or: Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1), 335-346.
or: https://en.wikipedia.org/wiki/Symbol_grounding
The Symbol Grounding Problem is related to the problem of how words get their meanings, and of what meanings are. The problem of meaning is in turn related to the problem of consciousness, or how it is that mental states are meaningful.
If you can't think of anything to skywrite, this might give you some ideas:In M. de Vega (Ed.), Symbols and Embodiment: Debates on Meaning and Cognition. Oxford University Press.
Taddeo, M., & Floridi, L. (2005). Solving the symbol grounding problem: a critical review of fifteen years of research. Journal of Experimental & Theoretical Artificial Intelligence, 17(4), 419-445.
Steels, L. (2008) The Symbol Grounding Problem Has Been Solved. So What's Next?
Barsalou, L. W. (2010). Grounded cognition: past, present, and future. Topics in Cognitive Science, 2(4), 716-724.
Bringsjord, S. (2014) The Symbol Grounding Problem... Remains Unsolved. Journal of Experimental & Theoretical Artificial Intelligence (in press)
Re: Harnard Scholarpedia
ReplyDelete“To be grounded, the symbol system would have to be augmented with nonsymbolic, sensorimotor capacities -- the capacity to interact autonomously with that world of objects, events, actions, properties and states that its symbols are systematically interpretable (by us) as referring to. It would have to be able to pick out the referents of its symbols, and its sensorimotor interactions with the world would have to fit coherently with the symbols' interpretations. The symbols, in other words, need to be connected directly to (i.e., grounded in) their referents; the connection must not be dependent only on the connections made by the brains of external interpreters like us. Just the symbol system alone, without this capacity for direct grounding, is not a viable candidate for being whatever it is that is really going on in our brains when we think meaningful thoughts (Cangelosi & Harnad 2001).”
Harnard’s explanation of the symbol grounding problem offered great clarification as to why T2 alone would not pass the Turing Test -- T3, which has the ability to pick out referents as they are consistent to the sensorimotor interactions with the world, is necessary to pass the Turing Test. A T3 can ground a symbol system, due its sensorimotor capacities, and demonstrating *understanding*, and thus consciousness. Since symbol grounding is necessarily linked to consciousness, consciousness is consequently central to cognition. Is symbol grounding, then, the closest we can get to skimming the hard problem? We know symbol grounding is necessary in cognition -- without it, we would not be able to generate meaning. But what is it that the brain is doing to generate meaning? And does it matter, if we are concerned with behavioral equivalence (“doing” the same thing?).
This all seems quite cyclical for the sake of discussion in our class: if consciousness is central to cognition, and we cannot study consciousness because within it exists “weasel words” like *feeling* and *understanding,* how can we speculate on cognition? And even further, how can we distinguish the hard problem from the easy problem, one necessarily requires the other?
Professor Harnad often stated in class that he believed that T2 could only be passed if T3 could also be passed. Thus, a robot would be necessary to pass both T2 and T3. After reading about the symbol grounding problem, I feel as though I have a better idea of why Professor Harnad would make this proposition. The Turing Test does not simply consist of asking and answering questions, it consists of interacting in a way that is indistinguishable from a human for a lifetime. T2 only requires verbal equivalence, but verbal capacities are quite powerful for humans. An important part of human verbal capacity is to understand what is being said (or communicated). An arguably large part of cognition is understanding, which requires meaning. In order to achieve meaning, it is necessary that sensorimotor grounding occurs (i.e. it is necessary to have the ability to pick out the correct referents for symbols). Only a robot (T3) with sensorimotor capacities would be capable of creating the link between a referent and it’s meaning by interacting with the world. T3 has the capacity to interact with the world as T2 does not. Meaning is not in play when looking at T2. T2 would be able to identify objects… but identifying objects is not the same as recognizing objects (which only T3 would be able to do). You cannot recognize an item/object without having sensorimotor capacities and interactions. In order to achieve verbal equivalence with humans, the meaning of symbols and the ability to pick out referents would be necessary.
DeleteI think Gili brings up a very interesting question by asking if symbol grounding is the closest we will get to skimming the hard problem. I was thinking the same thing while completing the reading. Knowing that grounding is a necessity for meaning and that understanding meaning makes us feel certain ways seems helpful when considering the hard problem of cognitive science…. But knowing that grounding is dynamic and implementation-dependent seems to prevent us from using this information to solve the hard problem.
Gili, Not so fast! Grounding is only part of meaning. For "the cat is on the mat" to mean that the cat is on the mat (to the one who is saying or thinking it, rather than just to the one listening to it), it is not enough that the speaker should be able to detect and interact with cats and mats and cats on mats in the world (grounding). It also feels like something to say and mean that the cat is on the mat:
DeleteMeaning = Grounding + Feeling.
(And here "meaning" does indeed mean "intentionality" in the sense that "that's what I mean, that's what I have in mind.")
If you have nothing in mind (i.e., you don't feel a thing), you're just a chatbot saying words, no matter how much sense they make to a listener with a mind (i.e., a listener who doesn't just process the words but understands what they mean: understanding, too = grounding + what it feels-like to understand the words: both of which Searle lacked, in the Chinese Room).
(If you could show how and why "symbol grounding is necessarily linked to consciousness (i.e., feeling)" you would have solved the hard problem.)
Feeling is not a weasel-word. The weasel-words are all the synonyms for consciousness (including "consciousness"!) except feeling:
Consciousness, awareness, subjectivity, intentionality, qualia, mind, 1st-person states, mental states, phenomenal states, subjective states, intentional states, representational states, etc. etc. etc. all the way to spiritual states and the soul.
That's all just felt states; feeling. The only innocent synonym is sentience (which literally means the capacity to feel (anything)).
Amanda, the T3 robot does not link the referent to its meaning; it links the symbol (word) to its referent. Some say that the meaning of "meaning" is the internal mechanism that links words to their trferents. But that is "only" grounding.
Meaning = Grounding + Feeling.
But that's not the solution to the Hard Problem. It's just symptom of it!
(For what it's worth, "Stevan Says" (1) only T3 can pass T2 and (2) T3 feels, but the other-minds problem prevents us from being able to verify it, and the hard-problem prevents us from being able to explain how or why it's true.)
Just out of curiosity, (and maybe it was stated in class at some point but I'm forgetting), why does "Steven Say" T3 feels?
DeleteCould you please elaborate more on what you mean when you say that if you have nothing in mind you don't feel a thing? I find that this is hard to believe because I think there is a feeling associated with having nothing in mind. For example, after a very stressful week full of assignments, midterms and tests there is nothing better than the feeling of having nothing in mind. Another example of this could be meditation. Often you are told to “stop thinking”. In this case, the goal of the meditative exercise is to rid your mind of any thoughts. Perhaps I am getting stuck on the colloquial use of the phrase “nothing in mind” so if you could clarify this that would be great.
DeleteAnna, "Stevan Says" T3 feels (1) because I accept Turing's "Bambi" criterion -- "If you cannot tell the difference between Gabe and a person who feels, don't" -- and (2) I don't think the Blind Watchmaker (Darwinian Evolution) can tell either. So I think feeling somehow comes with the T3 territory (but "how" is the hard problem); hence (3) I don't believe in T3 Zombies. (But that's all just "Stevan Says...")
DeleteAkhila, if it feels like something to have "nothing in mind," you have something in mind. "Have in mind" is synonymous with "feels like something." ("Mind" is a weasel-word.)
In response to Professor Harnad:
DeleteIf the other-minds problem prevents us from being able to verify that T3 feels, how can we ever possibly know that Gabe (T3) is feeling? Is there even an answer to this question? I feel as though we could argue about whether or not Gabe can feel and whether we feel that it is okay to kick him forever because there is no definite way of knowing if Gabe feels.
Hi! Your point is interesting, but I believe that it is specifically because there is no way of telling that we say he does feel. The other-minds problem does not technically only apply to T2-T3-T4 robots, but to every single intelligent organism we encounter. When I meet another human, I assume that this person feels a certain way because I know I can feel and I see that this person behaves in a way that is so close from the way I would behave in this situation, and that she looks so similar to the way I look, etc. I have no choice but to assume she feels, just like I feel. So, because the real answer is "I don't know whether anyone other than me feels.", I don't think it would ever be okay to kick Gabe, because he looks and behaves just like I do; I have no choice but to assume he also feels.
DeleteHi Amanda: Marianne has done my job!
DeleteRe: Harnad 2003
Delete“Here is where the problem of consciousness rears its head. For there would be no connection at all between scratches on paper and any intended referents if there were no minds mediating those intentions, via their own internal means of picking out those intended referents.”
It is now easy to grasp why a T2 robot with only linguistic capacities would not be considered a conscious being. Computationalists believe that a Turing machine can pass the Turing Test at T2. However, Searle exemplified in the Chinese Room thought experiment that understanding/intentionality is required in order for a robot to be able to do everything that a person can do. Harnad takes this argument further and proposes the Symbol Grounding Problem, positing that words get their meanings from the process of grounding the symbol/word to a sensorimotor referent. The meaning that we can extract from symbols comes directly from this grounding process coupled with the conscious awareness of having said something and meant it.
I find a lot of this incredibly difficult to articulate, because of the recursive nature of the words we use to explain this, in addition to the somewhat abstract concept of consciousness. I definitely feel like the SGP could be right (as in, could be a valid explanation for how we assign meaning to symbols), but I don’t think it gives more insight into the actual causal mechanism for much of non-linguistic cognition (although it could be argued that much of human cognition is linguistic). The SGP takes the more relevant approach of attempting to explain “how and why” rather than “when and where” (as with mirror neurons and neuroimaging) but I think that a full and accurate picture of cognition must include both – how we can do what we do, and what physical system makes these processes possible.
It's one thing to formulate the SGP, another to solve it. For that we have to solve the "Easy Problem," for which (since T3 Gabe is a fiction) there is as yet no solution in sight. But if Gabe really were an MIT T3, his causal mechanism would be the solution.
DeleteI think the section on “Iconic and categorical representations” can help us explain why associative learning is not the same as categorization. Associative learning seems to be the formation of iconic representations, such that they are the “analogs of the many shapes that horses cast on our retinas”. If we use a friend in class as an example (instead of a horse), each time we see said friend in class, they are wearing a different set of clothes as the last time we saw them. If this friend sits at a different seat each class (and/or if we sit at different seats each class), they form a different “shape” on our retina depending on what they are wearing and where they sit relative to us. However, despite these different shapes that this friend casts on our retinas, we still know that they are them. That is because we can assess what is consistent or different in this friend each time we see them. We form “analog copies of the sensory projection, preserving [their] ‘shape’ faithfully”. In other words, this this how we can associate a name with our friend.
ReplyDeleteBut to categorize our friend as a human (under the assumption that they are not an MIT robot!), we form a categorical representation of humans by utilizing icons that have “been selectively filtered to preserve only some of the features of the shape of the sensory projection: those that reliably distinguish members from nonmembers of a category.” I think this clarifies what we referred to as “feature extraction” in class when we discussed how we know that something is categorized as a table. In the case of our friend in class, we have disregarded the disparities between retinal projections of all humans and only considered shared features of their retinal projections to classify our friend as a member of the human category and not a member of the dog category, for example.
Iconic copies will not enable you to recognize a category, even when the category is all the possible variants on the shape of one individual. Not even with the help of "association." (Association of what with what?) One icon can be associated with another icon; or you can memorize all the icons (sensory projections) of an individual (f you have enough memory) and associate them all with one another. But then what happens the next day, when you encounter an icon of the same individual, but one you have never seen before, so it's not in your memorized set: memorize another one, and "associate" it with all the others?
DeleteNeither icons nor association will do the trick, noteven for recognizing individuals. You need a mechanism that will abstract the features shared by all the icons of the same named individual (or all the members of the same named category. That's what neural nets can do -- and it's not by "association," not even in "unsupervised learning," as we'll learn next week.
I urge everyone to read the (very) short story by Borges: Funes the Memorious. It will come up next week.
RE: Professor Harnad’s (1990) paper:
ReplyDeleteConsciousness might explain a larger part of meaning than computation alone does. I can agree with the fact that words are symbols that are arbitrarily assigned to objects that they refer to. In other words, “cat” is a shape that is essentially meaningless in the literal sense, such that “looking up the meaning of [cat] in a (unilingual) dictionary of a language that I do not understand would lead to an infinite regress”. But still, it feels like something to understand what a cat is referring to (in your own language, of course). When cat owners mention their cat, there is no denying that we know what a cat is. It feels like something to know what a cat is and what it represents. So, even though I agree that “cat” is meaningless in the literal sense, I don’t think it is meaningless in a conscious sense (so long as you understand the language of the symbol representation). Perhaps by looking more into consciousness, we can develop a better understanding of how symbols and symbol manipulations get their meaning. But since there is the other minds problem (since others might understand the word “cat” differently), trying to understand how symbols get their meanings seems unsolvable.
I agree with you that "cat" is not meaningless in the 'conscious sense' but I don't think anyone would argue against that case. The symbol 'cat' is meaningless, yes, (the letters and shapes in this symbol are arbitrary and do not mean anything on their own) however once their has been meaning attributed to the symbol 'cat' (through detection and categorization by means of sensorimotor experience) the symbol gains meaning.
DeleteI also wanted to address your point about the other minds problem because this is something I have been thinking about a lot as well when reading about symbol grounding and meaning attribution. It does seem difficult to ever understand how words (generally) get their meaning if every word's meaning may differ on a case by case basis. If people have different meanings for the same word (different things associated with them, a different feeling about that word etc.) then perhaps it is possible to say that people gain those meanings differently as well. If this is true, I wonder if there is any way to ever have a general understanding of how words get their meanings.
Regarding the same piece you’re talking about (Harnad, The Symbol Grounding Problem, 1990):
DeleteTo Erica: That’s the thing though, Harnad does propose a model (in the piece you’re referring to) for how symbols get their meaning, although you’re right that the actual process of it involves some pieces that are unsolvable).
To Flora: I don’t think that a different end point (ex. having a different thing associated with a word) actually necessarily says anything about a medium of acquisition. I recognize that this is a rather silly example, so bear with me. In Dirty Dancing, the main character is nicknamed “Baby” so much that I actually can’t recall her real name. Imagine being her and learning throughout your whole life that your name has essentially become “Baby”. You’d maybe even get more reinforcement for the meaning of “you” as a person (an odd way to put it if you know what I mean) than the more common meaning of “baby” which is a literal infant. Now me and that character probably have very different meanings that come to mind first when we think of the word “baby”, and likely different parts of our brains do light up if someone were to put us in an fMRI scanner and read the word to us. I don’t think that means that process of symbol grounding is any different. If you think of the model presented by Harnad, it works the same way (although I will agree that the fact that “Baby” in one of these cases is a proper name does complicate things). Maybe you are right, and it’s different (maybe, for example, in addition to learning that “Baby” refers to her, she also encodes that into many different types of memories), but the core process or grounding the symbol “baby” to a meaning and not just a referent is still the same, at least intuitively (right now I don’t believe we have any concrete evidence for this).
"Consciousness" (i.e., feeling) explains nothing: it is something waiting (probably in vain) for an explanation. Finding out how and why word meaning is grounded is, after all, an easier problem than finding out how and why it is felt!
DeleteThe other-minds problem is even more about whether others feel than about what others feel.
In week 8 we'll find out that, yes, the same word (and the category it refers to) can be grounded in (slightly) different ways in the history and experience of different people. Categories and meanings (and definitions) are approximate, not exhaustive or "exact." But when there is a misunderstanding, or a set of features that has never let you down fails for a new case, the approximation can always be tightened, either verbally, by adding more words (features) to the definition or description of the referent, or even by going back to direct sensorimotor learning of the missing features.
When in week 8 we ask "What is the smallest number of words that can define all the other words in the dictionary?" it looks like it's around a thousand -- but there are many sets of 1000 words that can do the job. Your vocabulary might be grounded in one of them and mine in another. It's like an N-dimensional vector space: You need N (linearly independent) points, and then all the rest of the points can be expressed as multiples of them. It cannot be less than N, and they do have to be linearly independent (i.e., not multiples of one another). But there is an infinite number of sets of N linearly independent points in any N-dimensional vector space.
Yes, the case of proper names of individuals is a very special case of the more general one of common names of categories. And yes, words may have different connotations (associations) for different users. But grounding just concerns all users' being able to recognize and agree on the name of the referent. Also, grounding concerns meaning for the user of the name, not for the bearer of the name (i.e., the referrer, not the referent)!
I realize that this comment is probably somewhat covered in week 8, but it's been nagging at me since the 1000 word concept was mentioned a few weeks ago.
DeleteAssuming two consciousnesses have been nurtured on the same 1000 base concepts, in the exact same way, can we be sure that they will perceive things the same way? Take the example of a colour: we can quantify a colour in a specific way (via the wavelengths of light) and that is presumably universal to perception (different people will perceive the number on the sensor as the same value). We've agreed that the spectrum of light living between 625–740nm is known as "red" and presumably everyone can notice that outside of those bounds it appears as a different colour. But despite being trained on the same 1000 concepts to know what "red" is (i.e. some apples are red, blood is red, etc...) the perception of what "red" is for others can't be communicated and verified.
Kyle, I don't know what you mean by "two consciousnesses have been nurtured". I guess you just mean two people (or T3s) have learned 1000 words from experience (from among the 1000-word smallest sets that can define all other words). To learn each category (which is what I guess you mean by "concept"), they need to sample members and non-members until their internal feature-learners detect features that distinguish the members from the non-members. This sample will differ between the two of them, and so may the features. And that means the categories will not look completely identical to both of them. Yet they will both be able to recognize the referent category (to a close enough approximation). And, yes, the other-minds problem still stands: there's no way to compare how similar looking at members of the same category feels to them. (More next week.)
DeleteRegarding the 1000 words, if the words have different grounding, would they also have different referents? If all of the words in the dictionary could be explained by 1000 words, I would assume that they could be used to to refer to all of the same referents. But I'm just curious about what it would mean for the 1000 "core" words?
DeleteIf the "same" word has a different referent, then it is either a homonym (sounds the same, but not the same meaning, such as in "mole" the animal and "mole" the birthmark) or polysemous (has multiple meanings, perhaps related, but different, as in "sharp" knife and "sharp" intellect).
DeleteThat said, all categories (except formal mathematical or logical ones) are underdetermined (different features may be work to recognize the same referent category, as in "weak equivalnce") as well as approximate (the features you currently use to (successfully) recognize a referent category may not be enough: you may eventually encounter a member for which you have to revise the features and tighten the approximation in order to categorize it correctly).
Two people referring to the same referent category (or person!) based directly on different sensorimotor features could be talking past one another until they discover it and revise their sensorimotor grounding to match. The same is true if their categories are indirectly grounded through a verbal definition or description, if their definitions are different and they were talking about a case where that matters. (This is related to the vocabulary-learning phenomenon of over- and underextension.)
"But whether its symbols would have meaning rather than just grounding is something that even the robotic Turing Test -- hence cognitive science itself -- cannot determine, or explain."
ReplyDeleteThis is the first quote that caught my attention. I thought at the beginning of the reading that you were referring to the hard problem. Although you can ground symbols in external referents, you can never truly tell if the machine can feel it. However, the more I read your article, I am beginning to question if that is what you truly mean. You take on this journey where you begin to question the difference between referent and meaning, and how 1 referent could have multiple meanings. I am a referent, but I can be stated in the context of being a McGill student, or the daughter of my parents. You then go to state that meaning, in the wide sense, is the referent along with its path to the referent. In this sense, the meaning is all that is required to bring the outer object to the inner word, as well as all the hard drive (head) and the software (process). Although I am aware that software is not enough to do the grounding itself (an argument against computationalism), I do believe that if enough words are grounded, a software is able to link the rest of the words together. So in the end, to have meaning, like we humans do, computers needs something in addition to symbol manipulation. It would require a bulk of these symbols to be grounded in external objects. You end the paper with a final question that is yet to be solved. We don't know if symbol grounding is enough to solve Searle's dilemma but it does seem like it's one step closer. Symbol grounding is yet to be deemed sufficient but it does seem to be necessary.
In the robotics categorization, symbol grounding is found to be a necessary condition, but not necessarily a sufficient one. Symbol grounding is required for there to be meaning, but just because there is symbol grounding, that does not mean that there is instantly meaning. This is why the example of the Turing Test passing "Zombie" could live among people but not have any "meaning meanings."
DeleteIn the Scholarpedia article, Professor Harnad argues that groundedness is a necessary condition for meaning, not a sufficient one, as Marine pointed out. Moreover, Professor Harnad suggests a second property potentially needed for meaning: consciousness. I don’t believe grounding “enough” words would not be sufficient to produce meaning because part of meaning is understanding. This point harks back to Searle’s Chinese Room Argument that we studied the third lecture. I don’t believe there can be meaning without understanding, and there cannot be understanding without consciousness. Unfortunately in the current state of cognitive science, we cannot know how consciousness facilitates the production of meaning even if we postulate that it’s integral.
DeleteRemember not to confuse the "other-minds problem" (of determing whether and what others feel) with the "hard problem" (of reverse-engineering and explaining causally how and why people (and other organisms) feel).
DeleteThe "easy problem" is "just" to reverse-engineer and explaining causally how and why organisms can do what they can do, including recognizing individuals and categories (and "doing the right thing with the right kind of thing") and (in the case of humans only) describing them in words.
Categorizing requires finding the features that distinguish the members from the non-members, whether through direct grounding in sensorimotor experience or indirectly, through (grounded) verbal definitions and descriptions. (In priniciple, though probably not in practice, only 1000 category names need to be grounded directly in order to be able to get all the rest through boolean re-combinations of the names these directly grounded categories.
Yes, Gabe might be Zombie, but I don't believe it: Do you? And Turing' point is that there's no way to tell (except T4, but even there the same question arises as long as the robot is synthetic. If it's indistinguishable biologically too, then we are almost at the level of the normal underdetermination between multiple explanations that all explain and predict all the data. But in the last week of the course we'll learn (or be reminded) that in the case of feeling, there is more to it than that -- in fact everything...
Meaning = Grounding + Feeling
I don't really know if the Gabe is a Zombie. For him not to a zombie, it would mean that Gabe is going through the underdetermination path to reach the same conclusion that humans arrive at meaning. In terms of the other-minds problem and the hard problem. Is this sequential? You have to determine that the other minds problem is saying that others feel and then you can go on to the hard problem to try to explain why and how others feel.
DeleteAmy, I don't quite understand your question but Marianne's reply to Amanda, above, sounds like the right one.
DeleteIn this either/or question, the (still undefined) word "ungrounded" has implicitly relied on the difference between inert words on a page and consciously meaningful words in our heads. And Searle is reminding us that under these conditions (the Chinese TT), the words in his head would not be consciously meaningful, hence they would still be as ungrounded as the inert words on a page.
ReplyDeleteFrom Searle’s point of view, as he refutes computationalism, it will never be possible to make a robot that has this capacity to ground symbols as they will be missing that “processing” which is required to connect the inner word to the outer object. I agree with Searle that there will never be the possibility of creating a computer that has this ability to ground symbols to the object. Even with a T3 Turing Test passing computer that is created, the meaning will still be missing and they will be a “Zombie.” This relates to the second property of consciousness mentioned in the paper that these Turing Test passing computers cannot be said to have.
Regarding the 2003 summary piece assigned:
DeleteJust a point of clarification: I don’t think that it’s that the core of the argument is that it will never be possible to make a robot that has the capacity to ground symbols, it’s that it will never be possible to make a robot that has the capacity to ground symbols based purely on computational abilities (i.e. computationalism is false).
You’re right, even a T3 Turing Test passing computer would still be a zombie. One thing mentioned in the piece though is that the second property that allows us to symbol ground (consciousness) is something cognitive scientists will never understand. I don’t know that I entirely believe that, although I buy that it’s probably a while away. Regardless, though, Turing is wrong – as the article states.
Searle’s argument does not lead to the conclusion that we can never make a robot that has symbol grounding abilities. As Veronique mentions, it follows from Searle’s argument that we can never have a robot do this based solely on computation because computation does not produce understanding. Understanding, or consciousness, is potentially a secondary property needed for understanding, as Professor Harnad argues in his Scholarpedia article.
DeleteIn class, we have discussed that a robot that passes T3 is indistinguishable from humans for sensorimotor and performance capacities. I don’t believe one can argue that a robot that passes T3 is a zombie, because of the problems of other minds. This is what Turing argues in his paper “Computing Machinery and Intelligence”. It is impossible to know for certain whether another human, let alone a T3 robot, can feel. I fail to see how Turing is wrong in regards to the Turing test and its implications on the symbol grounding problem.
A T3 Robot (Gabe) cannot be just a computer; hence Searle cannot use his periscope to penetrate the other-minds problem. If he just does it with the computer part of T3 (if any), then the System Reply (which was wrong for T2) would be right: Searle is not the whole of T3.
DeleteConsciousness (i.e., feeling) is not just a problem for computational explanations of cognition; it's a problem for any explanation of cognition.
The reason computation alone cannot pass T3 is not consciousness (feeling): it's grounding. Computers do not have any means to recognize and manipulate objects; only robots do.
Professor Harnad's article on the Symbol Grounding Problem asserts that T3 robots, with their sensorimotor capacities, are able to theoretically perform symbol grounding; however, as we had discussed in class, would they be able to derive meaning? The example of proper names as an example of the distinction between referents and meanings very clearly presents the issue: "(1) "Tony Blair," (2) "the UK's former prime minister," and (3) "Cheri Blair's husband" all have the same referent, but not the same meaning." I would like to expand on how meaning is closely correlated to feeling. Every one of those names, though they refer to the same entity, elicits different feelings. The meaning of symbols and words carries connotations and nuances that mere symbol grounding in referents does not.
ReplyDeleteFor instance, Gabe is evidently able to perform symbol grounding like every one of us; he reliably connects internal symbols to the external objects they refer to. However, would he be able to discern the emotive differences between "disabled" and crippled" or "conversational" and "jabbering"?
I wonder if, in addition to the categorization of instances and objects, is there categorization of different feelings? If so, I would assume that the categories of feelings would perhaps be less defined due to its nuances and subjectivity. There would be a spectrum of "rightness" in "doing the right thing with the right kind of thing". Having said that, though meaning itself is definitive (either you have feelings associated with referents or not), interpretations of *symbol* meanings would be subjective.
"For instance, Gabe is evidently able to perform symbol grounding like every one of us; he reliably connects internal symbols to the external objects they refer to. However, would he be able to discern the emotive differences between 'disabled' and 'crippled' or 'conversational' and "jabbering'?"
DeleteWhere does the line between connotation and feeling come in? "Crippled" is harsh and leaves a foul impression of whoever said it–that second part is the feeling associated with the word. I would say, though, that connotation can be known without understand the feeling associated with it. For example, when learning another language, one may be taught to avoid saying a certain phrase in a particular way because this way is offensive, but that way is more respectful. But the new language learner would only understand up until that point; they wouldn't feel the impact of the less respectful word like a native speaker would.
MIT me would still successfully pass the TT, which we agree on, but I just thought it would be interesting to expand on my point and to show that we also can relate to the MIT me in not feeling the meaning of words while understanding the tacit meaning.
The main problem of symbol grounding is the connection of the word with its referent in the world (naming the right category). Connotations and associations are important too, but easier than picking out the right referent.
DeleteAnd, yes, Gabe, there can be connotations and associations without feeling, hence without meaning, just T3 grounding: You could say "this reminds me of that," and it could be true -- but it doesn't feel like anything to you (even if you say it does!).
To add to what Vivian and Gabe have said, I’m interested in the ways meanings differ and are shared between agents.
DeleteTo Gabe: I’m not convinced that differences in connotations only refer to different felt states. It could be argued that the words Vivian used as example have different definitions. For example, perhaps “jabbering” could be defined as a specifically superficial colour of “conversational”, whereas the category “conversational” may be more general. Perhaps that the word “jabbering” feels like a hostile descriptor is explained by the fact that ‘superficiality’ is implicit in its grounding, and the concept ‘superficiality’ is something we’ve culturally agreed is negative.
I’m not sure I entirely understand Professor Harnad’s above proposition that meaning = grounding + feeling. For example, the statement “the cat is on the mat” is mildly pleasing to me, because I really like cats. I don’t expect this to be universal, as it would not be shared with someone who dislikes or is indifferent to cats. Is the ‘meaning’ professor Harnad refers to this very individualized meaning (everything that is felt when I hear or read ‘the cat is on the mat’), or is he talking about a more generalized, shared meaning (if that’s even possible), from words with shared cultural associations?
Further to this question, can the feeing elicited in an individual from an understood word be totally explained by one’s associations with that word (I’m imagining all of a person’s lived sensorimotor experience with the referents of the word + all instances of understanding of any statements one has heard/read regarding the word)? If so, theoretically, would the meaning be totally shared in the same mental state – that is, if all of one’s experiences with the word and all things associated with it, were perfectly duplicated in another individual, would the meaning be exactly the same?
What it feels like to understand or mean that "the cat is on the mat" has little to do with how you feel about cats, nor with the connotations of cats to you. It is about what it feels like to mean or understand that "the cat is on the mat." It feels different to mean and understand that "the mat is on the cat." And every different sentence you say (or hear or think) that you mean or understand feels different. It is not just that the words sound different nor that they have different "associations." It feels different to say and mean sentence "X" vs sentence "Y" as it does to see or hear one thing vs another, or to do one movement vs another.
DeleteI find your question on the categorization of different feelings quite interesting -- but perhaps it could be re-worded not as a separating feelings from things like "instances" and "objects," but rather *tied* to instances and objects. If categorization is doing the right thing with the right kind of thing, is there a hard-problem version of 'feeling' the right 'thing' with the right kind of...'feeling'? In other words, 'feelings' are linked to 'things,' because it 'feels' like something to 'do' -- it feels like something to categorize.
ReplyDeleteYou could continue on with a sort of meta-analysis of categorization, in which a 'feeling' is linked to feeling like you are categorizing; or, as I think you meant, an emotion/memory is linked to things/entities which are categorized. It is interesting to consider synonyms in general, as they relate to the symbol grounding problem. Why have synonyms at all, if they in fact, mean the same thing? I believe it is because they never truly mean the same thing, but rather they carry with them connotations and conventions that allow us to categorize them into the right kind of thing. Couldn't Gabe, like us, discern "disabled" and "crippled" just like we do: by choosing what is most appropriate (most 'right') based on learning, experience, and nuance in definition?
There is definitely something special about the category "feeling." Not that we don't all feel. But normally, to learn a category requires sampling both what's in it and what's not in it, in order to find the features that distinguish them.
DeleteBut none of us knows what it feels like to not feel: I'm not talking about feeling this vs that, where feeling this is not like feeling that, or feeling this is like not feeling that. I'm talking about not feeling anything at all.
(Think about it: we'll be getting back to it later in the course. Feeling is an "uncomplemented category." We know what's in it, but not what's not in it...)
You're right that there are no real synonyms. Othwerwise they're like two different proper names for the same person. But the differences between "synonyms" are not just connotations. You can define odd numbers as numbers that cannot be divided 2 or as even numbers plus 1 (as long as you know the meaning of all the words in the definitions). If a word's meaning is the means by which you pick out its referent, those two definitions differ in meaning, even thought their referent is the same.
Just improvising:
"Disabled" can be defined as "unable to do something because of something wrong with the body.
"Crippled" can be defined as "damage to the body."
That would not be just connotational difference but a definitional difference.
Of course pejorative connotations can be added to any word; all you need to is pronounce it in a mocking way...
Gabe can do all this, of course, whether or not he feels.
Something I am curious about, if we need to sample both what is in the category and what is not in the category, are we able to categorize feeling in general or are we only able to categorize emotions as we know what it feels like to not feel a particular emotion?
DeleteWe can learn and categorize what it feels like to feel X (e.g., hot, tired) and what it feels like to not feel X (not hot, not tired). We have positive and negative examples for both. But we have no negative examples for what it feels like to not-feel (at all, anything). Feeling what it feels-like to not-feel anything at all is self-contradictory. You can't feel what that's like because it doesn't feel like anything at all: you are not there!
Delete
DeleteHow does Gabe learn to categorize this? Can Gabe learn what it feels like to feel X or what it feels like not to feel X strictly using sensorimotor capacities and grounding? I would think that Gabe’s (T3) ability to categorize feelings would be attributed his sensorimotor capacities.
Also, is it really necessary to know what it feels like to not-feel? If we, as humans, do not know this, why would we consider this at all for T3? I believe this would not be very helpful… Although, would providing Gabe (T3) with negative examples of a category be necessary for Gabe to categorize things such as feelings?
UNCOMPLEMENTED CATEGORIES AND COULD THERE BE ZOMBIES? Part 1
DeleteAmanda: "How does Gabe learn to categorize [what it feels like to feel X]?"
Same way we do.
Amanda: "Can Gabe learn what it feels like to feel X or what it feels like not to feel X strictly using sensorimotor capacities and grounding?"
He would be able to do and say pretty much the same kinds of things we would do or say in the same situation, with the same history.
There is an interesting question, however, and a deep one, worth reflecting on (I certainly haven't figured out the answer): If Gabe were an insentient, Turing-indistinguishable T3 Zombie (i.e., doesn't feel a thing) why on earth would he talk about "feelings," exactly the same way we do?
One possible answer is a sort of behavioristic one: Like us, any grounded T3 must have sensors that detect external objects and their features (but without feeling a thing). He could also have detectors for internal states, such as the state of recognizing and categorizing an apple, or the state of being overheated. His words and descriptions of "what it feels like to see an apple" or to "feel hot" would (unlike ours) just be referring to those internal states, which would be detectable and nameable states, just like our felt states are, but they would not be felt, just detected, identified and named.
This may sound like idle sci-fi -- and maybe it is, if there can be no such thing as a T3 Zombie; i.e. if in order to have T3 capacity, any T3, whether a biologically evolved one like us, or an artificial T3 built in MIT like Gabe, would have to feel. Trouble is, that in order to know that, we'd have to solve the other-minds problem (which we cannot do); and in order to explain how and why a T3 could not be a Zombie, we'd have to solve the hard problem (which -- "Stevan Says" -- cannot be done). (This is also why "Stevan Says" (with no proof or evidence) that a T3 would feel; i.e., even if there can be a T3, there cannot be a T3 Zombie.)
UNCOMPLEMENTED CATEGORIES AND COULD THERE BE ZOMBIES? Part 2
DeleteAmanda: "Also, is it really necessary to know what it feels like to not-feel? If we, as humans, do not know this, why would we consider this at all for T3? I believe this would not be very helpful… Although, would providing Gabe (T3) with negative examples of a category be necessary for Gabe to categorize things such as feelings?"
You are right that in order to feel -- and to feel as if we know what it feels like to feel -- neither we nor Gabe need to know what it feels like to not-feel. After all, we do feel, we do know what it feels like to feel. And we don't feel that we would need to know what it feels like to not-feel in order to know what it feels like to feel. The lack of negative evidence "just" means that the category "what it feels like to feel" is somehow anomalous, because, unlike all of our other (learned and named) categories, it is "uncomplemented": it lacks "negative evidence," which is normally needed in order to learn the features that distinguish the members from the non-members. But the category "what it feels like to feel" has no non-members. Negative evidence does not exist; it would be self-contradictory.
So there's something odd about the category "what it feels like to feel." (Given all the other unsolved and perhaps unsolvable problems asscociated with feeling (consciousness), like the other-minds problem and the hard problem, it is perhaps not so surprising that this category turns out to be odd, among categories!)
But none of this is unique to Gabe: he needs negative evidence to learn categories, like anyone else. And on the behaviorist interpretation (detecting internal states rather than feeling felt states) Gabe could talk about it exactly the same way we do, but just (unbeknownst to him) referring to "detectable internal states" when he says "feeling felt states." His referent would be different from ours, but neither of us would notice, because our words would be Turing-indistinguishable. I myself could be the Zombie speaking these words, and no one would be the wiser, neither me nor you. (Yet one wonders why Zombies would ever talk about such things: Is this evidence that there could not be T3 Zombies?)
Nor would innate feature-detectors help: feature-detectors, whether learned or innate (because acquired through evolution) still distinguish category members from non-members. But with an uncomplemented category like feeling, there are no non-members and cannot be.
(Are there any other uncomplemented categories like this?)
Response to UNCOMPLEMENTED CATEGORIES AND COULD THERE BE ZOMBIES? Part 1
DeleteProfessor Harnad: “If Gabe were an insentient, Turing-indistinguishable T3 Zombie (i.e., doesn't feel a thing) why on earth would he talk about "feelings," exactly the same way we do?”
I have thought about this quite a bit! The immediate thing that comes to mind would be to pass the Turing Test (TT), but I do not believe this to be a satisfactory answer. I guess the goal of passing the TT would fall into the first possible answer that you gave about Gabe simply just being able to detect, identify and name felt states… and not actually feel the felt states. I believe Gabe would be able to pass the TT by simply being able to talk about felt states and would not be required to actually feel the felt states.
The second possible answer that you provide is quite intriguing, although I have to admit that it does quite seem like idle sci-fi (at least for now). I would like to imagine that I would be more likely to accept this as an answer to the above question you asked. I believe I would accept this answer more easily that the first answer proposed because if Gabe were an insentient T3 zombie, there would be no incentive for him to talk about or consider feelings in the exact same way that we do. As someone who feels things, this just wouldn’t make any sense. I agree that the hard problem seems impossible to answer and even if we did solve the hard problem, it would be impossible to tell if a robot from MIT like Gabe could actually really feel due to the other-minds problem. When considering the other-minds problem, I also believe that the hard problem cannot be solved with total certainty.
Response to UNCOMPLEMENTED CATEGORIES AND COULD THERE BE ZOMBIES? Part 2
DeleteProfessor Harnad: “Are there any other uncomplemented categories like this?”
After reading your response I tried to think about other examples of uncomplemented categories. I let this thought develop over a few days and came up with a few examples like: “what is it like to think?” or “what is it like to simply exist in the first place?”. Once I had come up with a few examples, I decided to do a little bit of research and ended up finding one of your papers titled “Uncomplemented Categories, or, What is it Like to be a Bachelor?” (here is the link if anyone else is interested as well: http://cogprints.org/2134/1/harnad87.uncomp.htm). In your paper you state that all uncomplemented categories are “experiential categories” and that because these categories are experiential all of their complements must be non experiential meaning that they are either “non-existent” or “self-contradictory”. One of the examples that I was able to come up with would indeed fit this criterion. To think is to experience something and its complement would be non-existent. It makes sense that the complement of an experiential category would not exist. No one could possibly know what it is like not to experience a certain experience.
You also stated in your paper that there are uncomplemented existential categories as well. The other example that I was able to generate “what is it like to exist in the first place?” would fit into this category. The questions why and how we exist are existential ones. The complements to these categories do not exist and would be impossible to try and extrapolate.
Lastly, you proposed that “some of the self-denial puzzles, such as the statement ‘This statement is false,’” may also fit into the uncomplemented categories. I have a harder time accepting self-denial puzzles as uncomplemented categories. They seem like much weaker examples than uncomplemented experiential and existential categories. I don’t quite understand how these types of statements are uncomplemented… would this mean we do not know what if it would be like if certain statements were not false (even though they are known to be false)?
I would like to ask one additional question with regards to this comment by Professor Harnad:
Delete"Like us, any grounded T3 must have sensors that detect external objects and their features (but without feeling a thing). He could also have detectors for internal states, such as the state of recognizing and categorizing an apple, or the state of being overheated. His words and descriptions of "what it feels like to see an apple" or to "feel hot" would (unlike ours) just be referring to those internal states, which would be detectable and nameable states, just like our felt states are, but they would not be felt, just detected, identified and named."
The idea that a mechanism could detect something and yet not feel it is tricky. What distinguishes detecting from feeling? I understand that theoretically this refers to the distinction between a T3 robot that is grounded yet does not understand meaning/does not feel and me (a person who is certain that I am a cognizer because I know that it feels like something to feel a certain way or understand a certain thing - although, of course, I have no way of proving that to you because of the "other minds" problem). However, the idea that detecting isn't feeling still isn't quite clear to me. What's the difference between detecting that there is a wall in front of me and sensing that there is a wall in front of me? Here it seems that the wording can be tricky as well because I am trying to pick words which are practically equivalent yet where one does not imply feeling. As I write this, it occurs to me that we must already have created a way in which a mechanism can detect its environment (with or without feeling), given that we have things like automated vacuum cleaners that redirect themselves when they hit a wall. Perhaps I could answer my own question if I was familiar with computer science and could deconstruct/reconstruct a robot like an automated vacuum cleaner to know what computation was inputted such that, when the vacuum bumps into an object, it reacts by moving in a different direction. But I am still left with the lingering uncertainty that there is a clear discontinuity between detecting and feeling. It becomes more confusing when considering the detection of internal states. If a computational mechanism is capable of grounding words in their meanings using a sensorimotor system - what keeps us from assuming that it does not also have the accompanying mental states? Then the "other minds" problem comes up again. I suppose this leads me to the same conclusion as the Professor: that T3 is sufficient to pass the Turing test.
"To be grounded, the symbol system would have to be augmented with nonsymbolic, sensorimotor capacities -- the capacity to interact autonomously with that world of objects, events, actions, properties and states that its symbols are systematically interpretable (by us) as referring to. It would have to be able to pick out the referents of its symbols, and its sensorimotor interactions with the world would have to fit coherently with the symbols' interpretations." (Harnad, 2003)
ReplyDeleteHarnad talks about how symbols in the brain possess the capacity to pick out their referents. Therefore, a symbol system alone cannot have this capacity because picking out referents is not just computational. In the above quote Harnad states that in order to be grounded the symbol system would have to be augmented with non-symbolic sensorimotor capacities. The system has to be able to interact with the world and interact with the things which the symbols are referring to - in order to ascribe meaning to those symbols. Additionally, there needs to be a mode or way by which these symbols are allowed to interact with the worldly entities which they refer to. The mode by which these (arbitrary) symbols are allowed to interact with the worldly entities is in our heads. We know that grounding has to be sensorimotor and that it is this sensorimotor capacity that augments a symbol system - so is there much more that we need in order to create a robotic mechanism that can do the same? There is obviously more to learn and understand - and knowing that grounding must be sensorimotor may not be the full story - however, I want to believe that this would be enough to create a robot that could ground symbols in a similar way as we do based on sensorimotor capacities which allow for the "capacities to detect, categorize, identify, and act upon the things that words and sentences refer to" (Harnad, 2003).
The main thing that's needed is a mechanism for learning what's the right thing to do with the right kind of thing. Deep-learning neural nets are the latest candidate, but it's not yet known whether they are up to the task. T3 capacity is a lot more than google's image- and word-processing power.
Delete“But if groundedness is a necessary condition for meaning, is it a sufficient one? Not necessarily, for it is possible that even a robot that could pass the Turing Test, "living" amongst the rest of us indistinguishably for a lifetime, would fail to have in its head what Searle has in his: It could be a Zombie, with no one home, feeling feelings, meaning meanings.”
ReplyDeleteSearle’s Chinese Room Argument showed that a computer can pass T2 without displaying any understanding of the meaning of the Chinese symbols it is manipulating – they are just meaningless squiggles. The rules Searle and the computer are following to manipulate these symbols are based on the symbols’ arbitrary shapes, not their meanings. What Searle and the computer are doing is just symbolic (computational) manipulation of ungrounded symbols.
According to Harnad, meaning requires symbols to be grounded, they need to be directly connected to their referents in the physical world. This makes the upgrade of the T2 digital computer to a T3 robot with hybrid symbolic and sensorimotor capacity inevitable. Since grounding links sensory inputs (vision, touch etc.) to objects in the world, it is necessary for a ‘grounded system’ (be it part of a human or a robot) to possess sensorimotor ability. Thus, grounding is necessary for meaning, but is it sufficient? Does a T3 robot that has the ability to ground symbols actually understand the meanings of these symbols? According to Harnad, there is a second property – consciousness – that may be required for meaning. I would say that this property could be equated to ‘feeling’, since it feels like something to understand the meaning of a word (or to not understand it). A T3 robot could pass the Turing Test due to its grounding ability (which in turn is rooted in its sensorimotor ability), but we still couldn’t tell whether it is actually understanding (i.e. conscious, feeling). However, as pointed out in one of the Skywritings above, we have no clue (and probably never will) as to how it is that consciousness facilitates the ‘production’ of meaning and how it is that it does this independently from the symbol grounding system.
This is a small point, but I think Searle's argument doesn't show that a purely computational symbol system could pass T2. He takes this premise and shows that, if Searle's Chinese room were fully capable, the system wouldn't be understanding Chinese on any meaningful level.
DeleteFurther, it seems that groundedness is a necessary condition for passing T2, as a system designed for proficiency in language built off of rules that only point to each other would be subject to the symbol grounding problem
As Kevin notes, Searle didn't show that a computer could pass T2, he assumed it, in order to show that even if it could, it would not understand.
Delete"the upgrade of the T2 digital computer to a T3 robot with hybrid symbolic and sensorimotor capacity [is] inevitable"
Maybe, but the upgrade is bottom-up, from sensorimotor function to language rather than top-down from language to sensorimotor function.
Re: Harnad (1990)
ReplyDeleteI'm a bit confused about the distinction between symbol manipulation and connectionism drawn here. Does symbol manipulation refer to systems/subsystems comprised of static, well defined rules on what operations to perform based on symbols, and connectionism to systems that learn patterns in data over time, without respect to hard coded rules? Because neural networks can be run on regular computers, doesn’t that mean connectionism is in some sense built on symbol manipulation? In addition, don’t well built connectionist systems generate semantically interpretable output (and have states of learning that are in some sense semantically interpretable), or is rule #8 of symbol systems more rigid than to count that?
Unless it is really a parallel and distributed system of physically interconnected nodes, a neural net is just a computational learning algorithm. It's a valuable component of a robot, but it is no more grounded that Searle's (and computationalism's) hypothetical T2-passing algorithm.
DeleteOn the other hand, if for some reason the parallelism and distributedness were essential to the power of neural nets, then nets would not be computational (because parallelism and distributedness are analog physical properties, not hardware-independent computational ones). This is unlikely to turn out to be the case, though, because all the powers of neural nets to date have been demonstrated with simulated neural nets (hence computation). Searle could execute their algorithm and still not understand Chinese if they were part of the T2-passing algorithm.
“The symbol grounding scheme just described has one prominent gap: No mechanism has been suggested to explain how the all-important categorical representations could be formed: How does the hybrid system find the invariant features of the sensory projection that make it possible to categorize and identify objects correctly?[20] Connectionism, with its general pattern learning capability, seems to be one natural candidate” Harnard, 1990.
DeleteAs Kevin, I was first confused by the distinction between connectionism and symbol-based manipulation systems. As I understand it, neural networks abstract relevant features from several encounters whit an object of a given category and dynamically adjust the “weight” of each features/feature combinations. It thereby reduces “the icons to the invariant (confusion-resolving) features of the category to which they are assigned”. However, in my mind this abstraction process has to be somehow probabilistic, leading us again to a computational system.
I then tried to take a grasp on how parallelism and distributedness could differentiate neural nets from simpler T2-passing computational algorithms.
From what I understand, parallelism is the idea that when seeing a horse, a horse shape is casted on our retina and after several encounters with horses, an invariant icon is abstracted. There would be a distinct icon for every category of things we encounter, stored at a basic level of abstraction into these neural nets we have.
Let’s say that seeing a triangle wherever in the receptive field lead to a specific pattern of activated adjacent neurons:
__X___XXX_
_XXX___X__
This pattern, regardless of its orientation, could be remembered and stored within the neural net as “triangle”. However, is this stored representation more analog to the object in the physical world or to the pattern of activated neurons? For vision, the retina more or less corresponds to the visual field so it might not cause a problem. But if we take the sense of smell, the representation that would hypothetically be stored in the neural net when encountering a horse is the pattern or set of neurons activated when smelling a horse. How is this pattern representative of horses in the physical world? It seems to be closely similar to what a computational model proposes with the idea of on/off switches. If that neuron and that neurons are activated similarly, then what I just smelled is a horse. I have difficulty picturing an iconic representation in any other sensory modality then vision. Is place coding in the cochlea an iconic representation? Place coding seems to be an analog representation, but that notion of "iconic representations" confuses me when considering different sensory modalities.
I think you might be mixing up the question of analog neural nets and analog sensory projections.
DeleteSensory projections don't have to be spatial to be analog. They could be frequency-based (as in color and hearing) or chemical (as in olfaction and taste). And even when the analog input is processed by the brain, it can remain analog, even all the way to motor output. If the input needs to be categorized (to learn to do the right thing with the right kind of thing), the icon can be filtered by a learned feature-detector (which could be a neural net, whether analog or computational). A symbolic representation would be a verbal description or definition of the categories, but the features it uses to describe the category have to be grounded category names too.
« […] there is an intrinsically dedicated symbol system, its elementary symbols (names) connected to nonsymbolic representations that can pick out the objects to which they refer, via connectionist networks that extract the invariant features of their analog sensory projections.”
ReplyDeleteHarnad is trying to tackle the “symbol ground problem”, i.e. how symbols (which are arbitrary by definition) get their connection to the real world (e.g. how is it that I know that “apple” refers to the fruit which is red, round, etc).
As a starter, Harnad proposes a hybrid model: 1) a sensorimotor section (dealing with iconic, analog “printed” representation of external objects) that would be able to tell things apart (discrimination) and know which things are within a category (identification) and 2) a symbolic level working as a physical symbol system, beyond an inert lexicon of labels.
As one could see here a “bottom-up vs top-down” debate here, Harnad pushes for the primacy of bottom-up processes that would build up, from the most basic to the most complex (we could even argue for abstract concepts latter on in the process). He also argues that the capacity to ground symbols is tributary to the capacity to abstract/extract general features from individual to create a “category”. When you say “zebras”, you’re not talking about an individual, but rather a kind.
I would add that abstraction (i.e. generalisation) is important for communication because it gives cognition some flexibility in regards of the context: you can think about a zebra even without being in presence of a zebra, and neither do you need to remember an particular instance of a zebra (even if that could help to think about). Without that, it would be only a reflex to say the sound “zebra” in presence of the animal, and it would be way less useful as a communication device. Obviously, this is not an explanation, and in this sense, I would agree with Prof. Harnad that a “features abstraction” mechanism is crucial for any grounded symbol theory.
(1) Analog sensory projections ("icons") are for discrimination. (2) Categorical representations are feature-detectors for categorization (identification). (3) Symbolic representations are verbal descriptions of categories (true/false subject/predicate propositions. Contemporary human communication is mostly verbal. Ancestral communication might have been more iconic (pointing, miming and acoustic imitation). (Important for later when we discuss the origin, evolution and function of language.)
Delete"In priniciple, though probably not in practice, only 1000 category names need to be grounded directly in order to be able to get all the rest through boolean re-combinations of the names these directly grounded categories."
ReplyDeleteI think that parallel to (but not isolated from) this quantitative question there is the qualitative ones: are there categories that are necessary to be grounded at first for any (grounded) symbol system? For example, do we need words for up, down, colors, textures, emotions, etc. Could any sufficiently diverse combination of grounded symbols do the trick?
If there is some primordial category that needs to be told apart, how could we discover which are they? I am wondering about the pertinence of experiment in the vein of the “bouba/kiki” effect (from Köhler 1947). In this experiment, people are asked to associate two fictive words (“bouba” and “kiki”) with two shapes: a very sharp star-like shape, and a round shape. People were more likely to associate “kiki” with the sharp shape and “bouba” with the round one.
While it would be ludicrous to argue that this is an explanation for grounding symbols, I’m wondering if there could be some hints about the basic mechanism at work. Could the brain have basic, hard-wired connections between auditory and visual modes? These connections could be the starting grounding point between any sound and visual features (even if we argue that communication was gestural at first), the point that allow human to understand that the sound **have** some meaning at all. It would be one of the basic mechanisms that an engineer would have to implement in a T3. It is not a very detailed nor convincing account, but I’m wondering if there is material to start from there.
Iconicity
DeleteYou are asking "which thousand words are needed, and in what order?"
Not yet clear, but among the huge number of potential 1000-word minimal grounding sets ("MINSETs") the words tend to be more frequent (in writing and speech) than the rest of the words in the dictionary, and tend to have been leaarned at a younger age.
There are many potential 1000 MINSETs, some differing only by a few words, but it is certainly not true that any directly grounded set of 1000 words is a MINSET that can define all the rest of the words in the dictionary. The likellihood that 1000 random words will be a MINSET -- or even directly grounded -- is near zero.
The "bouba/kiki" effect is just evidence of a tiny bit of iconiciy in words -- either because of multisensory resemblance or just convention. Iconicity makes the connection between icon and "referent," but to become language the iconicity must become irrelevant. That might have been part of the path from pantomime to propositions (week 8). The constraint on grounded words is not the shape of the word, or any resemblance to its referent but the feature-detectors that ground words through the capacity to recognize and categorize their referents.
"The necessity of groundedness...takes us from the level of the pen-pal Turing Test, which is purely symbolic (computational), to the robotic Turing Test, which is hybrid symbolic/sensorimotor (Harnad 2000, 2007)".
ReplyDeleteI found that this statement helped to clarify for me the main difficulty with creating a computer that can agreeably pass the Turing Test. Groundedness seems to be a way for everyone to be on the same page of what we're really looking for the word 'understanding' to mean, brought up in Searle's Chinese Room Argument. From this, if I understood Lady Lovelace's objection correctly, then perhaps she was correct in saying that a computer can never really do anything new because they're incapable of independent learning. Just based on the face that we know how computers work because humans have built and created them, I agree with her. A computer can only do whatever input is given to them. There is something unique in understanding and groundedness of symbols that requires more than just the symbolic/purely computational level, as Harnad said. So, this quite effectively points out the missing ingredient in creating the computer that will fully pass Turing's test, which is the full capacity of understanding/groundedness of symbols. To this, feeling is connected because it feels like something to really understand. So, as I see it now the difficulty comes with implementing a way for a computer to ground symbols for actual understanding to occur to get to passing T3 and beyond.
As discussed in Week 2, Lady Lovelace's objection is incorrect. A computer can do plenty that's new. This is not the reason grounding is needed. Grounding is needed in order to give words meaning.
Delete“But if groundedness is a necessary condition for meaning, is it a sufficient one? Not necessarily, for it is possible that even a robot that could pass the Turing Test, "living" amongst the rest of us indistinguishably for a lifetime, would fail to have in its head what Searle has in his: It could be a Zombie, with no one home, feeling feelings, meaning meanings (Harnad 1995).
ReplyDeleteAnd that's the second property, consciousness, toward which I wish merely to point, rather than to suggest what its underlying mechanism and causal role might be. The problem of discovering the causal mechanism for successfully picking out the referent of a category name can in principle be solved by cognitive science. But the problem of explaining how consciousness can play an independent role in doing so is probably insoluble, except on pain of telekinetic dualism. Perhaps symbol grounding (i.e., robotic TT capacity) is enough to ensure that conscious meaning is present too, perhaps not.”
If the problem itself is how symbols get their meaning/what meaning really is and Steels (2008) claims the ability to solve this problem in implementing a system of semiotic networks transpiring meaning through sensorimotor capabilities and categorization/negotiation of symbols, why does consciousness enter the equation? When Harnad refers to “meaning” versus “conscious meaning” (in the last sentence of the quote above) is there a difference?
On a separate thought - if consciousness in medicine is seen on a continuum based on arousal and awareness, would the ability to transpire meaning and connect your external surroundings to internal networks constitute awareness?
1. Steels has not solved the symbol grounding problem: he has not created a T3, not even close. (The MIT team that created Gabe would have solved the symbol grounding problem -- if they had really created Gabe...)
Delete2. Consciousness (feeling) comes in because meaning is not just grounding. It is grounding + feeling (because it feels like something to say and mean something.
3. Consciousness and awareness are just weasel-words for feeling. Important not to confuse the issue pf whether an entity can feel anything at all, from issues of degrees (or intensity or quality) of feeling.
You can have grounding without feeling. The best example discussed above is the insentient T3 zombie – an artificial agent that has sensorimotor capacities (therefore grounding), but doesn’t feel. The insentient T3 zombie detects internal states. Of course we don’t know what it is like to not-feel. As Harnad has mentioned: feeling is the only category that is uncomplemented (we don’t know what it is like to not-feel, we only know what it is like to not-feel X).
Delete“The problem of discovering the causal mechanism for successfully picking out the referent of a category name can in principle be solved by cognitive science. But the problem of explaining how consciousness can play an independent role in doing so is probably insoluble”
Does the above quote speculate that consciousness (feeling) initiates grounding independently? Can you have consciousness (feeling) without grounding? What would be an example of this? Does Searle 'feel' in the CRA? Clearly Searle doesn't understand. However, Harnad (and Searle) states Searle doesn't understand because he lacks symbol grounding.
Searle's response to the Systems Reply can be analyzed in comparison to Harnad's explanation on meaning in a wider sense. I could be incorrect in drawing this parallel -- in which case, I welcome corrections and an alternative explanation -- but I could not help but notice that the two both argue that meaning may be derived by internalizing the symbols/referents and the means by which they are assigned.
ReplyDeleteThe Systems Reply to the Chinese Room Argument asserts that while the man running the program (Searle) does not understand Chinese, he is only a part of the system and the whole system does understand. Searle refutes this counterargument by proposing: if he had internalized the entire system, memorizing all the instructions and symbols, Searle would still not understand because he is unable to attach any meaning to the symbols. In comparison, Harnad argues that meaning (or understanding) is generated within a head which holds "both the means that the entity uses to pick out the referent, and the referent itself". The obvious and key distinction here is that Harnad, by involving referents rather than symbols and computation, already implies that symbol systems must be "augmented with non-symbolic, sensorimotor capacities" in order to "pick out the referents of its symbols."
Ultimately, while the System Reply argues that knowing the computational rules and symbols will derive meaning, Searle and Harnad both recognize that understanding requires grounding of those symbols in referents. Furthermore, Harnad adds that in addition to "(1) a head, (2) a word inside it, (3) an object outside it, and (4) whatever "processing" is required in order to successfully connect the inner word to the outer object", there must be feeling; to reiterate the Professor's comment in this thread: meaning = grounding + feeling.
If the meaning of a word is the means of picking out its referent, the meaning is in the head -- but not the referentù; If I say or think of "apples," the means of picking them out is in my head (including my senses), but the apples are not.
Delete"So if we take a word's meaning to be the means of picking out its referent, then meanings are in our brains. That is meaning in the narrow sense. If we use "meaning" in a wider sense, then we may want to say that meanings include both the referents themselves and the means of picking them out. So if a word (say, "Tony-Blair") is located inside an entity (e.g., me) that can use the word and pick out its referent, then the word's wide meaning consists of both the means that that entity uses to pick out its referent, and the referent itself: a wide causal nexus between (1) a head, (2) a word inside it, (3) an object outside it, and (4) whatever "processing" is required in order to successfully connect the inner word to the outer object."
Sorry if the quote confused you. We are talking about cogsci's "narrow meaning" (meaning in the head) not philosopehrs' "wide meaning" ("meaning in the world"). In fact, forget about "wide meaning."
But no, the "System Reply" is not an invocation of "wide meaning." And since meaning is not just grounding, but also what it feels like to mean, I hope no one has any doubt about the fact that a feeling cannot be wider than a head (despite the nonsense that has been written about "distributed cognition," which used to be in week 11 of this course: I've since dropped it).
From the Scholarpedia article:
ReplyDelete"The symbols in an autonomous hybrid symbolic+sensorimotor system -- a Turing-scale robot consisting of both a symbol system and a sensorimotor system that reliably connects its internal symbols to the external objects they refer to, so it can interact with them Turing-indistinguishably from the way a person does -- would be grounded. But whether its symbols would have meaning rather than just grounding is something that even the robotic Turing Test -- hence cognitive science itself -- cannot determine, or explain"
The T3 robot has the ability to ground symbols to their referents because it has both symbolic and sensorimotor capacities, so it can interact with objects in the real world and assign the symbols to their referents. The problem regarding whether or not the symbols would have meaning to these sensorimotor robots can't be explained yet and is related to the hard problem of cognitive science.
I'm still confused about what it is to understand in relation to this topic. If meaning = grounding + feeling, where does "understanding" fit into this? Is it a weasel-word, or is it something more similar to grounding or the process of grounding?
Hi Anna,
DeleteI was also interested in how "understanding" is implicated with meaning. I think an important section of the Scholarpedia article that somewhat addresses this topic is when Harnad discusses the Chinese Room Argument. Searle's take home message is that he can compute proficiently enough to pass T2 without understanding a single word of what he outputs, but he never explains fully what this means to understand. We all innately know what understanding feels like and definitely as university students what it feels like to not understand something. So does Searle, "he understands the words of English, whereas the Chinese symbols that he is manipulating according to the program's rules mean nothing whatsoever to him". From this example, it seems to me that understanding, and meaning are synonymous. Although in general terms we usually say that "we understand what something means" understanding may also just be the combination of grounding and feeling because grounding already includes the process of linking a symbol with its referent.
Understanding is also Grounding + Feeling. To say and mean something feels like something; to hear/read and understand something feels like something. (And for Chinese, Searle does not feel that something!)
DeleteApart from the definitions given, "Understanding is also Grounding + Feeling." I feel it's also just not possible to have a counterpoint to it. To say "meaning = grounding + feeling" in a sentence would sound like "If you know the meaning of a word, you have grounded it and the feeling of what its like to say and think of its meaning as you say it." If you know the meaning, that means you understand. Thus, meaning = grounding + feeling = understanding.
DeleteIn Searle's Chinese Room, he has neither grounding nor feeling because they are just arbitrary symbols to him. In a regular T3 robot, there is grounding but not feeling, thus no understanding–basically if Searle's room had been given sensorimotor capabilities (I understand this is no longer computationalism and is beyond the scope of Searle's argument, but it's to illustrate). But the MIT Gabe somehow has grounding and feeling, giving him (it? I'm assuming he prefers male pronouns though, sorry for offending any T3s) understanding.
If you know the meaning of a word you can recognize its referent and you can define or describe it in words. It feels like something to know the meaning of a word, to recognize its referent and to mean or understand its verbal definition or description. (But a description is not just a long word: it is the combination of grounded words in a subject/predicate proposition, which you have to mean/understand too. We'll get to propositions when we get to language.)
DeleteYou (MIT Gabe) definitely have grounding. I cannot know (because of the O-MP) whether you feel, but Turing says if I have no (relevant) way to distinguish you from other people, then I have no basis for thinking they feel but you don't.
Regarding identification of objects, Harnad argues that "icons" are enough for discriminating between things, but not enough for identifying things. Discrimination is relative, it is seeing the difference/similarity between objects. Identification is an absolute judgement, which seems to be more difficult to explain.
ReplyDeleteAccording to Harnad (1990), "Icons of sensory projections are too unselective. For identification, icons must be selectively reduced to those "invariant features" of the sensory projection that will reliably distinguish a member of a category from any nonmembers with which it could be confused".
There's theories of how we identify objects, but I would have liked to seen more of an explanation of how we might do this and not just why icons are not sufficient.
I initially had the question that if a T3 robot can do everything we can do, is it able to autonomously categorize abstract things? I suppose yes, because our ability to identify and discriminate is mostly a learned one.
We'll discuss how the brain (might) learn to detect features so as to identify objects (categories) via neural nets next week.
DeleteGabe: are you able to learn to "autonomously categorize abstract things"? E.g., goodness, truth, beauty, and peekaboo unicorns...
I think that Gabe will answer yes to the question of whether he can autonomously categorize abstract things. Abstract things are still physically felt. It feels like something to find someone attractive, and we feel it through our sensory organs. A T3 robot that has the same sensorimotor capacities as we beings-that-think have has all it needs to learn to do so. We will never know ‘how’ it is that Gabe will categorize a specific person as being more attractive than another one. Abstract categories are subjective and can be grounded somewhat differently in different people. For the rest, I don’t think the categorization of abstract words differs that much from concrete words. If Gabe achieves to do one, he can surely do the other.
DeleteIn response to Chloe: I agree that both abstract and concrete concepts are felt. However, they are felt through different means. From what I understand, we are able to categorize concrete concepts through sensation (five senses). Categorization of abstract concepts requires feeling beyond our senses. Would you say that a sensorimotor complex attached to the T2 passing artificial agent is enough to categorize abstract language? Abstract and concrete language aren’t entirely separate. There is a gradient with general ideas or experiences (abstract) on the left and more specific physical objects or events (concrete) on the right. Harnad says “to categorize is to ignore the individual differences among members and abstract only what distinguishes the members from the non-members of the category.” In order to categorize you must be able to abstract the similar features, “increasing the weight of category-distinguishing features and reducing the weight of category-irrelevant features”. While abstract word categorization may be harder to ascertain, I agree with Chloe in that if you can categorize concrete concepts you will be able to categorize abstract concepts.
Delete“The expectation has often been voiced that "top-down" (symbolic) approaches to modeling cognition will somehow meet "bottom-up" (sensory) approaches somewhere in between. If the grounding considerations in this paper are valid, then this expectation is hopelessly modular and there is really only one viable route from sense to symbols: from the ground up. A free-floating symbolic level like the software level of a computer will never be reached by this route (or vice versa) -- nor is it clear why we should even try to reach such a level, since it looks as if getting there would just amount to uprooting our symbols from their intrinsic meanings (thereby merely reducing ourselves to the functional equivalent of a programmable computer).”
ReplyDeleteI’m somewhat confused by what Harnad is saying here. If I’m grasping this correctly, it’s something along the lines that a purely symbolic level of representation cannot be grounded, as grounding can only happen in a bottom-up fashion.
Applying these principles to Turing machines has interesting implications for questions surrounding meaning and consciousness. Say you have a Turing-scale robot consisting of both a symbol system and a sensorimotor system that reliably connects its internal symbols to the external objects that they refer to. This T3 robot would be able to ground symbols, but would it be able to derive meaning from this grounding process? I definitely don’t think so. As was shown in the Harnad paper, meaning can only be created from the combination of symbol grounding and the feeling of being in that specific mental state.
This must have come up before, but how does one ground words that refer to feelings? There is no physical referent for emotions, which suggests that the mechanisms that people use to define emotions is based on the conscious mental *experience* of an emotion/feeling, as opposed to sensorimotor categorization. For me, this is further evidence that even the most sophisticated T3 robot is not a conscious being.
A grounded symbol is necessarily hybrid symbolic/sensorimotor, hence it is no longer a "pure" (Turing) symbol system.
DeleteMeaning = Grounding + Feeling. (The other-minds problem prevents us from knowing for sure that Gabe feels.)
Gabe: A grounded symbol is necessarily hybrid symbolic/sensorimotor, hence it is no longer a "pure" (Turing) symbol system.
Meaning = Grounding + Feeling. (The other-minds problem prevents us from knowing for sure that Gabe feels.)
Gabe: Do you know what it feels like to feel tried, happy, see red, hear an oboe play middle C? Tell us. Show us...
I enjoyed reading the main article on symbol grounding problem. I find myself in agreement with the arguments in the main text.
ReplyDeleteI also enjoyed the idea of connectionism, linking a symbolic and non-symbolic\grounded system. From the main text, I understand that formal symbols do not have a grounded referent, while meaning is derived from grounding words in the physical world to a referent using sensorimotor information. I also recall that in conversations in class regarding why we need a biorobot to pass T3, Steven mentioned that it is not possible to transform all the dynamic analog information gathered by our senses into digital information. I understand that sensorimotor capacities enable one to collect various types of information about an object, categorise the object, manipulate it in space and form an association between an object and a referent. However, at the same time I wonder if it is actually "impossible" to transform all the analog information into digital information, or just inefficient. Given that there are machines that can guide a self driving car and smell substances, it seems possible that the information gathered through our senses can be turned digital. Are we then accurate in saying that this digital information is less "grounding" than the analog information? If so, why?
It is possible to transform sensory information to digital to as close an approximation as we like. That still does not mean that the digital computation, without the sensor from which it was transformed, is sensing. Ditto for moving. (This is even without asking about whether it is felt: a digital plane cannot fly; it just manipulates squiggles and squoggles.)
DeleteGrounding is the T3 robot's connection to the referents of its symbols.
“In contrast, the meaning of the words in my head – the ones I do understand – are “grounded” (by means that cognitive neuroscience will eventually reveal to us). And that grounding of the meanings of the words in my head mediates between the words on any external page I read (and understand) and the external objects to which those words refer.
ReplyDeleteIn this week’s readings, Dr. Harnard discusses the Symbol Grounding Problem, a consideration in which the question of the derivation of meaning from arbitrary symbols is questioned. In so doing, Dr. Harnard highlights that there is logically, a necessary presence of some sort of innate, or intrinsic understanding/knowledge present in T2+ level beings. Dr. Harnard argues that such innate meanings mediate external symbols with their corresponding referent (a process one can simply call categorization) and provide a “grounding” for the meanings of symbols within the brain: one understands. However, in attempting to accomplish the stated mission cognitive science, the Symbol Grounding Problem bears its head. How does a computer (ostensibly lacking any ‘innate’ capacities until the institution of rules) compute meaning from arbitrary symbols without running into overbearing paradoxes of meaning? Is meaning even present within the computer’s “brain” as it executes symbolic manipulation? Is meaning or understanding even necessary to accomplish T2 as in Searle’s Chinese room argument? And do our answers to those questions indicate the ability of the human brain to generate schema of arbitrary symbols whose relation to one another generate meaning?
How did questions of "innate meanings" get into this? All you need is sensors and effectors plus a learning mechanism.
DeleteFrom the Scholarpedia article:
ReplyDelete“The standard reply of the symbolist (e.g., Fodor 1980, 1985) is that the meaning of the symbols comes from connecting the symbol system to the world "in the right way." But it seems apparent that the problem of connecting up with the world in the right way is virtually coextensive with the problem of cognition itself. If each definiens in a Chinese/Chinese dictionary were somehow connected to the world in the right way, we'd hardly need the definienda!”
It seems very abstract and unclear to me how the definiens or any symbols “connect” with the world. Would Harnad say that this understanding is trivially important for the theory of symbol grounding? I also wonder why that definiens connecting the world in a right way would make definienda be hardly needed. I know that in a sense they can be substituted without changing the meaning in the language. But would simplicity also matters in we choose what symbols we use?
And I don’t know if that would relate to the Hesperus example, even if we have the right definiens of that star, to say it is the “Evening Star” or "Morning Star”, there seems to me still exists a difference between saying “the Hesperus is the Vesper” and “The venus is the venus”. I may be very wrong but I think that if we only have the definiens in the right way to replace the definienda, then some information is still left out.
I am not 100% confident in my response but I will try to address two different parts of your commentary. The first thing that I want to talk about if the "connection in the right way". From class discussion, I believe that this right way would refer to direct mapping between the symbol and the referent. The example prof Harnad gave in class was that of pantomime / mime. In this way, because the symbol is a direct reflection of the referent, there would not be any meaning lost. However, the substitution of a symbol (word) using the domain of the English alphabet by another symbol in that same domain of possible units would indeed not be very helpful in the attempt to connect the symbols to the world. The second thing I wanted to address was your example of the evening/morning star. To me, the difference between "the Hesperus is the Vesper" and "the Venus is the Venus" lies in the meaning. As we said in class, even though the Hesperus and the Vesper point to the same referent, they have a different meaning, they express a different reality, and they should therefore not be perfectly interchangeable. For example, I don't know which one is which, but if the evening star is the Hesperus, you would not be right to say you saw the Hesperus in the morning: you could have seen the Vesper and/or Venus, but definitely not the Hesperus. I hope you find this helpful!
DeleteThe words in Gabe's (T3) head are connected to their referents through his sensorimotor capacities, mediated by neural nets in his head that learn to detect the features of the referents. If they were connected innately, there would be no need for the learning.
DeleteMore than one word or description may connect to the same referent, but in different ways, based on different features.
Pantomime is not language; it only becomes language when category-names (including feature-names) can be combined in subject/predicate propositions. Pantomime and pointing can draw attention to an individual or an individual situation, but it predicates nothing; it just draws attention, based on imitating and pointing. Gestural language is propositional; the resemblance of gestural words (icons) to their referents in sign language is very partial, and inessential once language has evolved. But the link provided by miming and pointing might have helped language evolve. Once there, the shape of words may as well be arbitrary. And in the much more powerful and efficient modality of oral speech, there is no choice. There is very little scope for sounds to resemble their referent. Alphabets and writing come even later.
Yes, words and descriptions of the same referent do not have the same meaning.
Regarding Professor Harnad referring to Funes the Memorius above:
ReplyDeleteThis is my favourite Borges story, so I’m excited to discuss it in light of cognitive science. I’m sure my understanding of the story will change in class when we discuss it, but at this point it seems to me that this story is a thought-experiment about an agent who possesses consciousness, but becomes unable to ‘think’. This thought experiment is interesting, and the opposite of the thought-experiments we’ve been considering so far regarding agents that can ‘think’ (or at least perform computation) but who may be unable to feel (or possess consciousness).
In the story, Funes has a perfect memory, and can remember infinite details of every sensorimotor experience he’s had over his whole life. Because he cannot leave behind details, he stops performing categorization, as categorization stops being a useful tool to him insofar as to derive meaning. He begins more and more to think of every thing as an individual instead as of part of a larger ‘type of thing’, as he is unable to forget differences between individuals of the same category, and as he experiences very different meanings from individual instances of things. To rephrase, he stops finding grounding words that refer to categories as very useful, as each individual instance of things within a category possess very different meanings, as he experiences infinite differences between members of the same category.
Since he stops categorizing, he becomes incapable of generalizing and abstraction. At one point in the story, he re-writes numbers, and assigns them each an arbitrary/incoherent name not part of a system, (as opposed to developing a system, like the base-ten system). He rejects categories among numbers (rather than, for example, everything between 300-400 being named as 3— or in the category of between 3 and 4 hundred – I find it difficult to refer to this category of numbers without explicitly stating its definition), they would all have individual unpatterned names.
It’s posited in the story that Funes loses his ability to ‘think’: “To think is to forget a difference, to generalize, to abstract. In the overly replete world of Funes there were nothing but details, almost contiguous details.” And indeed, if we’re considering ‘thinking’ to be that which is described by computation, computation deals with formal symbols; if an agent can only consider discrete individual ‘things’ rather than dynamical formalized generalizations, the agent cannot perform algorithms that deal with formal symbols for input and output.
No, thinking (cognition) is not computation (as I hope the first five weeks have made you first believe, then doubt). Thinking is what goes on in the head of a thinking being (a cognizer). But to be able to do what cognizers can do, they have to be able to categorize -- to ignore (or forget) the individual differences among members and abstract only what distinguishes the members from the non-members of the category. And that's exactly what Funes lost the capacity to do. (In fact, if he were really the way he is described, he would not have been able to talk either, as words are category names, and Funes is not supposed to have them: just infinite, infinitely different individual instants (not even individuals).
DeleteBorges was not just a giant as a novelist, but as a cognitive scientist.
I do understand the logic behind why it is argued that “discrimination is independent of identification”, but I do have some concerns in terms of fully committing to this idea. “To be able to discriminate is to able to judge whether two inputs are the same or different, and, if different, how different they are”. In order to tell “how different they are”, it seems that we may need to know “how they are different” – it does not have to be exact, precise, or “correct” if there is “a correct way”, it can just be the iconic and categorical representations. Yes, these representations may not all have a “name”, but it does not mean we cannot assign a name to them, so the question is: Does this count as identification?
ReplyDeleteBetter to think of them as iconic projections. You just have to compare the projections. They have no names.
Delete“To be grounded, the symbol system would have to be augmented with nonsymbolic, sensorimotor capacities -- the capacity to interact autonomously with that world of objects, events, actions, properties and states that its symbols are systematically interpretable (by us) as referring to. It would have to be able to pick out the referents of its symbols, and its sensorimotor interactions with the world would have to fit coherently with the symbols' interpretations.”
ReplyDeleteI wonder how abstract words (e.g. ‘freedom’, ‘consciousness’, ‘love’, ‘moral’ etc.), which lack a clear external referent, could become grounded and meaningful to us. Do we construct the meaning of a word like ‘freedom’ by amalgamating already grounded words that are part of its definition, which, put together equate the meaning of ‘freedom’? That would be similar to how we look up the definition of a word we’ve never come across before in the dictionary (assuming that the definition consists of already grounded words). Given that ‘meaning = grounding + feeling’, it seems like there is also a feeling component that would contribute to how a word like ‘freedom’ gets its meaning. When we feel/experience situations that others describe as ‘freedom’ do we learn to associate that word to how it feels to experience it?
What about idioms like ‘the coast is clear’ where the meaning of the sentence is not predictable from the meanings of the words that constitute it?
Google "peekaboo unicorn" to see how abstract categories can be grounded without any direct sensorimotor experience."
DeleteIdioms are understood by convention, not context.
I am slightly confused by the end of the article that was proposed for this week's reading. Prof Harnad establishes a first condition to meaning, which is that the thing that is learning has to have sensorimotor properties. From what I understood, in the article, that's what the paper and the computer are lacking to attach meaning to symbols. Then, the professor states that he also believes consciousness has a role to play in meaning, and that’s what’s bothering me because I don’t quite see where the distinction is drawn between those two properties, because I believe we talked in class about "being conscious" as being the exact same thing as "feeling". My question is the following: Is there any difference between consciousness and feeling, or is there rather a difference between sensorimotor features and feeling? I would love to hear what you have to say about that, especially the latter question. I think I see sensorimotor abilities and the capacity of feeling as the same thing because of, among other reasons, the other minds problem and our class discussions about kicking-or-not-kicking Gabe because he can /sense/ and therefore presumably feel?
ReplyDelete1. Consciousness is synonymous with feeling. Consciousness is just one of many weasel-words for feeling.
Delete2. Sensorimotor function is, in reality, felt, in all organisms with nervous systems, but, because of the other-minds problem, we can't be sure (especially in T3s) and, because of the hard problem, we cannot explain how or why sensorimotor function is felt.
3. For T3 purposes, sensorimotor grounding merely has to take place, whether felt or unfelt. Felt is extra.
4. Hence Meaning = (T3) Grounding plus Feeling. Not just Grounding.
The necessary presence of sensorimotor capacities operating as the logical grounding mechanism for meaning is an incredibly interesting observation in my opinion.
ReplyDelete“So ultimately grounding has to be sensorimotor, to avoid infinite regress” (Harnard 2007)
From what I understand, the tangible ability to interact with a potential referent exists as the main method which allows for the construction/mediate of meaning but is not in it of itself sufficient for meaning to occur; hence the reference to a Zombie.
What I do not entirely understand is the role consciousness plays within the construction of meaning (a sentiment I imagine also felt by many others). If symbols can be grounded without infinite regress in sensorimotor capacities, and then resultantly can be operated upon and classified in relation to each other, then couldn’t T2, and perhaps even T3 be achieved without the presence of consciousness, and therefore prove as futile means of replicating/explaining human cognitive experience? Or perhaps, am I missing the essential(?) role (that we are mostly unsure about) that consciousness plays in organizing grounded symbols?
When we say things we mean and hear things we understand, we are not "speaking in tongues." It feels like something to understand; and it feels like something else to say or hear gibberish. We are not Siri -- not even (and this is the point) at grounded T3 scale. (Trouble is, we have no idea how or why feeling makes any difference.)
Delete“To be grounded, the symbol system would have to be augmented with nonsymbolic, sensorimotor capacities – the capacity to interact autonomously with that world of objects, events, actions, properties and states that its symbols are systematically interpretable (by us) as referring to.” (Harnad, 2007)
ReplyDeleteAfter reading this paper, I realized that this symbol grounding problem has implicitly been an essential part of our discussions during the last few lectures. About whether a robot can understand, we discussed how the symbol manipulation alone was not sufficient for understanding. And this statement of insufficiency was related to the fact that, understanding required more than manipulation, it required knowledge of the symbols that we are manipulating.
Then as I was reading the commentaries, I realized Professor Harnad gave the following expression: Meaning = grounding + feeling. I initially thought after reading the paper that if we were able to somehow create grounding and manipulation in a machine, that machine would have the ability to understand.
However, this expression (Meaning = grounding + feeling) states that there is more to understanding then just the grounding: we need feeling. But, it makes me question if feeling is even a thing to its own. Let me try to explain what I think:
We discussed quite a bit about feeling, and indicated that feeling was an essential component to a human, so brain cannot be all computations. Sure, when you do something, it is not just about you doing that thing as a physical execution, but there is a feeling associated to doing it. When you read a book, there is a feeling associated to it. BUT, at this point I think we are lost in our definition of the feeling.
To emphasize, I agree with the fact that it feels like to do something. But at the same time, how can we possibly “decide” that what we referred to as feeling is not just a side-product/side-effect of a brain having the combination of symbol manipulation, symbol grounding (ie. symbol manipulation + symbol grounding). I think obviously we cannot make a decision about this, since we cannot isolate either side and see if meaning arises. (we cannot have a human without feeling who has symbol groundings, and we cannot have a human with feeling states, but not symbol groundings)
I am inclined to say that, feeling is not sufficient nor necessary for meaning. But I want to propose: meaning = “using” symbol groundings. At the moment, I think that groundings are robust enough to lead to meanings. Sure, symbol groundings can exist without them being used (we don’t necessarily have all symbol groundings ready to be used in our heads), but what creates the meaning is just those groundings being in use, and explaining this “using” with “feeling” doesn’t help us in understanding how the meaning arises.
Onur: "what creates the meaning is just those groundings being in use, and explaining this “using” with “feeling” doesn’t help us in understanding how the meaning arises."
DeleteCorrect. And yet they are felt, we all know that, and yet have no idea how or why.
Onur: "how can we possibly “decide” that what we referred to as feeling is not just a side-product/side-effect of a brain having the combination of symbol manipulation, symbol grounding (ie. symbol manipulation + symbol grounding). I think obviously we cannot make a decision about this"
Welcome to the hard problem!
The idea of discrimination and identification and hence the idea of iconic and categorical representations behind them are pretty intriguing to me.
ReplyDelete"our ability to discriminate inputs depends on our forming "iconic representations" of them (Harnad 1987b) ... what about identifying them? Discrimination is independent of identification. ... For identification, icons must be selectively reduced to those "invariant features" of the sensory projection that will reliably distinguish a member of a category from any nonmembers with which it could be confused. Let us call the output of this category-specific feature detector the "categorical representation"."
This states that discrimination is independent of identification, which I totally agree with, since we don't need much of processing to distinguish two different or even similar things. But it seems to me that discrimination is the fundamental of identification, because in order to ultimately categorize a member into a category, "icons must be reduced to those invariant features" first, which requires the ability to discriminate everything from the others so as to extract the essential features of this specific category that this member belongs to.
Therefore it immediately leads me to the question that exactly how many features are needed for the categorization? In other words, when we try to discriminate two things, how many similarities and/or differences do we have to see in there when we decide to identify them as members of the same category or two separate categories? I could be wrong but as I'm writing this, I start to wonder that if the answer is, at least tangentially, related to the size of the minimal grounding set which is 1000 words as we discussed in class. Because within this set, everything cannot be defined in, but everything outside can be defined out by this set, does this entail the property that nothing inside this set shares the same category? If this is the case, it seems to me that the similarities and distinctions between categories should have a key impact on the size of this minimal set.
Discrimination starts as independent of categorization, but as we will learn next week, learning to categorize things (sometimes) changes how discriminable they are, making members of different categories more discriminable than before learning the category ("between-category separation"), and (sometimes) making members of the same category less discriminable than before ("within-category compression"). It is caused by the feature abstraction in category learning, increasing the weight of category-distinguishing features and reducing the weight of category-irrelevant features. This feature filter helps make the members of the grounded category "pop out" ("categorical perception"). It is analogous to a filter that reduces the dimensionality of a vector space by removing "irrelevant" dimensions.
DeleteCategorization is approximate. The distinguishing features are underdetermined. You may think you already know a category but then you discover that there are still candidates that you miscategorize, so you have to revise and update the features, tighten the approximation (still always approximate and underdetermined). This is true whether the category was grounded directly through sensorimotor learning by your brain's mechanism for abstracting features, or the category was learned indirectly, from a grounded verbal decription or definition (still always approximate and underdetermined) composed of already grounded words re-combined in subject/predicate propositions naming and describing the requisite features.
There is no way of knowing in advance how many features are needed to learn a category (let alone all possible categories!). And the features always remain approximate and underdetermined, so they may need to be increased or modified).
If there is any connection with the words in minimal-grounding sets, it is quite remote, especially since there are so many possible MINSETs. (But it's a good question.) Yes, probably two words for the same category (perhaps based on different features) are unlikely to be in the same MINSET (by analogy with collinearity in finding bases for vector spaces); nor the words that, together, would define one of the words in the MINSET.
"Connectionism, with its general pattern learning capability, seems to be one natural candidate (though there may well be others): Icons, paired with feedback indicating their names, could be processed by a connectionist network that learns to identify icons correctly from the sample of confusable alternatives it has encountered by dynamically adjusting the weights of the features and feature combinations that are reliably associated with the names in a way that (provisionally) resolves the confusion, thereby reducing the icons to the invariant (confusion-resolving) features of the category to which they are assigned."
ReplyDeleteThis part of description about how to find the categorical representations sounds a lot like the basic machine learning mechanism to me. However, as it was pointed out just several paragraphs above this, the categorical representations are "sensory and nonsymbolic", while computers and computer programs are purely formal and symbolic. Thus I'm a bit confused of how the machine learning categorization process - in essence, it's about feature extraction and assigning different weights to those features by trial and error, precisely as described above - reconcile with our nonsymbolic symbol grounding process. Is that what the machine is doing in no way resembles our symbol grounding procedure?
Feature extraction can be done by neural nets that are either analog (parallel, distributed units) or computational (a formal learning algorithm).
DeleteThe inputs to analog nets can be analog too, but for computational nets they have to be digitized.
Most of the nets in the literature are computational.
Following yesterday's class I had a question about the Whorf Hypothesis' and effects on language. The 'Strong Whorf Hypothesis' states that language *determines* how things look/seem and the way we think. In class we said that this was wrong - you would not see more colours in the rainbow if you knew more colour words. Additionally, there is the 'Weak Whorf Hypothesis' which states that language *influences* how things look/seem and how we think.
ReplyDeleteMy question is what about feeling? I was wondering if the language we use determines how we *feel*? Does the language have an effect in feeling at all? A few years ago I read a paper (Sex and Death in the Rational World of Defense Intellectuals - Carol Cohn) on the language used in nuclear strategic analysis and how the language used in those nuclear intellectual spheres has a large effect on how those individuals felt about the work they're doing on nuclear defense weapons. The language they used was extremely detached and took away all feeling and real life (person/personal) experience which allowed them to not feel any guilt when created and planning weapons which would kill and harms millions. My feeling is that here the language they were using to speak about nuclear strategies and weapons determined how they felt about doing such things. But perhaps this would be closer to a 'Weak Whorf Hypothesis' but I'm not sure.
Of course language can "determine" how you feel. If someone shouts "Run, fire!" that will make you feel panic.
DeleteThe Whorf Hypothesis (WH) is not about that. It is about how a language's vocabulary or syntax effects the way you perceive the world.
It feels like something to perceive: It feels like something to see blue, or see green. According to WH how green and blue look to people whose language has the words "blue" and "green" should be different from the way they look to people whose language has only "bleen." They don't. So that, and just about every other prediction of the WH -- the Hopi have no category of future because their language lacks a future tense; the Chinese could not invent western science because their language lacks a counterfactual conditional ("If Napolean had won at Waterloo..."), the Inuit do not have a richer perception of snow or more snow words, etc. -- have all failed.
And yet next week we will discuss learned categorical perception, a real (but weak) Whorfian effect.
But "feeling" does not refer only or especially to emotion. It feels like something to understand "2 + 2 = 4"
In the class, we talked about language determinism, which claims that the language and its structure would determine how the speaker constructs the worlds. For example, language determines how you perceive and categorize colors. We also talked about the weaker form that stresses the effect of language on human, which is language relativism. Language relativism argues that language has an effect on how the speaker perceives and constructs the world, rather than determination. As mentioned in the class, language determinism is false. But I wonder how would the language relativism could be tested and proved. Because language originates from the certain cultural background and also, in turn, affects the culture. Language and culture seem to perpetuate each other in a certain way. When we test out language relativism, how can we rule out the possibility that the speaker is influenced by the use of their language itself to act in a certain way. It just seems also possible to me that consider that sometimes the culture has a disposition for certain things that are thus reflected in the way language gets created and developed. If that is true, then I cannot tell then how much people can delineate the effect of language on this one from the culture. Is there any way to rule out the possibility? How is the theory of language get tested in a lab or real-world conditions
ReplyDeleteA test of the Weak Whorf/Sapir hypothesis is described in Week 6: Categorical perception induced by category learning. (Categorizing is doing the right thing with the right kind of thing; the right thing to do need not be verbal, but when it is naming the category, it's verbal. So if learning to categorize can alter perception, that's a weak Whorfian effect.
Delete"Even though these artificial systems now autonomously acquire their own methods for grounding concepts (and hence also symbols), it is still the human who sets up the world for the robot, carefully selects the examples and counterexamples, and supplies the symbol systems and conceptualizations of the world by drawing from an existing human language. So the semantics are still coming from us humans. Autonomous concept and symbol acquisition mimicking that of early child language acquisition is a very important step, but it cannot be the only one. The symbol grounding problem is not yet solved by taking this route and I believe it never will." (Steels, 2008)
ReplyDeleteThis seems very reminiscent of the Lady Lovelace Objection where she states that a robot cannot do anything new. Here, it is being argued that a robot is not learning anything new or 'on their own' because the world which they exist in is and the symbol systems they are subjected to are made by humans. We have already seen that the Lady Lovelace Objection is false - a robot *can* generate new things. Similarly, it doesn't matter if the world that the robot lives in and interacts with is created by humans - the robot is still creating meaning autonomously.
"Human representations are clearly incredibly rich and serve multiple purposes. They help us to engage with the world and share world views and feelings with others. Symbols or symbolic representations traditionally have this rich connotation. And I believe the symbol grounding problem can only be said to be solved if we understand, at least in principle, how individuals originate and choose the meanings that they find worthwhile to use as basis for their (symbolic) representations, how perspective may arise, and how the expression of different meanings can be combined to create compositional representations." (Steels, 2008)
I agree that the symbol grounding problem will only be solved once we understand how meanings are "chosen" or grounded in certain representations. However, I'm not sure if it's right to say that individuals choose and originate the meanings which they use as a basis for their (symbolic) representations. This seems to imply that individuals have the freedom to choose what meanings are grounded with what symbols. Additionally, if Meaning = Grounding + Feeling, and robots are able to autonomously ground concepts/symbols and acquire methods for grounding concepts (above quote), and if we can never truly know if anyone or anything else has feeling on account of the other minds problem then wouldn't we have to say that robots also have "rich" representations just as humans do? If a robot is able to ground symbols by means of sensorimotor experience which gives the robot the ability to "detect, categorize and act upon the things that words and sentences refer to" (Harnad, 2003) and if we can *never* know if the robot (or anything else for that matter) has feeling anyways, then can't we say that robots acquire meaning (perhaps not in the exact same way, but nevertheless, they acquire the same level of "rich" meaning) just as we do?
Your comment is spot-on, Flora!
DeleteThe question of whether and what T3 feels is not the hard problem, it's the other-minds problem (OMP).
ReplyDeleteThe hard problem (HP) is to explain how and why anything that feels (whether organisms or robots) feels.
The OMP cannot be solved with certainty, but Turing argues that the Turing Test (with organisms or robots) is as close as we can get (and it's close enough).
The HP may or may not eventually prove solvable; later in the course I will describe the reasons "Stevan Says" it's unsolvable.
I think Steels’s paper (2003) misses the main point about the symbol grounding problem, and then continues to speculate about a deemed solution. I also think that their solution is irrelevant to the symbol grounding problem.
ReplyDeleteSteels: “Even though these artificial systems now autonomously acquire their own methods for grounding concepts (and hence also symbols), it is still the human who sets up the world for the robot, carefully selects the examples and counterexamples, and supplies the symbol systems and conceptualizations of the world by drawing from an existing human language. So the semantics are still coming from us humans. “
Steels’ idea about why “a human-created AI would not be able to have groundings” bothered me. First, this kind of idea is similar to the Lady Lovelace Objection, as Flora Aldridge showed in the previous comment. A machine can, for sure, generate new things; it is not limited to what is given to it.
I also think that Steels’ argument is wrong for another reason. Steels supposes that the fact that an AI is created by humans (rather than generating itself) will lead to that AI not having groundings hence not having the understanding. Moreover, here, Steels specifically investigates how supervised learning will not allow the robot to have grounded meanings.
There are two things to point out here. First even if a machine is given random examples (instead of carefully selected examples that Steels is talking about) xi such that xi belongs to category X and, counterexamples x’j such that x’j doesn’t belong to category X, this machine can still have an ability to discriminate between the elements and non-elements of a category.
So, the reason why this machine will not have the meaning is not about the fact that “it was human created with carefully selected examples.” Here Steels should be talking about “if grounding of category X can be obtained” (not the ability to discriminate) even if this machine is provided with carefully selected or not-carefully selected examples.
The reason why the machine does not have an understanding is the fact that the machine does not have groundings. The machine’s inability to have understanding is not due to the type of examples given to it.
Steels falls into the same misunderstanding about what grounding is when he suggests the following solution to “the problem about supervised learning”:
“If an agent cannot make a distinction between the colour of the topic and the colour of the other samples in the context, it introduces a new prototype and will later progressively adjust or tighten the boundaries of the region around the prototype by changing weights.” (Steels, 2008)
Again, symbol grounding is not about being able to come up with new self-categories and fine-tuning them. It is about having “groundings”, it is about having the meanings for those categories. Being able to fine-tune a self-created category is not sufficient for the meanings to arise.
RE: Funes The Memorious
ReplyDeleteWere it possible to remember (in the way Funes can) each detail of one’s memories, one would not be able to abstract. Borges describes Funes’ capacity for memory such that he Funes can recall exactly what “the southern clouds at dawn on the 30th of April, 1882” looked like. Implicitly, Borges says Funes possesses the ability to say what the clouds looked like whenever he accesses a memory.
Borges has (knowingly) created an impossible situation (at the end of his tale, he says “to think is to forget differences, generalize, make abstractions”): if one were to ask Funes to describe clouds, he simply would not be able to do so. Funes, in all of his infinite memory, would be unable to abstract away from his specific memories of clouds (all the clouds he had seen), to discuss “clouds” generally.
When I ask you to describe clouds without looking at one, you say “a cloud is a billowy white thing that floats above our heads”. The reason you are able to give such a response is because you can go back into your memories, and access some internalized meaning that has been abstracted from all of your previous interactions with clouds.
When you say “cloud” to me, I do not envision a specific thing. Rather, I envision a category of things such that anything resembling a cloud would suffice for the purposes of your relaying meaning.
But, Funes would be unable to do such a thing; Funes cannot categorize: “Locke, in the seventeenth century, postulated (and rejected) an impossible language in which each individual thing, each stone, each bird and each branch, would have its own name; Funes once projected an analogous language, but discarded it because it seemed too general to him, too ambiguous”.
The word cloud has meaning to me because it is grounded in sensorimotor experience. Its meaning is in my head and picks out a specific (categorical) referent. But, Funes’ clouds, and indeed all of the words that Funes utters, are without meaning. There is no meaning of a Funes utterance because such an utterance picks out referents that are specific (grounded) instantiations, rather than abstract (grounded) categories.
While it is true what you’re saying regarding Funes The Memorious, I think Funes should have been able to categorize because, although he could not forget anything, that does not mean that he did not have the capacity, inherent to all humans, to differently weigh invariable features of certain things more so than their other, less important features, so as to discriminate members of that category from non-members of that category. If Funes had judged all features of all things equally, then that, in addition to his ability to remember everything (or inability to forget) would prevent him from categorizating/abstracting.
DeleteBorges message is “to think is to forget”, but the message should have been “to categorize is to forget”. In fact, Funes was categorizing whether he knew it or not. In Borges story, if I remember it correctly, Funes was obsessed with reading, he would ask for books, read them, then recite them. Since categorizing is doing the right thing with the right kind of thing, then Funes should not have been able to read because that is doing the right thing with the right kind of thing. He should not have even been able to distinguish capital letters from lowercase letters. Although, I hope I am not splitting hairs here, because everything you have written regarding Funes is correct, in the sense that it was the intended meaning of Borges as he wrote it.
A discussion last class sparked a question in my mind about the nature of symbols that must be grounded. Specifically, what can be grounded, apart from linguistic symbols (words)?
ReplyDeleteEdwin Hutchins, in 1995, wrote a paper called “How a Cockpit Remembers Its Speeds”. In his paper, Hutchins describes the ways in which airline pilots land planes, which (to no one’s surprise) turns out to be quite challenging. Pilots must keep track of their current speed, and maneuver the plane’s “flaps” and “slats” at the correct times to safely land. To make the pilot’s job easier, pilots place little indicators (“speed bugs”) on their speed dials, which mark the speeds at which “flaps” and “slats” need to be moved in order to avoid problems with landing.
Hutchins uses his 1995 paper in order to articulate an idea called “extended cognition”. Briefly, Hutchins says that the “speed bugs” take the place of having to actively remember the speeds at which the plane is at risk: the “speed bugs” reduce the cognitive load of the pilot. Thus, cognition does not occur solely within the brain; the part of the cognition that is responsible for remembering when the pilot has to move that “flap” or “slat” is being done by indicators on the pilot’s speed dial.
To these pilots, “speed bugs” have meaning and referent: each speed bug picks out the specific thing the pilot must do. But, to nonexperts like me (before having read Hutchins’ paper), these bugs would be meaningless – I would not understand what they refered to (in Searle’s terms, they would be as helpful as the Chinese squiggles and squoggles).
It appears that these speed bugs are like words on a page: they have meaning *only* if someone understands its referent (e.g. when I see the orange speed bug, I know that I need to move this lever down).
Is knowing what “speed bugs” do an example of categorization – of doing the right thing with the right kind of thing? Or is it an example of symbol grounding?
I think I understand what Paul is talking about and while I am not sure in saying this, I do not think that the speed bugs would be an example of either categorization or symbol grounding. I understand the analogy that he’s made between speed bugs and symbols for language such as Chinese in Searle’s Chinese room experiment, however, I don’t think this would qualify as cognition, cognition is (in the view of computationalists) just symbol manipulation, and while the speed bug may be interpreted as a symbol, it is not necessarily a part of a symbol manipulation system on its own.
DeleteThe idea of “extended cognition” seems to state that if any external thing is used to reduce cognitive load, then it would be considered as “extended cognition”, however, by this definition, a piece of paper would qualify as “extended cognition” if we were to use it to calculate mathematical equations. If cognition is just symbol manipulation based on rules, then reading the speed bug would only be symbol manipulation in the person’s head, based on rules in his head. The symbol may be external, but should not be considered cognition, no more than a book with words [symbols], whereby a reader must systematically manipulate external symbols, should be considered “extended cognition”.
Furthermore, I’m not certain that these speed bugs even have a referent and meaning in the same sense that a word does. I’m not certain what these speed bugs look like, but it seems like it may just be passive association rather than actual categorical perception, similar to classical conditioning whereby the speed at which the speed bug is would be associated with danger, however I may be wrong about this.
This comment has been removed by the author.
ReplyDeleteThis comment has been removed by the author.
ReplyDeleteRe: “Otherwise, I feel like there are loads of vegetative activities which would count as categorisation (e.g. motor activities)”
DeletePylyshyn (1999) points to cognitive impenetrability to differentiate between cognitive and vegetative functions. If goals and beliefs can shape your behaviour, then it’s a cognitive function, if not, then it’s a vegetative function.
To get back to your comment and following Pylyshyn’s argument, categorizing is cognitive because it impacts subsequent behaviour while vegetative activities are not cognitive because they remain those behaviours (including motor activities) will remain unchanged by categorizing.
“Perhaps symbol grounding (i.e., robotic TT capacity) is enough to ensure that conscious meaning is present too, perhaps not. But in either case, there is no way we can hope to be any the wiser -- and that is Turing's methodological point”
ReplyDeleteI agree with the premise that we can never really know whether simply grounding symbols with the help of sensorimotor systems is enough to truly understand it’s meaning or if something else (i.e. consciousness is required), but I do find it almost paradoxical too ask if consciousness would be required to have conscious meaning of words. This goes back to the other minds problem, which cannot be solved. If we can’t know whether other people are conscious, we would be equally unable to know whether a TT passing robot would be conscious.
However, to me, this seems like it’s a problem relating to how consciousness comes to be. If it were just a by-product of sophisticated brain activity (computation) coupled with sensorimotor systems, then it would be reasonable to expect symbol grounding to be sufficient for conscious meaning, but if consciousness arises some other way, not in the spiritual way, but through some biological mechanism we are unaware of, then it would be an acceptable conclusion that symbol grounding is not enough to have conscious meaning present.
Response to The Symbol Grounding Problem (Harnard, 1990):
ReplyDeleteWhat stood out to me was the differentiation Harnard makes between what underlies our human capacity to identify and discriminate. Specifically, the ability to discriminate involves making relative judgements on the sameness or differences between internal iconic representations of sensory input. Identification, or the assigning of unique arbitrary 'label' or 'name' to an object, instead relies on categorical representation where icons are compared to invariant features of a member of a specific category. From my understanding, it seems that categorical representation plays a more central role in understanding how we are able to connect a symbol to its referent (the symbol grounding problem). This especially came to my mind when trying to extrapolate how iconic and categorical representation plays a role in how our ability to manipulate, describe, produce and respond to descriptions of sensory inputs. Particularly, it seems that being able to name something via assigning it as a member of a category is a precursor to these capacities given that it comes along with it, associations with characteristics defining that category and outlines relation to other members / non-members. However, I do acknowledge that discrimination may be applied context-dependently as well. While reading this text, I also agree with Harnard that just uncovering mechanisms underlying grounding is insufficient to understand all of 'meaning'. Specifically since meaning also involves the feeling of knowing. It is clear that the hybrid connectionist model and categorical representation alone cannot answer this. Thinking of the other-minds problem however, it seems at the moment that it is not possible to ever know whether someone else (like we know we do individually) can truly also feel that they know.
Response to Barsalou (2010):
ReplyDeleteBarsalou gives an overview of the trajectories of research into grounded cognition in the past, present and future. In outlining the pitfalls of the current state of research, he highlights that there is an overemphasis on demonstration experiments (e.g., TMS effects) for evidence versus the existence of actual theoretical and computational frameworks. Along with this, he argues that using the top-down approach to using the brain to explain grounded cognition is difficult. Despite this however, Barsalou, unlike Harnard and Fodor, does not discredit entirely the meaningfulness of looking into the brain (or the when and what) to understand cognition, especially how we ground symbols to its referents. This is shown as he suggests that in the future, "environment, situation, bodies and simulations will be increasingly integrated and central [in grounded cognition theories]". Along with this, he also argues that, "If cognition heavily utilizes mechanisms for perception and action, then grounded accounts have potential to unify perception, action and cognition in the brain". While I can agree a majority of Barsalou's conclusions [in accordance to (Harnard, 1990)], such as that statistical and dynamic approaches will be increasingly integrated to how we understand grounding mechanisms in the future, I disagree that continuing to attempt to explain grounded cognition in terms of the brain (or consolidating considerations of the brain with computational/dynamical approaches) is useful. Not only may this lead to overdeterminance but as Harnard (1990) suggests, is an 'immature' prospect. This is because we do not yet know what 'brain-like' states of grounded cognition mean or look like (this applies to a majority of cognition as a matter of a fact).
Re: The Symbol Grounding Problem
ReplyDeleteT3 has sensorimotor capabilities, so if it was built in such a way that required activation of that system, then it would be grounding symbols. I think Harnad does a great job of demonstrating that grounding would be necessary to pass T3. We know that meaning requires grounding, but it also possibly requires consciousness which brings it back to the other minds problem. We know that T3 cannot be computational, so if there was some way of knowing that T3's had conscious mental states in addition to their ability to ground symbols then I think that T3 robots could cognize. But how could we possibly know that another thing is conscious? Searle's Chinese Room Argument is a computational system which we know would not understand meaning, but a non-computational T3 robot using some other means of intelligence might be able to cognize, but I think that we would probably just need to assume it does because it passes the TT. I am starting to see a general pattern in these cognition readings; a definition of a concept is first presented and then dismissed as incomplete.
As far as I understand it, the article on the symbol grounding problem how a T3 grounds arbitrary symbols because its interaction with the external world can connect symbols to their referents. Thus, a robot that is both capable of symbol manipulation (computation) and has a sensorimotor system is a necessary condition for groundedness. However, it is an insufficient condition given that referents do not represent the meanings of words. For this, the mind is required both to identify referents and to then know the rule for picking out the right referent for the symbol it relates to. The consequence of this argument is that it refutes computationalism because, if a brain is required in order to provide meaning, then cognition cannot be implementation-independent. Computation would lack the essential component of understanding when implemented by any other hardware than the mind.
ReplyDeleteRe: “For this, the mind is required both to identify referents and to then know the rule for picking out the right referent for the symbol it relates to. The consequence of this argument is that it refutes computationalism because, if a brain is required in order to provide meaning, then cognition cannot be implementation-independent.”
ReplyDeleteThis argument does not automatically and necessarily reject computationalism. If you follow the idea of strong AI wherein a correctly programmed computer really *is* the mind, then you could have a mind capable of picking a referent, and that would be the result of computation. I personally don’t believe in the claims of strong AI and computationalism but the argument you make about refuting is (to me at least) an over-reach.
“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.”
ReplyDeleteThe strong Whorf hypothesis states that language shapes how we view the world, meaning that previously learnt categorical representation influence our perception of newly learnt categories. Would this hypothesis apply to some extent to T3 sensorimotor system able to ground higher order symbols to elementary ones? Would it be biased by what it already knows and thus form wrong categories? Furthermore, how would it make the difference between grounding what is true from what is false? For example, let’s say a robot learns from previous association that Muslim=terrorist. Would it generalize this learning to any new member of the category Muslim? Where do the line is drawn and how is the distinction is made between something that is generally applicable and something that is not?
When talking about “The means of Picking out Referents” I don’t understand why in their definitions we have to separate the “meanings” from their “referents” since a “meaning” will never be able to be what it is without the referents. From what I understand, the meanings are dependent on their referents. Since, every word we eventually learn in our lives come from the outside they are automatically referents and meanings at the same time.
ReplyDeleteThe symbol grounding problem is about how words get their meaning and how those in turn get related to mental states. Following from computationalism, since all cognition is computation then language is as well, making it a symbol manipulation based on rules. Finding rules for every word’s retrieval seems like an impossible task. The argument here is that we don’t need to know the specific rules for words but instead just the way in which the brain uses and follows these rules.
ReplyDeleteThe issue of consciousness comes in when we try to introspectively search for the ways in which the brain comes to these conclusions.
To me it seems like grounding and the circular dictionary model are a good example of issues encountered when attempting to deconstruct the human brain. Every concept is created based on a web of definitions, previously existing concepts and theories. Without understanding one well, the rest become useless or unattainable. The minimal ground set should be seen as those few concepts we must first get a grasp on before being able to build on their foundations to truly understand how the brain works in order to engineer a similar machine.
(This is a pre-midterm skywriting that I didn't actually post but somehow included in midterm's skywrting file.)
ReplyDeleteIn Week 5, we discussed the innateness of categorization. I think my wording was a bit misleading. What I meant was not “if the category ‘chair’ etc. is innate”; what I meant was “if the ability to categorize is innate”. That is the reason why I felt that after I listened to your response (“doing the right thing with the right kind of thing”), I still feel that it is innate – no one told the (poor) kid to do the right thing with the right kind of thing. Also, I think it would make sense if the ability to categorize is innate because just by perceiving the world, the way we achieve that is categorized into “by seeing, smelling, touching, hearing ...etc.”. Although, another question arises, does it necessarily mean that the ability to categorize our perception will be generalized to everything else?
Another question that I had was about referent. It is said that functional words do not have a referent but only a syntactic role. I think in human language, there is nothing that is purely syntactic, and if it is semantically interpretable, it has a meaning, and if it has a meaning, why cannnot it have a referent? It may not have been a *concrete* referent, like “apple”, (but what referent is really concrete anyways?), but the referent would be “the feeling of using ‘and’” or “or” or past tense marker… It feels like something to “and” something – and in English, different “and’s” mean different things:
ReplyDelete“John and Mary are good friends.”
“Keep talking to me like this and I will report you to the police.”
“Go by air and you will save time.”
Another way to put this, I think, is to say that for the functional words, the "referent" is the abstracted context when the functional words are being used.
I agree that no word or part or combination is "purely" syntactic ("squiggles") in natural language. There is something semantic about if, not, and, for...
DeleteBut a referent is a category, something you can "point" to and name, and do the right or wrong thing with. I can see how we point to members and non-members of very abstract categories such as "justice" or "differentiable," but not how to point an "and" or an "or." I think that is more like a procedural category rather than a semantic one: We do this with an "and" and this with an "or."
But maybe the semantic/procedural category distinction breaks down at a fuzzy boundary (as "vegetative" vs "cognitive" does). After all, doing is doing, whether naming something or manipulating it. But it's odd to think of an "or" as a thing. "Or'ing" is an act, just like "running" is, but the referent there is the action, not the "object."
There's room for more thinking about this.