Harnad, S. (2009) Cohabitation: Computation at 70, Cognition at 20, in Dedrick, D., Eds. Cognition, Computation, and Pylyshyn. MIT Press
Zenon Pylyshyn cast cognition's lot with computation, stretching the Church/Turing Thesis to its limit: We had no idea how the mind did anything, whereas we knew computation could do just about everything. Doing it with images would be like doing it with mirrors, and little men in mirrors. So why not do it all with symbols and rules instead? Everything worthy of the name "cognition," anyway; not what was too thick for cognition to penetrate. It might even solve the mind/body problem if the soul, like software, were independent of its physical incarnation. It looked like we had the architecture of cognition virtually licked. Even neural nets could be either simulated or subsumed. But then came Searle, with his sino-spoiler thought experiment, showing that cognition cannot be all computation (though not, as Searle thought, that it cannot be computation at all). So if cognition has to be hybrid sensorimotor/symbolic, it turns out we've all just been haggling over the price, instead of delivering the goods, as Turing had originally proposed 5 decades earlier.
Subscribe to:
Post Comments (Atom)
-
Harnad, S. (2003) The Symbol Grounding Problem . Encylopedia of Cognitive Science. Nature Publishing Group. Macmillan. or: Harnad, S...
-
Bekoff, M., & Harnad, S. (2015). Doing the Right Thing: An Interview With Stevan Harnad . Psychology Today . Twelve years after steppi...
-
Extra optional readings: Harnad, S. (2011) Minds, Brains and Turing . Consciousness Online 3. Harnad, S. (2014) Animal pain and hu...
Although I agree that introspection and behaviourism cannot explain cognition, I do not think they should be disregarded in understanding cognition. In the first class, we defined information as the reduction of uncertainty. Since cognition entails many uncertainties, it might be possible that introspection and observable behaviours may partly reduce particular (and probably very specific) uncertainties. Sure, introspection disregards the thought processes we are unaware of, such as how we were able to mentally find a picture of our third-grade teacher; but the fact that we are able to have that mental image tells us that there is a process happening, which reduces the uncertainty that there is not. Moreover, the fact that we are able to name our third-grade teacher tells us that we have “computed” an answer, and that cognition could very well be the driving force of that computation. Even though introspection and behaviourism do not provide the answer for how we can do things, they at least contribute in tackling the “easy” question of cognition of what can be done.
ReplyDeleteThe task of cogsci is to explain how and why organisms can do all the things they can do. That is the “easy problem” of cogsci. Neither introspection nor behaviorism can answer those questions. (But it can tell you whether you feel hungry: that’s Descartes’ cogito/sentio. No uncertainty there — though your feeling of hunger could be a symptom of indigestion rather than low blood-sugar!)
DeleteIt is not at all clear yet whether (all) thinking is (just) computation.
“Searle thought the culprit was not only the insufficiency of computation, but the insufficiency of the Turing Test itself”. Searle’s Chinese room argument led him to the conclusion that cognition cannot simply be computation, because computation does not inherently imply understanding, and an entirely computational system leaves no room to explain how it does what it does. Searle can manipulate and execute problems (in this case, writing and responding in Chinese) without knowing how the procedures work (in this case, the definitions of Chinese characters and grammar). The reason this shows human cognition is not entirely computation is lost on me. It was stated in article that, “If we are to explain our cognitive capacities, we must somehow come up with explicit hunches about how we are able to do what we can do, and then we have to test whether those hunches actually work”. One is able to think without knowing how one thinks. One can know the definitions of Chinese characters without knowing how the mind learns, retains, and manipulates those characters. Is that not simply computing without understanding?
ReplyDeleteWhat is at issue in Searle’s Chinese Room Argument is not whether Searle knows how he does what he does (none of us do: we’re waiting for cogsci to solve the “easy problem” and tell us). What is at issue is whether Searle, in passing the Chinese Turing Test by executing the computer program that can pass it, is understanding Chinese
Delete“ Computation is rule-based symbol manipulation; the symbols are arbitrary in their shape (e.g., 0’s and 1’s) and the manipulation rules are syntactic, being based on the symbols’ shapes, not their meanings. (…) But if symbols have meanings, yet their meanings are not in the symbol system itself, what is the connection between the
ReplyDeletesymbols and what they mean? Harnad, S.
The arbitrariness of the sign, i.e., the idea that symbols have their shape AND their meaning because “it is so”, is at the same time a necessity and a difficulty of computation. It is necessary, because the machine, which is only a mechanism, cannot manipulate anything else than the physical aspects of the symbol. For example, a personal computer can only act on which of its electronical units are on/off (hence representing 0s and 1s, themselves representing complex commands in binary language, and so on to do what is useful to humans). But of course, there is something missing in this picture. Like painted by Searle’s Chinese room argument, the physical features of a symbol are not enough to provide its meaning. Even if I know every rules of Chinese symbols manipulation, in a way that I can imitate perfectly a Chinese speaker, if I don’t understand the symbols themselves, that I simply don’t understand what the symbols are doing/are meaning. They just appear to me as random scribblings.
The idea that cognition (i.e. human-animal general thinking) is computation is then put on jeopardy. Indeed, it seems doubtful that if a computer cannot understand the meaning of the symbols it’s using, then a computer inside your mind (or more exactly, if your mind is a computer) will do the trick. Except if you want to beg the question by imagining a homunculus inside your mind understanding the meaning of symbols (but how can he do it? The story rebegins). Maybe do we have to think differently. What if some of the meanings were non-arbitrarily associated with the shape of the symbols? Like the icon of a male/female on a bathroom door, representing (reasonably) a male/female. But then the difficulty is to explain how seemingly arbitrary symbols (like the shape of a 0 meaning the numeral concept of “nothing”) get their meaning, the very same symbol grounding problem.
Just pictures paired with symbols (words) certainly won't solve the symbol grounding problem. What will?
Delete"So Searle simply proposes to conduct the TT in Chinese (which he doesn’t understand) and he proposes that he himself should become the implementing hardware himself, by memorizing all the symbol manipulation rules and executing them himself, on all the email inputs, generating all the email outputs. Searle’s very simple point is that he could do this all without understanding a single word of Chinese."
ReplyDeleteTherefore, we may conclude that the Turing Test is not a viable option for determining intelligent cognition because the process, though computational, still does not encompass the "understanding" aspect of cognition. The man or machine translating is indeed, a weak AI, as he/it simulates cognition or "understanding" to users or any external entity. However, this all begs the question that many have posed and attempted to answer: how can we characterize and test understanding? What is understanding if not simply reacting appropriately to certain situations?
Just as the man or machine in the Chinese Room performs computations to derive the appropriate output, humans are more or less "programmed" to behave and react in certain ways. Societal expectation and cultural norms demand specific physical actions and emotional reactions to different scenarios. For example, as a Canadian living in an urban city, I would respond to someone raising their middle finger at me with anger; if current state is happy, and symbol is "person flipping me off", then go into state "feel angry" (and maybe "flip person back"). Contrastingly, perhaps an individual living in Japan, where according to a Quora forum the middle finger has no offensive meaning, would not react in the same way because they were "programmed" differently with a distinct computational process. Does this mean that the person in Japan does not understand? They still perform cognition and "understand" their surroundings.
The paper asserts that a full robotic version of the Turing test in which the machine can form and mediate connections between internal symbols and the external world may more accurately represent "understanding", and I agree. However, we must then consider the environmental influences or "program" the entity is subject to. What may be a correct response or answer in one context could be incorrect in another. Perhaps it is this point, that external variables may interrupt internal processes and thus produce the occasional, unexpected wrong answer, that one can identify an internal system adapting to different external inputs.
Searle does not show that the TT is not a valid test; he shows that passing it does not generate understanding; so cognition (which includes understanding language) cannot be just computation.
DeleteRead Searle (week after next) to understand what “Weak AI” means. (“Strong AI” is computationalism.)
If computationalism were true, it would not mean humans are “programmed” any more than DNA means they are programmed. It would mean that cognition (thinking) is just computation (the manipulation of symbols).
Computationalism (whether true or false) is just as compatible as any other theory of cognition with the fact that experience changes behavior.
Cohabitation: Computation at 70, Cognition at 20
ReplyDelete---“The answer in the case of syntax had been that we don’t really “learn” it at all; we are born with the rules of Universal Grammar already in our heads. In contrast, the answer in the case of vocabulary and categories is that we do learn the rules, but the problem is still to explain how we learn them: What has to be going on inside our heads that enables us to successfully learn, based on the experience or training we get, to identify categories, to which we can then attach a name?”
Behaviorism is wrong in that a reward/punishment mechanism is far too weak to account for the rapid language acquisition of children. Consequently, Universal Grammar, an innate language faculty, has been proposed. Similarly, there is also evidence showing that children may be genetically equipped to learn words and to identify categories. As young as the age of 18 months, children may show “shape bias” in extending the use of newly-learned word. That is, children prefer to extend words based on the same shape of objects, more than based on other properties including size, color and material. It is unclear how a child who can hardly speak may learn this bias by reward/punishment within such a short time, so this may be another piece of evidence against behaviorism.
---"We are unaware of our cognitive blind spots – and we are mostly cognitively blind.”
Just to add how wrong it may be to rely too much on introspection, from a linguist’s perspective: language is in many ways counter-intuitive. For every sentence that we speak without thinking, there exists a deep structure, which in many cases are drastically different from the surface one. Native speakers of any language are completely unaware of those structures.
Universal Grammar (UG) comes later in the course (Weeks 8 and 9). It is not relevant here yet, irrespective of whether it is innate or learnable or even exists. And UG is equally compatible with the truth or falsity of computationalism.
Delete“The gist of the Turing Test is that on the day we will have been able to put together a system that can do everything a human being can do, indistinguishably from the way a human being does it, we will have come up with at least one viable explanation of cognition.” I understand the explanation of cognition Turing would be getting at with the completion of this Turing Test, but what are these other viable explanations of cognition Harnard is referring to? Wouldn't the completion of a Turing Test that is indistinguishable from a human person necessarily support just Turing's hypothesis?
ReplyDeleteYes. But who says computation alone could pass the TT? (There are actually more than one version of the TT. Turing described only the verbal one, but there's a robotic one too: a solution to the easy problem has to explain how organisms do everything they can do, not just texting. "Stevan says": computation alone cannot pass even just the verbal TT. We'll get to that in the next few weeks.
DeleteSearle's thought experiment makes you think about what truly it means to understand and what is required to understand. Surely, it can't all be just computation because as Searle so cleverly made himself the machine, he himself did not understand Chinese. Although a solution to how we can understand is beyond my grasp and much of cognitive science at the moment, the idea of scaling the TT to all behaviour seems like the appropriate next step. As I stated in my comment on 1a, the differences between the machine and us lies the answer of understanding that we are looking for. A cochlear implant can be useful but, it doesn’t necessarily understand. It lacks the other sensorimotor capacities that the biological system contains.
ReplyDeleteSearle would not understand Chinese even if he executed the computer program that passes the Chinese (verbal) TT: Why?
DeleteIn reply to “Searle would not understand Chinese even if he executed the computer program that passes the Chinese (verbal) TT: Why?”, Searle does not understand Chinese because he lacks intention of his mental states. He cannot perceive the meaning of the words - they are incoherent symbols. As Searle states: understanding implies both the possession of intentional mental states and the validity of these states.
DeleteAmy, your comment compares a cochlear implant to Searle – they both act as the hardware for the system. While the cochlear implant (and Searle) does not understand, the system as a whole may. Someone with a hearing impediment can perceive sound and would claim they understand sound. The case for does the room (the system in CRA) understand is much more complicated. As Harnad points out in class, what if there was room… you were instead in an open field. What is the system? Is there any understanding?
"Intention"? "Mental states"? What it feels like to understand; felt states.
DeleteSearle's periscope works for hardware-independent computation, but not for cochlear function. For processes other than computation, all bets are off.
A section of the paper Cohabitation: Computation at 70 that I found particularly interesting was in the section Introspection Won’t Tell Us when Harnad points out that “we don’t notice what we are missing: We are unaware of our cognitive blind spots – and we are mostly cognitively blind.” (2005). By this I believe he means that most of us don’t even realize that there are mental processes between the input we receive from our environment and the output we execute. I think that this can be majorly attributed to the fact that society is for the most part solely concerned with the end product. In general, we don’t really care how our brain comes up with answers, similarly to how large corporations don’t pay any interest to whether their manufacturing processes affect the environment. We acknowledge that there are many different methods to get to the same answer, behavioural equivalence, but in reality, we only value the answer. For this reason, I agree that the study of cognition from a cognitive science perspective needs to not only ask more “how” type questions, but also just think about what questions we could ask and try to remove our “blinders”. This process may take a little time in the armchair, but it is not introspection. It is not thinking about how we are thinking, but rather asking questions about what we don’t know about cognition so that we can find a scientific method to figure the answer out.
ReplyDeleteI also found this section of the paper intriguing. However, I don’t think that by being ‘cognitively blind’ Harnad referred to our failure to notice mental processes. Moreover, we surely are aware of our thoughts, memories and emotions, for all we can be sure of is that we are feeling (cogito ergo sum, or more accurately sentio ergo sum). Thus, if by mental processes we mean the same thing, then aren’t we aware of them?
DeleteI believe by the term ‘cognitively blind’, Harnad meant to describe how, due to our brains serving us all the answers to our questions on a silver tablet, we fail to come up with the fundamental and non-trivial explanations. When Tom was asked to tell us the name of his 3rd grade teacher last class, we may have argued that he first pictured her in his head before associating this mental image with her name. This introspective explanation is incomplete though, it fails to account for the more fundamental ‘functional’ questions, which are part of the easy problem of Cognitive Science: “How did Tom come up with this mental image in the first place? How exactly did he identify his teacher’s picture? Because of our collective cognitive blind spots it took us 30 years to realize that we were missing a fundamental part of this puzzle.
While reading this chapter, the statement “the only way to do this…is if cognitive science hunkers down and sets its mind and methods on scaling up to the Turing test, for all our behavioral capacities” really summarized my view on cognition and computation. I agree with Searle’s idea that a correct answer may be given without actually understanding it, but for everyday problems where we are just concerned with getting the “right” answer, is a quick answer not all we need? If one wants to know what to wear when getting out of bed, for example, they do not need to think about the process of finding out what the temperature is like, they just know because they feel it. However, it is important to understand the processes that occur in order to reproduce some behaviors in the future, for example if one wants to create a robot that can emulate human behaviour to the extent of passing as a human being. Ultimately, I believe that in order to understand cognition, the Turing test would have to be reproduced for every behavior possible, thus showing all possible procedures and demonstrating the dynamic interaction between input and output. Such tests would demonstrate that cognition is a blend of both sensorimotor information and symbols, and that these not mutually exclusive.
ReplyDeleteWhen we are trying to decide what to wear, we are not doing cogsci (which asks how and why we can do it).
DeleteThe Turing Test and the easy problem are about explaining our behavioral capacities not "every possible [or actual] behavior." For the latter we would have to know the details of every difference between every individual, as well as their past, present and future experiences. That goes rather beyond the mission of cogsci!
This may be my misunderstanding of the Turing Test, but it seems to me that one could create a machine that passes the Turing Test but still not understand its software, ie the easy question of cogsci. This idea struck me when reading the Computation and Consciousness section, which says software can run on many different hardwares. This got me thinking about programmers and how, often, they will have a bug in their code, breaking their program. They fix the bug, breaking the program further. Then, finally they fix that bug, and their program now works! Hooray! But they don't explicitly know how it works; they don't know what they did to the code to make it work. So, again, the cognitive scientist is stuck wondering how does this machine get from point A to point B.
ReplyDeleteAt the same time, if these bugs never occurred, and the machine worked from the start, I believe it would still be possible for us to not be certain of how the machine's software worked. We would know the hardware inside and out, after all we designed it, but there might still be a degree of uncertainty in regards to the software. Hypothetically, we accounted for all the possible variables the software could encounter and put categories in places for the software to categorize them, but we wouldn't know how it categorized them. We could tell the software to categorize dogs and cats differently based on their faces, but we couldn't account for every variable for every kind of thing that exists.
Now, say there is an expert programmer who somehow did just that–program every possible variable and therefore knows every potential route their software could take to get from A to B. That's great! The problem is solved! Except there's always multiple ways to accomplish a task. Another master programmer could figure out a different way to code a TT-passing machine. This is all to say that there may be so many ways to pass the TT that knowing one way won't suffice. We'd know how that particular machine gets from point A to B but we'd still be left with every other human or machine that can pass the TT. How does every other consciousness work?
Good points. But before worrying about the details or alternative ways of doing the same thing keep in mind that computationalism is just a thesis. Finding the code that will pass TT (the email version) for a lifetime has not only not happened but it is not even clear whether it is possible.
DeleteIn any case, TT-passing code is not likely to be discovered through one huge accident, by luck, or through a dream.
And the proof of Fermat’s last theorem is a proof, whether or not you understand it…
Later in the course we will discuss “underdetermination,” which is the fact that data can often be explained in more than one way.
None of this says anything about consciousness (the HP); it’s all squarely in the EP.
What I think is wrong about the computational model of the mind is that it entails that cognition is something passive. As pointed out by Harnad (2009), limitations of computationalism can particularly be seen when looking at symbol-grounding. Let’s assume humans have a set of rules that link representations in the internal world to “things” in the physical world. The universality of these rules (applicability to every computer/Turing machines imaginable) depends on the ability to accurately recover real-world properties. However, evidences have shown that percepts are actively created and do not always reflect the state of the world as a universal rule would predict. The perfect example is color categorization. The same exact shade will be perceived differently depending on its surroundings even if the wavelength reflected by that color never physically changes (Purves 2014, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3977276/) . Sensorimotor dynamics are thus somehow embodied. The attempt to reduce them to computational algorithms takes us away from their biological nature and create this void we cannot explain. This leads us to Harnad’s conclusion that sensorimotor dynamics, and not just symbolic simulations of them, should supplement current computational models in under to arrive to a more satisfying functional explanation of cognition.
ReplyDeleteNeither computationalism nor the TT implies passivity. Verbal interactions are not passive. But what is missing in a purely verbal TT (which we will call T2), executed by computation alone (i.e. a computer) is interaction with the things in the world that words are about, rather than just interactions between words (symbols). For this you need the robotic TT (T3) which allows sensorimotor interactions with the world. But sensorimotor interactions with the world cannot be just computation (symbol manipulation).
DeleteHere are the passages I was referring to in Harnad 2009 :
ReplyDelete" The root of the problem is the symbol-grounding problem: How can the symbols in a symbol system be connected to the things in the world that they are ever-so-systematically interpretable as being about: connected directly and autonomously, without begging the question by having the connection mediated by that very human mind whose capacities and functioning we are trying to explain! "
" There is still scope for a full functional explanation of cognition, just not a purely computational one. As we have seen, there are other candidate autonomous, non-homuncular functions in addition to computation, namely, dynamical functions such as internal analogs of spatial or other sensorimotor dynamics: not propositions describing them nor computations simulating them, but the dynamic processes themselves, as in internal analog rotation; perhaps also real parallel distributed neural nets rather than just symbolic simulations of them. "
In other words, the interactions of a sensorimotor robot in the world, not just a symbol-manipulator (a computer, computing).
DeleteRegarding the 3rd grade teacher's name test:
ReplyDeleteI'm not convinced that the explanation "I pictured her and then I got to her name" is homunculus regression (a little person inside the head selecting the image from many) - can't this be explained from physiological processes of neurons firing from "name" and "third grade teacher" to bring, by association (without volition), the name and picture of one's third grade teacher?
Regarding Searle's Chinese Room Argument:
I have a hard time accepting that a person (a conscious, sentient being) could learn the rules of written Chinese to respond to any Chinese text with Chinese text in such a way that nobody would suspect this person does not understand Chinese, without actually (at least accidentally) learning to understand Chinese. Surely such a computational feat is too sophisticated for the human mind to accomplish.
If the person in the room were to follow, let’s say, a rulebook, written by a programmer that contains every possible message received in Chinese, and every possible response, there’s a distinction between the cognizer who matches and delivers a response (the person who doesn’t understand Chinese) and the cognizer who designs responses based on input (someone who understands Chinese – as the idea of a machine capable of this level of sophistication at analyzing and mimicking human responses to language would render this thought experiment redundant): so there are two cognizers responsible for the reply, one of whom understands Chinese. The deliverer of the message (the person who cannot understand Chinese) is performing a cognizant function not even as sophisticated, probably, as the computer he/she/they are using to email the response. The same could be said if this rulebook were memorized by the person in the room. But the rulebook was written by someone, surely, who must understand Chinese to design the replies! Unless a machine had access to every possible message and response in Chinese that a person would pose to the machine (impossible – and again, if such a thing existed, it would render this thought experiment redundant), or was good enough at analyzing the available written Chinese to induct the rules of how the language and symbol relationships work – but is this possible – even in a though experiment – without symbol-grounding? I’m not convinced that, if we assume it is possible, that renders the Chinese Room Argument worthwhile. I don’t see a compelling Chinese Room Argument scenario in which no part of the cognizer(s) responsible for the output do not understand Chinese.
I also don’t think it’s irrelevant that, during the exchange between the Chinese speaker and the person in the room, cultural context changes, and the machine does not (a static cognizant acting in a dynamic context) – and surely, the machine would eventually be found out, and would not be TT-passable. People adapt with change and culture, but programs do not. For it to approximate human responses closely enough to pass the TT, the machine would need to have constant human intervention & tweaking. Even if we assume a static culture for the purposes of the thought experiment, a unique and personal ‘culture’ (relationship) would be created between the two interactors. This cannot be ignored, as it is an essential trait of human interaction and language that it is dynamic.
1. Simple association would be the explanation -- if it worked.
Delete2. Sounds like you doubt that computation alone could pass the verbal TT (T2). I doubt it too. But why not?
3. A rule book for everything anyone could ever say to or answer anyone in Chinese (including this sentence, in Chinese)? Surely passing T2 through computation alone would require an algorithm, not an infinite input/output table! But if there were such an algorithm (whether designed by a person or by evolution and natural selection, like DNA), it would only require one entity to execute it, not two. (And task is just to execute the algorithm (symbol manipulation) not to use them to learn Chinese — whether the executing hardware is a computer or Searle.)
4. The TT is one step ahead of you: It has to be able to handle any (verbal) input, present and future, indistinguishably from a real Chinese-understanding person. (Think of Tom’s capacity in a lifelong email interactions.) “People adapt with change and culture, but programs (algorithms) do not. “ What makes you think that? There are algorithms that learn from their interactions, and are changed by them…
*"Beware of the easy answers: rote memorization and association. The fact that our brains keep unfailingly delivering our answers to us on a platter tends to make us blind (neurologists would call it "anosognosic") to the fact that there is something fundamental there that still needs to be accounted for. Memorizing single-digit sums, products and differences by rote, case by case, covers the trivial cases, but it does not generalize to the nontrivial ones. Surely we have not pre-memorized every possible sum, product and difference?"*
ReplyDeleteIn this, the implication is that brute memorization is not sufficient in order to process the arithmetic operations that are given as examples to allow processing of the trivial cases but not the nontrivial ones (presumably operations on numbers with several digits. I believe there is more nuance to this as the way students are often taught to work with larger numbers is to operate on them in their single digit forms then play tricks with the zeros. For example, when grade school students are taught addition for n-digit numbers, they are told to work from the least significant digit upwards and "carry the one" to the next digit. This means that their operations are always relying on the brute force memorization of single digit numbers. At what point does a calculation like this cross from the trivial brute force method of memorization to a non-trival computing effort beyond sheer memory?
The TT is about generating everything people are able to do (for T2, verbally), and computationalism tries to do it with symbol manipulation alone. But there is no reason to believe that computation alone is enough when the only thing the human is doing is maths, applying an algorithm (like extracting the roots of a quadratic equation): that people can do, and do by computation alone, even if they don't know what it all means.
DeleteBut understanding language is not like solving maths problems by applying a formula mindlessly! (Not even understanding maths is!)
“...if symbols have meanings, yet their meanings are not in the symbol system itself, what is the connection between the symbols and what they mean?” (Harnard 2005)
ReplyDeleteThe limits of computation in explaining cognition are clear when the question above is posed -- for while inputs lead to outputs in a computational understanding of how our minds “work”, without understanding the mechanisms which lead to the outputs a computational theory of cognition is incomplete. This connection between symbols and what they mean is not learned through reward/punishment systems, as theorized by behaviorism, either. And introspection of such a connection proves impossible, for in trying to “use cognition” to explain cognition leaves one in a circularity which obviously cannot address the “blindspots” of our mental processes. It seems to me that in order to understand cognition, then, there requires an almost transcendence of human thought process, a look into a process which we are unaware of or a mechanism overlooked. But how can one even begin to hypothesize such a process or mechanism using the very cognitive functions that limit us in the first place? Where do we look for such solutions, if not in our own heads?
I am tempted to look towards the non-human for where to start, for only perhaps in examining cognition beyond the human (perhaps in animal cognition studies, or even the study of semiotics as it relates to ecological systems) can we begin to pose questions that do not lock us, for lack of a better phrase, in our own heads. I believe that the study of cognition should not be limited to the study of human cognition -- for aren’t all systems on Earth (potentially the universe?) operating through a process between input and output whose mechanism is rooted in what we call ‘cognition’?
A computational understanding of how inputs lead to outputs requires an algorithm that generates the outputs, given the inputs. In computationalism, the algorithm (executed by hardware) is the mechanism. But manipulating words alone does not connect words to their referents in the world. That’s the symbol grounding problem.
DeleteNo circularity. Don’t let yourself be drawn into vague koans about not being able to know how we know:
If cognition were just computation, and we knew the algorithm, we could understand it. Trouble is, we don’t know the algorithm (that’s why we can’t introspect it) — and computationalism (“Stevan says”) is also wrong, so no algorithm alone can pass T2 anyway.
I’m all for studying cognition in all species (as long as you don’t harm them). But that’s not relevant to these questions about computation and computationalism so far.
"Memorizing single-digit sums, products and differences by rote, case by case, covers the trivial cases, but it does not generalize to the nontrivial ones."
ReplyDeleteIt is true that we have not pre-memorized every possible sum, product, and difference, but for the ones that we have never memorized, it seems that our approach for these problems is to break them down into familiar, smaller problems. This ability to break down harder problems into smaller ones and, at the same time, recognize the prototype of a familiar problem seems to be the core. Yet, I am wondering if it is possible that this is a more abstract version of rote memorization and association. If that were the case, could I say that the main problem would be how we abstract things?
One thing that troubles me is that for a question such as “how do we breathe?”, a physiological answer would certainly satisfy us, yet for a cognitive problem that happens in this flesh too, neuroscience/physiology cannot answer that anymore. It creates too much a gap within one unity. This seems to bring us back to the question of “the existence of zombie”. If there were such an “artificial human” that has every material a human has, with the input that it receives, will it not create human-like “intelligence” in itself?
Up to about 12 x 12, we do multiplication by rote. Past that, we do it by recipe. Yes, the main problem is finding the recipe. We'll talk about abstraction in Week 6.
DeleteA T3 Zombie (e.g., Gabe if he did not feel) would not be cognizing: just doing everything a cognizer can do. ("Stevan Says" there cannot be a T3 Zimbie.)
“Searle thought the culprit was not only the insufficiency of computation, but the insufficiency of the Turing Test itself”
ReplyDeleteSearle’s chinese room argument serves as the basis of his claim that the turing test (or the path towards making the test more comprehensive) should be abandoned in favor of studying the dynamics of the brain. This argument rests on the assumption that a set of rules alone would be sufficient for proficiency in chinese -- without some interface with the world, or at least Searle’s mental representation of the world, wouldn’t it be impossible for such an ungrounded system to even pass the text turing test or is the jury still out on that? Although it might present similar roadblocks, it feels like an updated test around a fully robotic, sensory-motor capable machine is a step in the right direction.
The symbol grounding problem suggests that it would not be possible to pass the verbal version of the TT (T2) unless the candidate could also pass the robotic version (T3), because word meaning has to be grounded in sensorimotor links the the referents of the words.
DeleteIn the paper Cohabitation: Computation at 70, Cognition at 20:
ReplyDeleteIt is mentioned that Chomsky’s ramification calls Skinner’s theory into question not only by looking into syntax but also vocabulary learning. The argument maintains that vocabulary is learned and recalled through categorization, as “Naming things is naming kinds.” Learning categorization is through learning the invariant features extracted from the irrelevant ones.
I am wondering if Skinner could make a case to say that naming kinds is essentially naming things to resolve this issue in behaviorism. Is it possible that Skinner argues that vocabulary learning is still rote memorization in the way that people associate “beak“ and ”wings” with birds and whenever seeing something with recurring stimuli beaks and wings, the speaker says “bird”?
If vocabulary learning is just memorizing every member of the category, we'd not only have to have monstrously powerful rote memories, but we would only know the members of the category that we had already encountered. But a learned feature-detector can generalize to members it has never encountered. (This will be covered in Week 6.)
DeleteI guess I'm still lost on how exactly computation does not require understanding of the meaning. In the 2+2=4 example, it was said that arithmetic would tell us the rules of how to check whether it's true or false, without ever knowing what 2 and + and = and 4 actually mean. However, if I were telling a baby how to examine whether 2+2 amounts to 4, I guess the very fundamental I would set at the first beginning would be what 2 and + and 4 is. Say I'm teaching him how to count these numbers with his candies. Then at least he needs to know that 2 can be represented by 2 candies on the table, and when we say we're adding another 2 to it, it means we can take another 2 candies from the box and put them together on the table too. Then how to arrive at the answer, namely what we will get at the end? We can get that by counting the number of candies that are currently on the table, which leads us to 4. So 4 is represented by the 4 candies presented on the table, and 2 + 2 = 4 works in this way. Without even understanding what 2 and 4 and + and = represent, even we know that +, by definition in math, is to combine two folds of things together and then counting the total, we cannot really compute this math equation, right?
ReplyDeleteI also found that cognitive impenetrability very interesting. According to Pylyshyn, cognition happens on the "functional architecture of the virtual machine" and above which is cognitive penetrable. And below that there's cognitive impenetrability, which means what we sees, or perceives, cannot be changed by anything that someone else tells us. This argument is piquant to me as it leads me to the question that "is it true that all our cognition can be changed by external circumstances, while some of the 'lower-level' biological function in us simply cannot be altered no matter what?" The so-called "lower-level", from my understanding, refers to our biological instincts, which is kind of fixed and limited by our organs. So they are not cognition, but more like simple reflection, which sounds even more like computation to me, since computation, at least in a narrow sense, is about simply following a set of rules to ultimately get the output from the input. So if reflection is also computation, or at least a subset of computation, and according to Pylyshyn, cognition = computation, then we can arrive at that reflection is at least a subset of cognition. But he also thinks biological reflection is exactly what separates from cognition. This is a contradiction to me.
Hi!
DeleteWhat you’re describing in your first paragraph is probably the human instinct to want to instill mathematical basics that are easily computable and that will create a universal form of understanding of a given mathematical concept (in this example, we’re talking about what numbers inherently are meant to represent, and how this makes all future additions, divisions, subtractions, etc. easier and more intuitive). However, a computer doesn’t need an intuitive way to generate answers to new mathematical problems. In their case, it is probably actually more efficient to embed them with a computing strategy that doesn’t rely on the complex representations that make it easier for us humans to understand and apply them to new problems).
The bottom line is that the reason this argument works as it relates to your concern, is likely that computation (or, that recipe the computer is going to use to solve a problem) doesn't need to be intuitive, it just needs to work every time.
Yes, children are taught how to count and add with candies. I was talking a formal, axiomatic maths course in which you are taught to prove that 2 + 2 = 4. Proofs in mathematics are formal and symbolic, relying ultimately on the (formal) law of noncontradiction: "P & not-P is necessarily false."
Delete"Cognitive penetrability" as a criterion for what is and is not "cognitive" looked promising for a while. But since the notion of above/below the level of the "virtual machine" is homuncular (who is the "user"?) and computationalism is wrong (Searle's Argment and the Symbol grounding Problem), we can forget about it. (Besides, the distinction between cognitive and vegetative function is a fuzzy one, and breaks down at the boundary.)
I found it to be a useful and clarifying strategy that theories would initially be presented while acknowledging their intuitive logic before proceeding to draw attention to the flaws in their reasoning. For instance, I can see how easy it would be to buy into the mental imagery theory. However, once it is pointed out how this serves to question-beg and does not address the root of the how we retrieve any information, including images of past school-teachers. Beginning with the superficially intuitive responses aids in the recognition of blind spots, thereby priming readers to go beyond these to search for the underlying functions.
ReplyDeleteIs “symbol grounding – the ultimate question about the relation between the computational and the dynamical components of cognitive function” specifically/only referring to the mind-body problem, with the computational as the cognitive functions and the dynamical as the explicit physical components? I understood that the theory that the hardware/software distinction could resolve the mind-body problem is said to be false, however, I do not entirely understand whether or not an equivocation is being made – If computation is not in fact referring to cognitive functions and dynamical to physiological/explicit operations.
Also, I am not convinced that Searle’s theoretical experiment of becoming the hardware to complete the Turing Test in Chinese is sufficiently reasonable as a premise. I doubt those symbol manipulation rules could be memorized.
I am not clear on what dynamical functions are. I do not know what internal analog rotation is. An example of a physiological process?
Is the final proposed solution to build a T4 and then use backwards engineering to uncover the functional explanation of cognition?
Dynamic means anything physical and changing: hardware, e.g. sensors, motor effectors, heaters, the heart, waterfalls, stars, molecules. Computation is purely formal (symbolic). although it has to be implemented physically (hardware-independent): Countless different hardwares can execute the same computation (or any other computation).
DeleteSymbol grounding (weeks 5 and 6) will be about why cognitive systems have to be hybrid computational/dynamic -- not just computational, in which the hardware is irrelevant except that it needs to execut the right computation.
For Searle's point -- that symbol-manipulations alone cannot generate meaning -- does not depend on whether there is enough time or capacity for him to memorize and execute the T2-passing program.
Internal analog rotation means spinning a shape around to see if it matches another shape. If this is done in your head instead of by your hands externally, it is "internal rotation." Google "shepard rotation."
T4 includes T2 and T3, so if you can build a T4 you can explain T4, as well as T3 and T2. But is T4 necessary to explain T3?
“The answer in the case of syntax had been that we don’t really “learn” it at all; we are born with the rules of Universal Grammar already in our heads. In contrast, the answer in the case of vocabulary and categories is that we do learn the rules, but the problem is still to explain how we learn them: What has to be going on inside our heads that enables us to successfully learn, based on the experience or training we get, to identify categories, to which we can then attach a name?”
ReplyDeleteBehaviorism can’t explain the generative property of language, how we learn semantic categories, and the exponential nature of language-learning in children. Taking all of this into consideration, I still don’t agree that computation by itself can fully replace behaviorism as the primary theory of categorization and language learning.
Computation offers an easy answer to the ‘easy problem’ of cognitive science, but it lacks in its ability to explain the intuitive understanding that humans have regarding syntactic structures, in arguably the same way that computation does not imply understanding in a Turing machine (via Searle’s Chinese Room argument).
One cannot solely look to Universal Grammar to solve this problem, either. Recent research has posed the question of the validity of Chomsky’s theory, the hypothesis being that if rule-based approaches and pattern recognition are widely used methods for children to learn semantic categories, then these methods most likely play a role in the acquisition of syntactical knowledge as well.
Computation not solve the easy problem if it cannot do everything T3 can do.
DeleteWe'll discuss Chomsky in Week s 8 and 9. Grammar is only a (small) part of T3.
“And kinds cannot be identified by just rote associating names to stimuli. The stimuli need to be processed; the invariant features of the kind must be somehow extracted from the irrelevant variation, and they must be learned, so that future stimuli originating from things of the same kind can be recognized and identified as such, and not confused with stimuli originating from things of a different kind.”
ReplyDeleteWhile I do agree that behaviorism isn’t sufficient to explain how things can be identified, the statement feels dismissive towards behaviorism. I do agree that rote association is not the whole story, but I would argue that rote association plays a crucial role. It is my understanding that rote learning is learning through repetition, thus, after encountering a stimulus enough times, you will eventually come to recognize that stimulus for what it is. The logical argument that follows would be: if a person were to encounter and recognize a stimulus, for example a chair, that’s only one specific instance of that stimulus, he should not be able to recognize other kinds of chairs as being a chair by rote associating names to stimuli, because all chairs are different stimuli. However, I think even a behaviorist would agree that stimuli must be processed in our brains. If you acknowledge that stimuli are stimuli and perceptions are perceptions, then you can make the argument that you can identify things through rote association. For instance, if a person encounters a chair, that chair would need certain core features that would make it vaguely resemble that person’s idea of a chair, and if those requirements are met it will be perceived as a chair. So, every time a person encounters a chair, if it meets the requirements of a chair, it will be recognized as such through rote association. Although, in my example, there is more than just rote association, rote association nevertheless is used to identify the chair.
I agree with you that rote association is needed in the process of naming - but as you have described here, there needs to be more than just rote association. The example with the chair involves more than just rote association. You must be able to process the features that are characteristic of a 'chair' in order to create categories or 'kinds' to which a 'chair' would belong. Rote association would be the process of merely seeing one chair and labelling that as 'chair' - this wouldn't allow you to recognize other chairs because they wouldn't fit the exact picture of the original chair. Therefore, there needs to be a process of categorization (and a process wherein you create the category of 'chair'). "To learn to name kinds, you first need to learn to identify them - to categorize them" (Harnad, 2009). I don't think anyone would argue that there is no association at all involved in naming, it is just that rote association is not the only process that leads to successful naming - there must be other processes involved to be able to successfully name.
DeleteRe: “While I do agree that behaviorism isn’t sufficient to explain how things can be identified, the statement feels dismissive towards behaviorism.”
DeleteI did not find the article or the quote to be dismissive towards behaviourism, I actually felt like it if gave it the recognition it deserved. Behaviourism played a crucial role in taking psychology away from armchair introspection and towards a tangible science. It also acknowledges the fact that we don’t have access to some of the processes going on inside out head. In this case, we don’t really have access to the processes responsible for us knowing that a chair is a chair. Borges really drives the point home with his story of Funes the Memorious. We do not have conscious access to the processes responsible for abstraction, but these are still necessary in order for us to categorize and name things.
As I understand it, the part that rote association plays in learning to categorize and name objects is more centered around giving a series of examples, and from that we can extrapolate the inclusion and exclusion criteria relevant to specific categories. The role of rote association is basic but fundamental.
I agree with Harnard that the only way to effectively address the easy and hard problem of consciousness is to scale up on the Turing Test. However, where to begin on achieving a TT-passing robot really fathoms me. If we cannot prejudge the proportion (at least have some governing estimate) at which the internal structure and processes underlying the robot will be computational or dynamic, should the focus then be on creating implementation-independent programs around the aspects of cognition "as cognition can do"? While the halting problem does reveal that there are some uncomputable programs that the Universal Turing Machine cannot resolve, how are we to be sure that our TT-passing robot 'can do' and similarly, 'cannot do' all that we can? Does this involve, extensively testing this TT robot with a mass majority of the population? Behavioral equivalence? What specific indicators will allow us to prove that the TT-robot has passed?
ReplyDeleteThe only way to reverse-engineer what organisms can do is to observe what they can do and then try to model it. The Strong Church/Turing Thesis suggests that computer simulation is one way to test models, but the real test is eventually to build a T3 robot. Gabe is the test. And, yes, it's weak (input/ouput) equivalence. Gabe need not be T4.
DeleteThe most interesting part in “Cohabitation: Computation at 70, Cognition at 20” for me was where computation and consciousness are discussed. I do understand that if the mind turns out to be computational then the mental states could be seen as computational states -or at least in theory it makes sense- but I think mental states are much more complicated than that.
ReplyDelete“The gist of the TT is that on the day we will have been able to put together a system that can do everything a human being can do, indistinguishably from the way a human being does it, we will have come up with at least one viable explanation of cognition.” This reminds me of weak equivalence where the outputs of two systems are the same. Even such day as described above arrives, just because computation gives us the same results or is indistinguishable from the way humans do things, it doesn’t mean that it could explain cognition.
We can get computers or robots do what we do but not how we do it, especially without understanding human cognition by itself. Introspection might not be a requirement to explain or model human cognition but 0s and 1s are simply not enough.
I'm not sure I agree that 0s and 1s are not enough to explain human cognition. It certainly feels weird and it’s far from how we intuitively imagine the human brain functions and how thoughts come about, but there is at least some evidence that neurons are either in a firing state (1) or inactive (0) at any given moment. What we perceive as nuanced or coming in degrees is actually explainable by the amount of 1s or 0s that are received by our system in a certain time frame. For example, a stimulus that makes a neuron fire at a rate of 300 times/second will be perceived as extremely intense stimulus, whereas a rate of 60 times/second will be a weak stimulus. It might not explain all of cognition, but we also don’t know that it doesn’t. I don’t think we should make any assumptions about what cognition is and how it works. Therefore, I’m not sure that we could dismiss so easily that a system that is able to do what humans do indistinguishably from humans is not a system that actually does what humans does. Indeed, I feel like if they are able to do what we do, how could we say they for sure don’t do it how we do it when we have no idea how we do it?
DeleteKarin, you are underestimating how challenging it is to create T3.
DeleteMarianne, yes, because of the other-minds problem, we can never know whether any entity other than oneself feels. But T3 is close enough. But just computation cannot pass T3.
Post that I accidentally published in last year's blog:
ReplyDeleteWhat is computation?
Although there is no clear answer at the moment, I thought that Professor Harnad brought up some great points about what our definition of computation should include. Firstly, Professor Harnad pointed out that “computations need to be dynamically implemented in order to run…”. I believe this is an obvious, yet important point to highlight. Computation is the basis of every action or thought that we produce. If we are not actively thinking or doing something, no computations will be performed. We cannot compute the answer to “Who is was your third-grade teacher” without being aware of the question being asked and without an active mind.
Secondly, Professor Harnad writes that “computation is rule based symbol manipulation… and the manipulation rules are syntactic, being based on the symbol’s shapes, not their meanings”. Computation is something that is done automatically. Humans are unable to come up with an answer when asked how they were able to remember the name of their third-grade teacher because the meaning of the symbols being manipulated to conjure up the answer are not important for the computation of this answer. Humans are only able to remember and state the name of their previous teachers, meaning that once computations are complete, they are interpretable (Fodor & Pylyshyn).
Looking back at the question at hand, computation can be described as meaningless rule based manipulations of information that are dynamically implemented when we are in thought or doing something to produce meaningful answers and actions. But, to put it in a more kid-sib friendly way… Computation is an active process. Computation is something that happens automatically when we think or do something. Computations manipulate the information we receive from the world and produce meaningful thoughts and actions.
You are mixing up computation and cognition somewhat, in your definitions...
DeleteThe sensorimotor interactions of T3 with referents of its words (usually categories) are what grounds its symbols.
ReplyDelete" If the TT-passing computational states are really the cognitive
ReplyDeletestates, they will be the right cognitive states in every implementation...Searle’s very simple point is that he could do this all without understanding a single word of Chinese. And since Searle himself is the entire computational system, there is no place else the understanding could be. So it’s not there."
Does this mean that understanding doesn't fit into cognition? I don't think understanding can be explained by answering the easy question of cognitive science (why and how do we do what we do), so the hard question (why and how we feel the way we do) is what we're dealing with once we realise that understanding is a process outside of our cognitive state, shown with Searle's Chinese Room Argument and Harnad's additions to it.
If understanding is beyond the computational system, this is when Harnad's symbol grounding problem comes into play. Symbolic capacities, which I understand to be understanding, is grounded in sensorimotor capacities, because sensorimotor is beyond the cognitive state and it takes our senses into consideration to explain passing the TT. But, to me it seems that sensorimotor capacities is a blanket term like cognitive state. We don't really know what goes into both, we're just working with theories and ideas. Although it adds some clarity to the discussion as a whole, I feel like it has the same number of questions connected to it as other theories about 'understanding'.
I do not think that Searle’s conclusion’s mean that understanding does not fit into cognition; I think quite the opposite. Searle is saying that understanding is such an integral part of cognition that if computation does not yield understanding, then cognition can not be seen as computational.
DeleteI also think that symbolic capacity is not the same as understanding; symbol manipulation is syntactic, yet becomes useful through grounding. This is Harnad’s major addendum to the Turing Test (TT): the necessity of sensorimotor capacity for symbol grounding to truly pass a stronger TT (T3).
Also, I think we do have a wealth of knowledge regarding sensorimotor capacities. The models we have for our visual, auditory, tactile, gustatory and olfactory senses seem pretty comprehensive, at least until interpretation of stimuli comes into play (which enters the cognitive domain). While our knowledge about these systems is most likely growing and changing, I do not think they have the same level of mystery as consciousness and understanding.
As Professor Harnad points out in this paper (Cohabitation: Computation at 70, Cognition at 20), Searle appears to have gone too far in his concluding that computation cannot be any part of the explanation of cognition. There does appear to be room for an explanation of cognition that is a combination of both the computational understanding, along with a dynamically realized machine that has sensorimotor capabilities. But, this new machine would be what Harnad calls T3. It would, in fact, be able to pass Searle’s Chinese T2 test because, by its nature, T3s can do all the things that a T2 can do *and more* (the “more” being this sensorimotor capacity equivalent to human sensorimotor capacity).
ReplyDeleteThus, the question of whether it is possible to create a T2 machine that *understands* (understanding being the indication of the machine’s cognizing) appears to have an answer: no. We have to get into the realm of T3s before we can have a T2 that understands.
"So the TT-passing program is no more cognitive than any other symbol system in logic, mathematics or computer science. It is just a bunch of symbols that are systematically interpretable by us -- by users with minds. It has again begged the question of how the mind actually does what it does -- or rather, it has failed to answer it." (Harnad 2009)
ReplyDeleteAs Searle points out with the Chinese Room Argument, a system that can merely compute is not a sufficient for answering the hard problem of cognition. A system that passes T1 and is only indistinguishable on a syntactic level from humans does not truly understand the meaning of the symbols it is manipulating, because the symbols are not grounded to things in the external world.
Harnad proposes an alternative, and that "… the full robotic version of the TT, in which the symbolic capacities are grounded in sensorimotor capacities and the robot itself (Pylyshyn 1987) can mediate the connection, directly and autonomously, between its internal symbols and the external things its symbols are interpretable as being about, without the need for mediation by the minds of external interpreters."
If we want to reverse engineer cognition in order to understand it, a T2 passing system is not enough. Harnad argues that T3 system (one with sensorimotor capacities), is sufficient.
This comment has been removed by the author.
ReplyDeleteHi Nathan,
ReplyDeleteI agree that introspection does offer some unique insights into cognition that we take for granted. Nevertheless, I think you are nit-picking the details and not looking at the bigger picture Harnad paints. Cognition cannot be deciphered using introspection. Of course we can introspect some aspects of how we identified our 3rd grade teacher, but so much along that path is invisible to us - "we are unaware of our cognitive blind spots".
"So Searle simply proposes to conduct the TT in Chinese (which he doesn’t understand) and he proposes that he himself should become the implementing hardware himself, by memorizing all the symbol manipulation rules and executing them himself, on all the email inputs, generating all the email outputs. Searle’s very simple point is that he could do this all without understanding a single word of Chinese."
ReplyDeleteReading about Searle's Chinese Room Argument, I think there is a difference between what humans consider to be "understanding" versus responding to input. My interpretation of the reading was that understanding is uniquely human while symbol manipulation is computation and can be replicated by a machine. Therefore, as is stated in the readings, cognition is not just computation. The machine could compute Chinese without having a human-like understanding of it.
Searle, having no understanding of the Chinese language, must have had to get the symbol manipulation rules somewhere (ie: a programmer). The programmer of this Chinese Room had to understand Chinese to be able to provide Searle with all the appropriate symbol manipulation rules in the first place.
A question came to mind, though. Since language is an ever-changing thing, wouldn’t Searle's Chinese Room one day fail the Turing Test if Searle was told to never break the symbol manipulation rules given to him from the start? Certain words could have new meanings and certain new words may be added to the Chinese language. A machine would have to learn and adapt accordingly with the times, and Searle would have to be updated with new symbol manipulation rules.
Re: “But if symbols have meanings, yet their meanings are not in the symbol system itself, what is the connection between the symbols and what they mean?” (Harnad, 2005)
ReplyDeleteRetrospective thinking about this reading makes me have further questions about the hybrid sensorimotor/symbolic robot. We have been talking about many things that give rise to cognition, but let me, for instance, keep it down to three main elements which I think creates the cognition: computation, symbol grounding, feeling.
Let’s start with the one that I am not certain about: feeling. We talked that “it feels” like to do a thing; there is a feeling attached to it. However, I can’t find adequate opposition to the idea that the feelings do exist, but they are not a part of cognition. The feeling might just be a side-effect of computation + groundings. So, I will switch to computation and groundings since I cannot further investigate the feeling part.
I think that computation can be used to obtain cognition. I am not saying that computation is necessary for human cognition; for instance, maybe human cognition has different components that we cannot even observe. BUT, I think that “a cognition” can be in part created by computations. Creating a type of cognition is related to under determinism as well. Obviously, this way of creating cognition would not be the only way.
To continue with the idea of creating a cognition with computation, I think that we need one more element: groundings. But then the hybrid system becomes a bit unclear. If the feeling is a side effect and we do not need reverse-engineer the feelings, the cognition then can be created by computation and groundings. And regarding groundings, we talked about the example of how to ground 1000 words so that we are later able to ground other words even if we don’t have sensory-motor input. The groundings for 1000 words are the learning of the features of those 1000 words. As far as my understanding goes, a similar kind of feature extraction can be achieved by neural networks with the use of sensory modalities (like computational vision, etc.). If we were to obtain a similar type of feature extraction for each of the 1000 words, then does it mean that we can just combine these groundings with computation, and as a result we have the cognition? (I assumed that feelings do exist, but they are just side-effects)
Obviously, I am not suggesting that this is a direct method to achieve cognition. But if feelings are irrelevant to creating cognition, what would stop us from creating cognition only from computation and groundings? And what would it mean concerning this hybrid system? Is it still hybrid? Computation is computation for sure, but what about neural network way of creating groundings? Would we call it computation? Or not, since it also includes sensory modalities?
In, the reading, “Cohabitation: Computation at 70, Cognition at 20”, Harnad offers a good definition of cognitive science: equipment and processes we need in our heads in order to
ReplyDeletebe capable of being shaped by our reward histories into doing what we do. I also like the fact that “successful cognitive theory must make it explicit, so it can be
tested (computationally) to see whether it works”. However, cognitive science as it is today seems to suffer from the problem of underdetermination. There are myriad explanations offered for a single observed cognitive phenomenon, based on different models in physics or computer science. It is hard to understand which of these models actually bear more resemblance to the mechanisms by which the brain functions (actuality). Many models seem to rather accurately predict a certain behaviour. In this scenario, how do we know which one is closest to how brains function? Is it theoretically possible to know with high accuracy? Certainly various neuro-imaging techniques, currently existing and planned for the future, might provide further data to zoom in on the correct explanations. But will we ever be fully accurate remains to be seen.
Additionally, while I agree that the division between cognitive and subcognitive as Zenon argues is rather arbitrary, it makes me wonder whether it is possible and is close to actual neural implementation, to argue that cognitive processes happen across different layers. For instance, in studies comparing neural nets and fMRI signals, it has been shown that the bottom layers seem to encode more basic features of the stimuli while higher layers correlate more with higher order cognitive processes (such as categorisation). This makes me wonder whether there is a possible hierarchy of neuronal networks in the brain underlying all cognition in which the complexity of the cognitive process increases as we move up the layers. The encoding of sensory motor information could be the first layer, and the cognitive process directly related to the behavioral outcome could be the highest layer.
Regarding the arguments by Zenon that mental states are computational states which are hardware independent – this doesn’t seem to have a one-to-one correspondence with human cognition, as each individual might have grounded objects or words within different sensorimotor inputs. And as such the neuronal firing associated with a “cat” for me and you might be different, and hence the mental state too. In that sense, human cognition might not be hardware independent? Or another way to ask the same question is, if I were to somehow extract my mental states (the state of neuronal firing at each junction) associated with the word “cat” and reproduce it in another person’s brain, it might not need to the recognition of “cat” in the new person?
“Searle thought the culprit was not only the insufficiency of computation, but the insufficiency of the Turing Test itself; he thought the only way out was to abandon both and turn instead to studying the dynamics of the brain. I think Searle went too far.”
ReplyDeleteFor this part, I tend to agree with Searle’s idea that we shouldn’t only focus trying to fit everything into the Turing Test and into computation. We should go where the problem leads us and not try to fit the problem into the mold, we think is most appropriate for it. It might come up as a strong opinion, but I think that focusing only on computation is being narrow minded and not seeing the bigger picture. Some words mean more than others, often because they are emotionally charged. For example, hearing the name X for the first time, then person with the name X becomes a friend and it means something else, then we stop being friends for some reason, then the meaning of X becomes something else. Meanings change and, in this case, they depend on feelings. Can feelings be explained by computations? Also, feelings come as a continuum so can the degree of a feeling be computed? If we don’t take all the factors into account, we will never find the answer. I don’t believe that one thing can explain everything. I don’t believe in a universal answer for everything. I don’t believe in a mathematical formula for everything. Trying to explain a complicated phenomenon such as the human brain with only one concept is failing to realize the complexity of it.
Ultimately what can be garnered from the discussion on Searle is that in some ways the Turing Test can be fooled and still be passed.
ReplyDelete“Searle thought the culprit was not only the insufficiency of computation, but the insufficiency of the Turing Test itself”
So, when Searle is able to embody all the rules and algorithms of the Chinese robot, and essentially become it functionally, he is theoretically able to respond in exactly the same way as the robot. Therefore, he can fool the Turing Test by passing it but not actually understanding anything. This speaks to the conditions of cognition, in that understanding is a necessary requirement for it, and being indistinguishable to a human, at least in written conversation, is not sufficient for cognition either. At this point it seems like this level of the Test (T2) is indeed insufficient.
“The only way to do this, in my view, is if cognitive science hunkers down and sets its mind and methods on scaling up to the Turing Test, for all of our behavioral capacities. Not just the email version of the TT, based on computation alone, which has been shown to be insufficient by Searle, but the full robotic version of the TT, in which the symbolic capacities are grounded in sensorimotor capacities and the robot itself (Pylyshyn 1987) can mediate the connection, directly and autonomously, between its internal symbols and the external things its symbols are interpretable as being about, without the need for mediation by the minds of external interpreters.”
ReplyDeleteIt feels like something to understand and we don’t know how we do it, however, could we replicate the process of understanding in the future with T3 hybrid sensorimotor robots? By combining what we know with minimum grounding sets, create neural nets that associate minimum grounding sets with their referents and grow from there. Harnard goes into it in his symbol grounding problem article when he explains that “Symbols are manipulated not only on the basis of the arbitrary shape of their tokens, but also on the basis of the decidedly nonarbitrary "shape" of the iconic and categorical representations connected to the grounded elementary symbols out of which the higher-order symbols are composed”. Moreover, by combining what we know about mirror neurons for example and the fact that they allow replication, imitation and learning, a sensorimotor hybrid robot could interact with the world and learn from experience by making new associations with what it already stored. We would miss the feeling of course, but again feeling is a part of the hard problem and maybe grounding is enough for a T3 sensorimotor hybrid robot.
“In essence, Hebb's point was about question-begging: behaviorism was begging the question of "how?" How do we have the behavioral capacity that we have? What makes us able to do what we can do? The answer to this question has to be cognitive; it has to look into the black box and explain how it works. But not necessarily in the physiological sense. Skinner was right about that. Only in the functional, cause/effect sense.”
ReplyDelete“For this, Hebb had a companion example to his 7/2 task. He would ask the intro class: "What was the name of your 3rd grade school-teacher?" When we triumphantly produced our respective answers, he would ask "How did you do it? How did you find the name?" He might have asked the same of addition and subtraction: "How is it that your head, having been told Ôsubtract', manages to come up with '5' in response to '7'/'2'? "
And what if how we can do all of this is also part of the hard problem of consciousness? It feels like something to remember the name of my 3rd grade school-teacher, I can do it, yet I can’t explain how I did it. In parallel to the fact that we can’t explain the hard problem nor consciousness, what if trying to explain “how” we do what we do enters the same range of questions for which we may never have the answer because they are not accessible for the minds of human beings?
On “This sorting and labelling of functional “modules” is more an exercise in semantics or taxonomy than anything substantive functionally speaking”. I agree with the point made that the basics should be covered before the details are. While organising functions is important, the basics allow a structure on which theories can be built. Setting firm grounding is important.
ReplyDeleteThe Turing test is supposed to essentially be a system that can do everything a human can do, undifferentiated from us in an email exchange. If a computer were able to be as a human is, this would indeed provide a potential framework for cognition. However, as we cannot fully understand what human cognition consists of (what exactly triggers thoughts and feelings and how they follow through) it would be difficult to gage whether or not the Turing test could really mirror it in the exact same way.
I agree that beginning by robotically recreating our behavioural capacities is important, understanding the basic processes and their step by step sequences. There is enough scientific evidence on sensorimotor capacities to start there and focusing research where we already have a base. As mentioned in the paper, the robotic version could then make its own connections, as humans do. It is known that there is an additional element past simple computation in human cognition (computationalism is false), accounting for the human aspects like feelings. Human aspects being the things implied when talking about the difference between the mind and the brain, the yet un-explicitly-known components contributing to the human experience. The provision of dynamic components to the Turing test robot would allow it to more closely mimic the human brain. This ultimate goal requires beginning from the basics and understanding them in detail first thought.