Is Artificial Intelligence Genuine Intelligence?

[Peter Soderqvist]

Wikipedia Daniel Dennett
Darwin's Dangerous Idea: Evolution and the Meanings of Life (1995) is a controversial book by Daniel Dennett which argues that Darwinian processes are the central organising force in the Universe. Dennett asserts that natural selection is a blind and algorithmic process which is sufficiently powerful to account for the generation and evolution of life including the ins and outs of human minds and societies.
http://en.wikipedia.org/wiki/Darwin%27s_Dangerous_Idea

Wikipedia Alan Turing
Turing machines are to this day the central object of study in theory of computation.
He went on to prove that there was no solution to the Entscheidungsproblem by first showing that the halting problem for Turing machines is undecidable: it is not possible to decide, in general, algorithmically whether a given Turing machine will ever halt. While his proof was published subsequent to Alonzo Church's equivalent proof in respect to his lambda calculus, Turing's work is considerably more accessible and intuitive. It was also novel in its notion of a "Universal (Turing) Machine", the idea that such a machine could perform the tasks of any other machine.
http://en.wikipedia.org/wiki/Alan_turing

Soderqvist1: It seems to me that Dennet’s proposition is not consistent with Turing’s!
Because Turing has proven that the Entscheidungsproblem have no Algorithmic answer. Thus Turing’s Mind cannot be a machine becuase he has proven something which is undecidable by a Turing Machine! It is a contradiction in terms to say that; “it is not possible to decide, in general, algorithmically whether a given Turing machine will ever halt. Can be accounted for by the evolutionary algorithm!”

[Beerina]

Are you asking if evolution could solve the halting problem?

No, it could not, with one caveat. We presume reality doesn't have infinite granularity or access to methods to perform infinite processes (speed or quantity.)

If that is possible, then all bets are off.

In any case, evolution is not an algorithm guaranteed to find the best solution to a problem, or even any solution. It is a rule of thumb, a heuristic, i.e. something that works pretty well most of the time in practice to give decent results.

Note further that the human mind, a product of evolution, and everything the human mind can figure out count as "wins" for evolution, as well as for the mind itself.

[rocketdodger]

Soderqvist1: It seems to me that Dennet’s proposition is not consistent with Turing’s!
Because Turing has proven that the Entscheidungsproblem have no Algorithmic answer. Thus Turing’s Mind cannot be a machine becuase he has proven something which is undecidable by a Turing Machine! It is a contradiction in terms to say that; “it is not possible to decide, in general, algorithmically whether a given Turing machine will ever halt. Can be accounted for by the evolutionary algorithm!”

No.

The halting problem only applies to algorithms that are supposed to determine whether all program-input pairs will lead to termination. It says nothing about algorithms used to determine specific ones.

In particular, the halting problem itself is a specific program-input pair -- so determining the decidability of the halting problem (and we know it to be decidable, because turing proved it) is not precluded by the halting problem.

[drkitten]

Soderqvist1: It seems to me that Dennet’s proposition is not consistent with Turing’s!
Because Turing has proven that the Entscheidungsproblem have no Algorithmic answer.

You misunderstand both Dennett and Turing.

Thus Turing’s Mind cannot be a machine becuase he has proven something which is undecidable by a Turing Machine! It is a contradiction in terms to say that; “it is not possible to decide, in general, algorithmically whether a given Turing machine will ever halt. Can be accounted for by the evolutionary algorithm!”

No, it's possible for a Turing machine to prove, in general, that a particular class of problems is unsolvable. Such a proof is not inconsistent with an inability on the same part of the machine to solve any specific instance of the class. Or, for that matter, an ability to solve any specific one.

Case in point : It's easy to prove (and a computer could easily do it) that an assembly language program with no goto statements in it will halt. (It would be equally easy to prove that an assembly language programs with no backwards goto statements will halt.)

It's easy to prove that no computer can determine whether or not a given statement is a consequence of a given set of premises (this is basically Godel's theorem in disguise), but automatic proof generators exist, and they have generated proofs.

[rocketdodger]

It's easy to prove that no computer can determine whether or not a given statement is a consequence of a given set of premises (this is basically Godel's theorem in disguise),


Did you mean to say "no computer can determine in general whether or not" instead?

I was under the impression that specific instances may be decidable -- that is why theorem provers work?

[drkitten]

Did you mean to say "no computer can determine in general whether or not" instead?

Yes, exactly.



I was under the impression that specific instances may be decidable -- that is why theorem provers work?

Yup. And the difference between Goedelian undecidability and the effectiveness of theorem proving is precisely the issue of the OP. Give that man a cigar!

[cyborg]

No, it could not, with one caveat. We presume reality doesn't have infinite granularity or access to methods to perform infinite processes (speed or quantity.)

There is always a bigger infinity.

[blobru]

In Darwin's Dangerous Idea, Dennett actually devotes a chapter to something like the above -- arguing against Penrose in The Emperor's New Mind: that Godel undecidability means AI will never match human intuition.

As I recall, he uses the example of a chess program: while it's not feasible to search every node in its decision tree, and it's unlikely there is a feasible algorithm to guarantee checkmate, yet there are algorithms that are very good at achieving checkmate; so a human player who's very good at achieving checkmate may be using an algorithm too.

In the same way, even without an algorithm to generate every truth in arithmetic, there may be heuristic algorithms that are great at guessing arithmetic truth, and which account for human mathematicians' intuition. Likewise for Turing Machines: even without an algorithm for determining whether every program / input pair will halt, there may be algorithms that are great at guessing which will halt, for machine and human.

[Complexity]

If and when a few smart people are able to create the circumstances under which an artificial intelligence emerges, then it will be real intelligence.

Next question?

[CardZeus]

By definition 'artificial' intelligence cannot be 'genuine'.

[Complexity]

By definition 'artificial' intelligence cannot be 'genuine'.


You are wrong, I'm afraid.

I do not think that these words mean what you think they mean.

[Moochie]

Oh dear! Is there an "ultimate programmer"?


M.

[PixyMisa]

Soderqvist1: It seems to me that Dennet’s proposition is not consistent with Turing’s!
As others have noted, this is not correct.

Because Turing has proven that the Entscheidungsproblem have no Algorithmic answer.
Yes.

Thus Turing’s Mind cannot be a machine becuase he has proven something which is undecidable by a Turing Machine!
And welcome to non-sequitur of the month club!

Turing's proof that the halting problem is not generally solvable is in itself reachable algorithmically. The existence of the proof proves this.

Turing cannot solve the halting problem. The proof proves that.

It does not necessarily follow that Turing's mind is a machine, but it certainly does not follow that Turing's mind is not a machine.

It is a contradiction in terms to say that; “it is not possible to decide, in general, algorithmically whether a given Turing machine will ever halt. Can be accounted for by the evolutionary algorithm!”
No. In fact, Turing proved that this is actually true.

[69dodge]

Turing cannot solve the halting problem. The proof proves that.

I don't think the proof proves that.

The proof applies to Turing machines, not minds, unless minds are equivalent to Turing machines in their capabilities. I believe they are, and I think Turing did too, but how could this possibly be proved?

[Skeptical Greg]

By definition 'artificial' intelligence cannot be 'genuine'.

By definition, genuine ' intelligence ' cannot be artificial ...

Once a computer ( program ) is indistinguishable from human consciousness ..
It would be.. Uhh, well ... Indistinguishable from human consciousness ..

[69dodge]

Because Turing has proven that the Entscheidungsproblem have no Algorithmic answer. Thus Turing’s Mind cannot be a machine becuase he has proven something which is undecidable by a Turing Machine! It is a contradiction in terms to say that; “it is not possible to decide, in general, algorithmically whether a given Turing machine will ever halt. Can be accounted for by the evolutionary algorithm!”

No algorithm correctly decides, for every Turing machine, whether it halts.

Turing gave a proof of this fact.

Such a proof is not an algorithm that correctly decides, for every Turing machine, whether it halts. (It's not an algorithm at all. It's a proof.)

So, where's the contradiction?

[PixyMisa]

I don't think the proof proves that.

The proof applies to Turing machines, not minds, unless minds are equivalent to Turing machines in their capabilities. I believe they are, and I think Turing did too, but how could this possibly be proved?
That's a valid point. But are there rigorous algorithmic sytems that are more general than the Turing machine? I recall reading a discussion on that subject, but I can't recall the conclusion.

[rocketdodger]

I don't think the proof proves that.

The proof applies to Turing machines, not minds, unless minds are equivalent to Turing machines in their capabilities. I believe they are, and I think Turing did too, but how could this possibly be proved?

I suppose you could start with the properties of neurons and show that any combination of them can always be reduced to a turing machine.

That would prove it for materialists, at least. Does anyone else matter?

[drkitten]

I suppose you could start with the properties of neurons and show that any combination of them can always be reduced to a turing machine.

Except that you can't. In fact, neurons are provably NOT Turing-equivalent, since neurons are continuous and Turing machines are discrete.

Depending upon how you generalize Turing machines, it's actually fairly easy to solve the halting problem for a discrete Turing machine with a continuous one --- but the program (being an infinitely long lookup table) would take a lot of time to write.

[drkitten]

That's a valid point. But are there rigorous algorithmic sytems that are more general than the Turing machine? I recall reading a discussion on that subject, but I can't recall the conclusion.

Yes there are, but we can't build any of them.

Not that it really matters, since the mind is demonstrably not a rigorous algorithmic system (if it were, we wouldn't make logical errors). That's the basic fallacy of the Penrose/Lucas argument. Since humans are demonstrably capable of "deriving," "learning," and "believing" things that are not only false, but actively contradictory, they wouldn't be a "consistent" system, and therefore the Goedelian limitations about not being able to prove their own limits doesn't apply.

[rocketdodger]

Except that you can't. In fact, neurons are provably NOT Turing-equivalent, since neurons are continuous and Turing machines are discrete.

I think it would be pretty hard to prove that there isn't any granularity at which a neuron can effectively be considered to be an aggregation of discrete entities. You would have to delve into QM for that, wouldn't you?

[Modified]

Except that you can't. In fact, neurons are provably NOT Turing-equivalent, since neurons are continuous and Turing machines are discrete.

But at some level of granularity, the difference between continuous and discrete will make little difference in function.

ETA: Oops, I'm not sure why rocketdodger's post is just showing up for me now.

[Modified]

Yes there are, but we can't build any of them.

To give a bit more of an explanation, such systems include an "oracle", which can compute things which are not Turing computable.

[drkitten]

To give a bit more of an explanation, such systems include an "oracle", which can compute things which are not Turing computable.

Or alternatively, a real-valued RAM is provably more powerful than a fixed-point RAM, which is known to be Turing-equivalent. But building machines with an infinite amount of memory gets expensive.

[joobz]

I think it would be pretty hard to prove that there isn't any granularity at which a neuron can effectively be considered to be an aggregation of discrete entities. You would have to delve into QM for that, wouldn't you?
I don't think you need to go down to the QM level to show that. I think the receptor/ligand pairing systems are discrete enough. Further, all membrane potential transmissions are regulated by aggregates of discerete ion channels.

[rocketdodger]

I don't think you need to go down to the QM level to show that. I think the receptor/ligand pairing systems are discrete enough. Further, all membrane potential transmissions are regulated by aggregates of discerete ion channels.

Yeah but someone could claim that if the channels have different numbers of atoms in just one of the many proteins that make them up that the channels are in fact different.

They could also claim that the distance between channels, or whatever, affects the probability of something somewhere happening.

Of course, you and I know it wouldn't make any difference -- just like the exact arrangement of atoms in a transistor is irrelevant as far as a computer is concerned -- but someone can always claim "well, it just just just just just might."

[drkitten]

Of course, you and I know it wouldn't make any difference -- just like the exact arrangement of atoms in a transistor is irrelevant as far as a computer is concerned -- but someone can always claim "well, it just just just just just might."

Well, that's a problem. I don't see how you "know" that, especially since we know that lots of other biological and physical processes appear to be chaotic (in the technical sense) and therefore succeptible to things like the butterfly effect. In point of fact, the exact arrangement of atoms in a transistor is quite relevant as far as a computer is concerned, as anyone who is trying to figure out just which component blew can and will attest.

Now, of course, you can always claim that the neurons were "designed" (like a properly built circuit) with a certain amount of tolerance at each component (remember the +/- 20% resistors we used to play with) and that quantization below the tolerance level will therefore have no effect on overall system performance. That's more or less the same argument as the (true) statement that you can always run a finite RAM program on a real-valued RAM. In either case, you've got a system capable of greater representation and you're not making use of the full capacity of the system. Aside from the theological implications of "design," we've actually got very little evidence of such error-correction in the neuroanatomy, so it's at best a very weak claim.

But it doesn't affect the fundamental argument that neurons are inherently continuous, and therefore closer in their theoretical computational capacity to a real-valued RAM than a discrete/fixed-point one or a TM.

[rocketdodger]

But it doesn't affect the fundamental argument that neurons are inherently continuous, and therefore closer in their theoretical computational capacity to a real-valued RAM than a discrete/fixed-point one or a TM.

I agree.

But to be fair, joobz and I are thinking of a static collection of neurons -- grown and wired -- as being discrete. You seem to be thinking of the dynamic collection, able to re-grow and change their wiring.

[Vorpal]

Well, on the positive side, we can say that the quantum phase space is separable, and therefore any localized wavefunction can in principle be computed to any desired accuracy, as can any smooth one (regardless of localization, although this would be perverse in the first place). On the common-sense side regarding AI in general rather than idealized Turing representation in particular, we can say "your neurons are quantum systems? so are our circuits."

[69dodge]

Except that you can't. In fact, neurons are provably NOT Turing-equivalent, since neurons are continuous and Turing machines are discrete.

Do you mean that neurons are provably stronger than Turing machines, assuming that neurons are continuous? Or do you mean that neurons are provably continuous? I don't see how it could be proved that neurons are continuous: we've only done, and can only ever do, a finite number of finite-precision experiments on them.

[drkitten]

Do you mean that neurons are provably stronger than Turing machines, assuming that neurons are continuous? Or do you mean that neurons are provably continuous?

Primarily the first, but secondarily the second. If neurons are provably continuous, then a neural network is provably equivalent to a real-valued RAM, which is more powerful than a Turing machine.

But since neurons are known to use the time domain for their computations (we know this from the experiments that we've done; time and duration of pulses are actually more important than magnitude, and to the best of our physics knowledge, time is continuous, we've got what amounts to scientific proof that neurons are continuous.

So we've got a mathematical/theoretical proof for the first, and a scientific/empirical proof for the second.

The alternative would be to throw out more or less every neurophysiological experiment since Hebb or re-write physics in full since about 1905 to make time discrete.

I don't see how it could be proved that neurons are continuous: we've only done, and can only ever do, a finite number of finite-precision experiments on them.

We've also only done a finite number of experiments in a very local area showing that gravity exists, or that the speed of light is constant, or that charge is conserved -- and yet no practicing physicist would consider them to be anything other than "proved" or take seriously an argument that started with "Since we don't know that charge is not conserved...."

[drkitten]

Well, on the positive side, we can say that the quantum phase space is separable, and therefore any localized wavefunction can in principle be computed to any desired accuracy, as can any smooth one (regardless of localization, although this would be perverse in the first place).

That's exactly the difference between real-valued RAMs and discrete ones; we can similate real-valued RAMs on our (discrete) computers to any desired accuracy, but the representational errors multiply (cf. the butterfly effect alluded to earlier, or the problems with simulating the three-body problem) over extended time scales. Without true infinite precision, we can't actually compute the true values.

You've basically proven my point for me.

For which I thank you.

[69dodge]

So we've got a mathematical/theoretical proof for the first, and a scientific/empirical proof for the second.

Yes, that's a good way of putting it. But, then, any conclusion that depends on both has at most the status of being empirically "proven". And, on the other hand, it seems to me that the distinction between a Turing machine and something more powerful is meaningful only mathematically. How could you distinguish between them empirically?

We've also only done a finite number of experiments in a very local area showing that gravity exists, or that the speed of light is constant, or that charge is conserved -- and yet no practicing physicist would consider them to be anything other than "proved" or take seriously an argument that started with "Since we don't know that charge is not conserved...."

How does the hypothetical argument end? If the conclusion really depended on whether charge is conserved exactly, rather than on whether charge is conserved to within the experimental error of the best experiments done to date, I would take it seriously. I'm not a practicing physicist, but I hope they would take it seriously too.

[Modified]

That's exactly the difference between real-valued RAMs and discrete ones; we can similate real-valued RAMs on our (discrete) computers to any desired accuracy, but the representational errors multiply (cf. the butterfly effect alluded to earlier, or the problems with simulating the three-body problem) over extended time scales. Without true infinite precision, we can't actually compute the true values.

In the real world, there is noise. At some level of granularity, a digital simulation will be as powerful as a brain. To better simulate an actual brain, we might have to go to an even lower level of granularity and simulate the effects of noise as well.

[drkitten]

In the real world, there is noise. At some level of granularity, a digital simulation will be as powerful as a brain.

Unless the noise is actually an integral part of brain activity (for example, in symmetry breaking).

[Modified]

Unless the noise is actually an integral part of brain activity (for example, in symmetry breaking).

But the noise can be simulated as well. It is unlikely that true randomness would provide any advantage.

[Robin]

I recall attending a lecture some time ago by a researcher in the field of neural networks and he suggested that artificial neural nets should not use discrete computers at all, but should be constructed from scratch using components that have continuous values.

I am not sure whether any practical device of this type has ever been built. But if so then it could be called intelligent in the same way that a human is intelligent.

But even if a simulation on a discrete system could not model any actual brain, I don't see any reason why we wouldn't call it 'intelligent' if it could produce the same type of results, for example adaptive pattern matching.


PS didn't Chaitin conjecture (quoting Feynman) that there may be no continuous values in reality? That would add another interesting dimension.

[Robin]

But the noise can be simulated as well. It is unlikely that true randomness would provide any advantage.
Yes, in fact the level of noise provided by the randomness of our sense data would certainly swamp any randomness inherent in our brains so I imagine there would not be any practical difference between a discrete system processing actual sense data and a continuous system doing the same.

[Nogbad]

Sooo... if we build a machine that enjoys watching Big Brother have we created artificial intelligence?

[Robin]

Sooo... if we build a machine that enjoys watching Big Brother have we created artificial intelligence?
Hmm.. I imagine that the whole field of artificial stupidity is stymied by the abundance of the naturally occuring product.

[Nogbad]

Hmm.. I imagine that the whole field of artificial stupidity is stymied by the abundance of the naturally occuring product.

One suspects artificial stupidity would be the more challenging discipline to work in.

[cyborg]

Enjoyment is an emotional trait, not an intellectual one.

[Modified]

I recall attending a lecture some time ago by a researcher in the field of neural networks and he suggested that artificial neural nets should not use discrete computers at all, but should be constructed from scratch using components that have continuous values.

That may be a way to create brain-like processing that operates much faster than a digital simulation of brain-like processing.

[Nogbad]

Enjoyment is an emotional trait, not an intellectual one.

http://onthebox.netfirms.com/Articles/SheepInCity/sitbcHCms03.gif