It is possible to build a superhuman computer?

[ClaysGhost]

Did you know that a single neuron has as much processing power as 5 of the worlds most advanced supercomputers?

I kind of doubt it. Can you support this?

[Ford Prefect]

Did you know that a single neuron has as much processing power as 5 of the worlds most advanced supercomputers?

I kind of doubt it. Can you support this?

If it was at all possible, I can certainly see a market in human computers.

[Captain tycho]

I thought neurons were more analagous to a single processor in a multi-processor computer? Forgive me if I'm wrong, though.

[Dahak]

Quite wrong. Today's computers are Turing's machines and are therefore capable of performing any algorithm. Of course there is a question if human "mind" is some kind of algorithm ( just to simplify ).

That's not completely correct. There are problems, for which you will not find any computable algorithm, for instance the halting problem.

And what is not completely correct ? Of course there are many problems that are not computable ( Post cor. problem , or problems shown by Rice's theorem ) . But that does not change the fact that computers are not specialists. And i said they are capable solving any problem that can be solved by algorithm. And we even do not know if our brain can solve those problems , so i am really not sure what is not completely correct ?

It may have been your wording, but I felt you implied that a Turing-equivalent computer could compute anything, which they clearly cannot.

And the brain won't, either. If it is, as many say, a Turing-equivalent computer, than our brain will only be able to do as much as a Turing machine. So a human won't be able to go around, for instance, the halting problem.

[Dooey Jo]

Did you know that a single neuron has as much processing power as 5 of the worlds most advanced supercomputers?

How do you measure something like that? In terms of processing things, a single neuron is pretty useless. It's basically a glorified logic gate (with thousands of inputs). And a very slow one at that. Good luck using that as a super computer. In fact, the very reason we have super computers is that our brain, even though it has hundreds of billions of neurons, really sucks in comparison to the pure number-crunching abilities of a super-computer.

[anybody_mcc]

That's not completely correct. There are problems, for which you will not find any computable algorithm, for instance the halting problem.

And what is not completely correct ? Of course there are many problems that are not computable ( Post cor. problem , or problems shown by Rice's theorem ) . But that does not change the fact that computers are not specialists. And i said they are capable solving any problem that can be solved by algorithm. And we even do not know if our brain can solve those problems , so i am really not sure what is not completely correct ?

It may have been your wording, but I felt you implied that a Turing-equivalent computer could compute anything, which they clearly cannot.

And the brain won't, either. If it is, as many say, a Turing-equivalent computer, than our brain will only be able to do as much as a Turing machine. So a human won't be able to go around, for instance, the halting problem.

Well i really didn't want to imply anything like that , because i know enough about undecidable problems to not say something like that. I just wanted to show that the sentence that computers are specialists is wrong. And for that Turing machine capabilities are enough IMHO.

[sketerpot]

To who ever said that, computers may have the ability to calculate trillions of numbers per nanosecond or crunch floating point numbers but computers are no more smarter than lets say the lowest form of life that uses a nervous system.

I'd like to see an earthworm design a low-pass filter. Computers have done so, given only knowledge of the expected inputs and outputs and how to test the circuits using an FPGA (a reprogrammable microchip, essentially).

Did you know that a single neuron has as much processing power as 5 of the worlds most advanced supercomputers?

The nut I have in my hand can perform far more accurate simulations of a falling nut-shaped body in my dorm room than the world's most advanced supercomputers, simply by being chucked into the air. That still does not make it intelligent or useful.

Did you know that computers have designed some new, patentable electric circuits through evolution? That's human-competitive design "intelligence", right there.

Guess who had to program them? who had to show them how to design them? Humans right?

Humans had to specify the inputs and outputs, the testing method, and the evolutionary method. Once that was done, the computers were on their own. And they did it. Sometimes they even cheated, using unforeseen subtleties of the testing equipment to simplify their circuits. That seems intelligent, and in a way that many people thought computers could never match: creative design.

[WyrdNyrd]

Don't be silly, Sketerpot. Computers aren't smart. Only humans are smart. Computers aren't human. Therefore computers aren't smart.

That's the kind of reason I saw underlying all the anti-AI arguments we covered in my (antiquated) AI course. Hopefully the anti- brigade have come up with better ones in the last 10 years, otherwise they sure-as-hell aren't going to convince me.

[Peregrin Toker]

On the other hand , when we will try to construct computers with emotions , we may succeed , and they may become a threat.

How possible is that? I think that psychiatry has already found out which emotions are caused by which reactions in the brain...

@Peregrin Toker: I am sorry, but Vernor Vinge simply puts forth claim without any supporting arguments ("There will be a singularity", "It will be between 2005 and 2030", etc.). This may be nice for SciFi, but that's not honest discussion that I find helpful in any way. Hofstadter and Penrose both go to great lengths to make a case for why they believe what they believe, and how they got to that conclusion. Both build on the body of math and physics. Whereas Vinge only seems to be a "mathematician" and spouts out his phantasies.

I thought that Vernor Vinge also was a computer scientist which meant that he would know something about what he was saying.


Stephen Hawking, though, has said that such a Singularity will very likely occur if computer technology continues to advance with the a pace equal to or greater than now - unless if we through genetic engineering create a race of hyperintelligent superhumans. (source: "The Machine Horror")

[Shroom Man 777]

You can always have a computer so advanced that it can simulate the brains functions, copy the chain reactions and so forth. It's all in the programming, if you ask me.

[sketerpot]

@Peregrin Toker: I am sorry, but Vernor Vinge simply puts forth claim without any supporting arguments ("There will be a singularity", "It will be between 2005 and 2030", etc.). This may be nice for SciFi, but that's not honest discussion that I find helpful in any way. Hofstadter and Penrose both go to great lengths to make a case for why they believe what they believe, and how they got to that conclusion. Both build on the body of math and physics. Whereas Vinge only seems to be a "mathematician" and spouts out his phantasies.

I thought that Vernor Vinge also was a computer scientist which meant that he would know something about what he was saying.

A degree in CS means jack shit when it comes to predicting the pace and direction of increases in computer processing power. You should try talking to electrical and computer engineers if you want good predictions on that.

Right now, many of the speed increases in computers are already from exploiting parallelism, with graphics cards and hyperthreading and CPU architectures which use instruction-level parallelism. Our dependency on parallel computation is often predicted to grow in the future, and it's unclear what this will mean for people who wish to make computers superhuman in more areas. Will it require huge supercomputers? Will the problems to be solved simply be too fine-grained for efficient execution on many parallel processors?

Additionally, AI research in computer science doesn't yet have a magic bullet which will produce self-enhancing intelligence if you just throw enough computer power at it. Vinge should know that.

[sketerpot]

You can always have a computer so advanced that it can simulate the brains functions, copy the chain reactions and so forth. It's all in the programming, if you ask me.

Okay, let's think about that for a moment. Now, neurons fire in a horrendously complicated and not entirely understood way (as in, we are still discovering important aspects of the process). To oversimplify a lot, neurons recieve a number of signals called excitatory and inhibitory post-synaptic potentials, which I'll abbreviate as IPSPs and EPSPs. These act in local areas and do not propagate very far along the neuron's membrane. If enough EPSPs are recieved in a certain area at close enough times, and if they are not counterbalanced by enough IPSPs, then a chain reaction will start in which ion channels open, ions rush across the membrane, the pulse travels across the neuron's membrane. This sets off some processes in the end of the axon (the long tail of the neuron) which release chemicals called neurotransmitters and neuromodulators into the gap between the axon and another neuron, and some of this leaks into the interstitial fluid in significant ways.

Unless you've studied physiology, you probably had trouble following that---and that's a hugely simplified version of our incomplete understanding of this process. Suppose that you wrote a program which simulates this, and then the neurologists come up with something new: suddenly there are some important neuromodulators called endocannabinoids which can actually go from a neuron to the neurons that ordinarily send signals to it. You'd have to change things around a lot, and you still might not have your computer simulation right.

Also, there's the problem of what parts you can simplify. Do you have to use finite-element methods to simulate the opening and closing of ion channels on the membranes of each neuron (and the electric potentials across those parts of the membrane), or can you simplify it? If you simplify it, what effect does that have on the accuracy of your simulation of temporal and spatial summation of EPSPs and IPSPs? And if you don't simplify it, how could you possibly simulate a human brain in any reasonable length of time? There are lots of similar issues that even my meager knowledge of neural physiology lets me see.

It's probably theorectically possible to simulate the brain on a quantum level using Monte Carlo techniques, but the practicality of this, or any simplification of it, leaves a lot to be desired.

[R. U. Serious]

@sketerpot: Nice explanation, though I admit I didn't understand most of it. But it reminded me a little bit of Feynman's Cargo Cult Science.

In the South Seas there is a cargo cult of people. During the war they saw airplanes with lots of good materials, and they want the same thing to happen now. So they've arranged to make things like runways, to put fires along the sides of the runways, to make a wooden hut for a man to sit in, with two wooden pieces on his head to headphones and bars of bamboo sticking out like antennas--he's the controller--and they wait for the airplanes to land. They're doing everything right. The form is perfect. It looks exactly the way it looked before. But it doesn't work. No airplanes land.

So we're plugging in all these chips and cables, and writing software that uses statistical learning and it looks exactly like what we know of the brain. But it refuses to become scentiests. We probably just need more fires along the runway, err, I mean computing power.

[R. U. Serious]

Wow, I can't spell at all. Really need to get used to the fact, that posts cannot be edited here. Please replace and insert as appropiate: sentient, programming tricks, AI techniques, processing power.

[Dennis Toy]

I'd like to see an earthworm design a low-pass filter. Computers have done so, given only knowledge of the expected inputs and outputs and how to test the circuits using an FPGA (a reprogrammable microchip, essentially

i was talking about processing power and not the ability to use it. An earthworms nervous system has more processing power than even the most advanced supercomputer.

[sketerpot]

I'd like to see an earthworm design a low-pass filter. Computers have done so, given only knowledge of the expected inputs and outputs and how to test the circuits using an FPGA (a reprogrammable microchip, essentially

i was talking about processing power and not the ability to use it. An earthworms nervous system has more processing power than even the most advanced supercomputer.

How do you define processing power? Allow me to quote myself for emphasis:

The nut I have in my hand can perform far more accurate simulations of a falling nut-shaped body in my dorm room than the world's most advanced supercomputers, simply by being chucked into the air. That still does not make it intelligent or useful.

I accidentally dropped the nut on the floor earlier today and I lack the wherewithal to look for it, but my point still stands: what sort of processing power are you talking about? How are you measuring it? Without units, the term "processing power" is useless when comparing systems with completely different mechanisms.

[Alan Bolte]

What's typically used, I believe, is the capacity of the collection of neurons in the retina to detect edges and motion. This is assumed to be equivilent to the capacity of the neurons in the rest of the brain, and so is scaled up from the number of neurons in the retina to the number in the brain. This is compared against, say, $1000 worth of computing equipment one might put in a robot to process whatever it's using for vision. Apparently the number of instructions per second required to detect edges and motion hasn't changed significantly over the years.

[Dennis Toy]

I'm talking about the ability to input, calculate, process and then output when i discuss processing power. Computer may be able to process cold data but they cannot process such things as sight or taste. Todays super-computers may have visual processing, but all it does is detect movement and shape and colors using algorithims. They can no more experience the sensation of sight than a camera does.

A rudimentary life form like a one-cell animal with light processing ability can make quick-time reactions based on stimuli, this power alone requires a number of processors to be parallel processing. This in the animal only requires DNA filaments to be wired to take the stimuli and process it and send it to the cells.

Wonder why we don't have robots in our kitchens. Well we can make the parts but we can't create the programming, why? because it would take almost billions of code of data just to get the robot to pick up an orange and put it down safely. This would be impractical for home because this would cost millions and would be time-consuming. Neural Networks are being sought as a method of not having to do all this programming, all you have to do is show the robot what to do like a child and it will mis-step but it will learn.

[Dennis Toy]

How do you measure something like that? In terms of processing things, a single neuron is pretty useless. It's basically a glorified logic gate (with thousands of inputs). And a very slow one at that. Good luck using that as a super computer. In fact, the very reason we have super computers is that our brain, even though it has hundreds of billions of neurons, really sucks in comparison to the pure number-crunching abilities of a super-computer.

oh really...


http://vadim.www.media.mit.edu/MAS862/Project.html


http://www.scism.sbu.ac.uk/inmandw/revi ... .html#kw58

Both sites which are valid all state that the neuron is essentially a small computer that is connected to thousands of small computers.

[Lord of the Abyss]

I'm talking about the ability to input, calculate, process and then output when i discuss processing power. Computer may be able to process cold data but they cannot process such things as sight or taste. Todays super-computers may have visual processing, but all it does is detect movement and shape and colors using algorithims. They can no more experience the sensation of sight than a camera does.

A rudimentary life form like a one-cell animal with light processing ability can make quick-time reactions based on stimuli, this power alone requires a number of processors to be parallel processing. This in the animal only requires DNA filaments to be wired to take the stimuli and process it and send it to the cells.

Wonder why we don't have robots in our kitchens. Well we can make the parts but we can't create the programming, why? because it would take almost billions of code of data just to get the robot to pick up an orange and put it down safely. This would be impractical for home because this would cost millions and would be time-consuming. Neural Networks are being sought as a method of not having to do all this programming, all you have to do is show the robot what to do like a child and it will mis-step but it will learn.

I think you're confusing processing power and architecture. Neural networks don't process data faster; they process it differently. You'll note that neural networks are being proposed as a solution for such things as visual recognition ( which standard computers do badly ), but not math ( which standard computers do well ).

For that matter, you'll note that humans ( who are all neural network of course ) do much worse at the things standard computers are good at, like pure logic, mathematics and so on. Frankly, one could use the same reasoning you are to "prove" that humans have less processing power than a PC; they easily do things we can't manage, just as we do things they can't manage.

[Dooey Jo]

Both sites which are valid all state that the neuron is essentially a small computer that is connected to thousands of small computers.

Perhaps you can quote the parts that you feel support your case? Because I can only find one vague statement in the second link, that a neuron has a processing power "equivalent to a medium sized computer". Your average neuron either fires, or it does not, based on the inputs. Not only that, but it takes time, at least in the milliseconds, for the neuron to react to the input. It's not a medium sized computer (or five of worlds most advanced super-computers for that matter), it's a transistor with lots of pins (and a really complicated function for determining wether to fire or not).

So, again, how can you compare a single neuron to a CPU, which has hundreds of thousands of very fast transistors?