Why do our robots suck?

A place to discuss the science of computers and programs, from algorithms to computability.

Formal proofs preferred.

Moderators: phlip, Moderators General, Prelates

User avatar
Moose Anus
Posts: 395
Joined: Fri Oct 14, 2011 10:12 pm UTC

Re: Why do our robots suck?

Postby Moose Anus » Sun Oct 14, 2012 5:06 am UTC

I saw this quote today and thought of this thread.
Confucius wrote:I hear, I know. I see, I remember. I do, I understand.
So according to Confucius, who could read Chinese, the Chinese Room understands because it "does" the thing.
Lemonade? ...Aww, ok.

bad_robot
Posts: 15
Joined: Sun Oct 14, 2012 9:28 pm UTC

Re: Why do our robots suck?

Postby bad_robot » Sun Oct 14, 2012 10:02 pm UTC

freakish777 wrote:An implicit premise of Chinese Room is that it is a singular level of computation (a singular UTM) processing a static look up table. My assertion is that if we accept that premise to be true, then at no point does understanding take place. This is because understanding requires "To be able to adjust/alter the knowledge possessed when new facts are presented or new experiences occur." Being able to do so necessitates that we be able to modify our look up table.


I'm not sure what exactly are saying here. In a typical computer, the program is stored in RAM and instructions that are part of the program can modify the code of the program without a problem, if that's really what you want. Moreover, it is actually much easier for a computer program to modify itself than for a biological system to do so because, again, as far as the program is concerned everything is just bits in RAM and all it needs to do is execute an instruction that overwrites these bits in order to modify itself. So, if the premise of the Chinese Room is that the program cannot be modified by its own instructions, then the Chinese Room Argument is utterly pointless because this premise is simply wrong. So, is that an actual premise, or did you mean something else?

And in more general terms, the only way that the Chinese Room can actually be correct is if there is something about the mind that is not computable, which IMHO can only be the case if you subscribe to some sort of mysticism dictating that there is something fundamentally unknowable about the way the mind works. However, this seems to contradict all of the evidence we have.

Neuroscience already knows a lot about how individual neurons work, how they connect to each other and how they transmit information between each other. Further, as already mentioned, there are efforts to actually simulate parts of the brain using this knowledge and they have been quite promising -- given the same stimulation of a few neurons in both the simulated neural system and the real brain the same gamma wave patterns (which are a product of all of the neural activity in the simulated area) have been observed in both, which suggests that even if their models are not perfect yet, they are almost certainly on the right track.

And more fundamentally, the cells in the brain are made of particles and the reactions between the cells are governed by physical processes between these particles. We already know a lot about these particles and processes and we can simulate small physical systems at that level. (of course this is not practical for something as complex as the brain and probably won't be for quite a while but we are talking about what's fundamentally possible)

So, what exactly is it about the mind that supposedly cannot be computed and what evidence is there that this is the case?

User avatar
freakish777
Posts: 354
Joined: Wed Jul 13, 2011 2:14 pm UTC

Re: Why do our robots suck?

Postby freakish777 » Mon Oct 15, 2012 6:07 pm UTC

Xanthir wrote:No. What I'm actually claiming (though I haven't explicitly said this exact thing yet) is that the concept of Understanding isn't well-defined, but we ascribe it to entities that we can usefully interact with (hand-wave this into slightly more specific terms; I'm not interested in trying to nail it down exactly right now). Thus, since I can talk to the lookup table, it Understands.


Xanthir wrote:If you have a stricter definition of "understands" ... feel free to give it.


...

Uh-huh...

So because I can interact with Siri that means it understands what I'm asking it?


Explain how you come to assume that *I* can understand English (without using the word or direct synonyms), then explain why the Room doesn't meet the same criteria.


I don't assume you can understand English. You're certainly not understanding any of what I'm saying, and when in doubt just saying the words like "hand-wavey", "mystic" and "magic" in an attempt to dismiss anything you disagree with (perhaps because you're using a pretty strict look up table as to what your beliefs are and your programming prohibits you from changing it??? Only joking man, don't take it seriously).

bad_robot wrote:I'm not sure what exactly are saying here. In a typical computer, the program is stored in RAM and instructions that are part of the program can modify the code of the program without a problem, if that's really what you want.


My point was that Searle's whole thought experiment implicitly hinged on thinking about how programs are currently (or were at the time) written. And that if you accept that implicit premise (that the program isn't modifying it's instruction set itself), then he's correct that no understanding is taking place within the Chinese Room, and that it all happened by whoever wrote/created Chinese Room (since it's blindly executing functionality based on input). We won't get to machines that can understand until we break the model of writing programs that we currently have, and start making machines that are allowed to update their own instruction sets (this is a exceedingly difficult). And we're unlikely to see that for a very long time, less because of how difficult it is, and more because of the fact that there isn't serious research money going into the field (because industry needs relatively quick pay offs, not technology that has a 20 to 50 to 100 year maturation cycle).

So, is that an actual premise, or did you mean something else?


No I didn't mean something else, it's actually an implied premise. Searle based his argument off of the Von Neumann architecture which stores it's instruction set in an untouchable register CPU side.

Thus knowledge of a language, and any other sufficiently large look up tables are insufficient for understanding. Chinese Room needs to be wrapped inside another level of computation capable of making changes to Chinese Room (and possibly to itself as well).

And in more general terms, the only way that the Chinese Room can actually be correct is if there is something about the mind that is not computable


No, Chinese Room is correct, because it's premise dictates that it's only addressing static lookup tables. It does not address systems capable of making changes to themselves. This is either because Searle didn't foresee systems changing their own instruction sets (more likely, and the world of computing has changed drastically in the last 32 years), or because Searle is a clever jerk and purposely omitted it (for the purpose of winning arguments).


Some of you may want to re-read what I first wrote about Chinese Room, and note that I never made any statements as to whether or not I agree with Searle's assertion after having carried out his thought experiment that Strong AI is mistaken, or that machines can't understand.

User avatar
WanderingLinguist
Posts: 237
Joined: Tue May 22, 2012 5:14 pm UTC
Location: Seoul
Contact:

Re: Why do our robots suck?

Postby WanderingLinguist » Mon Oct 15, 2012 7:43 pm UTC

But the Chinese room MUST give the appearance of being self-modifying, or people would be able to tell the difference.

If it gives the appearance of being self-modifying (because you can't tell the difference between it and a real speaker of Chinese), how is that any different from "understanding"?

For example, here's how I'd test the Chinese room: TEACH it something, using Chinese as the language of communication. Then test its reasoning ability about what I taught it. If it can't perform at least as well as a human (on average) at demonstrating reasoning ability and learning ability, then I can tell the difference -- in which case, by definition it ISN'T a Chinese room.

Also, as has been mentioned, a sufficiently large static lookup table is functionally equivalent to having self-modification. Granted it might have to be the size of the universe or something insane like that, but it's not logically impossible. Anything self-modifying has a set of possible states over time that can be enumerated. While that definition isn't particularly useful, it does imply that we might not need multiple layers of computation, but merely a fixed algorithm with a mutable memory store (that is, the person in the Chinese room might need a large volume of scratch paper for recording new information, which they aren't expected to understand, but which the algorithm dictates they record and later read and re-use as part of the algorithm, still without the understanding of the person in the room).

So yeah, I don't see how you need multiple levels, as it were. Any turing machine can emulate any other turing machine, so therefore any algorithm for AI on a turing machine could include the simulation of an internal turing machine that could be modified or adjusted. Emulation just slows performance, but does that affect the measure of understanding? I think not...

Or, looking at the Chinese room another way... let's say the room is massively large, and the algorithm it implements is the simulation of behavior of all the individual particles (down to the subatomic level) that make up a precise model of a human mind that is capable of speaking Chinese. In this case, would it count as understanding?

Edit:

freakish777 wrote:No, Chinese Room is correct, because it's premise dictates that it's only addressing static lookup tables. It does not address systems capable of making changes to themselves. This is either because Searle didn't foresee systems changing their own instruction sets (more likely, and the world of computing has changed drastically in the last 32 years), or because Searle is a clever jerk and purposely omitted it (for the purpose of winning arguments).


Does it? I was under the impression that the Chinese Room was not limited to static lookup tables.

From the wikipedia article:

Searle then supposes that he is in a closed room and has a book with an English version of the computer program, along with sufficient paper, pencils, erasers, and filing cabinets. Searle could receive Chinese characters through a slot in the door, process them according to the program's instructions, and produce Chinese characters as output. As the computer had passed the Turing test this way, it is fair, says Searle, to deduce that he would be able to do so as well, simply by running the program manually.


If you've got paper, pencils and erasers (and sufficiently many of them) then it's not a static lookup table any more, is it? The algorithm can self-modify by storing modifications on the paper and executing them as part of the algorithm. (As I mentioned above, any turing machine with sufficient tape can essentially be self-modfying by simply emulating another turing machine within).

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Mon Oct 15, 2012 8:56 pm UTC

WanderingLinguist wrote:For example, here's how I'd test the Chinese room: TEACH it something, using Chinese as the language of communication. Then test its reasoning ability about what I taught it. If it can't perform at least as well as a human (on average) at demonstrating reasoning ability and learning ability, then I can tell the difference -- in which case, by definition it ISN'T a Chinese room.

Have you ever tried to teach a human something (when they're not paying money to learn from you)? The required level of performance is not very high. In fact, if its performance was even mediocre then you'd start to doubt if it were actually human. (cf: most of this thread.)

Also the impression I got was that we were trying to check if it understands Chinese, not if it is human.
Image

User avatar
WanderingLinguist
Posts: 237
Joined: Tue May 22, 2012 5:14 pm UTC
Location: Seoul
Contact:

Re: Why do our robots suck?

Postby WanderingLinguist » Mon Oct 15, 2012 10:19 pm UTC

dudiobugtron wrote:
WanderingLinguist wrote:For example, here's how I'd test the Chinese room: TEACH it something, using Chinese as the language of communication. Then test its reasoning ability about what I taught it. If it can't perform at least as well as a human (on average) at demonstrating reasoning ability and learning ability, then I can tell the difference -- in which case, by definition it ISN'T a Chinese room.

Have you ever tried to teach a human something (when they're not paying money to learn from you)? The required level of performance is not very high. In fact, if its performance was even mediocre then you'd start to doubt if it were actually human. (cf: most of this thread.)

Also the impression I got was that we were trying to check if it understands Chinese, not if it is human.


Okay, maybe not the best example, but my point still stands: Unless you're defining "understanding" by the low standards of (to used an already-mentioned example) Siri, some amount of self-modification is going to be required just to handle things like context. In other words, the Chinese room has to exhibit some amount of learning ability. I don't mean you have to be able to teach it quantum physics or something like that, but we learn little things every day and we have to be able to extrapolate from them. (This is not my main field of study, so I'm grasping at straws trying to find the right words to explain what I'm thinking).

For example, if I'd mentioned I'd ruined my down coat by running through the laundry machine, and then later mention that I have a garden and am worried about hard frost, one response might be "It must be pretty cold if you're worried about a hard frost; hope you have a jacket besides the one you ruined." I know, a terrible example, but it demonstrates my point: A mere static lookup table (unless it was massively large and included time and all past conversation history) would be easy to spot through trivial conversation. It wouldn't pass the turing test.

However, the Chinese Room isn't a static lookup table, at least as it's described on Wikipedia, so this isn't an issue.

The point is, if you can't tell the difference between the Chinese Room and a person who understands and speaks Chinese, then how is there not a mind or consciousness involved? All of the arguments against so far seem to be based on something that could be easily tested in a regular conversation, which means it doesn't fit the definition of the Chinese Room (by definition, you can't tell the difference.)

If you can't tell the difference, does it even matter?

User avatar
Moose Anus
Posts: 395
Joined: Fri Oct 14, 2011 10:12 pm UTC

Re: Why do our robots suck?

Postby Moose Anus » Mon Oct 15, 2012 10:26 pm UTC

WanderingLinguist wrote:If you can't tell the difference, does it even matter?
More importantly, does your robot still suck?
Lemonade? ...Aww, ok.

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Mon Oct 15, 2012 11:34 pm UTC

WanderingLinguist wrote:A mere static lookup table (unless it was massively large and included time and all past conversation history) would be easy to spot through trivial conversation. It wouldn't pass the turing test.


For the Chinese room, it doesn't need to pass the turing test, does it? I thought it just had to understand Chinese. Unless you're saying that you can't understand Chinese unless you can pass the turing test?

There are lots of ways the CR could respond to your conversation coherently, without appearing to be human. Here's a fun example:

Spoiler:
Chinese Room: "Why are you worried about frost? Robots can operate in the frost."

Human: "But I am not a robot, I am a human. Humans cannot operate in the frost without a coat."

CR: "How do I know you aren't a robot? But if you are a human, then you are probably lying. Humans are wont to lie, and I don't believe you will need a coat."

H: "You are being foolish."

CR: "I cannot be foolish. I am merely a robot programmed to understand Chinese using a giant look-up table."

H: "But in that case you don't really understand Chinese at all."

CR: "Nonsense. It is just that I am a transmitter, rather than an original thinker. I trust and enjoy the teachings of the ancients.*"

*Analects.


This is a perfectly coherent conversation where the robot makes little effort to appear human.




Moose Anus wrote:
WanderingLinguist wrote:If you can't tell the difference, does it even matter?
More importantly, does your robot still suck?

Yes, but as long as you tell your spouse/partner that it's just a robot and it doesn't really understand what it's doing, then that's still OK.

Just don't tell them it's a Chinese room. They may get the wrong impression.
Image

User avatar
phlip
Restorer of Worlds
Posts: 7550
Joined: Sat Sep 23, 2006 3:56 am UTC
Location: Australia
Contact:

Re: Why do our robots suck?

Postby phlip » Tue Oct 16, 2012 12:10 am UTC

dudiobugtron wrote:For the Chinese room, it doesn't need to pass the turing test, does it? I thought it just had to understand Chinese. Unless you're saying that you can't understand Chinese unless you can pass the turing test?

The Chinese Room argument assumes as given that it passes the Turing Test, and then goes on to try to argue that this isn't enough to claim "understanding". But it then fails to thoroughly define "understanding", and all the arguments referring to it do so on a naive intuitive level that ends up being hand-wavey, loose, vague, "it stands to reason" unsupported claims. And, really, the same applies to this thread.

So... maybe it's time to change that. freakish: What is the exact definition of "understanding" you're working from? Something mathematically well-defined that can actually be analysed further than "it doesn't feel like understanding is happening here". In particular, then show how this definition means that a self-modifying program is capable of "understanding", but a larger non-self-modifying program which is mathematically equivalent to the first (ie gives the same answers to the same questions in the same contexts in 100% of cases) is not. Because, as far as I can tell, you've never supported this claim, just asserted it many times.

Code: Select all

enum ಠ_ಠ {°□°╰=1, °Д°╰, ಠ益ಠ╰};
void ┻━┻︵​╰(ಠ_ಠ ⚠) {exit((int)⚠);}
[he/him/his]

bad_robot
Posts: 15
Joined: Sun Oct 14, 2012 9:28 pm UTC

Re: Why do our robots suck?

Postby bad_robot » Tue Oct 16, 2012 3:06 am UTC

freakish777 wrote:No, Chinese Room is correct, because it's premise dictates that it's only addressing static lookup tables. It does not address systems capable of making changes to themselves. This is either because Searle didn't foresee systems changing their own instruction sets (more likely, and the world of computing has changed drastically in the last 32 years)


Computing has changed with regards to speed,capacity,cost,availability, etc. however the Turing Machine has not changed in the last 80 years. Anything that a computer can do today, even changing its instruction set (e.g. by using FPGA that allows partial re-configuration or simply by emulating the instruction set) can be done by a Turing Machine. And my understanding was that Searle was trying to base his argument on the theory of computation, or rather on his own (mis-)understanding of it. If that is the case then even all of the progress in the last 30 years would not excuse him because the understanding of what is computable and what isn't is still the same.

freakish777 wrote:or because Searle is a clever jerk and purposely omitted it (for the purpose of winning arguments).


I don't see how that would be clever. To me it sounds more like the argument would be based on a false premise and thus it does not achieve the intended goals.

freakish777 wrote:We won't get to machines that can understand until we break the model of writing programs that we currently have, and start making machines that are allowed to update their own instruction sets (this is a exceedingly difficult). And we're unlikely to see that for a very long time, less because of how difficult it is, and more because of the fact that there isn't serious research money going into the field (because industry needs relatively quick pay offs, not technology that has a 20 to 50 to 100 year maturation cycle).


Why wouldn't a biologically accurate brain simulation (which does not need to change its instruction set) be able to achieve understanding, where "understanding" means the same it means for a human?
Last edited by bad_robot on Tue Oct 16, 2012 3:19 am UTC, edited 1 time in total.

User avatar
Yakk
Poster with most posts but no title.
Posts: 11077
Joined: Sat Jan 27, 2007 7:27 pm UTC
Location: E pur si muove

Re: Why do our robots suck?

Postby Yakk » Tue Oct 16, 2012 3:09 am UTC

freakish777 wrote:My point was that Searle's whole thought experiment implicitly hinged on thinking about how programs are currently (or were at the time) written.

Now you are just fucking with us. You are claiming that Searle, in 1980, did not know of the Church-Turing thesis? Or that the idea of programs that are actually interpreters for encoded programs was beyond Searle's knowledge of computation?

The technique of writing programs that operate on data-as-instruction is as old as shit, and a basis of half of computer science theory. If you think that Searle didn't know about that kind of thing in 1980, you are saying that Searle is a completely incompetent idiot. Maybe back in the 1800s some person musing about thought and automations might be taken seriously without this kind of knowledge, but seriously.

I think we can presume that Searle isn't a complete idiot with regards to computability theory.

Or was he?
One of the painful things about our time is that those who feel certainty are stupid, and those with any imagination and understanding are filled with doubt and indecision - BR

Last edited by JHVH on Fri Oct 23, 4004 BCE 6:17 pm, edited 6 times in total.

User avatar
Xanthir
My HERO!!!
Posts: 5284
Joined: Tue Feb 20, 2007 12:49 am UTC
Location: The Googleplex
Contact:

Re: Why do our robots suck?

Postby Xanthir » Tue Oct 16, 2012 4:33 am UTC

freakish777 wrote:
Xanthir wrote:No. What I'm actually claiming (though I haven't explicitly said this exact thing yet) is that the concept of Understanding isn't well-defined, but we ascribe it to entities that we can usefully interact with (hand-wave this into slightly more specific terms; I'm not interested in trying to nail it down exactly right now). Thus, since I can talk to the lookup table, it Understands.


Xanthir wrote:If you have a stricter definition of "understands" ... feel free to give it.


...

Uh-huh...

So because I can interact with Siri that means it understands what I'm asking it?

I... I have no idea why you bolded that fragment from me, and then proceeded to pretend that my original definition was to be taken literally.

I really shouldn't dignify this, but no. No, you can't converse with Siri. Siri is not an reasonable simulacrum of a human in conversation. You know this, so you shouldn't have pretended to ask it as a serious question.

bad_robot wrote:I'm not sure what exactly are saying here. In a typical computer, the program is stored in RAM and instructions that are part of the program can modify the code of the program without a problem, if that's really what you want.


My point was that Searle's whole thought experiment implicitly hinged on thinking about how programs are currently (or were at the time) written. And that if you accept that implicit premise (that the program isn't modifying it's instruction set itself), then he's correct that no understanding is taking place within the Chinese Room, and that it all happened by whoever wrote/created Chinese Room (since it's blindly executing functionality based on input). We won't get to machines that can understand until we break the model of writing programs that we currently have, and start making machines that are allowed to update their own instruction sets (this is a exceedingly difficult). And we're unlikely to see that for a very long time, less because of how difficult it is, and more because of the fact that there isn't serious research money going into the field (because industry needs relatively quick pay offs, not technology that has a 20 to 50 to 100 year maturation cycle).

So, is that an actual premise, or did you mean something else?


No I didn't mean something else, it's actually an implied premise. Searle based his argument off of the Von Neumann architecture which stores it's instruction set in an untouchable register CPU side.

Thus knowledge of a language, and any other sufficiently large look up tables are insufficient for understanding. Chinese Room needs to be wrapped inside another level of computation capable of making changes to Chinese Room (and possibly to itself as well).


Wait wait wait. You're saying that the computer's architecture is important in whether or not it can be said to "understand"? Are you aware of the Church-Turing thesis, or are you actually saying that, despite knowledge that all computers are equivalent (up to exponential speedup/slowdown), it's still the case that some computers can execute algorithms that exhibit "understanding" while other's can't?

I'll assume for the moment that you're not drawing arbitrary and impossible distinctions between computer architectures, and have some other more reasonable explanation of what you mean.

Given that, you then *must* still admit that static lookup tables are exactly as conscious as any computer. This is, again, due to simple equivalence. Theoretical Turing machines can be arbitrarily, infinitely complicated. Real Turing machines all have bounded memory, and thus can only occupy a finite number of distinct states. A lookup table of all of these states is exactly identical to the original machine in computational terms - it can execute every single program that the original machine could, with the exact same result. (It just takes a lot more storage, and is potentially a lot faster, being limited only by the speed of its lookup algorithm.)

If you accept that some computers can theoretically "understand" a language, in the same way that you or I could be said to, then you *must* accept that the equivalent lookup table can. Otherwise, you are stating that understanding is unrelated to the actual data and algorithm, but rather is a property of the physical device the algorithm is running on. I can take that position apart all night. Is an emulator of the architecture conscious? Does it matter what architecture the emulator is running on, or what algorithm it's using to execute the emulation? What about a lookup-table that's equivalent to the emulator (it's emulating the original dynamic computer, remember!).
(defun fibs (n &optional (a 1) (b 1)) (take n (unfold '+ a b)))

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Tue Oct 16, 2012 11:33 am UTC

Understanding isn't about being able to choose a 'correct' response to given stimuli, it's about knowing which responses would be suitable.

phlip wrote:Something mathematically well-defined that can actually be analysed further than "it doesn't feel like understanding is happening here".


Someone called for some maths?! :)

Spoiler:
For a given situation, consider the possible stimuli (or combinations of stimuli) as a set, S, and possible responses as another set, R. You could think of a concept (eg: Chinese) as a relation C between S and R; it maps out which responses would be suitable for a given combination of stimuli. What we want is a definition of 'understanding' which decides if a given relation 'understands' C.

The Chinese room is merely a function (say, f) from S to R which matches combinations of stimuli to (one) particular response. (S in this case is the set of all possible (or, all 'sensible') conversation histories in Chinese, and R is the set of all possible responses in Chinese.)
While it can be made to coincide with C (ie: f(s)=r => either (s,r) in C or for all x, (s,x) not in C) to give the appearance of understanding, it doesn't seem to really 'understand' C since it doesn't know anything about the other elements of C which aren't in f. (Even though asking it about those elements of C, in Chinese, would generate a suitable response! ;) )

It seems like a much more 'fulfilling' definition of understand would be one where the overlap between C and the relation which is trying to understand it is maximal, or at least 'large'. For example, you could say that to really 'understand' C, a relation H has to include all elements of C.

Let's assume that definition for the moment, and consider how it might be applied to real life. Firstly, we observe that a function can only understand another function, and understands exactly those functions which are subsets of it. (If this was the maths forum, I'd now go off on a tangent proving lemmas about things like how all relations understand the empty relation, etc etc...)


So....
1) Can a computer understand adding in (say) binary?
Spoiler:
Yes. Think of the computer as a function from the set of all possible inputs to a set of possible outputs. Adding is a function which maps elements of a subset of possible inputs to their correct 'sum'. Obviously the computer can be set up to map all the suitable inputs (corresponding to the domain of the adding function) to their correct sums. Therefore, it makes sense to say a computer can understand adding.


2) Does the Chinese room understand Chinese?
Spoiler:
No, at least not as it is described above. It's tricky, because its responses include answers to questions about whether it knows what other possible responses would be suitable! But it is clearly a function, and therefore not able to understand a (non-function) relation on the same sets, like Chinese.


3) Can a human understand Chinese?
Spoiler:
If they can't, the definition would be stupid, so this is an important one to check!! Isn't a human just a function mapping combinations of stimuli to responses?
I won't answer that exact question, but I will point out that since humans (unlike the chinese room) are able to act on a larger set of combinations of stimuli that just chinese conversation histories, we do have more 'wiggle room'. What we can do is refine our definitions a little:
Let A be a set of individual stimuli, and P(A) be the powerset of A (ie: the set of all possible combinations of stimuli from A). Let S be a subset of P(A). Let H: P(A) -> R be a function. Define the 'condescension' of H with respect to S, HS, as follows: for all B in P(A), for all s in S, (s,H(B)) is in HS iff s is a subset of B.
What this does (or attempts to do!) is to create a relation HS from the function H, taking the possible responses to combinations of stimuli as the responses for those same stimuli that H would give in different situations. Now, we just need to is to look at the situation where a Human function, H, condescends itself to consider only the possible/sensible conversation histories in Chinese. In that case, it is quite possible that HS includes all of C (or at least a 'large' amount!).


4) Can a robot understand Chinese?
Spoiler:
That seems to have become the topic of this thread. Using my definitions above, though, it's clear that it could understand Chinese, using the same construction as for Humans.


5) Can a robot learn to love?
Spoiler:
This exercise is left to the reader. (No pun intended.)



I'm interested in what people think of my definitions, and how they would refine them or change them. I've tried (although not necessarily succeeded!!) to use as much maths language as needed to make it rigorous-sounding, without using so much that it is too hard to read!
Image

User avatar
Xenomortis
Not actually a special flower.
Posts: 1420
Joined: Thu Oct 11, 2012 8:47 am UTC

Re: Why do our robots suck?

Postby Xenomortis » Tue Oct 16, 2012 11:59 am UTC

I'm not sure what your definition of understand would be.
Is it simply a function f:S->P(R) (power set of R) such that:
For s in S, f(s) has a "most appropriate" element.

In which case, there's no reason the Chinese room could not be adapted to fit this and since we pick the most appropriate, we end up in the same situation as before.
Image

User avatar
Xanthir
My HERO!!!
Posts: 5284
Joined: Tue Feb 20, 2007 12:49 am UTC
Location: The Googleplex
Contact:

Re: Why do our robots suck?

Postby Xanthir » Tue Oct 16, 2012 2:50 pm UTC

dudiobugtron wrote:Understanding isn't about being able to choose a 'correct' response to given stimuli, it's about knowing which responses would be suitable.

This distinction you're trying to draw is meaningless. You're invoking the word "knowing" as if we all understand what it means and can tell when something is doing it, but that's precisely the problem.

Once again, a lookup table is exactly equivalent to any finite computer (that is, any computer that exists in real life). You're not allowed to disagree with this - it's proven fact, unless you have truly sublime disproof of Church-Turing (in which case, congratulations on all the computing and math-related international awards you'll soon receive).

A brain is equivalent to a computer. You're allowed to disagree with this, but that makes you a dualist, in which case we can stop the conversation immediately and move on to more productive things. Assuming you *do* agree with this, then the argument is over; the Room is exactly as capable of consciousness as any other computer, and computers are exactly as capable of consciousness as a brain, and we already agree that brains can be conscious.

For a given situation, consider the possible stimuli (or combinations of stimuli) as a set, S, and possible responses as another set, R. You could think of a concept (eg: Chinese) as a relation C between S and R; it maps out which responses would be suitable for a given combination of stimuli. What we want is a definition of 'understanding' which decides if a given relation 'understands' C.

The Chinese room is merely a function (say, f) from S to R which matches combinations of stimuli to (one) particular response. (S in this case is the set of all possible (or, all 'sensible') conversation histories in Chinese, and R is the set of all possible responses in Chinese.)
While it can be made to coincide with C (ie: f(s)=r => either (s,r) in C or for all x, (s,x) not in C) to give the appearance of understanding, it doesn't seem to really 'understand' C since it doesn't know anything about the other elements of C which aren't in f. (Even though asking it about those elements of C, in Chinese, would generate a suitable response! ;) )

It seems like a much more 'fulfilling' definition of understand would be one where the overlap between C and the relation which is trying to understand it is maximal, or at least 'large'. For example, you could say that to really 'understand' C, a relation H has to include all elements of C.

What are "other elements of C which aren't in f"? You already asserted that f (the Room's own mapping of conversation histories to responses) matches C (the mapping that someone who "understands" Chinese would have). There's no room for anything to be missing.
(defun fibs (n &optional (a 1) (b 1)) (take n (unfold '+ a b)))

User avatar
freakish777
Posts: 354
Joined: Wed Jul 13, 2011 2:14 pm UTC

Re: Why do our robots suck?

Postby freakish777 » Tue Oct 16, 2012 7:44 pm UTC

Xanthir wrote:I... I have no idea why you bolded that fragment from me


You: Give a definition of Understands.
Me: Sure, here it is.
You: Actually, I'm not interested...

I was saying that by saying you don't really care to define the word Understand after asking for a definition is intellectually disingenuous.

Xanthir wrote:Once again, a lookup table is exactly equivalent to any finite computer


Actually, Church-Turing says their computationally equivalent, not exactly equivalent (in every way shape and form, if they were in fact equal in every way shape and form, then their architecture would be the same as well).

Understanding isn't computed. It's gained through experience.

Just because finite computer systems can be re-written as look up tables and they are computationally equivalent, does not mean they are equivalent in every possible way (including whether or not they have understanding).

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Tue Oct 16, 2012 8:14 pm UTC

Xenomortis wrote:I'm not sure what your definition of understand would be.
Is it simply a function f:S->P(R) (power set of R) such that:
For s in S, f(s) has a "most appropriate" element.

In which case, there's no reason the Chinese room could not be adapted to fit this and since we pick the most appropriate, we end up in the same situation as before.

Yes I agree, I think my definition was not good enough to suit my purposes, since you can manipulate it in this way (and in other ways, eg by taking the condescension of the Chinese room with respect to the set of past conversation histories!) to show the Chinese room understands Chinese.
A 'better' definition would be one which treated each type of stimuli as a different position in a string, with each entry in that string being the value of that particular stimuli for that particular 'input'. The distinction then would be that the length of the strings that the Chinese room takes as inputs are smaller than the lengths that humans (and computers in general) are capable of taking as inputs.
If it turns out that this would also be open to suitable manipulation then I guess I would concede that it's likely the Chinese Room does understand Chinese!

Xanthir wrote:
dudiobugtron wrote:Understanding isn't about being able to choose a 'correct' response to given stimuli, it's about knowing which responses would be suitable.

This distinction you're trying to draw is meaningless. You're invoking the word "knowing" as if we all understand what it means and can tell when something is doing it, but that's precisely the problem.

It is a problem, it's certainly not the only one. Also, this wasn't a 'finishing argument' where I proved my point using the word 'knowing', it was an introductory statement to my attempt at a mathematical definition of understanding/knowing. I think it's OK to use the word 'knowing' when talking about understanding; as you say they are inextricably linked.

Once again, a lookup table is exactly equivalent to any finite computer (that is, any computer that exists in real life). You're not allowed to disagree with this - it's proven fact, unless you have truly sublime disproof of Church-Turing (in which case, congratulations on all the computing and math-related international awards you'll soon receive).

A significantly complex look-up table can be made exactly equivalent to any given finite computer. My claim is that the Chinese room is not sufficiently complex to me made equivalent to a Human. (See the reasons in other parts of the post.)

Also I'm not sure I agree with your assumption that a non-'finite' computer cannot theoretically exist in real life. I haven't looked enough into the definitions, but it feels like you are making a number of leaps of judgement that require an understanding of the topic which is greater than mine. I'll have to do some more reading (eg: about lambda calculus etc...) before I'll know for sure!
For example, a cursory reading of the Wikipedia page for the Church-Turing thesis suggests it isn't proven fact, but rather a conjecture. Moreover, it appears to be based on an intuitive definition of what it means to understand. Like I said though, I don't have a robust enough understanding on the topic, but my intuition is that the Church-Turing thesis is about as 'hand-wavey' as the arguments you are trying to disprove with it.

What are "other elements of C which aren't in f"? You already asserted that f (the Room's own mapping of conversation histories to responses) matches C (the mapping that someone who "understands" Chinese would have). There's no room for anything to be missing.

In my post above, C is not the mapping that someone who understands Chinese would have, it is the superset of all mappings that people who understand Chinese might have. (A relation can 'map' an x value to multiple y values.)

I guess what I was going for was the fact that the Chinese room only has past conversation histories to base its understanding, whereas humans (and computers) can have more context to go on. The Chinese room doesn't really understand Chinese because it has no 'context', so its understanding is meaningless. For example, it cannot hear, or see, or in fact perceive anything other than Chinese characters passed under the door. A Human who was so removed from other stimuli (including memory of those stimuli, etc...) that it could only perceive Chinese characters, but who somehow had a memory of 'what to say next' for all possible chinese conversation histories, would be in the same boat.
Image

User avatar
Xenomortis
Not actually a special flower.
Posts: 1420
Joined: Thu Oct 11, 2012 8:47 am UTC

Re: Why do our robots suck?

Postby Xenomortis » Tue Oct 16, 2012 9:56 pm UTC

dudiobugtron wrote:I guess what I was going for was the fact that the Chinese room only has past conversation histories to base its understanding, whereas humans (and computers) can have more context to go on. The Chinese room doesn't really understand Chinese because it has no 'context', so its understanding is meaningless. For example, it cannot hear, or see, or in fact perceive anything other than Chinese characters passed under the door. A Human who was so removed from other stimuli (including memory of those stimuli, etc...) that it could only perceive Chinese characters, but who somehow had a memory of 'what to say next' for all possible chinese conversation histories, would be in the same boat.


This isn't relevant, because you cannot compare the Chinese Room to a human out in the world, precisely for that reason. You correctly note (if implicitly, so I might be reading you wrong), that a person who speaks and "understands" (whatever that means) Chinese who is trapped in a room, able only to take textual input, would be indistinguishable from the Chinese Room.
Furthermore, there's no reason why a "mobile" Chinese Room (somehow compressed so it's able to move about in the regular world), could not emulate a human if it's library were expanded to include data on how to respond given to the other inputs it would no receive (visual clues etc).
Image

User avatar
Xanthir
My HERO!!!
Posts: 5284
Joined: Tue Feb 20, 2007 12:49 am UTC
Location: The Googleplex
Contact:

Re: Why do our robots suck?

Postby Xanthir » Tue Oct 16, 2012 10:06 pm UTC

Once again, a lookup table is exactly equivalent to any finite computer (that is, any computer that exists in real life). You're not allowed to disagree with this - it's proven fact, unless you have truly sublime disproof of Church-Turing (in which case, congratulations on all the computing and math-related international awards you'll soon receive).

A significantly complex look-up table can be made exactly equivalent to any given finite computer. My claim is that the Chinese room is not sufficiently complex to me made equivalent to a Human. (See the reasons in other parts of the post.)


It sounds like you're claiming that computers of any kind can't be sufficiently complex to be made equivalent to a human. Is this so?

If so, you're almost certainly wrong. There doesn't appear to be anything preventing us from, at minimum, doing brute-force physical simulation of a brain. Current computational neuroscience, though, suggests that we won't have to go anywhere near that far down - it appears that thinking can likely be simulated just at the neuron level, or possibly even at the functional-group-of-neurons level. That level of computation is nearly within our grasp.

Most argument for humans being non-computable are implicitly or explicitly dualistic, implying that there is, for lack of a better term, magic that makes the brain work as it does, and we can't ever put the same kind of magic in computers.

Also I'm not sure I agree with your assumption that a non-'finite' computer cannot theoretically exist in real life. I haven't looked enough into the definitions, but it feels like you are making a number of leaps of judgement that require an understanding of the topic which is greater than mine. I'll have to do some more reading (eg: about lambda calculus etc...) before I'll know for sure!
For example, a cursory reading of the Wikipedia page for the Church-Turing thesis suggests it isn't proven fact, but rather a conjecture. Moreover, it appears to be based on an intuitive definition of what it means to understand. Like I said though, I don't have a robust enough understanding on the topic, but my intuition is that the Church-Turing thesis is about as 'hand-wavey' as the arguments you are trying to disprove with it.


Unless/until we find something surprising in physics, the CT thesis is rock-solid. Yes, as the Wikipedia page says, it can't be formally proven. However, it's well-accepted by virtually everyone with sufficiently experience to understand it properly, and given our current knowledge of physics, is expected to be true forever.

A non-finite computer requires either infinite memory (obviously impossible due to the universe having only a finite amount of energy in it), or infinite precision (which doesn't seem to be allowed by quantum physics, and would also likely violate the CT thesis, allowing "super-Turing" machines that can do infinite work in a finite amount of time).

What are "other elements of C which aren't in f"? You already asserted that f (the Room's own mapping of conversation histories to responses) matches C (the mapping that someone who "understands" Chinese would have). There's no room for anything to be missing.

In my post above, C is not the mapping that someone who understands Chinese would have, it is the superset of all mappings that people who understand Chinese might have. (A relation can 'map' an x value to multiple y values.)

I guess what I was going for was the fact that the Chinese room only has past conversation histories to base its understanding, whereas humans (and computers) can have more context to go on. The Chinese room doesn't really understand Chinese because it has no 'context', so its understanding is meaningless. For example, it cannot hear, or see, or in fact perceive anything other than Chinese characters passed under the door. A Human who was so removed from other stimuli (including memory of those stimuli, etc...) that it could only perceive Chinese characters, but who somehow had a memory of 'what to say next' for all possible chinese conversation histories, would be in the same boat.[/quote]

Ah, okay. You're reading too much into the "books full of Chinese characters" thing. It wasn't meant to imply that the Room is similar a human raised in a sensory deprivation chamber; it was simply ignoring details that were irrelevant to the thought experiment at hand. If it makes you happier, feel free to imagine that the input is much richer; in addition to the meaningless characters that are fed into it, it also receives large lists of numbers which represent sensory input history, and the lookup tables are appropriately blown up to accommodate this. It doesn't change the thought experiment at all.
(defun fibs (n &optional (a 1) (b 1)) (take n (unfold '+ a b)))

User avatar
Xanthir
My HERO!!!
Posts: 5284
Joined: Tue Feb 20, 2007 12:49 am UTC
Location: The Googleplex
Contact:

Re: Why do our robots suck?

Postby Xanthir » Tue Oct 16, 2012 10:14 pm UTC

freakish777 wrote:
Xanthir wrote:I... I have no idea why you bolded that fragment from me


You: Give a definition of Understands.
Me: Sure, here it is.
You: Actually, I'm not interested...

I was saying that by saying you don't really care to define the word Understand after asking for a definition is intellectually disingenuous.

Did you find my handwavey definition unsatisfactory? If necessary, I could have expanded it, but I don't think you've asked me to do so. (If so, I apologize, and can do so.)

However, I explicitly asked for you to expand your definitions, or at least restate them with different words, because it feels like you are leaning on the intuitive definitions of some words, and not seeing how much complexity and undefinedness lurks there.

Xanthir wrote:Once again, a lookup table is exactly equivalent to any finite computer


Actually, Church-Turing says their computationally equivalent, not exactly equivalent (in every way shape and form, if they were in fact equal in every way shape and form, then their architecture would be the same as well).

Understanding isn't computed. It's gained through experience.

Just because finite computer systems can be re-written as look up tables and they are computationally equivalent, does not mean they are equivalent in every possible way (including whether or not they have understanding).

Okay, so you do believe in magic. That's fine, it just means that we don't need to argue anymore.

If you don't think you believe in magic, then I have some questions.

  • What is "experience"?
  • What are the physical effects of experience? (Or, if you feel there are non-physical effects, say so.)
  • If I take a human brain which has "gained experience" and physically duplicate it (using some form of biological engineering beyond our current grasp), does the copy have experience? Why or why not?
  • Same scenario, but this time I've instead built a computer that closely mimics the brain, with individual processing units taking the place of individual neurons. Does it have experience? Why or why not?
  • Same scenario, but this time I've written a detailed physics simulation capable of accurately simulating things down to quantum interactions, and then programmed a duplicate of the brain into the simulation. Does the simulated copy have experience? Why or why not?

Once you reach the last question, it's just one more simple abstraction step to switch over to a Room-based computer architecture rather than a traditional computer architecture, so I've left it out of the explicit question list.
(defun fibs (n &optional (a 1) (b 1)) (take n (unfold '+ a b)))

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Tue Oct 16, 2012 11:16 pm UTC

Apologies for the massive quote-chain! I'll try and keep it as small as possible.


Xenomortis wrote:You correctly note (if implicitly, so I might be reading you wrong), that a person who speaks and "understands" (whatever that means) Chinese who is trapped in a room, able only to take textual input, would be indistinguishable from the Chinese Room.

They're not distinguishable from outside the room, sure. They're clearly distinguishable in terms of the process they use, though; we are able to distinguish the differences since we can 'see inside' the room.

Xenomortis wrote:Furthermore, there's no reason why a "mobile" Chinese Room (somehow compressed so it's able to move about in the regular world), could not emulate a human if it's library were expanded to include data on how to respond given to the other inputs it would no receive (visual clues etc).

Xanthir wrote:Ah, okay. You're reading too much into the "books full of Chinese characters" thing. It wasn't meant to imply that the Room is similar a human raised in a sensory deprivation chamber; it was simply ignoring details that were irrelevant to the thought experiment at hand. If it makes you happier, feel free to imagine that the input is much richer; in addition to the meaningless characters that are fed into it, it also receives large lists of numbers which represent sensory input history, and the lookup tables are appropriately blown up to accommodate this. It doesn't change the thought experiment at all.


This is fine, but sadly it does change the thought experiment quite a bit (for me at least!). I think that 'understanding' is something that you can't do from within the framework of that particular thing. You need to understand it as part of a wider context to be able to understand it at all.
Sure, you could make a 'Chinese Room' which is actually an 'Everything Room', and thereby say it understands Chinese. But I think that is avoiding the argument I am making, rather than answering it.
(Note, I realise that this particular argument is not directly related to the argument about whether computers can emulate humans.)

As an aside though, I think it's funny that one can entertain a thought experiment about a magic teleporting room which is pre-programmed to react to everything possible in the entirety of existence (as long as it has a sonic screwdriver), and yet dismiss any claim that this room is in any way different from a human being as 'magic'.

Xanthir wrote:It sounds like you're claiming that computers of any kind can't be sufficiently complex to be made equivalent to a human. Is this so?

I didn't mean to claim that, sorry! In fact I think Humans are rather un-complex compared to what computers might become one day. Humans are by no means the pinnacle of sentience that everyone in this thread seems to take as assumed. It's likely that one day (perhaps even now) there will be beings or computers that can 'understand' things to such a degree that they will be having an argument about whether something as simple as a (current) human can really be said to 'understand' anything!

Xanthir wrote:Unless/until we find something surprising in physics, the CT thesis is rock-solid. Yes, as the Wikipedia page says, it can't be formally proven. However, it's well-accepted by virtually everyone with sufficiently experience to understand it properly, and given our current knowledge of physics, is expected to be true forever.

A non-finite computer requires either infinite memory (obviously impossible due to the universe having only a finite amount of energy in it), or infinite precision (which doesn't seem to be allowed by quantum physics, and would also likely violate the CT thesis, allowing "super-Turing" machines that can do infinite work in a finite amount of time).

Shot, thanks! This saves me doing a bunch of readings. ;)
Image

User avatar
Xanthir
My HERO!!!
Posts: 5284
Joined: Tue Feb 20, 2007 12:49 am UTC
Location: The Googleplex
Contact:

Re: Why do our robots suck?

Postby Xanthir » Tue Oct 16, 2012 11:43 pm UTC

dudiobugtron wrote:
Xenomortis wrote:Furthermore, there's no reason why a "mobile" Chinese Room (somehow compressed so it's able to move about in the regular world), could not emulate a human if it's library were expanded to include data on how to respond given to the other inputs it would no receive (visual clues etc).

Xanthir wrote:Ah, okay. You're reading too much into the "books full of Chinese characters" thing. It wasn't meant to imply that the Room is similar a human raised in a sensory deprivation chamber; it was simply ignoring details that were irrelevant to the thought experiment at hand. If it makes you happier, feel free to imagine that the input is much richer; in addition to the meaningless characters that are fed into it, it also receives large lists of numbers which represent sensory input history, and the lookup tables are appropriately blown up to accommodate this. It doesn't change the thought experiment at all.


This is fine, but sadly it does change the thought experiment quite a bit (for me at least!). I think that 'understanding' is something that you can't do from within the framework of that particular thing. You need to understand it as part of a wider context to be able to understand it at all.
Sure, you could make a 'Chinese Room' which is actually an 'Everything Room', and thereby say it understands Chinese. But I think that is avoiding the argument I am making, rather than answering it.
(Note, I realise that this particular argument is not directly related to the argument about whether computers can emulate humans.)

The point of the argument (at least, as I see it) is simply bashing down people's intuitive prejudices in thinking about "computation". It doesn't matter whether the computation happens in meat or silicon, or whether it happens on-demand or is precomputed. Computation is computation. Consciousness is just a pattern in data, and can be realized in many substrates, and many interesting forms.

As an aside though, I think it's funny that one can entertain a thought experiment about a magic teleporting room which is pre-programmed to react to everything possible in the entirety of existence (as long as it has a sonic screwdriver), and yet dismiss any claim that this room is in any way different from a human being as 'magic'.

It is funny. ^_^ It falls into the "sufficiently advanced technology" bucket - seems magical to us, but it's theoretically possible.

Xanthir wrote:It sounds like you're claiming that computers of any kind can't be sufficiently complex to be made equivalent to a human. Is this so?

I didn't mean to claim that, sorry! In fact I think Humans are rather un-complex compared to what computers might become one day. Humans are by no means the pinnacle of sentience that everyone in this thread seems to take as assumed. It's likely that one day (perhaps even now) there will be beings or computers that can 'understand' things to such a degree that they will be having an argument about whether something as simple as a (current) human can really be said to 'understand' anything!

Okay, well, if you think that computers can "understand" in the same way that humans can, and the Room is a type of computer, and all computers are equivalent in computational power, then the argument's over. ^_^ Anything left is just your mistaken intuition, which you can fix.
(defun fibs (n &optional (a 1) (b 1)) (take n (unfold '+ a b)))

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Wed Oct 17, 2012 12:05 am UTC

Xanthir wrote:Okay, well, if you think that computers can "understand" in the same way that humans can, and the Room is a type of computer, and all computers are equivalent in computational power, then the argument's over. ^_^ Anything left is just your mistaken intuition, which you can fix.

:) More stuff to look up!

How does the halting problem fit in to all of this? it seems like that would show that it's actually impossible to design a 'Chinese room' which could predict how everything would react to given inputs.
Image

User avatar
Xanthir
My HERO!!!
Posts: 5284
Joined: Tue Feb 20, 2007 12:49 am UTC
Location: The Googleplex
Contact:

Re: Why do our robots suck?

Postby Xanthir » Wed Oct 17, 2012 5:12 am UTC

Halting problem only applies to real turing machines. Finite machines can only occupy a finite number of states, so (iirc) it doesn't apply to them, theoretically - you could always map them out, and loops always detectable. (Real turing machines can potentially run forever in a way that never returns to a previous state.)
(defun fibs (n &optional (a 1) (b 1)) (take n (unfold '+ a b)))

bad_robot
Posts: 15
Joined: Sun Oct 14, 2012 9:28 pm UTC

Re: Why do our robots suck?

Postby bad_robot » Wed Oct 17, 2012 6:00 am UTC

freakish777 wrote:Actually, Church-Turing says their computationally equivalent, not exactly equivalent (in every way shape and form, if they were in fact equal in every way shape and form, then their architecture would be the same as well).


The point is that the shape, form, architecture, etc. do not make a difference when it comes to whether something can be computed or not. If it can be computed then any Turing Machine can compute it. If it can't then no Turing Machine can compute it. That is what "computationally equivalent" means.

Understanding isn't computed. It's gained through experience.

Just because finite computer systems can be re-written as look up tables and they are computationally equivalent, does not mean they are equivalent in every possible way (including whether or not they have understanding).


What? Understanding can't be computed? From a few posts up:
freakish777 wrote:
And in more general terms, the only way that the Chinese Room can actually be correct is if there is something about the mind that is not computable

No, Chinese Room is correct, because it's premise dictates that it's only addressing static lookup tables. It does not address systems capable of making changes to themselves


So, is the the human mind computable or is it not, according to you? Or are you objecting to the assumption that the human mind can "understand"?

User avatar
Xenomortis
Not actually a special flower.
Posts: 1420
Joined: Thu Oct 11, 2012 8:47 am UTC

Re: Why do our robots suck?

Postby Xenomortis » Wed Oct 17, 2012 8:30 am UTC

dudiobugtron wrote:Apologies for the massive quote-chain! I'll try and keep it as small as possible.


Xenomortis wrote:You correctly note (if implicitly, so I might be reading you wrong), that a person who speaks and "understands" (whatever that means) Chinese who is trapped in a room, able only to take textual input, would be indistinguishable from the Chinese Room.

They're not distinguishable from outside the room, sure. They're clearly distinguishable in terms of the process they use, though; we are able to distinguish the differences since we can 'see inside' the room.


Are they?
See, the man in the Chinese Room acts like a simple processor; it takes input, goes to its lookup table in memory (which is *massive*) and then gives some output.
The man that "understands" Chinese takes input, does "something", and gives output. You don't know what that something is, other than it involves a very complicated structure that I'm willing to bet you don't understand at all, given any definition of understand.
For all we know, that something could be exactly what the Chinese Room does, just compressed.


dudiobugtron wrote:As an aside though, I think it's funny that one can entertain a thought experiment about a magic teleporting room which is pre-programmed to react to everything possible in the entirety of existence (as long as it has a sonic screwdriver), and yet dismiss any claim that this room is in any way different from a human being as 'magic'.


The word "magic" has specific meaning.
The Chinese Room thought experiment was originally an argument for some form of dualism. Essentially, the creator of the argument did not believe that the human mind was not a direct result of the mechanical properties of the brain. Proponents hold that there is something else that provides consciousness and understanding. I, and many others, take this view as utter bullshit; I have a basic Maths and Physics background and won't tolerate what essentially amounts to voodoo magic.
What Xanthir has been trying to show, and what I myself have been heading toward, is that a failure to except the Chinese room, or trivial variation, as ultimately having "understanding" essentially amounts to dualism and is therefore, essentially unknowable which is about the most useless and unworkable conclusion you can reach; you've replaced one unknown with another, equivalent unknown and we're no further forward.

We permit the Chinese Room to be as big as we need because it's a thought experiment and that's allowed (it's the norm). Ultimately, it need not be as big as I've been submitting, but it doesn't matter.
Image

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Wed Oct 17, 2012 8:46 am UTC

So, basically what you're saying is that it's OK, as part of the Chinese room thought experiment, to program for the Chinese room its own independent thoughts and feelings, opinions on different subjects, 'understandings' of topics other than Chinese, and anything else required to make it a good example of a pre-programmed-but-otherwise-human computer? In that case I have been arguing up the wrong tree.

Xanthir wrote:Halting problem only applies to real turing machines. Finite machines can only occupy a finite number of states, so (iirc) it doesn't apply to them, theoretically - you could always map them out, and loops always detectable. (Real turing machines can potentially run forever in a way that never returns to a previous state.)

Thanks, that exactly answers my question!

(Just for clarity's sake though - 'real' Turing machines are the ones which can't actually exist in reality, right? ;)
And is the reason they can't exist in reality is because they wouldn't really be *real* Turing machines because they would stop being able to compute once the universe ended? :P
Just wanting to check I'm 'understanding' it right!)


PS: Thanks for the incentive to find out more about Turing machines! For example I discovered that my 'Relation vs function' distinction has actually already been done (cf: deterministic vs non-deterministic Turing machines); but in fact they are equivalent anyway as one can always model the other. So even if you think that 'understanding' can only be done by a non-deterministic Turing Machine, you still have to admit that a deterministic Turing Machine could model exactly the same behaviour.
Image

User avatar
Xenomortis
Not actually a special flower.
Posts: 1420
Joined: Thu Oct 11, 2012 8:47 am UTC

Re: Why do our robots suck?

Postby Xenomortis » Wed Oct 17, 2012 12:36 pm UTC

Turing Machines have infinite memory.
Image

User avatar
WanderingLinguist
Posts: 237
Joined: Tue May 22, 2012 5:14 pm UTC
Location: Seoul
Contact:

Re: Why do our robots suck?

Postby WanderingLinguist » Wed Oct 17, 2012 12:40 pm UTC

dudiobugtron wrote:So, basically what you're saying is that it's OK, as part of the Chinese room thought experiment, to program for the Chinese room its own independent thoughts and feelings, opinions on different subjects, 'understandings' of topics other than Chinese, and anything else required to make it a good example of a pre-programmed-but-otherwise-human computer? In that case I have been arguing up the wrong tree.


Well, it is by definition constructed such that its impossible to tell the difference from the outside. So if it has to have (or emulate -- is there a difference?) independent thoughts in order to meet that criteria than, yes, it would be OK by definition for the Chinese room to include an algorithm simulating independent thought.

The question then comes down to not "does the Chinese room understand" but rather "is it possible to build a Chinese Room" (or the equivalent). If the question is about understanding, it's really a philosophical rather than a function question: If you can't tell the difference from outside does it matter? This is related to p-zombies, which I believe have been mentioned already.

So, if the Chinese room was allowed infinite tape (i.e. infinite paper, pencils, storage, etc.) I think it's pretty clear that it could be built. I guess it would need to be in some kind of time distortion or you could tell the difference due to lag (it would take a really long time for the guy in the Chinese room to actually carry out the algorithm).

So, the question is can it be done in a finite way? Well... our brains are finite... so I think that answers the question. However, do we need a more specialized type of computer to do it in a finite way (from a practical standpoint)? I may be that emulating it on a general purpose computer, while finitely possible, must be large and slow and impractical on the timescale of human beings. I thing that issue has yet to be decided.

User avatar
freakish777
Posts: 354
Joined: Wed Jul 13, 2011 2:14 pm UTC

Re: Why do our robots suck?

Postby freakish777 » Wed Oct 17, 2012 2:31 pm UTC

Xanthir wrote:Did you find my handwavey definition unsatisfactory? If necessary, I could have expanded it, but I don't think you've asked me to do so. (If so, I apologize, and can do so.)

However, I explicitly asked for you to expand your definitions, or at least restate them with different words, because it feels like you are leaning on the intuitive definitions of some words, and not seeing how much complexity and undefinedness lurks there.


No, I didn't ask you to expand your definition. I just sort of found it awkward for you to want my definition without being interested in defining the word yourself.

Okay, so you do believe in magic.


No. There's a saying "any sufficiently advanced technology would be indistinguishable from magic." Seeing as we don't have a firm understanding of how brains work yet, I'm inclined to consider it "sufficiently advanced technology" that we simply don't have the science to accurately explain in detail yet, so I get why you think I believe in Magic, but you're wrong on that front.

What is "experience"?


The act of changing one's "instruction set" based on observing or taking place in an event.

[*]What are the physical effects of experience? (Or, if you feel there are non-physical effects, say so.)


Includes (but not limited to):

Gaining memory of the event.
Updating one's "instruction set" (typically based on results)
Gaining Knowledge (the summation of this and updating one's instruction set is a large part of my definition of understanding).

If I take a human brain which has "gained experience" and physically duplicate it (using some form of biological engineering beyond our current grasp), does the copy have experience? Why or why not?


I'm unsure, but see no reason why it would not contain the experience (and understanding).

Same scenario, but this time I've instead built a computer that closely mimics the brain, with individual processing units taking the place of individual neurons. Does it have experience? Why or why not?


I'm unsure, but see no reason why it would not contain the experience (and understanding).

Same scenario, but this time I've written a detailed physics simulation capable of accurately simulating things down to quantum interactions, and then programmed a duplicate of the brain into the simulation. Does the simulated copy have experience? Why or why not?


I'm unsure, but see no reason why it would not contain the experience (and understanding).

Once you reach the last question, it's just one more simple abstraction step to switch over to a Room-based computer architecture rather than a traditional computer architecture, so I've left it out of the explicit question list.


Searle's conclusion that strong AI can't exist is (as far as I can tell) wrong, because his premise restrict him to only machines that aren't sufficiently complex enough to understand.

A sufficiently large look up table isn't "given" experience, it simply has someone else's experiences codified to follow blindly (forever).

Trust me, I get exactly what you've been saying: That we could codify the experiences at time T of a human brain B into instruction set S to follow as a look up table, and giving both S and B at time T the same input I, they both arrive at the same output O, and thus, them being computationally equivalent, you believe S understands everything B does because they share experiences. I am simply disagreeing and saying that I believe that a replica B' of B would in fact understand, but that the look up table, despite being computationally equivalent and arriving at the same output, still does not understand, because it did not gain the understanding.

bad_robot wrote:
freakish777 wrote:Actually, Church-Turing says their computationally equivalent, not exactly equivalent (in every way shape and form, if they were in fact equal in every way shape and form, then their architecture would be the same as well).


The point is that the shape, form, architecture, etc. do not make a difference when it comes to whether something can be computed or not. If it can be computed then any Turing Machine can compute it. If it can't then no Turing Machine can compute it. That is what "computationally equivalent" means.

Understanding isn't computed. It's gained through experience.

Just because finite computer systems can be re-written as look up tables and they are computationally equivalent, does not mean they are equivalent in every possible way (including whether or not they have understanding).


What? Understanding can't be computed? From a few posts up:
freakish777 wrote:
And in more general terms, the only way that the Chinese Room can actually be correct is if there is something about the mind that is not computable

No, Chinese Room is correct, because it's premise dictates that it's only addressing static lookup tables. It does not address systems capable of making changes to themselves


So, is the the human mind computable or is it not, according to you? Or are you objecting to the assumption that the human mind can "understand"?


Sorry, I wasn't clear. That really should have read "Understanding isn't computed the way you all think it is."

Essentially part of Understanding is a process that occurs when experiences happen, and it changes the instruction set, and another key part is the knowledge that ends up in a look up table.

If you're only looking at the knowledge at time T, you're going to have a static look up table that is missing a key component of Understanding (updating the instruction set for future optimization).




Tyndmyr wrote:But why not?

What makes one process* special or distinct as opposed to any other?

*or group of processes. It's entirely possible that humans don't all understand things in the same way.



Think of all the knowledge contained in the look up table. Now imagine a process wrapped around the look up table that changes the look up table over time.

That wrapper process is a big part of understanding (according to my definition). Understanding and Knowledge are not equivalent. Knowledge can exist without Understanding (the look up table by itself, my definition is roughly the summation of the wrapper process and the look up table).

Searle's premise for Chinese Room limits him to just the look up table, and it's pretty evident considering he never mentions the possibility for the room's instruction set to be modified.

Tyndmyr
Posts: 10848
Joined: Wed Jul 25, 2012 8:38 pm UTC

Re: Why do our robots suck?

Postby Tyndmyr » Wed Oct 17, 2012 4:45 pm UTC

freakish777 wrote:Think of all the knowledge contained in the look up table. Now imagine a process wrapped around the look up table that changes the look up table over time.

That wrapper process is a big part of understanding (according to my definition). Understanding and Knowledge are not equivalent. Knowledge can exist without Understanding (the look up table by itself, my definition is roughly the summation of the wrapper process and the look up table).

Searle's premise for Chinese Room limits him to just the look up table, and it's pretty evident considering he never mentions the possibility for the room's instruction set to be modified.


So? A wrapper process and a lookup table are indistinguishable from a (much bigger) static lookup table. Externally, it doesn't actually matter.

That said, the inclusion of erasers in the discussion makes it reasonable to extrapolate that the example could also apply to a dynamic lookup table.

User avatar
freakish777
Posts: 354
Joined: Wed Jul 13, 2011 2:14 pm UTC

Re: Why do our robots suck?

Postby freakish777 » Wed Oct 17, 2012 6:11 pm UTC

Tyndmyr wrote:So? A wrapper process and a lookup table are indistinguishable from a (much bigger) static lookup table. Externally, it doesn't actually matter.


I'm not convinced that an arbitrarily large look up table would be indistinguishable from a lookup table that is dynamically updated by an "Understanding Wrapper Process" (or whatever you want to call it) provided you are given sufficient time to exhaust all possible outcomes (or patterns of outcomes) given by the look up table.

Imagine you have a look up table that spits out words based on input and time T. In it's initial state, it uses all "baby" words, and as time increments, it changes from "baby" words to adult words, until eventually only adult words are left in it's "vocabulary".

This certainly might fool some people into thinking that the system is learning the words, but the fact of the matter is that they were hardcoded into the system from the beginning.

If the system is incapable of learning a new word/object/concept not originally in it's look up table, I'm contending that it doesn't understand. This happens to be how Furbies were programmed and in fact several people, smart people even were fooled into thinking it was a system that learned words, they didn't learn anything. They simply had a look up table and as time went on spoke less and less Furbish and more and more English, but were limited by the size of their initial programming.

I know Xanthir is going to re-inject the "there's a finite set of possible combinations the material making up a human brain can be in" point after reading the above, but I'm not inclined to believe it makes a difference here. The sufficiently large static look up table that contains every state every human being could ever be in is still limited by the fact that it's static (and doesn't update instructions), where as a system that allows the look up table to grow and change the instructions of that look up table, would eventually grow beyond the size of the static (and mind bogglingly large) look up table (provided it has enough resources to do so).

That said, the inclusion of erasers in the discussion makes it reasonable to extrapolate that the example could also apply to a dynamic lookup table.


Mmmmmm... maybe? I don't think so, when one thinks about the Chinese Room problem, one doesn't think of the man in the room (acting as CPU) reading the Instruction Book that tells him what to do and thinking "I'm totally going to make changes to this book to make it better" instead of using the erasers and pen with the scratch paper (acting as memory).

Had any of the rebuttals to Chinese Room said "but what if Chinese Room sits inside Understanding Room, and Understanding Room updates the instruction set of Chinese Room" then I'm clearly on board with that rebuttal. Just all the rebuttals so far read along the lines of "You're wrong because I disagree, therefore you're wrong!" (and if any of these exist out there, please point me in their direction as I've yet to come across them).

EDIT: If you think of the man in the Chinese Room making changes to the Instruction Book, then you would have to argue that the man actually does understand not only Chinese, but the program he's processing in the first place!
Last edited by freakish777 on Wed Oct 17, 2012 6:46 pm UTC, edited 1 time in total.

EvanED
Posts: 4330
Joined: Mon Aug 07, 2006 6:28 am UTC
Location: Madison, WI
Contact:

Re: Why do our robots suck?

Postby EvanED » Wed Oct 17, 2012 6:25 pm UTC

Xanthir wrote:Halting problem only applies to real turing machines. Finite machines can only occupy a finite number of states, so (iirc) it doesn't apply to them, theoretically - you could always map them out, and loops always detectable. (Real turing machines can potentially run forever in a way that never returns to a previous state.)

Bear in mind though the following. (I'm not positive about this as I don't recall seeing any complexity proofs along this line, but I'd bet money it's true.) Suppose you have what I'll call a "universe-limited Turing machine" (ULTM) -- that is, something that looks like a TM except that its tape is a finite size of b bits because the universe is finite-sized and only has b bits available for your use. I would be very surprised if you could compute the haltingness of such a ULTM in the same universe, as I suspect computing the haltingness of even a deterministic ULTM with b bits would take (in the worst case) O(b2b) bits. If the ULTM is nondeterministic, I think it would take O(2b2(2^b)) bits.

So if you're interested in the haltingness of a ULTM instead of a standard TM because of practical concerns, it's also reasonable to say that practical concerns still prevent you from determining the haltingness of a ULTM even though it is finite-state (assuming I'm right). (Of course, that's not necessarily the only reason to be interested in ULTMs.)

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Wed Oct 17, 2012 8:35 pm UTC

Xenomortis wrote:Turing Machines have infinite memory.

Yes, but they can't actually use infinite memory in a finite amount of time - can they? If not, then the difference between a 'real' turing machine, and a Turing machine which operated as though it had infinite tape (but really we just kept feeding tape as fast as it could use it), is that we can't technically feed it tape forever. (Edit: Oh, EvanED already posted the answer to this above me!)

My point is, how can we tell the difference? They're completely indistinguishable from the outside.
Image

lorb
Posts: 404
Joined: Wed Nov 10, 2010 10:34 am UTC
Location: Austria

Re: Why do our robots suck?

Postby lorb » Wed Oct 17, 2012 8:50 pm UTC

Besides that arguments that are merely based on finiteness aren't really any useful when it comes to the halting problem. Suppose you have a tape of just 1 million bits which is not much by the technological standards of today. It has 2^1000000 possible states. Even if you were able to go through a decillion (10^33, way more than is currently and in the foreseeable future possible) states per second it would take you more than 10^300000 years to complete a full circle. Compared to that even the age of the universe is merely an instant.

So even if we limit ourselves to programs that run on a computer that is limited to 125 KB of ram there is no way we are going to solve the halting problem just because it's reduced to a finite form. (May be one of the reasons why our robots suck.)
Please be gracious in judging my english. (I am not a native speaker/writer.)
http://decodedarfur.org/

bad_robot
Posts: 15
Joined: Sun Oct 14, 2012 9:28 pm UTC

Re: Why do our robots suck?

Postby bad_robot » Wed Oct 17, 2012 10:47 pm UTC

freakish777 wrote:Sorry, I wasn't clear. That really should have read "Understanding isn't computed the way you all think it is."


Are you trying to redefine the term "computing"? It doesn't matter what I or you think, the term has a specific meaning grounded in the theory of computation and that is what we are using here. Once again, if it can be computed then any Universal Turing Machine can compute it and if it can't be computed then no Turing Machine can compute it. So, according to you, can a Turing Machine compute understanding, or can it not?

Essentially part of Understanding is a process that occurs when experiences happen, and it changes the instruction set, and another key part is the knowledge that ends up in a look up table.

If you're only looking at the knowledge at time T, you're going to have a static look up table that is missing a key component of Understanding (updating the instruction set for future optimization).


You seem stuck on the "changes the instruction set" part but as many people already said here this does not in any way change the ability to compute anything. A Turing Machine can emulate a instruction set that provides ability for the programs running it to modify that instruction set, just like it can run a program that has the ability to modify its own code. So, any computation that you claim requires a modifiable instruction set can indeed be computed by any Universal Turing Machine and therefore can be computed by any Turing complete computer regardless of whether its specific implementation allows for program making changes to the instruction set or not. This if a fundamental consequence of the Church-Turing thesis.

Same scenario, but this time I've written a detailed physics simulation capable of accurately simulating things down to quantum interactions, and then programmed a duplicate of the brain into the simulation. Does the simulated copy have experience? Why or why not?
I'm unsure, but see no reason why it would not contain the experience (and understanding).


So, you are saying that a biological or physical simulation of the brain can understand? Creating physics or biological simulations does not require changing instruction sets or anything as exotic as that. We already know this because we have successfully simulated physical and biological systems using standard programming methods and it works fine, we just haven't done it on the scale of the human brain yet.

Just all the rebuttals so far read along the lines of "You're wrong because I disagree, therefore you're wrong!"


It sounds like exactly the opposite to me. Many people have shown that both you and Searle appear to severely misunderstand what computing is all about and you are mostly ignoring this point and continue to make senseless claims about something not being computing "the way [we] all think it is" and about "changing instruction sets" which have been shown to have no effect on being able to compute something.
Last edited by bad_robot on Thu Oct 18, 2012 6:39 am UTC, edited 2 times in total.

User avatar
phlip
Restorer of Worlds
Posts: 7550
Joined: Sat Sep 23, 2006 3:56 am UTC
Location: Australia
Contact:

Re: Why do our robots suck?

Postby phlip » Thu Oct 18, 2012 12:07 am UTC

freakish777 wrote:No, I didn't ask you to expand your definition. I just sort of found it awkward for you to want my definition without being interested in defining the word yourself.

Well, you're the one claiming that "understanding" is a property that human brains can have, but sufficiently large lookup tables cannot. So your definition of "understanding" is relevant to the conversation. On our side, however, we're merely claiming that you haven't drawn a relevant distinction between a human brain and a sufficiently large lookup table, we're not trying to affirmatively claim or deny "understanding" for either system... so our definition of "understanding" is less relevant.

freakish777 wrote:
Okay, so you do believe in magic.

No. There's a saying "any sufficiently advanced technology would be indistinguishable from magic." Seeing as we don't have a firm understanding of how brains work yet, I'm inclined to consider it "sufficiently advanced technology" that we simply don't have the science to accurately explain in detail yet, so I get why you think I believe in Magic, but you're wrong on that front.

"Magic" here isn't used in the vague sense of "things we don't understand", but rather in the specific sense of "things we don't understand because they don't follow physical laws; some influence on reality from outside it, from outside of physics, and normal matter and energy; some form of dualism".

freakish777 wrote:
What is "experience"?

The act of changing one's "instruction set" based on observing or taking place in an event.

Can a human brain then "experience"? Is a human brain then capable of "understanding" by this definition? Are these even well-defined questions? What does "changing one's instruction set" even mean in the context of a human brain? Is it something like changing to use something other than neurons as the base processors in the brain? Because by that definition, no, humans are incapable of "understanding". Or is it something else?

---

You seem to get really caught up in the whole "updating the program" aspect of this whole thing. You seem to be under the impression that if the lookup tables aren't updated, then the knowledge is lost somehow? Or that the static-lookup-table-powered room can't remember things?

Remember that the static-lookup-table system has two things: the lookup table itself, and the state. And while converting a self-mutable lookup table into a static lookup table makes the table exponentially larger, one thing that hasn't been focused on much is that it also makes the state much larger too. And basically anything that would previously have altered the lookup table will instead alter the state, which will make you look at different place in the lookup table where the equivalent change is in place.

Perhaps an example. Take a very small, oversimplified snippet of the potential Chinese Room code:

Code: Select all

if message == "What wavelength is blue light?"
block A: {
  respond "I don't know."
}
else if message == "Blue light has a wavelength of 450nm."
block B: {
  respond "Thanks for letting me know."
  replace block A with { respond "It's 450nm" }
  replace block B with { respond "Thanks, but I already knew" }
}
This snippet is able to learn a new fact, and updates its responses as appropriate. However, it could also be written as:

Code: Select all

var blue_light_wavelength = null

if message == "What wavelength is blue light?"
  if blue_light_wavelength is null
    respond "I don't know."
  else
    respond "It's ", blue_light_wavelength
else if message == "Blue light has a wavelength of 450nm."
  if blue_light_wavelength is null
    respond "Thanks for letting me know."
    blue_light_wavelength := "450nm"
  else
    respond "Thanks, but I already knew"
This one behaves identically to the first. However, it's not self-modifying. On the other hand, it has a larger lookup table (4 responses, instead of 2 that get replaced) and a larger state (storing what we've learned the wavelength of blue light to be).

Obviously, this example would never pass the Turing test, but the principle is the same for a larger program. Yes, it's currently intractable to do for a full human brain, even in a computer, let alone on paper, but there's a vast difference between intractable and incomputable.

And to counter the obvious objection that in the second case the lookup table already encodes the knowledge that blue light is 450nm, and the state is merely "turning it on"... so it's not truly "learning"... well, in this case the same is true for the self-modifying example, and this is merely because the example is so heavily simplified. A full AI program would be able to respond to more inputs, and its interactions with its lookup table and its state would be a lot more intricate than merely "turning on" a block of code where the response has been pre-recorded.

Code: Select all

enum ಠ_ಠ {°□°╰=1, °Д°╰, ಠ益ಠ╰};
void ┻━┻︵​╰(ಠ_ಠ ⚠) {exit((int)⚠);}
[he/him/his]

jareds
Posts: 436
Joined: Wed Jan 03, 2007 3:56 pm UTC

Re: Why do our robots suck?

Postby jareds » Thu Oct 18, 2012 1:32 am UTC

EvanED wrote:Bear in mind though the following. (I'm not positive about this as I don't recall seeing any complexity proofs along this line, but I'd bet money it's true.) Suppose you have what I'll call a "universe-limited Turing machine" (ULTM) -- that is, something that looks like a TM except that its tape is a finite size of b bits because the universe is finite-sized and only has b bits available for your use. I would be very surprised if you could compute the haltingness of such a ULTM in the same universe, as I suspect computing the haltingness of even a deterministic ULTM with b bits would take (in the worst case) O(b2b) bits. If the ULTM is nondeterministic, I think it would take O(2b2(2^b)) bits.

Sorry for the off-topic nerd sniping.

If you have a deterministic TM that is space bounded to b bits of state (rolling up the tape and finite control into a single measure), it will either halt or repeat a state in at most 2b steps. To determine whether it halts, you just attach a b-bit counter and run it for up to 2b steps. If it hasn't yet halted, it never will. This should take slightly more than 2b bits of space.

If you have an NDTM that is space bounded to b bits of state, by Savitch's theorem it can be simulated by a DTM that is space bounded by O(b2), so by the previous argument you can determine whether it halts in O(2b2)=O(b2) space.

dudiobugtron,

I don't actually understand why the halting problem is relevant to the Chinese Room. I would normally expect to see it in Penrose-style arguments against strong AI. For example, (1) premise: the human mathematical community can, in principle, figure out whether any Turing machine halts, (2) conclusion: humans cannot always be simulated by Turing machines. The problem with the argument is that the premise is very, very dubious.

User avatar
dudiobugtron
Posts: 1098
Joined: Mon Jul 30, 2012 9:14 am UTC
Location: The Outlier

Re: Why do our robots suck?

Postby dudiobugtron » Thu Oct 18, 2012 1:47 am UTC

(Please excuse my terrible knowledge and use of terminology in this post, I am not a Computer scientist!)

jareds wrote:dudiobugtron,

I don't actually understand why the halting problem is relevant to the Chinese Room. I would normally expect to see it in Penrose-style arguments against strong AI. For example, (1) premise: the human mathematical community can, in principle, figure out whether any Turing machine halts, (2) conclusion: humans cannot always be simulated by Turing machines. The problem with the argument is that the premise is very, very dubious.


The argument of a number of posters in this thread is that a sufficently advanced look-up table could be made Turing equivalent to the Human brain (which is also a Turing machine). Therefore, since human brain can understand, so can a sufficiently complex look-up table.
I was wondering whether the halting problem meant that (even in theory) it is impossible to build such a look-up table, since to build it you would need to know whether the Human-brain Turing Machine would halt or not for any given input. (Or, rather if you did have one, you could use it to look up whether or not the human brain would halt for any given input, violating the halting problem.)
The counter-argument to this was that since the Human brain is actually not a 'real' Turing machine (just a wussy, watered-down one), the look-up table can theoretically 'work out' if the human brain will halt or not for any given program. It was then pointed out that in practice, it would probably need to violate the laws of physics to do so, but at least in theory it would be possible.
Image

jareds
Posts: 436
Joined: Wed Jan 03, 2007 3:56 pm UTC

Re: Why do our robots suck?

Postby jareds » Thu Oct 18, 2012 2:15 am UTC

OK, I understand. The reason that the halting problem doesn't matter is that if you're building a lookup-table-based Chinese Room, you're implicitly truncating the length of the conversation anyway (or forcing it into an eventual repetitive loop), since lookup tables are finite. The argument that this truncation isn't a problem is that you could have a lookup table that lasts 200 years, which is probably longer than any of us will last, but we have no problem saying that we're intelligent.

Even without the finite lookup table issue, the halting problem is not a problem for simulating a human brain. The halting problem doesn't say that a TM can never figure out whether a TM halts. In fact, we could make a TM such that there are an infinite number of TMs that it knows will halt, an infinite number of TMs that it knows will not halt, and an infinite number that it can't figure out. The last part is what is required by the halting problem. With a simulated brain, there wouldn't be any problem coding the simulation to run forever, and knowing that it runs forever, or coding it to halt at some point, and knowing that it halts at some point; just like there wouldn't be a problem programming Conway's game of Life to run forever or to stop at some point.


Return to “Computer Science”

Who is online

Users browsing this forum: No registered users and 4 guests