change?
Visit Dave's Psoriasis Info, too. |
|
|
|
Ricky
SFN Die Hard
USA
4907 Posts |
|
|
Dr. Mabuse
Septic Fiend
Sweden
9688 Posts |
Posted - 09/11/2004 : 14:10:44 [Permalink]
|
quote:
Originally posted by Ricky
If you wanted to implement this, just go to KaZaA. They have in their license user agreement to be able to use any extra system resources for any purpose they want.
Really? I didn't know that. So that's what the officer from the Swedish security police (SÄPO) was rambling about on TV-news a while back, when he talked about the secutiy-risk high-tech companies experience when they let their employees have programs like SETI@Home and similar programs installed.
I always figured web-pages could have applets running distributed computing-tasks, but I never learned enough to work out how it could be done. |
Dr. Mabuse - "When the going gets tough, the tough get Duct-tape..."
Dr. Mabuse whisper.mp3
"Equivocation is not just a job, for a creationist it's a way of life..." Dr. Mabuse
Support American Troops in Iraq:
Send them unarmed civilians for target practice..
Collateralmurder. |
|
|
 |
|
|
|
Topic Locked
