John Searle: Consciousness in Artificial Intelligence [video]
youtube.com
youtube.com
Kurzweil, in summary, asks "You say that a machine manipulating symbols can't have consciousness. Why is this different than consciousness arising neurons manipulating neurotransmitter concentrations?" Searle gives a non-answer: "My dog has consciousness because I can look at it and conclude that it has consciousness."
How is it even an argument? It doesn't illuminate anything, and it's not even clever. It seems like the most facile wrong-headed stab at refutation, by begging the question. As far as I can tell, the argument is, "well you can make this room that manipulates symbols like a computer, and of course it's not conscious, so a computer can't be either"? There are so many problems with this argument I don't even know where to begin.
The fact that he appears to think that changing a "computer" to a "room" has persuasive power just makes it all the more antiquated. As if people can't understand the idea that computers "just" manipulate symbols? Changing it to a "room" adds nothing.
I mean, agree with it or not, but I think that's a bit stronger than just, making it seem intuitively worse because its a room instead of "a computer"?
I think the important part isn't the swap of "room" for "computer", but instead the swap of "person" for "cpu"?
And because the cpu is the part that does stuff, like the person, then if the system w/ the person doesn't understand chinese, then the computer w/ the cpu doesn't understand chinese.
Because there's nothing to make the cpu more understand-y, only things to make it less understand-y, and otherwise the systems are the same.
A better example would be saying something like - does this company have the knowledge to make a particular product? We can say that no individual member of the company does, but the company as a whole does.
Which, well there's a whole series of responses back and forth, with different ideas about what is or is not a good response.
One idea describes a machine where each state of the program is pre-computed, and the computer steps through the states one by one, but in each state, if the next of the pre-computed states was wrong (i.e. would not be the next step of the program, following from the current state), if (e.g.) a switch was flipped, it would cause the program to be computed correctly despite the pre-computed states being wrong, and if the switch is not flipped, then it would continue along the pre-computer states. If the switch is flipped on, or if its flipped off, and all the pre-computer states are correct, the same things happen, and it does not interact with the switch at all. If all the pre-computed states are nonsense, and the switch is flipped on, then it runs the program correctly, despite the pre-computed states being nonsense.
So, suppose that if the pre-computed states are all wrong, and the switch is on, that that counts as conscious. Then, if the pre-computed states are all correct, and the switch is on, would that still be conscious? What if almost all the pre-computed states were wrong, but a few were right? It doesn't seem like there is an obvious cutoff point between "all the pre-computed steps are wrong" and "all the pre-computer steps are right", where there would be a switch between what is conscious. So then, one might conclude that the one where all the pre-computed steps are right, and the switch is on, is just as conscious as the one which has the switch on but all the pre-computed states are wrong.
But then what of the one where all the pre-computed states are right, and the switch is off?
The switch does not interact with the rest of the stuff unless a pre-computed next step would be wrong, so how could it be that when the switch is on, the one with all the pre-computations is conscious, but when it is off, it isn't?
But the one with all the pre-computations correct, and the switch off, is not particularly different from just reading the list of states in a book.
If one grants consciousness to that, why not grant it to e.g. fictional characters that "communicate with" the reader?
One might come up with some sort of thing like, it depends on how it interacts with the world, and it doesn't make sense for it to have pre-computed steps if it is interacting with the world in new ways, that might be a way out. Or one could argue that it really does matter if the switch is flipped one way or the other, and when you flip it back and forth it switches between being actually conscious and being, basically, a p-zombie. And speaking of which you could say, "well what if the same thing is done with brain states being pre-computed?", etc. etc.
I think the Chinese Room problem, while not conclusive, is a useful introduction to these issues?
The states were (probably) produced by computing a conscious mind and recording the result.
Follow the improbability. The behavior has to come from somewhere. That somewhere is probably conscious.
Similarly, authors are conscious, so they know how conscious characters behave.
Pre-computation is another one that seems to obfuscate the actual issue. No, I don't think anyone would think a computer simply reciting a pre-computed conversation had conscious thought going into it; but that same is true for a human being reciting a a conversation they memorized (which wouldn't be that different from reading the conversation in a book). But that's a bit of a strawman, because no one is arguing that lookup table-type programs are conscious (you don't see anyone arguing that Siri is conscious). And the lookup table/precomputations for even a simple conversation would impossibly large (run some numbers, it's most likely larger than the number of atoms in the universe for even tiny conversations).
So I don't see these arguments as bringing up anything useful. They seem more like colorful attempts to purposefully confuse the issue.
Consciousness resides in the minds that created the lookup tables: they were constructed by conscious beings to map queries to meaningful responses.
The lookup tables are the very sophisticated part of Searle's Chinese Room.
The recent emergent semantic vector algebra discovered in human languages by Mikolov's word2vec [Mikolov et al] demonstrate that some of the computational power of language is inherent in the language ( but only meaningfully interpretable by a conscious listener.
Meaning requires Consciousness but language is unexpectedly sophisticated and contains a semantic vector space which can answer simple queries ("What is the Capital of...") and analogise ("King is to What as Man is to Woman") algebraicly. [Mikolov et al]
This inherent semantic vector space is discoverable by context encoding large corpii.
Language is a very sophisticated and ancient tool that allows some reasoning to be performed by simple algebraic transformations inherent to the language.
-
[Mikolov et al] : Efficient Estimation of Word Representations in Vector Space Mikolov, Chen, Corrado, & Dean http://arxiv.org/abs/1301.3781
This then devolves into semantics. Can a person locked in a room really come to "understand" chinese culture, for example, if only non-experiential learning were used as data inputs? I think we have to say the answer is yes. I am a chemist. I have never seen an atomic orbital with my bare eyes, yet I can design chemical reactions that work with my understanding of chemistry. Because I have not experienced an atomic orbital does that mean I do not understand? Even, when I set up my first reaction, I did not have any experience, and knew what I was doing only through what could be described as sophisticated analogy. I would say my understading was low, but it was certainly non-zero. Where does one draw the line?
If Searle is right then we should be able to perform a MRI on a blind person while they are talking to someone and spot the point where their brain switches into "symbol manipulation mode" when the conversation subject becomes something visual.
He could have an entire conversation with you in Chinese, and only know that what you said merits the response he gave. He doesn't if he's telling you directions to the bathroom, or how to perform brain surgery.
Searle does try to explain why there's a difference. Although the person in the Chinese Room might be conscious of other things, he has no consciousness of understanding the Chinese text he's manipulating, and will readily verify this, and nothing else in the room is conscious. Chinese speakers are conscious of understanding Chinese.
Yet here I am, I am the one who is seeing what my eyes see and I am distinct from you. Science still has no idea how that happens, as far as I know.
So how knows, maybe all computer programs are in fact conscious in some way.
Its always distressed me that some people take it seriously as an argument against AI consciousness; it does a better job of illustrating the incoherence of the fuzzy concept of consciousness on which the argument is based.
The Chinese Room argument is that because computers can't "learn", everything has to be taught to them directly, whereas humans are able to take knowledge given and apply it to other situations. While some computers can "learn" enough rules to follow patterns, the argument is that computers can't "jump the track" and that humans can.
Another way to think of it is: a fetus, or a sperm, or ova is not conscious. Some might argue that a newborn isn't really conscious. Somewhere along the line it becomes conscious. How does that happen? Where is the line? We have no idea.
You can't assert that meaning can't arise from "dumb symbol manipulation" without understanding how meaning arises in the former case. We simply don't know enough to make any sort of judgement. The Chinese room argument is trying to make something out of nothing. We don't know.
I think of it this way:
I have two rooms: one has a person who doesn't speak Chinese in it, but they have reference and books that allow them to translate incoming papers into Chinese perfectly.
The second room just someone who speaks Chinese, and can translate anything coming in perfectly.
Searle says that AIs are like person in room one: they don't know Chinese.
I would argue that is the wrong way to look at things. A better comparison is that an AI is like the system of room 1, which does know Chinese, and from observation is indistinguishable from the system of room 2. What's going on inside (a human with Chinese reference books vs a human who knows Chinese) doesn't matter, it's just internal processing.
If it walks like a duck and quacks like a duck, then it's a duck.
If a machine claims to be conscious, and I can't tell it apart from another conscious being, who am I to say it isn't conscious?
The correct question to ask: How is a machine manipulating symbols (that someone says is conscious) different from any other complex physical system? Is New York city's complex sewer system conscious. What about the entire world's sewer & plumbing system?
Does a machine have to compute some special function to be conscious? Does the speed of computation matter? If so who measures the speed? (let us not bring in general relativity as speed of computation can be different for different observers.
Kurzweil et al's definition of consciousness is exactly as silly as Searle saying "My dog has consciousness because I can look at it and conclude that it has consciousness."
IMO Kurzweil's response is actually spot on, although not that hard to come up with. You could make the argument: in your brain, there are just atoms following the laws of physics. The atoms have no choice in the matter, and know nothing about Chinese, or hunger, love, life goals, etc. Your brain is entirely composed of atoms so you can't be conscious.
Obviously "meaning" arises from mindless syntax or mindless physics somewhere in the process. We just don't know where. The Chinese room doesn't bring us any closer to that understanding, and doesn't refute anything.
I think that -- from memory -- this is approximately what Searle believes himself to be saying:
1. Imagine that you're having a conversation with a person in another room, in Chinese. You're writing stuff down on little scraps of paper and getting little scraps of paper back. It's 100% clear to you that this is a real live person you're talking to.
2. Except here's the thing, it isn't. There's actually just this guy in the other room, he doesn't speak read or write Chinese at all. He just has a whole bunch of books and ledgers that contain rules and recorded values that let him transform any input sentence in Chinese into an output sentence in Chinese that is completely indistinguishable from what a real live person who spoke Chinese might say.
3. So, it's ridiculous to imagine that someone could actually simulate consciousness with books and ledgers. There's no way. Since the guy doesn't understand Chinese, he isn't "conscious" in the sense of this example. So we can't describe him as conscious. And the idea that the books are conscious is ridiculous, because they're just information without the guy. So there actually can't be any consciousness there, even though it seems like it. Since consciousness can't be simulated by some books, it's clear that we're just interacting with the illusion of consciousness.
Meanwhile, this is what people like myself hear when he tries to make that argument:
1. Imagine that you're having a conversation with a person in another room, in Chinese. You're writing stuff down on little scraps of paper and getting little scraps of paper back. It's 100% clear to you that this is a real live person you're talking to.
2. Except here's the thing, it isn't. There's actually a system made up of books and ledgers of rules and values in the other room. There's this guy there who doesn't read or write Chinese; he just takes your input sentence in Chinese and applies the rules, noting down values as needed, transforming it until the rules say to send it back to you. That's the sentence that you get back. It's completely indistinguishable from what a real live person who spoke Chinese might say.
3. So, it's ridiculous to imagine that someone could actually simulate consciousness with books and ledgers, but we're doing it for the sake of argument because it's a metaphor that we can hold in our heads. No one would claim that the guy following the rules in the other room is the "conscious" entity that we believe ourselves to be communicating with. And no-one would claim that static information itself is conscious. So either the "conscious" entity must be the system of rules and information as applied and transformed, or else there is no conscious entity involved. If there is no conscious entity involved, and since this is a metaphor, we can substitute "books and ledgers" with "individual nerve cells with synaptic connections" and "potentials of activation", and the conclusion will still hold true; there will still be no consciousness there.
4. However, we feel that there is a consciousness there when we interact with a system of individual nerve cells, synaptically connected with various thresholds of potentiation: even if it's a system smaller by an order of magnitude or so* that the one in our skulls, like our dog has. Thus we must conclude that the "conscious" entity must be the system of rules and information as applied and transformed, or we must conclude that notion of consciousness is ill-founded and inarticulate, that our understanding of consciousness is incomplete, and that our sense of "knowing" that we or another person are conscious is likely an illusion.
*I am fudging on the figure, but essentially we're comparing melon volumes to walnut volumes, as dogs have thick little noggins.
I think I mostly agree with you, but I would argue that if your notion of consciousness is ill-founded and inarticulate, you can't really decide whether it's an illusion either. After all, the subjective experience quite definitely does happen/is real, thus obviously not an illusion, while the interpretation offered for that subjective experience is incoherent, thus there is no way to decide whether it's describing an illusion or not.
Searle would look at that and conclude it had consciousness.
It is important to distinguish between "understanding Chinese" and "knowing what it's like to understand Chinese". We immediately have a problem: knowing what it's like to understand Chinese involves various qualia, none of which is unique to Chinese speakers.
So I'll simplify the argument. Instead of having a room with a book containing rules about Chinese, and a person inside who doesn't Chinese, we have a room, with some coloured filters, and a person who can't see any colours at all (i.e. who has achromatopsia). Such people (e.g. http://www.achromatopsia.info/knut-nordby-achromatopsia-p/) will confirm they have no idea what it's like to see colours. If you shove a sheet of coloured paper under the door, the person in the room will place the different filters on top of the sheet in turn, and by seeing how dark the paper then looks, be able to determine its colour, which he'll write on the paper, and pass it back to the person outside. The person outside thinks the person inside can distinguish colours, but the person inside will confirm that not only can he not, but he doesn't even know what it's like. Nothing else in the room is obviously conscious.
A propos of the dog, this is the other minds problem. It's entirely possible that I'm the only conscious being in the universe and everyone else (and their pets) are zombies. But we think that people, dogs, etc. are conscious because they are similar to us in important ways. Kurzweil presumably considers computers to be conscious too. Computers can be intelligent, and maybe in a few years or decades will be able to pass themselves off over the Internet as Chinese speakers, but there's no reason to believe computers have qualia (i.e. know what anything is like), and given the above argument, every reason to believe that they don't.
>But we think that people, dogs, etc. are conscious because they are similar to us in important ways.
Specifically, mammals have mirror neurones. More complex mammals also seem to have common hard-wired links between emotions and facial expressions - so emotional expression is somewhat recognisable across species.
I'm finding the AI debates vastly frustrating. There are basic features of being a sentient mammal - like having a body with a complicated sensory net, and an endocrine system with goal/avoidance sensations and emotions, and awareness of social hierarchy and other forms of bonding - that are being ignored in superficial arguments about paperclip factories.
It's possible that a lot of what we experience as consciousness happens at all of those levels. The ability to write code or find patterns or play chess floats along on top, often in a very distracted way.
So the idea that an abstract symbol processing machine can be conscious in any way we understand seems wrong-headed. Perhaps recognisable consciousness is more likely to appear on top of a system that models the senses, emotions, and social awareness first, topped by a symbolic abstraction layer that includes a self-model to "experience" those lower levels, recursively.
But bit by bit, you learn to control your eye's focal length independent of "where" in space you want to look. It really is astonishing.
It made me think of consciousness as a measure of ability to integrate information, because this process is truly fascinating to anybody who tries it (and I really think you should!) Perhaps that's because with this trick, you were able to integrate more information, and thus tickle your brain more?
The filters are just pieces of transparent coloured plastic. How are they capable of forming associations?
Also, associations on their own (e.g. blue with sky, red with blood, green with grass) don't give you any idea what colours are like. Knut Nordby (and many other people with achromatopsia) knew these associations as well as you or I know them, but made it quite clear that he had no idea what it was like to see in colour.
So what about those original experiences? How are they important at all if there is nothing to associate them with?
I think this is an excellent point. I like your example with colors, which shows that there is a difference between seeing (i.e. experiencing) colors and producing symbols which give the impression that an entity can see colors.
I don't follow any argument that proposed that computers can be conscious but other machines (e.g. car engines) cannot. In the end, symbols don't really exist in physical reality - all that exists is physical 'stuff' - atoms, electrons, photons etc. interacting with each other. So how can we say that one ball of stuff is conscious but another is not? And why isn't all of the stuff together also conscious? Why not just admit we don't know yet?
Consciousness may be hard to define, but lets take something simpler - experience, or even more specifically - pain. I can feel pain. While I can't be 100% sure, I believe other humans feel pain as well. However I don't believe my laptop has the capacity to feel pain, irrespective of how many times and in how many languages it can say 'I feel pain'.
Perhaps the ability to experience is the defining characteristic of consciousness?
It really lives up to its title. Suddenly computational complexity is not just a highly technical CS matter anymore, and the Chinese Room paradox is explained away successfully, at least for me.
1) Syntax without semantics is not understanding. 2) Simulation is not duplication.
Claim 1 is a criticism of old-style Symbolic AI that was in fashion when he first formulated his argument. This is obviously right, but we're already moving past this. For example, word2vec or the recent progress in generating image descriptions with neural nets. The semantic associations are not nearly as complex as those of a human child, but we're past the point of just manipulating empty symbols.
Claim 2 is an assertion about the hard problem of consciousness. In other words, about what kinds of information processing systems would have subjective conscious experiences. No one actually has an answer for this yet, just intuitions. I can't really see why a physical instantiation of a certain process in meat should be different from a mathematically equivalent instantiation on a Turing machine. He has a different intuition. But neither one of us can prove anything, so there's nothing else to say.
Neural nets are somewhat starting to escape that dynamic but there still isn't a neural net that reliable pulls in a continuous stream of randomness to generate meaningful behaviour like our consciousness does.
Now to be honest; I'm not entirely sure if John Searle would agree that that is consciousness when we do get there but I do agree with him that deterministic consciousness is essentially a contradictio in terminis.
So the question, to Searle, is not "about what kinds of information processing systems would have subjective conscious experiences", but "what kinds of electrochemical interactions would cause conscious experiences".
The intuition/assumption of his questioners seems to be that whatever electrochemical interactions are relevant for consciousness, they are relevant only in virtue of their being a physical implementation of some computational features, but plainly he does not share this assumption and favours the possibility that the electrochemical interactions are relevant because they physically (I think he'd have to say) produce subjective experience - and that any computational features we attribute to them are most likely orthogonal to this. Hence his example of the uselessness of feeding an actual physical pizza to a computer simulation of digestion. His point is that the biochemistry (he assumes) required for consciousness isn't present in a computer any more than that required for digestion is.
Another example might be: you wouldn't expect a compass needle to be affected by a computer simulating electron spin in an assemblage of atoms exhibiting ferromagnetism any more than it would be by a simulation of a non-ferromagnetic assemblage.
To someone making the assumption that computation is fundamental for explanations of consciousness, these examples seem to entirely miss the point, because it's not the physical properties of the implementation (the actual goings on in the CPU and whatnot) that matter, but the information processing features of the model that are the relevant causal properties (for them.)
But to Searle, I think, these people are just failing to grok his position, because they don't seem to even understand that he's saying the physical goings on are primary. You can almost hear the mental "WHOOSH!" as he sees his argument pass over their heads. In an observer-relative way, of course.
As you imply, until someone can show at least a working theory of how either information processing or biochemistry can cause subjective experience the jury will be out and the arguments can continue. I won't be surprised if it takes a long time.
(Edited to add the magnetic example and subsequent 2 paragraphs.)
> but we're past the point of just manipulating empty symbols.
We've now reached the point where we can manipulate large matrices containing floating point numbers. I don't see how this makes systems any more conscious.
I think people fall for Searle's argument the same way people fall for creationist arguments that make evolution seem absurd. Complex systems that evolve over long periods of time have enormous logical depth complexity and exhibit emergent properties that really can't be computed analytically, but only but running the simulation, and observing macroscopic patterns.
If I run a cellular automaton that computes the sound wave frequencies of a symphony playing one of Mozart's compositions, and it takes trillions of steps before even the first second of sound is output, you can rightly ask, at any state, how is this thing creating music?
I could be an empty shell imitating a human perfectly, all other humans would buy my lack of consciousness, and nothing would be different, from their perspective I exist, from mine, I don't have mine.
How does one know that I really understand something? Maybe I can answer all the questions to convince them?
The audience question about intelligent design summed up my frustration nicely (or rather the amoeba evolving part of it).
I think studying the brain (as opposed to philosophical arguments) is the thing that will eventually answer these kinds of questions, though.
Nothing Searle says stands in the way of creating intelligent or super-intelligent entities. All Searle is saying is those entities won't be conscious.
No can prove this claim today. But more significantly I think it's extremely likely no one will ever prove the claim. Consciousness is a private subjective experience. I think it's likely you simply cannot prove it exists or doesn't exist.
Mankind will create a human-level robots and we'll watch them think and create and love and cry and we'll simply not know what their conscious experience is.
Even if we did prove it one way or the other, the popular opinion would be unaffected.
Some big chunk of people will insist robots are conscious entities who feel pain and have rights. And some big chunk of people will insist they are not conscious.
It might be our final big debate. An abstruse proof is not going to change anyone's mind. Look at how social policies are debated today. Proof is not a factor.
It's amazing to me that people with such a pathological disconnect between mind and intuition can get so far in life. He's incredibly smart, has a great intuition, but when exposed to some problems he simply can't CONNECT his REASON with his INTUITION. This is a MENTAL ILLNESS and we should invest in developing ways to treat it, seriously!
Of course that "the room + person + books + rule books + scratch paper" can be self conscious. You can ask the room questions about "itself" and it will answer, proving that it has a model of itself, even if that model is not specifically encoded anywhere. It's just like mathematics, if you have a procedural definitions for the set of all natural numbers (ie. a definition that can be executed to generate the first and the next natural number), you "have" the entire set of natural numbers, even if you don't have them all written down on a piece of paper. Same way, if you have the processes for consciousness, you have consciousness, even if you can't pinpoint "where" in space and time exactly is. Consciousness is closer to a concept like "prime numbers" than to a physical thing like "a rock", you don't need a space and time for the concept of prime numbers to exist in, it just is.
His way o "depersonalizing" conscious "machines" is akin to Hitler's way of depersonalizing Jews, and this "mental disease" will probably lead to similar genocides, even if the victims will not be "human" ...at least in the first phase, because you'll obviously get a HUGE retaliation in reply to any such stupidity, and my bet it that such a retaliation will be what will end the human race.
Now, of course the Chinese room discussion is stupid: you can't have "human-like consciousness" with one Chinese room. You'd need a network of Chinese rooms that talk to each other and also operate under constraints that make their survival dependent on their ability to model themselves and their neighbors, in order to generate "human-like consciousness".