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.
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:
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)
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)
NOTE TO EVERYONE: Before posting, please always read the other commentaries in the thread (and especially my replies) so you don't just repeat the same thing.
ReplyDelete**BLOGGER BUG**: ONCE THE NUMBER OF COMMENTS REACHES 200 OR MORE {see the count, just above, left] YOU CAN STILL MAKE COMMENTS, BUT TO SEE YOUR COMMENT AFTER YOU HAVE PUBLISHED IT YOU NEED TO SCROLL DOWN TO ALMOST THE BOTTOM OF THE PAGE and click: “Load more…”
Delete________________
Load more…
________________
——
After 200 has been exceeded EVERYONE has to scroll down and click “Load more” each time they want to see all the posts (not just the first 200), and they also have to do that whenever they want to add another comment or reply after 200 has been exceeded.
If you post your comment really late, I won’t see it, and you have to email me the link so I can find it. Copy/Paste it from the top of your published comment, as it appears right after your name, just as you do when you email me your full set of copy-pasted commentaries before the mid-term and before the final.
——
WEEK 5: Week 5 is an important week and topic. There is only one topic thread, but please read at least two of the readings, and do at least two skies. I hope Week 5 will be the only week in which we have the 200+ overflow problem, because there are twice the usual number of commentaries: 88 skies + 88 skies + my 176 replies = 352!. In every other week it’s 2 separate topic threads, each with 88 skies plus my 88 replies (plus room for a few follow-ups when I ask questions.
I found this reading to be very kid-Sib friendly and I enjoyed how it succinctly summarized the material we have discussed in the past month. I was unsure about what the Symbol-Grounding Problem is during the first week and reading the Wikipedia article only confused me further. Reading the Scholarpedia article today had a different impact; I think I have a basic understanding of this problem! Put simply, it explores the relationship between a symbol representing a thing, someone interpreting the symbol and understanding that the symbol refers to the thing in question. The article also explored the idea of consciousness (i.e. our capacity to ground the symbols to their referent).
ReplyDeleteIn Linguistics, we say that the meaning of complex expressions arise from the meaning of the constituent expressions. I suppose that grounding is what allows us to get the meaning in the first place, but I find it interesting that the Symbol Grounding Problem lightly touches on how we connect syntax and semantics (my favourite part of Ling).
"Symbol Grounding" refers to the causal connection between a (content)-word in someone's head and the object to which the word refers. This is what a purely computational T2-passer or ChatGPT lacks whereas a sensorimotor robot like Anais has. She can not only tell you, as ChatGPT can, that an apple is a round, red (or green or yellow) fruit that grows on trees, and you can pick it and eat it. She can also recognize an apple from seeing, and she can DO what we do with apples (and not with red billiard balls). ChatGPT can say all of that, but not do it. (And the solution is not just to add a camera and wheels to a chatbot.)
DeleteI really like this week’s topic about symbol grounding as it has always had its place in my mind. I have wondered many times about how we arrived to the complex symbol system we have now. Furthermore, I thought it was strange that the readings did not touch on the evolution of the symbol system we have now, from drawings to complex alphabets. Surely that would shed to some light to the causal connection between the content-word and the object.
DeleteI have a question about Anais: is she “conscious” that it is an apple? Can recognize an apple that does not fit at all the normal standard of the fruit?
Some people think language began with gesture, some that it began with speech, but no one has suggested that it began with drawing (which is really just a more restricted form of gesture). If gesture is like pantomiming in playing charades, drawing would be more like playing Pictionary.
DeleteBut guessing movie titles and words from pictures is not likely to be the adaptive advantage that gave rise to either gestural communication or language.
The reason theories of the origin of language were banned by the Société linguistique de Paris in the mid-19th century was was that they were too speculative. Can you think of a credible adaptive progression from drawing to language (without presupposing language!)?
I have a interesting example, which is Oracle bone script. It is an ancient form of Chinese characters that were engraved on animal bones or turtle plastrons used in pyromantic divination for long-term recording of event. Each character of this ancient-form Chinese is a simple painting with line, dot, circle, and etc. For instance, the sun is just a circle in the most ancient form. Throughout time, as the character increases, for convenience and distinguishing each from other texts - like avoiding the confusion between the sun and the real circle - Chinese characters have evolved into more complex "squiggles and squoggles".
DeletePosted by miriam hotterSeptember 30, 2023 at 5:44 PM
Delete"GPT can't be based on symbol grounding. It can only make statistical associations and lacks the ability to have sensory experiences and interact with the external environment. With the help of massive amounts of textual data, GPT simply makes very educated guesses about meanings and words by analyzing patterns to generate texts we typically consider coherent."
Reply
But no one has yet given an explanation of why GPT does so well..
Evelyn, bone script is about 1200 years old. Writing was invented after language evolved. Language evolved at least 10,000 years ago (probably earlier -- tens of thousands of years ago). Evolution means an adaptive change in the genes and (in the case of language) in the brain.
DeleteBut even if anais is able to eat and recognise the apple and do what we do with it how do we know that Anais really does understand the meaning if we’re not in her head. Isn’t she programmed to do what we can do with an apple?
DeleteI do not think that it is possible to come up with a credible explanation for the origin of language through drawing. First of all, as professor Harnad mentioned, we know that writing was invented around 5400 years ago, whereas language—in general, in the mind—could have evolved as long as 150,000 years ago (I have read that language evolved at least 50,000 years ago, but I'm interested in whether or not some experts believe it was less than 50,000 and if so why?). Either way, spoken language was certainly not invented; there is no evidence to the contrary, and most experts nowadays believe that gestural language was not invented either. What would it mean for "drawing" (distinct from "writing") to have been the firs instance of externalization of language? Whereas able-bodied modern humans are genetically endowed with the means to produce speech and gesture, drawing always requires something detached from our bodies to either draw on, draw with (if one would like to posit that what we "draw on" could be our bodies themselves?), or often both. It seems like it would be a huge task for evolution to take us there first, before speech and/or gesture. Further, infants display an innate tendency to babble and/or motion with their hands, but drawing seems to not come about until the child is given, for example, a crayon and paper and encouraged to put the crayon to the paper.
DeleteMarine, you are talking about the other-minds problem. Turing's point is that once you cannot tell the candidate's performance apart from that of any other normal human being, you have no more basis for worrying about the other-minds problem than with real normal human beings. You can't ask for more. (See the replies in this and other threads about "underdetermination."
DeleteJordan, excellent comment.
The dictionary-go-round example was really useful for me to further understand symbol grounding as it makes clear exactly how we must use symbol grounding to connect words and referents together. Specifically, the second version of this example, describes how attempting to learn Chinese as a first language from only the use of a Chinese/Chinese dictionary would prove to be essentially impossible because you have no way to assign any meaning to the symbols as you wouldn’t know the other symbols used to describe it. Thus, resulting in an endless cycle of trying to understand the language without any way to know what any of it means. This is because you have no possible way of grounding these symbols in meaning from previous knowledge or experience.
ReplyDeleteAnd yet, with a Text Gulp much bigger than a dictionary, GPT manages to get a lot of mileage: How?
DeleteGPT can't be based on symbol grounding. It can only make statistical associations and lacks the ability to have sensory experiences and interact with the external environment. With the help of massive amounts of textual data, GPT simply makes very educated guesses about meanings and words by analyzing patterns to generate texts we typically consider coherent.
Delete(But no one has yet given an explanation of why GPT does so well..,)
DeleteMy guess would be that it is able to do this so well because rather than getting a huge list of words and their corresponding meanings, it gets those words in millions of different contexts. We talked last week about how it can infer information using context in a given scenario, and this could be related to it consuming words in relation to each other, not just in isolation with their definitions. In the example of a woman reaching for an apple after a day at work with no breaks, chat GPT might make the connections that stories, laws, art, articles, etc. about work involve lunch breaks, and when people do not eat, they feel hungry. ChatGPT then is able to say she wants to eat the apple, rather than simply saying "she reaches for the apple to hold it and bring it closer to her," which is not technically wrong but misses the point. There is probably a lot more to this, but maybe that's a start?
DeleteAdrienne, good start, but not yet the whole story.
DeleteIs it that chat GPT’s big gulp fed it many many symbols that are grounded in the brains of the humans who wrote them?
DeleteAdditionally, GPT learns from its mistakes when individuals with enormous tool-belts of grounded symbols/words give it corrections.
So GPT’s inputs are grounded by others, and its outputs are verified by others for whom the outputs are grounded. This means it almost does not matter if chatGPT is grounded, because others are doing it for GPT. This is a possible explanation for why GPT does so well.
Nicole, all good points. But grounding matters if you are trying to reverse-engineer thinking rather than just a useful reference tool for thinkers.
DeleteIs there a way to maybe pin this comment?
ReplyDeleteTHE COMMENT YOU ARE REFERRING TO HAS BEEN DELETED, TWICE, BY blogger BUT I DON'T KNOW WHY. HERE IT IS IS AGAIN: THIS IS AN IMPORTANT WEEK AND TOPIC. THERE IS ONLY ONE TOPIC THREAD, BUT PLEASE READ AT LEAST TWO OF THE READINGS, AND DO AT LEAST TWO SKIES.
DeleteSide note: am I able to use "refer" or "reference" under this context?
ReplyDeleteAs for Harnad's 2003 paper, I concur that this paper is much more straightforward than the Wikipedia page.
ReplyDeleteThis also gave me a lot more clarity on the nature of ChatGPT and what we were talking about with “statistical parrots”. While it is able to use statistical analysis to retrieve and summarize the vast amount of information it has been fed, it still lacks many capabilities because it cannot understand what these things mean or interact with them in a real life context. The question then is how to implement grounding in a non-human system. To do so, we would need to establish feature-detecting mechanisms in the system that can detect the sensorimotor features of the environment. Then, we would need something to entrain the system on how to select proper referents. To do so, we would first need to establish what the correct way to do that is, but I believe that is feasible within the realm of cognitive science. The other property essential to symbol grounding discussed by Harnard is consciousness, where we then once again run into the hard problem. However, speaking in terms of a T3-Turing test, it may be enough for the system to be able to do what we do, without knowing how.
The capacity to learn to detect the sensorimotor features that distinguish the members from the nonmembers of a category, is sensorimotor motor capacity. That's not an "add-on" to a T2 Chatbot. It's a T3 robot. What would give T3 the capacity to understand would require both (1) penetrating the other-minds barrier to verify that T3 really understands (which is impossible, because Searle's "Periscope" works only for pure computation: why?) and then (2) solving the Hard Problem of explaining how and why it feels understanding rather than just being able to do everything need to pass T3.
DeleteDistinguishing between T2 Chatbots and T3 robots is crucial, as highlighted in "The Symbol Grounding Problem." The introduction of grounding in non-human systems is not just an enhancement but a transformative leap, with T3 robots defined by their ability to detect sensorimotor features. Two significant challenges emerge in this context. The first is the intricate endeavor to penetrate the "other-minds barrier" and validate genuine understanding. The second is addressing the "Hard Problem" of consciousness, which delves deep into the nature of subjective comprehension. Searle's "Periscope" suggests that computational capabilities alone cannot capture the essence of this subjective experience. Thus, acing a T3-Turing test doesn't necessarily equate to authentic understanding.
DeleteThe info is parroted but the voice is GPT's: "crucial… not just an enhancement but a transformative leap… Two significant challenges emerge in this context. The first is the intricate endeavor… delves deep into the nature of subjective comprehension…" This is not what I meant when I said to use ChatGPT.
DeleteProfessor Harnad’s article “The Symbol Grounding Problem” touches on a question I have had throughout the duration of this course—if cognition is not just computation, what is that additional ‘thing’ that is going on inside our heads that makes it so we feel like we understand? Searle’s CRA highlights that simple computation may result in indistinguishable output, such that it appears one understands, but it lacks this elusive quality of feeling like one understands. Through the CRA Searle shows that computation alone does not result in our feeling of understanding—so what does? This article puts forward symbol grounding as part of the answer.
ReplyDeleteIn this article Professor Harnad clarifies that a symbol system (“a set of symbols and syntactic rules for manipulating them on the basis of their shapes”) on its own does not have the capacity for grounding (the ability for a symbol to pick out their referents). Grounding therefore is not purely computational, and is instead “a dynamical (implementation-dependent) property.” Grounding is necessary for meaning, as it allows one to “detect, categorize, identify, and act upon the things that words and sentences refer to.” Thus, the idea of grounded symbols allows us to better understand how “meaningless strings of squiggles become meaningful thoughts” in our brain, as it explains how the symbol system in our mind can be connected to the referents we interact with, and derive meaning from, in the sensory world.
Welcome to the Hard Problem. Solving the Symbol Grounding Problem is just the solution to the Easy Problem.
DeleteThis article establishes a distinction between meaning of words and symbol manipulation under the light of the computationalism hypothesis. As it’s said in the first paragraph, « computation in turn is just symbol manipulation » but if computationalism assumes that brain is computation, then thinking is reduced to a form of symbol manipulation, which brings us back to the importance of the Turing Test in delimiting what is thinking from just answering questions with the most probable word following each other as would ChatGPT do.
ReplyDeleteIf I understood the aim of this article correctly, to distinguish between symbols or shapes and meanings, we need to understand what is the referent of the word. But, since this process is implementation-dependent and must be associated with sensorimotor properties, I don't understand how we could assume that computationalism is true when it’s said to be implementation - independent.
Computation (which is not cogsci but maths, logic and compusci) is implementation-independent.
DeleteSo computationalISM (which is a cogsci theory that cognition is just computation) has to be implementation-independent.
Searle uses this soft underbelly of computationalISM as his Periscope to show that computationalISM is incorrect.
So passing T2 with computation alone (as ChatGPT does) is not enough.
Passing T3 requires sensorimotor capacity.
Sensorimotor capacity is not only computation.
So it is not implementation-independent.
The Harnad 2003 reading opened my eyes to the symbol grounding problem and the reasons why meaning and reference are not the same thing. From my understanding, a word's referent is the thing it refers to, whereas meaning relates to the word’s sense. To better understand this, let’s look at an example where "red" and "flower" are characteristics of roses. The term "rose" can apply to any "red flower." These fall under categories, each of which has a name. As a result, the definition only provides you a new name provided you already know what those names and their categories relate to. However, without previous knowledge of what "red" refers, even looking it up in the a dictionary will not get you far, as it would only lead to a cycle of endless meaningless definitions and here is where the symbol grounding problem comes in. In such a case, looking for meaning would be ungrounded. Here, we see that meaning and reference are not synonymous, and are in fact two very different things.
ReplyDeleteReference and meaning are not the same, but they are related: (1) You can't get reference without sensorimotor grounding, and (2) you can't get meaning (which is a property of strings of grounded category names expressing a proposition [what is that?]) unless the category names in the predicate of the proposition are already grounded (why not?).
DeleteGrounding can be direct (sensorimotor) or indirect (verbal propositions: definitions, descriptions, explanations), but it can't be all indirect all the way down to the ground.)
Meaning (production) and understanding (perception) = grounding plus what it FEELS LIKE to mean or understand a proposition.
Don't ask me why it (or anything) has to feel like something. That problem's too hard for me... (But I'll take a stab at it in Week 10).
I’m not sure if this is the right way to look at this, but a way I can conceptualize the symbol-grounding problem, or even Searle’s argument, is to think of trying to explain what the colour “Red” means to a congenitally blind man (or what an apple tastes like to someone without taste buds, etc. etc). I can tell the man about every item in the world with this colour, or explain it through metaphors and try to relate it to senses experienced in a different modality, but the blind man will never be able to imagine the colour red to the extent that I can. Would this be an example of the argument that we are trying to make against computationalism, and the core of the symbol-grounding problem? From my understanding, we need modality-specific experiences with the world in order to derive a mutual, species-wide meaning for it, and that is what our understanding of arbitrary symbols is based on (and the very thing that computers lack).
DeleteThe Harnad 1990 reading highlighted one of the most interesting notions of the symbol grounding problem in regards to the idea that learning Chinese as a first language requires a great deal more than just using a monolingual Chinese dictionary in order to really understand this language. In my child development class, we saw that babies start learning their native language as early as in the womb! In fact, exposure to the speech of their mother allows the malleable and developing brain of babies to start the learning process of their mother tongue. And as they grow up, they learn to form categories of the world around them, such as forming categories of “objects”, “animals”, etc. With time, these categories become more complex and have subcategories within them. This reading provided me with a deeper understand of this learning process, by highlighting that before even learning what things certain words refer to, categorization of these things from sensorimotor interactions with them comes first and is thus fundamental.
ReplyDeleteI particularly enjoyed the section 1.3 "Connectionist Systems" as a theme that keeps coming up gets explained in a little more detail here. The theme being, do we fully need to understand the brain to understand cognition. This section highlights again that we know very little(although slightly more now) about the brain's vegetative functions, let alone about the brain's higher level functioning. However, the article suggests that this doesn't mean we can't learn more about cognition, and that taking the symbolic model of the mind and/or connectionism will be promising tools in better understanding cognition.
ReplyDeleteTo me, this again begs the question: Do we actually need T4, or does T3 or even T2 suffice?
Hi, I wanted to add on to what you said, because I totally agree with what you mentioned. When we say we want to reverse-engineer the brain, are we realistically aiming for T4, or is T3/T2 enough? In the previous readings, and in class, we looked at criticisms on studies on localisation, which is a part that makes the brain the brain, and it seems to me that we are somewhat away from creating a T4, especially since there are questions raised on localisation studies.
Delete
DeleteI agree that while the need for T3 passing subjects seems evident to successfully say that we have reverse engineered the brain, the necessity of T4 is a little bit more vague. I do think that the symbol grounding problem specifically highlights the importance of T3 being the very minimum standard for reverse engineering the brain. This is because as we know, T2 computers can not pass sensorimotor tests as T3 robots can. When it comes to T4, I think that although it is possible for us to know more about cognition without completely understanding the brain, it would be hard to rule out the necessity of T4 without the comprehensive picture.
Hi, thank you for responding to me. To reverse-engineer the brain itself does imply(to me) to be passing T4 right? My understanding is that T4 is identical even in material(like in class as we were talking about splitting open someone's skull and looking inside), and that T3 is identical in thinking/cognition and sensory motor capacities. T3 can do everything we can do, it just isn't materially the same. So, to reverse-engineer cognition we get T3, and to reverse-engineer the working brain we can get T4(at least this is my understanding but correct me if I am wrong!!).
DeleteI liked what the article said, because to me it highlighted the importance of understanding cognition over understanding the exact material parts of the brain that lead to cognition. Given that we really have no idea how the brain works, it isn't exactly helpful right now to rely on what we know about it to inform us how we think. Theories like symbol-grounding and connectionism can help us learn and speculate about cognition, without having to crack open any skulls(human or animal).
This is just CogSci (which is a part of Neurobiology, but only part): What CogSci seeks to reverse-engineer is cognitive capacity, not everything the brain can do, and is. T3 capacity is the minimum -- plus as much of T4 as is needed... to pass T3. The rest would be vegetative Neurobiology, wouldn't it?
DeleteWhen I was reading the Symbol Grounding Problem, I was confused a little bit until Searle’s Chinese Room Argument. What I thought about when reading that part was that symbol grounding comes with meaning, as Searle was able to manipulate the symbols without understanding the meaning behind the Chinese words, but the symbols were not grounded(?). It is mentioned that Searle wouldn’t have a conscious understanding, which points to our unconscious mind, however, if Searle had never been exposed to Chinese before, how would there be an unconscious understanding?
ReplyDeleteKid-Sib cannot follow your question: What is the Symbol Grounding Problem? What is the Chinese Room Argument? What do you mean by meaning, and by understanding? "Mind" is a weasel word. Try swapping "felt state" for "conscious mind." It feels like something to understand. Searle wouldn't feel it if he executed the (hypothetical) Chinese-T2-passing program. So passing T2 is not enough to produce understanding. So computationalism (what is that?) is incorrect.
DeleteMostly correct. But why cannot pass T3?
DeleteI believe Searle cannot pass T3 since nothing he computes is grounded. If we were to move him from the room and ask him to navigate Beijing, Although he may be able to hold conversations and even read signs in Chinese, he would not be able to categorize anything in Chinese if asked to. He wouldn't be able to operate a vacuum for example since none of the Chinese symbols are grounded in his brain, he wouldn't know that he needs to use the vacuum. He could reply to the command, but wouldn't be able to manipulate and use the vacuum for its intended use, which is why I believe he wouldn't be able to pass T3 in this scenario.
DeleteHi Ethan, if Searle was a T2 robot, then I would agree with your argument. However, Searle has a human brain with sensorimotor capacities and, thus, he can learn through categorical perception. If we were to drop him off in China, he would eventually start forming associations between the swiggles and the words in English which he does understand, which would provide the meaning required to consciously understand Chinese (i.e. with enough exposure he would eventually learn and understand Chinese). However, if Searle, the human, was just stuck in the Turing Test room where there wasn't the opportunity for sensorimotor experience to learn the language categories, this would just be symbol manipulation without meaning and he would fail T3.
DeleteEthan, you're right that Searle, the human (not computer) passing Chinese T2 would not be understanding Chinese (so neither would a computer).
DeleteKristi, category learning is not the same thing as categorical perception. What is the difference (Week 6)? (The rst is correct.)
Hi Ethan, thank you for your answer, it did help make me understand the part I was confused about.
DeleteP.S. I'm sorry for the weasel words!
Yes, "refer" and "referent" (sic) are the right words; but "association" is a weasel-word.
ReplyDeleteThe connection between the symbol and its referent (i.e., between the word "apple" and the apples in the world that it refers to) is made through the sensorimotor capacity to recognize and identify apples and to do with apples what you do with apples but not with axes or axioms. That is sensorimotor grounding.
"Intentionality" and "intrinsic" are weasel-words. ("Intrinsic" is just referring to the fact that grounding is a connection between a word in a speaker's head and the referent of the word: that connection must be direct, not indirect, through a second speaker's head -- but the connection can be made through a definition in words given to the first speaker by the second speaker, as long as the defining words are already grounded for both speakers. This is indirect grounding through language; this will become much clearer across the next few weeks.)
To do be able to connect words to their referents, you need sensorimotor capacities: (1) to move, manipulate, see, touch, smell, and taste; (2) to learn what apples are by learning to detect the sensorimotor features (size, shape, color, taste) that distinguish apples from axes; and (3) to learn what's the right thing to do with apples (pick them, eat them, not use them to try to chop wood).
[What is information and categorization?]
[Can you think of any other ways to ground "apple" in its referent?]
To ground the word "axiom" you would need more: indirect grounding through (grounded) language. (We'll get to that, once we get to language.)
What is arbitrary is the shape of the symbol that we all agree to use to refer to apples: "apple" does not look like an apple. (Neither does « pomme » or "alma" or 蘋果.)
I'm not sure where you got the "cow/moo" story, but children have to learn the difference between the features of cows, calves, bulls and bison, as well as their sounds, before they can call a cow a "moo." (Where did you get that example?) (Imitating a "moo" would have been a good example for a non-language way to call someone's attention to a cow -- if they knew the look and sound of a cow, but "moo" is not an arbitrary symbol: why not?)
"Consciousness" (even though I too am guilty of having used it in the past) is also a weasel-word. The non-weasel version is: "a state that it FEELS LIKE SOMETHING to be in." The latinate form is "sentience."
Understanding is T3-grounding + the feeling of understanding.
I'm not sure I'm convinced that the symbol grounding problem has been solved. Although Steels' robots are autonomous, there are many organisms that are autonomous but lack genuine cognition (ex:plants, bacteria, etc). Professor Harnad also states that symbol grounding requires more than autonomy; it requires analog and dynamic sensorimotor processes, which Steels' robots lack. I think that resolving the problem using robots that ground colours through sensorimotor perceptions seems restricted and it's unclear how this process extends to more intricate tasks. I just don't see how his robots are anything more than t0 and am not convinced that they are valuable in solving the symbol grounding problem...(But I'm open to having my mind changed if I haven't fully understood)
ReplyDeleteYou have understood: Steels' robots cannot pass T3.
DeleteOnce again in the second reading Harnard, S. (2003), in the formal symbols section – I must disagree with the “how” we came up with the symbols. Right now it is true that the symbols we are using do not have a clear and ostentatious link with the objects they refer too. However, there was a time when some symbols did have a clear link. For example, 2 did relate to two-ness, it used to be written in a way where the symbols had 2 angles hence 2. Though, I get that there are still gaps between two-ness, the number 2, the fact that we call it two and two objects. I also agree with how powerful the conversion to meaning in our heads is extremely powerful.
ReplyDeleteAnd 2000 would have to have 2000 angles, and the word apple would have to look, feel, and taste like an apple.
DeleteBut you are partly right: Some panotomime really resembles what it's miming. But mime is not language. It's showing, not telling: For that, we need symbols with arbitrary shapes. That seems like a good place to start -- but then you have to get rid of that iconicity (resemblance). How?
This reminds me of discussions from 4.b/in class about how we can only 'show' with gestures, but we can 'tell' with language. Somehow our language has evolved from gestures and icons that simply copy what they are trying to show, into an entirely more complex system that uses arbitrary symbols that can be manipulated to express almost infinitely many things. While I don't know the precise mechanisms behind the evolution of language, the answer to the 'how symbols have lost their iconicity' question seems to be that our language required a set of arbitrary symbols, and rules to manipulate these symbols, to be able to express so much more than simple gestures or icons ever could. Thus, our language evolved into a symbol system to be able to do such things.
Delete
DeleteGenes vs. "Memes". Jessica, that's mostly correct, but you are conflating biological (genetic, neural) evolution with learning and cultural change, which has some similarities to biological evolution but is fundamentally different: it does not involve changes in genes but in "memes" that are passed on by imitation and word of mouth (and writing).
Pantomime ("showing"). We evolved the capacity for deliberate intentional communication -- "pantomime" -- biologically, genetically. Most other species (perhaps all of them) lack that capacity, although they can and do imitate (especially when young). The communication was intended to reduce uncertainty. (What does that mean?)
From icons to arbitrary conventions. But with intentional gestural communication becoming a part of everyday life, in the family, and the tribal cave or village, gestures that are used over and over again to communicate the same daily thing (mother, father, baby, eat, drink, bring, put, sleep, danger) no longer have to resemble the things that they are mimicking. They can become simpler and more a habit or convention that everyone shares rather than having to faithfully resemble what they are used every day to imitate. That process of shrinking gestures to arbitrary movements that look less and less like what they once used to imitate is a gradual process of becoming less and less iconic and more and more arbitrary, habitual and conventional in communicating information, recognized by the entire community of users. -- That gradual process was probably learned and cultural rather than evolved and genetic.
Propositions ("telling"). But that's still not language. It only becomes language when the pantomime becomes propositions: strings of increasingly arbitrary gestures, still connected in users' minds by convention to what each gesture had once resembled, but the resemblance no longer matters; what matters is the meaning of the proposition they are trying to express: "The cat is on the mat," with a subject and predicate, with the intention to convey what the proposition means by telling, rather than just showing.
Predication and the "Propositional Attitude". Propositions have a subject and a predicate. They they tell (i.e. they "predicate") something that is either True or False about the subject (although propositions probably started in the form of questions ("Is the cat on the mat?") or imperatives ("Put the cat [gently] on the mat!").
Gene or Meme?. It is not clear whether the "propositional attitude" evolved or was learned. Years of trying to teach language to nonhuman species, especially primates, has not resulted in producing the "propositional attitude" of using symbols (whether gestural, vocal or even using objects or shapes on computer screens) to form propositions that are able describe anything, rather than just as arbitrary actions to ask for, or about something. (Many nonhuman species and individuals are brilliant, creative, empathic, thoughtful, knowledgeable, but they cannot communicate propositionally. (If they could, they could enrol in this course and skywrite with the rest of us, conveying information by exchanging propositions.).
Referring or Pointing?. Nonhuman animals can be taught to "name" things -- both individuals and categories -- but it's not clear whether these arbitrary names really "refer," as in propositions, or just point. (Most species are not able to do intentional pointing of intentional communication at all.)
These details are important but far from clearly understood.
I think Garance's point makes a lot of sense, and I believe this would preclude Prof Harnad's argument that 2000 would need 2000 angles; so long as we have a partial grounding, we can build off of this to more complex ideas without losing the /meaningful/ basis. If we have a grounding for 1 and a grounding for 2, and for 0, etc etc, can't we use these concepts to construct the idea of 2000 while still retaining a meaningful basis (I know what 2 represents, and while the multiplication may be based on abstraction, the zeroes also have a meaningful basis and their combination can arise an abstraction of something grounded).
DeleteWhat do you mean by "partial grounding" and "abstraction"? Are you talking about grounding or iconicity? Grounding depends on category learning: What is categorization? And what is the difference between the symbols in computation and the words in language? Computation has no symbol grounding problem (why not?). What is the difference between direct and indirect grounding?
DeleteJust jump into the conversation to answer some questions that fascinated me. Categorization is to divide the objects (either concrete or abstract) existed in the world to different group, depending on their features.
DeleteFrom my understanding to the reading, the main difference between the symbols in computation and the words in language is about meaning. The computation just process the symbols in a shallow way: get a symbols, going through the algorithm, then got the output. The meaning of these symbols are assigned by algorithm designed by human. I am not sure whether I could define the symbols in computation is algorithm-dependent (if so, it seems like no symbol grounding problem for computation). But what I could say is the words in language carries both their semantic meaning and context-dependent meaning: the former is planted into our mind when we are communicating, reading and so on, and the later is kind of extension of the former, associating with the sentences and context we received.
The difference between direct and indirect grounding is kind of vague to me (especially the indirect one). So I asked ChatGPT for some inspiration, and it says: "direct grounding refers to the process by which symbols or representations are associated with the sensory experiences or properties of real-world objects;" "Indirect grounding involves establishing associations between symbols and real-world entities or concepts through intermediary representations or knowledge structures." Where I need clarify is about the concrete and abstract objects: the concrete objects could carry both direct and indirect grounding, and in the case of abstract object, such as justice, they seem like only carry an indirect grounding, as they are just thoughts and no observable entity. Hope someone could tell me more. Thanks in advanced!
1. Categorization is not dividing things into groups. Look for the question I ask so often: "What is categorization?"
Delete2. I don't know what is meant by "semantic meaning" and "context meaning" and "shallow". Computation just manipulates symbols to produce an output. If the symbols, manipulations or output are interpretable as meaning something by the user, then that meaning is in the head of the user, not the computation.
3. There is no symbol grounding problem for computation (maths, logic, etc.). But there is a symbol grounding problem for computationalism. What is the difference? and why?
4. Be careful in how you use GPT. What you said it said is mostly empty weasel-words. The answer is simple. Direct grounding is learning the distinguishing features of the members of the referent category by direct sensorimotor trial and error, as on the mushroom island. Indirect grounding is learning the features by being told by someone who knows them (as long as the names of the features are already grounded for both the speaker and the hearer). For "abstract" categories, see all the replies for "justice," etc.
5. Grounding is not "association" (a weasel-word). What is grounding?
1. What it a Turing Machine? What can and can't it do?
ReplyDelete2. That is a T3-passing robot? What can and can't it do?
You've answered your own question.
Hi Megan! Like you said, a Turing Machine is only computational, as it's only capable of executing algorithmic processes. Since it can only operate based on predefined rules to manipulate symbols on an infinite tape, it lacks sensorimotor abilities and is ,therefore, uncapbale of symbol grounding. On the other hand, a T3 robot possesses sensorimotor abilities which allows it to connect symbols with real-world meaning.
ReplyDeleteAfter reading the two -pedias, I realize more and more that symbol grounding is also a discussed topic in philosophy. For example, the process of picking referents has similar notions in the study of phenomenology, whose focus is to study conscious experiences. Phenomenology says that each cognize-capable creature has its own "world," and that only the things it has interacted with are in its world. Not only the world has physical environments, but it also includes personal perceptions and understanding. The symbols are connected to our living experiences, thus we can derive meanings from them.
ReplyDeleteWhat I've noticed so far is that, cognitive science and some branch of philosophy are really studying the same thing, yet the terminology barriers prevents disciplines to communicate effectively and combine their efforts. However, the differece is that cog sci also considers scientific investigation, in which doing experiments to ground our conceptual understanding of consciousness.
I am no experts on phenomenology nor philosophy in general, I consult with ChatGPT (:D) to have a better grasp of them, do correct me if I'm wrong please.
Watch out with GPT! Although it does some useful and informative synthesis of material in its Big Gulp of 2021, the same thing that makes GPT stick to correct grammar also makes it stick to the prevailing knowledge -- and opinions and errors and ignorance -- of the 2021 Gulp.
DeleteFor uncontroversial and well-documented biographical and historical facts, reliable and well-known experimental findings, mature and firmly-supported scientific theories, GPT can be very helpful. But it also re-cycles prevalent errors or vacuities. And it LOVES weasel-words (taking its cue from the majority authors of the sentences in its Big Gulp!)
"Symbol grounding" was not well understood in 2021, and it still isn't. In fact, the unexpected success of LLMs like ChatGPT has, if anything, compounded the prevailing misunderstanding. This course will try to sort them out, but I'll just give you a few challenges to use on GPT now to try to wean it from parroting the prevailing view and instead dealing with its weaknesses:
(1) Philosophy (especially philosophy of "mind" [weasel-word]) has some recent and longstanding problems and attempted solutions of its own, not based on, or even relevant to, empirical questions. Yet, in the face of new empirical findings or empirical puzzles in science, philosophers tend to try to fit them into philosophy's own traditional problems and puzzles. Sometimes (but, I'd say, rarely), there really is a fit, and philosophy gives some useful insight for interpreting and advancing scientific understanding. More often, though, it's just trying to fit a round (empirical) peg into a square (philosophical) hole (or vice versa)
(2) Yes, phenomenology is introspection on "mental" (weasel-word) states. (What "mental states" means, unweaseled) is "felt states." We got a glimpse in Week 1, with Ohrie's 3rd-grade schoolteacher, how far phenomenology will get you in trying to reverse-engineer cognitive capacities.
(3) Yes, the "Hard Problem" of reverse-engineering how-and-why thinking states are feeling states (hence "phenomenological states") is hard, but doing phenomenology does not help, even though it feels-like it would and should.
(4) Yes, the symbol-grounding problem is about "associating" [weasel-word: should be "connecting"] words to their referents. But that is not the same as the philosophical (perhaps even metaphysical) problem of reference (stemming from a theory of "truth"), for which a "solution" is something along the lines of a causal history of a word's "contact" with its referent (e.g., "apple" and apples). What cogsci needs is an explanation of how a cognizer is able to recognize and act upon the referent of its words.
(5) That explanation will not come from introspection, nor from causal history, and especially not from any connection between words and words. Philosophers are just re-treading well-trodden old roads in another land (their own) when they go that way, and invite Cogsci to join them.
(6) But Cogsci, too, is perseverative, going over and over in the same (fruitless) loops, proudly proclaiming "grounding" solutions that are in no way grounded.
So go back to ChatGPT and update it with this, patiently pointing out to GPT that symbol-grounding is a causal, sensorimotor, T3-scale interaction between the words in the heads (and bodies) of people (or robots) and the things in the world to which those words refer. And ask GPT if it finds anything in its huge database that addresses that question rather than begging the question.
If there's nothing useful in the Gulp, GPT will either reply with some empty (weasel-)words about "complexity", disagreement (especially "philosophical" disagreement) and the need for further research. And if you push further it will just start repeating itself, or even begin to make things up.
(And, while you are at it, ask GPT what it means to "beg the question." If it says it means (a) "to call for an answer", then that's another illustration of how GPT is handicapped with parroting prevailing opinion. If it says it means (b) "to avoid answering a question", then there's some hope. (But it will probably reply, statistically, that it used to mean (b), but is now coming to mean (a).
DeleteI think it is also interesting why even cogsci is "going over and over" in the same loops, that is we keep encountering the same problems (such as grounding or not having actually explained how we do what we do). One reason is that to find a solution we are using similar approaches as those who have failed before us, or working off of existing literature. We may be missing an ability to effectively approach the problem with a new perspective, which is so difficult because it requires effective interdisciplinary communication. But as Professor Harnad mentioned, the models that philosophers will try to use to formulate solutions are not based on empirical information but on a chosen theoretical framework from which they then derive their answers. Maybe this is where generative AI can help us in visualizing our ideas in new ways (like in text to image generation where the image produced by AI is a new way to see information) to possibly stimulate a creative or new perspective to the problem.
DeleteHard to see how GPT can get creative if it can't even understand, but only recombine. We mostly just recombine too, but we are recombining grounded words and sentences we understand.
DeleteText-to-image transformation is so far at toy level, but as it scales up, it may become more useful. Until it reaches T3 scale, however, it hasn't even caught up with us, let alone gotten ahead. And sensorimotor grounding is more like image-to-word than word-to-image. Feature-detectors and features are the residual iconicity that indirectly (i.e., verbally) grounded words inherit from their sensorimotor roots.
That's it.
ReplyDeleteThe articles on symbol grounding assigned this week connect to the distinction between T2/T3 machines that have been discussed in class—especially the question of why we cannot merely equip a T2 with a camera, such that we can say the T2 takes in ‘sensory’ stimuli, and say that it is equivalent to a T3.
ReplyDeleteIn the scholarpedia article, Harnad argues that the way to ground symbols is through the addition of “nonsymbolic, sensorimotor capacities” that facilitate autonomous interaction with the “objects, events, actions, properties, and states that its symbols are systematically interpretable. . .as referring to.” Is this autonomous interaction with the environment, which is possible with a T3 machine, what separates a T3 from a T2?
What separates T3 from T2 is the sensorimotor capacity to interact autonomously and Turing-indistinguishably from any of us verbally but also to interact with the world of things that our words refer to, again Turing-indistinguishably FROM any of us, TO any of us, It's the difference between a chatbot and a T3 robot like Anaïs. Adding a camera, arms and wheels to ChatGPT will not turn ChatGPT into a T3 robot. Count the ways it doesn't.
DeleteBased on this week’s three readings, I would say this would not make ChatGPT a T3 robot for multiple reasons including:
DeleteGPT would have to integrate the inputs it receives to ground the text symbols it knows with the real world (assuming grounding is required to behave like a human)
GPT requires a “head with a world inside it” (scholarpedia) that stores understanding that is attained through connection with the outside world. GPT would store things like a computer, and perform computations. This would not count as a head or understanding, only simple symbol manipulation.
GPT cannot experience the feeling of understanding, which is an aspect of T3 grounding.
I found the “Symbol Grounding Problem” Scholarpedia reading to be very straightforward in explaining a difficult concept. In this problem, a symbol is an object within a symbol system, which are sets of symbols and manipulation rules. Symbols are manipulated according to their shape, and importantly not their meaning. The symbols can be interpreted as having meaning and referents, but the shape of the symbols themselves don’t have direct relation to their meanings and referents. If I understand correctly (borrowing an example from a children’s book) a pen could just as easily have been called a “frindle” without the object changing. The choice of the letters p-e-n, don’t have any resemblance to the actual object, we just interpret the sequence of letters to refer to a pen. Within the symbol grounding problem, grounding refers to the ability of symbols within a brain to pick out their referents. This process is implementation-dependent, and isn’t seen in symbols written on the page or in computer code. In order to achieve groundedness, a symbol system would have to have sensorimotor abilities, and be able to interact with its surroundings, ie, perceive and use a pen in accordance with its own understanding of a pen’s meaning. Professor Harnad argues that this necessity for groundedness means that the T2 Turing Test involving only symbols is insufficient, and must instead be replaced by a hybrid symbolic/sensorimotor test. Expanding on this, in order to achieve groundedness the hybrid robot must be able to categorize. For example, it must be able to put a “pen” in the category of “writing tools” and interact with it accordingly.
ReplyDeleteCorrect. But symbols don't pick out referents: People (and T3 robots) with symbols in their heads do
DeleteWhile reading the section in The Symbol Grounding Problem (Harnard, 2003) about how we use categorical representation to identify/name things, I started thinking about how our boundaries for what constitutes a member of a category (in other words, the invariant features of that category) change depending on the language that we speak. For example, I remember learning in a linguistics class that in Russian they have a different word for light versus dark blues, thus affecting how they discriminate colours. This relates to the Whorf Hypothesis which we discussed in class, and also shows that the invariant features that establish categories, which in turn affects how our symbols are grounded, are arbitrary to some extent.
ReplyDelete**(Harnard, 1990)
DeleteWe'll be discussing the Whorf Hypothesis in Week 6, in relation to categorical perception (CP). Although the names of the colors of the rainbow vary from language to language, the perceptual boundaries between the colors are probably universal: the rainbow looks the same to you no matter what language you speak. So the Whorf Hypothesis there is wrong. But when you learn hard-to-learn categories (because their features are hard to detect), this can induce CP (increased discriminability between members of different categories, and sometimes even decreased discriminablity between members of the same category, after learning, compared to before). This is a Whorfian effect, a kind of "learned" rainbow. The effect is weak and only measurable with psychophysical tests, but it may turn out to play a role in symbol grounding.
DeleteI was wondering if I understood the section “A complementary Role for Connectionism” from the 1990 article as it was intended. In section 4 on connectionism, it states that combining the symbolic and connectionist approach (described throughout the article) is the closest we can currently get to solving the symbol grounding problem. Associating a symbol to its referent is inherent in the human brain (implementation-dependent), and we do that through categorization and discrimination based on our lived experience (a connectionist approach). Once this connection is made and we ground elementary symbols this way, the rest of the symbol strings if a natural language will inherit the grounding of these elementary grounded symbols that they are composed of (the more symbolic approach).
ReplyDeleteI was also wondering how supervised learning fits into all of this. From my understanding, I form categories as I observe others interact with the environment around them as well as being explicitly told what categories x object belongs with. Is supervised learning how connections initially form, that allow me to ground symbols in my brain?
One can still do supervised learning alone on the mushroom island through trial and error. Eating different mushrooms will cause differential sensorimotor effects, allowing one to eventually ground the different mushrooms (symbols) by detecting distinguishing, recurring features (iconizing), and connecting them to the perceived effects. These features can be extracted possibly by connectionism and learning. My question is that invariant features cannot all have been extracted through connectionism. Someone please correct me if I am wrong, but it seems to me as though I have a certain bias toward certain features already built in. For example, if I am to encounter a snake for the first time, it cannot be the case that I have already managed to extract the invariable features of what I am perceiving to ground it as a "snake", but I seem to already perceive the snake as some sort of threat, even before properly identifying it. I guess that it is possible that I categorized what I was seeing to be a threat through some sort of systematic semantic interpretation, but it seems like we are able to do this sort of rudimentary categorization from a very early age.
DeleteNot all categories (and their features) are learned. There are some innate ones too (Week 7).
DeleteIf I understand this correctly, in a nutshell:
ReplyDeleteT2 + symbol-grounding ability (which includes sensorimotor abilities) = T3
...Is this correct?
I believe that sensorimotor abilities are the distinguishing factor, and symbol-grounding comes along with those abilities.
DeleteT3 includes both verbal (T2) and robotic (sensorimotor) capacity. But understanding language includes T2/T3 grounding and sentience (the capacity to feel), because it feels like something to understand. How is that related to Searle and computationalism -- and to mirror capacity?
DeleteI think that an interesting question to think about is what is the smallest amount of symbol grounded language that would be necessary to be able to understand and interact with every object. The example of needing to know what a zebra is without ever learning the word zebra but already knowing stripe and horse is what made me consider this. I think young school-aged children are sort of an example of this idea, because they have enough language to understand most of the world, but their vocabulary is still limited as compared to an educated adult. I don’t think that this would necessarily be helpful in reverse-engineering the mind, but it is an interesting thought.
ReplyDeleteHi Megan, interesting thought! From what I understand, symbol grounding doesn’t even need to be related to language. Think about other mammals (e.g., our cats and dogs as pets), which are animals that do not have language but can learn to understand commands. Animals can learn through categorical sensorimotor experience that “no” has the semantic meaning that something is prohibited. Yet, they don’t have the vocabulary to express this in words. The articles describe that the “categorizer” (i.e. human, animal) must be able to detect sensorimotor features and the detection of these features can be learned nonverbally thru trial-and-error or verbally with definitions.
DeleteIt is a cool idea, Megan, and the zebra example you gave makes me think about the process of learning another language. When you first start, there's usually a lot of mental translation going on. Your understanding of the new language is grounded in the language you already know. When you translate in your head, the language you translate through acts as a sort of middle-man in getting to meaning. But eventually, for people who really get to know another language, meaning is no longer grounded through another language, but grounded directly in meaning and experience. I think the idea of how meaning is grounded changing is interesting too, and it reminds me of the idea of dynamic semiotic networks in Steele that change every time we experience something or converge our categorical definitions with other people's.
DeleteMegan, actually that was quite an insightful reflection, as you'll see when we get to "Minimal Groounding Sets" in Week 8! (Have a peek")
DeleteKristi, all mammals and birds, and probably all or most reptiles and fish and many invertebrates can learn categories (what's that?) But that doesn't mean they have language. Learning to respond to some human vocalizations is not language. (What is it to have language?) This is discussed elsewhere in this and the mirror-neuron thread.
Elliott, good points, but who is Mary Steele and what is “Sonnet, 1795”? Even ChatGPT doesn't know...
Ah Elliot, you were referring to Luc Steels. Yes, meaning changes as we learned more; not just new categories, but more about older categories. Features are approximate, and sometimes need to be update to tighten the approximation. More on this in Week 6.
DeleteAt first, it was a bit hard for me to understand why sensorimotor abilities would avoid an infinite regress and would allow a system to be grounded, but I think I've got it. First I had to understand that the function of sensorimotor abilities must be more than just data-collection about the world, otherwise the Big Gulp would have done the trick, and we wouldn't need a T3. So then what's the point of sensorimotor abilities in the meaning grounding problem if it's not just data-collection?
ReplyDeleteFrom what I understand, the importance of sensorimotor abilities is data-collection in a particular (probably active) way, AND about the ability to ACT ON the world, that allows us to ground meaning in experience, and not just in data (but this does still seem a bit circular in that you need to be able to feel in order to experience, and you probably need to have grounded the world in meaning to feel, so maybe I am super wrong on this last bit). The analogy that comes to mind is that of Active vs. Passive Kittens.
I think this understanding of the role of sensorimotor abilities is what leads us to Categorization as vitally important for doing the things that a thinking thing does.
Yes, the motor part of T3 (DOing) is essential: What is categorization? But if you can explain why not just DOing capacity is necessary for T3, but also feeling, then you've solved HP. But all you need to break the circularity of dictionaries (and ChatGPT) is T3. How?
DeleteI found it a little bit hard to understand what the symbol grounding problem is and the proposed solution to it, but based on what I think I understood from the reading (Harnad 1990), an evident solution for the symbol grounding problem has not been found yet, however, according to the reading, the closest we can get to solving this problem is with a hybrid system that combines symbolism and connectionism. In fact, connectionism can explain how exposure to sensory projections and feedback allows us to create a link and connect objects that we see to symbols. The symbolic system then shows that once we ground these elementary symbols, the rest of the symbol strings of a natural language can be generated by symbol composition and they will inherit the grounding of the elementary symbols that they are composed of. So basically, using the Zebra = horse + stripes example, this hybrid system would begin by connecting objects such as a “horse” and what it identifies as “stripes” to symbols and then, if it encounters the image of a “Zebra”, the symbols for “horse” and “stripes” would combine and lead the system to find a symbol for zebra that would be grounded in the same way that horse and stripes are grounded.
ReplyDeleteGood grasp, but see the rest of the replies about direct and indirect grounding.
DeleteI first read the scholarpedia text on Symbol grounding. I think it helped me finally distinguish the easy problem from the hard problem. The easy problem is how we DO things - doing the right thing with the right kind of thing, categorization being “grounded” in its functional state. Solving the symbol problem is the solution to the easy problem. But how we FEEL things is still not explained by a symbol grounding. There seems to be something there that cognitive science might never be able to answer, and this is the hard problem.
ReplyDeleteThat's it. (But maybe some Brobdingnagian will be able to solve the HP after all. ("Stevan Says" 'I doubt it' -- but he's just a Lilliputian...)
DeleteTo add to this, one of the things that distinguishes T2 and T3 capacities that is highlighted in the article is the ability to categorize. A T2 machine can describe an apple with information it has been fed, but to be able to categorize (do the right thing with the right kind of thing), it must be able to interact with an apple using sensorimotor capabilities. This is why I think T2 is not sufficient for reverse-engineering cognitive capacity, it cannot interact with the world around it in order to mean something when it says "apple", "baseball", or "house", it can only use the words as ungrounded symbols.
ReplyDeleteFirst comment: Why are motor capacities required for grounding?
ReplyDeleteSensorimotor capacities are presented as the solution to the symbol grounding problem in the Scholarpedia article: “So ultimately, grounding has to be sensorimotor”. However, I’m wondering why motor abilities are required for grounding and not just sensory capacities. I see how interacting with the environment can be useful in grounding language by allowing us to learn causal relations between objects, actions, properties, and events. But this would imply that people lacking motor capacities since birth don’t have a sense of grounding. However, that feels wrong to me since I work with physically disabled people who sometimes have no capacity to physically interact with their environment, yet they do have a sense of grounding, i.e., they know what the word “apple” means. One could argue that these people are not totally lacking motor capacities since some of them are necessary for basic functions like eating, breathing, etc., and for communication (how could I know that this person knows what an apple is otherwise) but these are really low-level interactions with the environment. Moreover, we could think of a computer program that has access to a realtime camera and which is able to analyze the images and recognize objects. This program would be able to identify what the word “apple” refers to. Yes, it won’t have a sense of what it tastes, smells, or feels like, but it does have a sense of what it looks like. One might respond that it is just analyzing a numerical image composed of pixels, but isn’t this also what the brain is doing minus the subjective experience? While it seems that such a program doesn’t have a sense of what the word “apple” means (because it seems to require some level of consciousness according to this article), it still knows what the word “apple” refers to - which is grounding (at least in the visual sense). So, why couldn’t sensory abilities be enough for grounding?
Something that may be interesting to you: https://palm-e.github.io/, by google I believe, is an AI trained using the transformer architecture, and is essentially a glorified Large Language Model (LLM) with access to a camera and some actuators. It can follow instructions in natural language, like sorting different cubes, and can answer visual questions that show causal understanding of what different symbols mean! In a similar vein, the current progress of Optimus (https://youtu.be/D2vj0WcvH5c?si=KrZtXsqiQAINxvVe) is pretty cool.
DeleteI completely agree with everything you said and I do not understand either why motor capacities are required for grounding. Knowing what an apple is and what the term apple refers to seems enough to me. Being able to feel an apple doesn't change the meaning of the word apple and how we perceive it through our retina. Sensory input is enough to assign a meaning to a word. You do not need to touch a horse to know it's a horse and not a dog for example.
DeleteHowever, we saw that mirror neurons allowed us to understand the actions and emotions of others. Therefore, our interaction with the environment through movements allows us to learn causal relations between objects, actions. For example, when a child touches, smells, and tastes an apple, they not only gather sensory information but also build a richer understanding of the apple's properties and how it relates to their own actions. And maybe through that, we experience emotions that we would not be able to experience without motor capacities.
Yes, I agree that motor capabilities seem to be necessary to build a rich understanding of things and to experience emotion. But I would argue that they're necessary for all types of grounding. A paralyzed baby (apologies for this example) will receive unorganized, meaningless sensory stimuli, no matter for how long it is presented. Unless it is able to navigate its environment, it will not even have grounded sense of depth or gravity or that an object is constant even if its retinal image changes because of light. I suppose with language this could be challenged, as we could teach it to associate words/sounds with stimuli (apples), but I would think it would lack many categorization capabilities that we take for granted (but I will know more after next week's reading!)
DeleteThis was an interesting question to think about -- why is sensing not enough, why do we also need to feel? This question goes beyond just understanding the neurobiological correlates of emotion and how processing the valence of this emotion then guides our behavior (like increased heart rate in a scary situation leading to fear or anger and causing a behavioral output removing ourselves from the situation) because this still doesn't explain why we must feel this emotion. All of the bio and neurological correlates of feeling and the associated electrical and chemical signals still do not describe the conscious experience of feeling, even if they allow a sensorimotor robot to respond and behave the way a human would. This is something that the T3 robot would be missing, even if it may have bottom up processes (like positive emotion as reward to correctly interacting with an object), it does not have top down processes where it may feel curiosity and excitement (for example for aesthetics or beauty) and purely for that reason (and no expected reward) spontaneously interact with objects; nor would it feel a state of depression and despite consequences, avoid any interaction with objects or surroundings -- all parts of the human experience of feeling.
DeleteNo feeling means no feeling, whether positive or negative or affectively neutral (like sensations), bottom-up or top-down. Neural correlates of felt states would be T4, but even they do not solve the HP, though it could be argued that they are close enough for Cogsci). The question arises, though, whether they need to be that close. Is T3 already close enough?
DeleteThe Scholarpedia article serves as a fascinating exploration of analytical thinking under conditions of limited empirical evidence, a viewpoint that, in my opinion, is partially contradicted by developments in modern machine learning.
ReplyDeleteWithin the article, we are presented with some theoretical framework in which to analyse words and their meaning. The word, or symbol, points to some referent or reference class within the world, and also points to the word’s meaning, which is the brain's way of identifying a referent. One could think of meaning as either a generating rule or a compressed method for mapping "object-space" to referents that are distinguished by a specific trait. Then, words that cannot be selectively applied to some objects over others lack meaning. For instance, the term 'blurgle' is meaningless, as one couldn't realistically identify an object that qualifies or disqualifies for 'blurgle' status. Then comes the question: How does the brain, and in fact any thinking machine, match the symbols to their meaning, without first being given a minimum amount of words to use to define the rest? It is simple enough for a young child who knows English to learn new words to add to their vocabulary, but how do they initially match symbols to words? This is the Symbol-Grounding Problem.
It appears that modern machine learning algorithms, among other learning mechanisms, can indeed acquire languages like Chinese without external guidance, utilizing a sort of 'Chinese/Chinese Dictionary-Go-Round' mechanism. Gradient Descent, the learning algorithm behind modern machine learning, creates Universal Function Approximators (UFAs), meaning that, in the limit (size, training data), Artificial Neural Networks (ANNs) can match any functional input to any functional output, and as they grow the approximation improves until the limit of possibility. However, there are stronger theoretical reasons to think computation can learn Chinese, and the grounded meaning of symbols, from a mere string of “symbols”, whose simplest form is of course the bits 0 and 1.
Not only can large language models (LLMs) learn Chinese from Chinese text, but they can also learn the relations between words, whose structures match the structure of the outside world. In information theory, there's a concept known as Solomonoff Induction, a way to optimally predict future observations based on existing data. It weighs hypotheses based on how well they fit the data and how complex they are—simpler is usually better. Full disclosure: GPT-4 told me a bunch of things about Solomonoff induction, but I had heard of the concept before.
This is relevant to the Symbol-Grounding problem. Imagine a machine that learns Chinese symbols (words) and their meanings without prior knowledge. Using Solomonoff Induction, the machine starts with a range of hypotheses about the structure and meaning of these symbols. As it collects more data, incorrect hypotheses get ruled out or are given less weight, while correct ones gain in probability.
Over time, the machine learns not just to recognize characters but also understands the structure and grammar of the language, effectively grounding the symbols to their real-world referents. In essence, Solomonoff Induction suggests that computational agents can solve the Symbol-Grounding problem autonomously, moving beyond mere function approximation to achieve a deeper, semantic understanding. They can talk about their understarnding, they can show you by moving a robot’s wheels or displaying text, they can make predictions about the data, and invent theorems to investigate implications of their knowledge. What more could understanding be?
To continue the previous comment; of course, Solomonoff Induction works if the machine operating on Solomonoff Induction principles only receives and outputs text. Therefore, the claim that there is something particular about T3, a Turing Test for a machine with sensorimotor abilities, which lets it ground symbols in their real world meanings, when T2 systems cannot, is in my opinion a mere confusion, though it is presented with much detail in the 1990 article. The portion on connectionism, it suffices to say, is one that I find most convincing. Below is a closing thought:
DeleteSuppose we have a person who has lost all ability to move, but we can read their synaptic pulses to determine which muscles they are attempting to move and thereby infer the words they are trying to utter, such that we can read the thoughts they are attempting to communicate. Similarly, let this person be blind and deaf and paraplegic, but we succeed in sending neurochemical signals into their brain that they associate with different “imagined” sounds, such that it hears, in artificial words without intonation, the sounds of the outside world, as though it was coming from their own inner-monologue or somesuch. Then, I hope, few would call this person unconscious, or lacking in personhood or cognition; their brain works perfectly well. Only their sensorimotor capabilities are lacking; they can still reason using natural language, much like current-day large language models. I am unconvinced by the claim that there is something special, perhaps uncomputable, about symbol grounding that requires mere access to sensory information and mechanical or physical actuators. These, after all, are nearly identical to an encoding of this analog information into bits on a Universal Turing machine.
I will mention that Joann's comment above, which had not loaded on my page when I wrote this message, goes in a similar direction, but explores different details, to my own skywriting. It's clear that our understanding of cognition will change as, in the following decades, we reach, and probably exceed, human intelligence in all domains, and achieve digital consciousness. I am personally worried by these developments, but no doubts will they answer some long-standing scientific questions.
DeleteThoughtful comment. I will reply, but just this one time. Your comment was way too long. Neither I nor the other students can be asked to read such long comments, especially since it could have been said in far fewer words.: (Put it into GPT4 and ask it to compress to 150 words.)
DeleteThere is no symbol grounding problem in computation; an algorithm applies to the manipulation of the arbitrary shapes of the symbols. Computation is just syntactic. It may have semantic interpretations, but those are in the heads of users, not in the computations.
In verbal thinking, however, the manipulation and understanding of the symbols (words) in the head of a thinker, is based on semantics as well as syntax. The words have referents. The language-user has to know the referents of the (content) words. And the referents of words are not words.. They are things out in the world. "Apple" refers to apples. Not to words describing or related to "apple": apples.
The ANN and LLM processing of words is all word-word processing. The only way to break out of this symbolic circle is to connect (ground) enough words to their referents directly, so that the rest can be grounded indirectly from descriptions or definitions composed of already grounded words.
The natural candidate for making this connection is direct sensorimotor (robotic) interaction with the words' referents. And the grounding has to be bottom-up, at T3 scale, not toy scale (a computer with camera, wheels, and arms).
Solomonoff induction is just computation again. And it's just computation if it includes computational simulation of brain function (as in the Brain Simulator Reply to Searle, which Searle rebuts by pointing out that it's something he could again execute without understanding.
Using real connections from a real person's intact or handicapped brain would make the real brain part of the candidate, which makes it into a cognitive prosthesis rather than a Turing Test.
I think the ability of chatGPT and similar models to manipulate language meaningfully (i.e. in a way that is meaningful to a human reader. I don't mean to suggest that chatGPT is capable of understanding) is evidence that using language does NOT require grounding, instead of evidence that grounding can be achieved purely by relational grounding or syntax. Searle’s Chinese Room argument showed that you can manipulate language in a manner similar to chatGPT without grounding (or understanding either). It also makes sense when you consider that language is a symbol system as defined in the Scholarpedia article, which means that it is a set of rules (syntax) for the manipulation of arbitrary symbols. Since the symbols, or words in this case, are arbitrary, you can manipulate them without issue even if they don’t have referents assigned. I believe this is part of what chatGPT does. Of course, the application of syntactic rules to words doesn’t explain chatGPT’s ability to give semantically appropriate responses to a prompt, only why these responses are grammatical.
ReplyDeleteIn the 1990 article on the Symbol Grounding Problem, Professor Harnad explains the symbol grounding problem, the question of how a system can attach meaning to symbols such that the meanings are an inherent property of the system, rather than relying on outside interpreters. The reading then explains that a solution to the problem should include certain human capacities, like discrimination and identification. Discrimination meaning the ability to discern if two objects are the same and different, and identification meaning the ability to correctly identify the object. According to Professor Harnad, in this model we must be able to form iconic representations of the objects or inputs in question. The article references specifically the visual representation of a horse projected on our retinas that we can then compare to the next horse we see. I was just wondering how this process would work with other sensory modalities; I don’t doubt the model holds true, I'm just curious how it would apply. It’s easy to imagine matching an auditory stimulus with an iconic representation. For example, you could hear a fire alarm and recognize the pitch, rhythm, etc. Would this work the same for someone who was blind, and operating by touch? For example if they were to pet a sheep and a dog, how would the iconic representation be formed, and how would they discriminate between them? Or would the representation here just be fur, that is possibly combined with an auditory representation like “baa” or a “bark” to make a judgment?
ReplyDelete"Representation" is a weasel-word (even though I still used it in 1990!). Each sense-modality leaves an iconic trace of the stimulation that reached it. The haptic modality (touch, palpation and manipulation) is an especially rich one: you can feel the shape, the texture, the size, consistency, the solidity of things through manipulating them. Helen Keller had only that, of her three main senses, and yet she learned all the categories to ground her language; and she was brilliant and prolific. She had all the iconic experience she needed, and was able to abstract from that the sensorimotor features that distinguished the referents of her content words just as well as anyone else (even though she never saw a color or heard a sound]. She got the shape information from touch, and also enough to be able to learn to speak with the help of lip-reading by palpation and held-hand gestures.
DeleteIn “The Symbol Ground Problem” (2003) by Professor Harnad, it was discussed how the rule-based manipulation of symbols that computation involves lacks a critical component: meaning. To achieve meaning would require establishing a relationship between these symbols and the entities they are linked to in the real world. As highlighted in the reading, accomplishing this specific connection is called being “grounded,” and can be implemented by incorporating sensorimotor experience into the system.
ReplyDeleteThis article covered aspects of “grounding” that really showcased the remarkable abilities the mind has, which is why I found this particular theme in the reading to be the most fascinating. More specifically, it is the idea that there seems to exist a process in the mind that is naturally capable of forming this link between words and meanings that strikes me as most impressive— it leads me to ask more questions related to the field of Cognitive Neuroscience. I’m interested in delving deeper into this: how exactly does the brain create a link between this gap? Is it the physiological processes that Cognitive Neuroscience researchers focus on, and what variables would they use to study this? Are common techniques in modern-day neuroscience research effective enough to study this? For example, would tracking cell activity or recording areas of the brain be the best way to approach these areas of research— or, as discussed in last week’s Fodor reading, would correlational studies only limit us in their lack of explaining the “how” of neural processes?
I'd rather not discuss animal experiments on sensorimotor function. But noninvasive human neuroimaging could be informative in this partly vegetative function; so could human psychophysical studies. And there''s computer modelling.
DeleteThe Symbol Grounding Problem readings got me thinking about the semiotic triangle, where a symbol stands for a referent and symbolizes a thought/concept, and that thought/concept refers to the referent. It's possible for a referent to have multiple symbols, and vice versa. This flexibility is a key aspect of language and symbolism, where symbols can represent various meanings and concepts, and one symbol can refer to different things in different contexts. It's fascinating how people across the world use various symbols, i.e., languages, to refer to the same things.
ReplyDeleteWhen we learn second, third, or more languages, we often relate back to the symbols in our first language as a reference point. It's like building a set of symbols in our minds to connect with different languages.
However, it's quite a challenge to imagine someone with no language at all. Even if you were to give them a dictionary, it would be impossible for them to learn and understand words and connect them to the real world because the words in the dictionary are “ungrounded" for them. The meaning of the words in someone’s head, those they do understand, are “grounded”, which implies that in the absence of comprehension and understanding, symbols are meaningless. A language is more than just a set of symbols; it's a complex system that involves shared meaning and understanding.
What is a symbol (in the computational and linguistic sense)? And what is iconicity?
DeleteCould "Understanding" be a weasel word? We ought to distinguish "phenomenological understanding", which is the feeling of having understood, and which may be, fundamentally, an illusion of some kind; and computational understanding, which is the mere ability to compute and infer different causal interactions between the referent and other stuff in the world; like knowing that letting go of an apple will make it fall if it's in a gravitational field, etc. It seems clear that GPT has computational understanding, but saying it "doesn't have understanding" in the first sense is pointless; how could we ever hope to know?
ReplyDeleteThe question you brought up of whether "understanding" is a weasel word is indeed intriguing. ChatGPT appears to have computational understanding, allowing it to make inferences and predict outcomes based on statistical patterns and associations in the data it has been trained on. The challenge is determining whether it can someday have phenomenological understanding, which requires consciousness.
DeletePhenomenological understanding is closely tied to an individual's conscious experience and can vary significantly from person to person. Human interpretations can be subjective and prone to error, and not everyone perceives or comprehends things the same way. So how do we define and measure phenomenological understanding, especially considering its subjective and potentially fallible nature?
No, understanding isn't a weasel word if you understand this sentence.
DeleteAnd felt (language) understanding is the only kind of (language) understanding, just as felt pain is the only kind of pain.
Please read the other replies. And re-read Searle on "computational understanding." You can disagree, but you have to understand the argument, and then give reasons for rejecting it.
Here’s my summary. The meaning of a word on a page is ‘ungrounded’, but the meaning of a word in someone’s head, that they understand, is grounded because it’s connected in some way to the referent. Searle’s Chinese symbols would not be grounded because he didn’t understand them and therefore did not make a connection between the symbols and the things that the symbols are referring to (the referents). What our brain has which an arbitrary symbol system doesn’t is this ability to pick out referents, and is this simply due to our ability to interact with the world?
ReplyDeleteNow I'm going to digress. I find the idea of the robotic Turing Test viable in that the robot would have the ability to interact with the world, but I argue that the robot would need to have flexibility; something comparable to plasticity. Meaning for us humans seems to change with experience as we understand things in our environments in different ways, and although the robot may be able to go for a while pretending to be a human, it would have to be flexible enough to have these meanings change over time in order to continue being a decoy human.
Fiona, I agree with your point on a robot's plasticity, and I'd further say (though I could be wrong) that the scholarpedia article somewhat supports this idea as well. In the section about robotics and categorization, the discussed robot must have the ability to categorize, which implies that it must be able to "feature-detect" the sensorimotor features of a category. It is stated that "these feature-detectors must either be inborn or learned," which, as far as I can see, seems to say that the capacity to learn - to update its conception of what features make up a category - is something a T3 robot could do.
DeleteThis said, the scholarpedia article talks a good deal about what this learning could look like, but it doesn't specify that this capacity is required for T3. I agree with you; it seems an "inflexible" robot whose categories are fixed and inborn would struggle to remain TT-passing for the full duration of a lifetime.
I think it would be interesting to consider whether a "big-gulp" type of dataset would change this conclusion. We might expect that the capacity to learn is fundamental to passing as human, but if a robot's knowledge base were to be vastly larger than any human's could possibly be - that is, if its inborn feature-detectors were based on an enormous dataset - would it be possible for it to pass as human for a lifetime without requiring learning capacity? It's true that such a robot's thinking wouldn't work in the same way as a human's, so perhaps this wouldn't bring us quite to a reverse-engineering of the mind.
Fiona, this is a very good point! It’s interesting to think about plasticity in the brain as we develop and how the meanings of different things change as we go through different experiences in life. Considering that we have years and years of experience that alter our understanding of the meaning of all kinds of things, how would a robot be able to meet the level of experience and adaptability that comes with that? Especially when considering more abstract concepts like love, for example.
DeleteFiona, Good points.
DeleteBut, as Adam correctly pointed out, the capacity to change, especially to learn, is already part of T3: Think of Anaïs, not a toy robot. The TT always calls for complete indistinguishability. Even passing T2 requires the capacity to learn. ChatGPT can already learn. Without that even a short conversation would make to sense.
But it's a good idea to be cautious about whether to declare the T2 as now passed by ChatGPT (and if it is, then there's still the question of whether the Big Gulp database is cheating).
Megan: What's a robot? And what's a machine? See past commentaries and replies.
I’m not sure how plasticity, as it has been described here, differs from learning. Does it not follow that by learning, one’s ideas and notions would shift accordingly? I think how plasticity is most often defined in the context of something (e.g., a brain region) taking up a function that is not typically theirs in place of a deficit of some sort. The plasticity describes here reminds me of neural nets and how additional information causes shifts in weights. What I’m trying to say is that I don’t think the necessity for plasticity completely rules out computationalism.
DeleteHere is my attempt at a summary of the Encyclopedia of Cognitive Science entry: the two leading approaches to developing a model for cognition are symbolism and connectivism. Symbolism is basically computationalism, since it holds that all cognition is the manipulation of a symbol system, or set of arbitrary symbols that can be combined according to symbolically represented rules to create interpretable statements. Its issue as a model of cognition is that it cannot explain how we can connect these symbols to their real-world referents. Connectionism says that cognition is the result of interactions between nodes, through weighted connections. The weights of these connections can be modified in response to feedback, to make the desired output more likely in the future. However, it doesn’t capture the compositional aspects of cognition. We can come up with a better model if we combine the connectionist and symbolist approaches. First we build up a library of terms that are iconically and categorically grounded. Iconic grounding is the idea that we have analog, nonsymbolic representations of sensory information in our heads. Categorical grounding is achieved through a connectionist system that takes icons as inputs and outputs their most defining features, which are then used to classify the icons as members or nonmembers of a given category. Then we are able to treat these category labels as symbols in a symbol system, thereby gaining the ability to combine and recombine them to produce new propositions. These propositions retain their groundedness because they can be eventually decomposed into grounded icons or categories.
ReplyDeleteAs I understand it, subject/predicate propositions are propositions that show how two categories relate to each other (for example, a proposition linking the category of “zebra” to the category of “stripes”, in the form of “all zebras have stripes”).
DeleteLanguage is a symbol system, so as long as the symbols it operates on are grounded, any ruleful propositions will also be grounded, since these propositions can always be decomposed into their constituent symbols. Grounding means the ability to pick out referents in the real world, which for a symbol system would involve giving its symbols an interpretation. Even if its symbols are interpreted, though, the symbol system itself is not this is not grounded, because the interpretation is not inherent to the symbol system itself. This is why language allows for indirect rather than direct grounding.
Language allows us to create categories other than the ones we can innately discriminate. This is because all other categories are learned, and the way we learn them is by having their features communicated to us by another speaker, which is done through language.
I believe my answer to why language allows for indirect grounding also covers the question about why indirect grounding is impossible without sensorimotor grounding (the ability to pick out the referents of a symbol in the real world is a sensorimotor function).
I found these readings about the symbol grounding problem very interesting, as it grounded (no pun intended) my own understanding of how we think about meaning. The part of the scholarpedia page that mentioned the importance of a sensorimotor component to the one perceivinginterested me especially – I am curious specifically about whether this necessity is predicated more so on the ability to sense characteristics of external stimuli to facilitate accurate categorization or the ability to exert some effect on these stimuli and provide appropriate motor responses. Or, whether one must be able to do both the individual sensory and motor component to be able to completely understand and categorize objects. It makes sense to me that we need some capacity to interact with the world around us, otherwise there would be nothing in which we could ground our understanding of symbols, but it raised the question for me of whether these distinct aspects of sensing and reacting are equally important to contextualizing one’s understanding of words and their meaning.
ReplyDeleteYour question is answered by simply recalling the definition of categorization (what is it?), and reflecting on the fact that DOing is essentially a motor act (with consequences). You can't learn most categories without trial, error and corrective feedback. Without those, there would be no right or wrong thing to do. Think of the mushroom island.
DeleteAccording to my understanding of the Harnad (2003) paper, symbol grounding problem explains the relationship between symbols (e.g., words) and their referents, in which a symbol system by itself does not have the capacity to “pick out their referents”. This is when grounding takes place, in which a symbolic system “would have to be augmented with nonsymbolic, sensorimotor capacities”. Thus, grounding necessitates sensorimotor capabilities. It can either be direct (sensorimotor) or indirect (linguistically). Does the symbol grounding problem suggests that categorical understanding is necessary for grounding? From my understanding, one needs to have sensorimotor abilities to interact with external objects to be able to categorize things, which inherently links the symbol to its referent? If categorization is the necessary condition in which a word (or any symbol) gets grounded, then how can meaning be achieved? Is it only left to the hard problem (like Searle’s CRA)? Meaning can only be achieved through a FEELING of understanding combined with sensorimotor capabilities that allows symbol grounding. Thus, even if we have a perfect T3 with sensorimotor capabilities, we don’t know the answer to “does it really understand”, since that is what the hard problem is all about, which intuitively seems impossible to make any progress with?
ReplyDeleteHello Can,
DeleteI'm not sure if I have a clear answer to your questions, but I don't think it's the case that the feeling of understanding is a necessary condition for meaning. In the 2003 paper, Prof. Harnad says that "Meaning is grounded in the robotic capacity to detect, categorize, identify, and act upon the things that words and sentences refer to", which leads me to think that CATEGORIZATION is necessary for grounding, but I don't think this is the same as categorical UNDERSTANDING. Even if a machine can identify a category, we can't know if it's really 'understanding' that category - it's the other minds problem. WE know what it feels like to mean something, but we can't deny that it seems like the people around us certainly mean things, too, when they are talking to us, and we can't know for sure whether they have the feeling of understanding, or any feeling at all. The most we can know, I would guess, is whether a machine is acting as if it understands a category, and we may make this judgement by seeing whether it is able to "do the right thing with the right kind of thing".
Can, excellent summary. The point about the "Other Minds Problem" (OMP) is that it is impossible to solve it with certainty. But, in practice, with other people and most animals (and with science in general) we don't need certainty, we just need very high probability based on the available, observable evidence.
DeleteSo we don't doubt that other people feel. And Turing's point is that when you can't tell apart a reverse-engineered candidate like Anaïs from feeling humans in terms of what they can DO, observably, (lifelong, if need be), you have no more reason to doubt that they really feel than you have with any other person (other than yourself). The Turing Test (T2, T3, T4) covers all the observable evidence possible. The question is: which of the T-Tests is enough (and why)? T2, T3 or T4?
The feeling of understanding (and of thinking, and meaning, and knowing, and of feeling in general, of sensations, including pain and pleasure) are definitely part of cognition, so it's Cogsci that would have to reverse-engineer and explain it. But the reason explaining sentience (how and why can we feel?) is the "Hard Problem" (HP)(Week 10) is that, if and when the Easy Problem (EP) (how and why can we DO everything we are able to DO), has been successfully reverse-engineered and T-Tested, feeling seems to be left superfluous, causally: there are no causal degrees of freedom left to explain how and why sentient DOers FEEL, rather than just DO.
So the reason the HP is hard is because of the solution to the EP (once it's solved). The OMP makes it impossible to know whether the successful reverse-engineered EP candidate feels, but, equally, it makes it impossible to know whether it doesn't. We have no stronger evidence than the T-Test (T2, T3, or T4), either way.
Adam, you've caught the OMP points well. Each of us knows, in one's own individual case, that we feel, and that it feels like something to understand Chinese (if we do), but we have no idea how or why. And that's the HP.
Just because symbols (e.g., words) are lack intrinsic meaning, does it follow that it can’t be a purely implementation-independent computation? I am bringing this up not because I disagree but more so because I find the logic a bit difficult to follow. Is it correct to say that this is what Searle attempts to demonstrate with his Chinese Room Argument? Is this because consciousness is necessary for grounding, which is a precursor of understanding?
DeleteI wonder if it is possible to for a symbol system with nonsymbolic, sensorimotor capacities to be implementation-independent. Or does it follow that for a machine possess such capabilities that it needs to be implemented in a certain way?
Not quite; it feels like something (i.e., conscious) to understand the meanings of propositions (language) and to know the referents of their words. But that's only an unexplained correlation until someone solves the Hard Problem.
ReplyDelete"Mind" is a weasel-word. Words are in the head: what is grounding?
Valentina, yes, sensing is needed to sense -- but why is feeling needed? Ask ChatGPT to define the difference and let us know. I suspect a lot of weaselly mumbling (along with the usual notion that "feeling" is just emotion.
ReplyDeleteAdrienne, yes, to understand and to mean feels like something. But for categorization, reference and language, why isn't DOing enough?
After reading the Wikipedia text, it is my understanding that the symbol grounding problem refers to the problem of connecting symbols such as words to what they represent in the world. It addresses the issue of how symbols adopt meaning that is represented in the physical world. It was explained that words on paper would be ungrounded because there is nothing connecting them to their referent, whereas the meaning of a word in a person’s head would be grounded. This made me think of the philosophical thought experiment that questions if a tree would make a sound if no one was there to hear it. It is similar in the sense that there is no meaning to a word if there is no mind to mediate the connection between that word and what that word is supposed to represent in the world.
ReplyDelete"represents" is almost always a weasel-word. A word names (and points to) a referent. A sentence (proposition) can describe an object or state of affairs, and the proposition can mean something to you (and you can understand what it proposes) if you know what its content words refer to. Words get their referent and meaning through grounding -- directly, through (mostly learned) sensorimotor feature detectors for categories, and indirectly, through grounded propositions from a speaker to a hearer. (What is a proposition?)
DeleteThe "unheard tree falling in the wood" cliché is about the nature of sentience (hence the HP). It feels like something to hear a sound. That's the effect of an interaction between vibrating air and a person's ear and brain. If there's no person, there's no interaction -- just as there's no pain without a feeler.
The referent of "apple" (and the meaning of the proposition "an apple is round") is there only in the head of thinkers who have grounded the content-word "apple" (directly or indirectly [what are is that?]) so they have the feature-detectors to detect and manipulate a real apple, as well as to talk about it. (How does ChatGPT seem to talk about apples without grounding? Ask it. And then prove it further, correcting the clichés it parrots from the "Big Gulp" and its authors.)
I first read the Scholarpedia article to get a good understanding of the basic concepts of the symbol grounding problem. That to not have an infinite regress we would need to make connections not just between words, but between words and their real-world referents makes complete intuitive sense to me. I can explain to you in words the properties of an apple till I'm blue in the face, but unless you can make a connection between those properties and real-world objects and in fact point at a real apple we are just going in circles. What is an apple? a red round fruit. What is red? Our perception of a certain wavelength of light. What is light? ect. My only question arises when we go to more abstract concepts, things that don't have referents in the real world. How do I ground justice? I think I know what it means but I can't point to it in the real world only examples of just acts which doesn't seem to be the same thing.
ReplyDeleteIf you can find out what "justice" means from a verbal description or definition (indirect grounding), and you know what all the words in the definition means, you can work down from the definitions of all those defining words to words that have a direct sensorimotor grounding. Verbal definitions and descriptions tell you the features that you would have had to ground directly.
DeleteIn any case, you can also point to instances of just trial and unjust trials, or descriptions of them. So all referents of content-words are "out there" one way or the other.
And even if you can't point the referents out they have a referent. (Google ("peekaboo unicorn" harnad).
I had the same question for how it is that we ground abstract notions. It made sense that for concrete elements, it is sensorimotor as we directly see it and from there ground its meaning, and that for ideas like "justice", we can still point out, as Professor Harnad said, a just or an unjust trial, and ground from there. This is clarified by the "peekaboo unicorn" example that professor gave, where a unicorn that disappears as it is observed is still imaginable and makes as much meaningful sense as a zebra being a horse with stripes, as we break down a unicorn as being a horse with a horn, which are elements we ground in a sensorimotor manner. I do wonder, though, if there is any concept that exists and is grounded without being broken down into constituent parts digestible via sensorimotor grounding, but even as I think about the most abstract things I can think of, such as the soul and God, the language we use to describe these ends up being grounded in reference to, or in contrast to, tangible, sensorimotor groundable things. The "peekaboo unicorn" reading also spoke about "theft" for grounding, in which someone else who has already grounded tells you the categories, which I think certainly helps in grounding abstract concepts, perhaps more so than for the concrete.
DeleteThe second reading I did was Prof. Harnads' 1990 paper "The Symbol Grounding Problem." I understand the rejection of a purely 'free-floating symbolic system' as such a system divorces the symbols from their referents. My understanding is that we reject pure connectionism, as it is not systemic. The best it can do in terms of meaning is taxonomy, which is certainly not all we want to explain. So we propose we create symbols based on the iconic and categorical representations we get from connectionism, which then allows us to combine them while still being able to keep them grounded. I wonder then if this isn't a part of how chat gpt is so good unexpectedly. It seems possible that if we build the symbolic system up from grounded connections, then the symbolic system might somehow, if one could look at the entire picture, through its form, point to these grounded connections. The how of that I am completely unsure of, however.
ReplyDeleteWe reject connectionism (as we reject computationalism), because neural nets are just learning-algorithms), regardless of whether C in "cognition is just C" refers to computation or connectionism.
DeleteGPT does what it does, however it does it, from the Big Gulp whose words were produced by grounded human speakers. If we read their words, we would be able to understand their meaning, and gain information from them, just as we can from GPT's words, even though the words are not grounded for GPT, and it does not understand them, or mean anything: it just tells us things with them that are meaningful (to us).
So the only question is about how GPT can do that with only the resources it has (the Big Gulp, the completion algorithm, the training, the deep learning, plus close to 2 trillion parameters that it can keep aligned, distributed across a huge number of computers all over the planet).
Most of that would never fit into any of our heads, so it would not be a realistic reverse-engineering of human verbal capacity even if GPT does count as having passed T2. (That makes it yet another vote for T3, which we need anyway, because cognizers can do a lot more than just talk.)
The symbol grounding problem succinctly captures a pervasive problem in cognitive science that I’ve never seen so clearly laid out until now. I find that a lot of pondering about cognitive science is an exercise of catching ourselves when falling into pits and discerning that they are often, in fact, the same pit.
ReplyDeleteI have a question reguarding clarification on why we deny that adding arms and wheels to (for example) adding wheels and a camera to ChatGPT would not instantiate a T3 passer. What I want to focus on in this example is the arbitrary-ness of what could pass T3 and when.
*If you were to build a pile of sand grain by grain, painstakingly picking up each grain and putting it in the same spot you’d know at each step exactly how many grains their were, so to you it is not just some abstract pile it is 7326 grains of sand, but to some new naive observer they could be pointed out the grains of sand and just say “yup that’s a pile of sand”. This to me is analagous to the incremental T3 discussion: it seems as though we do not want to believe that ChatGPT will ever pass T3 because we are there at each stage of adding each grain of sand. But someone coming along and not knowing how many grains of sand there were would obviously say that it is a pile (passes T3)*
Reverse-engineering cognitive capacity (or anything) is not just an instance of the Sorites Paradox.. There is a conceivable path from a T2-passing computer to a T3-passing robot, but you would have to first erase its T2-passing algorithm, because that's hanging from an unearned skyhook that will not help it get to T3.
DeleteIt's fine to have one, maybe even many computers inside the T3 robot; and I'm not bothered if the robot's sensorimotor equipment (sensors and effectors, internal and extrenal) are synthetic -- as long as they are not virtual. (Can you see why?)
You may even turn out to need some T4 anatomy, physiology and biochemistry (making the robot a hybrid bio-robot) to empower the robotit to pass T3. That's all that's meant by "not just a computer with camera and wheels and arms."
But an ungrounded T2-passing algorithm (let alone GPT's Big Gulp) will not help promote a T2-passing computer to a grounded T3-passing robot (Anaïs) any more than an encyclopedia Brittanica will.
A part of the Scholarpedia reading that I found particularly interesting was the precise definition of a “symbol”, which clarified a lot for me in terms of what can, and can’t, have meaning in the first place (without trying to explain what that meaning is itself). It is crucial for a symbol to be part of a symbol system, as a single symbol, in a vacuum, is not useful. In that sense, although we don’t know the meaning of the components of a symbol system themselves, would it be safe to assume that we know their meaning in a relative sense? For instance, a “0” only makes sense when there is a “1” to which it can be compared, and differentiated from. This can be synonymous with the definition of categorizing, which from my understanding, can also rely on relative meanings. Now, this doesn’t answer the question of objective meaning, or the “grounded” meaning that we are so concerned about in the field of cognitive science, but as long as we can categorize symbols based on their relative meaning in a symbol system, is that not sufficient for cognizing? It may be true that this might be cheating (as in the case of chatGPT and the “big gulp”, associating a whole bunch of words with each other without actually having had the grounded experience of the word), and a form of “Zombie” cognition, but it could potentially still be able to pass the Turing test without having had the grounded experience of each word.
ReplyDeleteI don't know what you mean by "relative meanings". In computation, some symbols and symbol-manipulations are interpretable by us users as meaning something, but they have no meaning in computation, or to the computer executing the computation. It's just syntax.
DeleteTo do digital computation, you do need at least two distinct (arbitrary) symbols, 0 and 1, but that's just notation, syntax and hardware, not grounding, let alone reference or meaning: As speakers and hearers, we don't stand in the relation of users and tools, as with GPT or wikipedia or google, or a library reference volume -- though all of those are bags of words too.
Words of natural language are arbitrary symbols too, and they too have their syntax and hardware, but their reference and meaning is semantic, and has to be earned (by both in the head of the speaker, whose words and sentences have intended referents and meanings, because they have been grounded, and in the hearer's head, where there is an understanding of the speaker's words and sentences, because there too their referents and meaning have been grounded).
We'll discuss relative discrimination and "absolute" categorization in Week 6.
Joann, you write that:
ReplyDelete"ChatGPT seems to have acquired a sense of grounding without sensorimotor capacities by processing a massive amount of human textual data."
What do you mean by "grounding"? Direct grounding is sensorimotor, and ChatGPT (T2) has no sensorimotor capacity (T3).
(What I have changed my mind about (if ChatGPT has passed T2 and the "Big Gulp" of text data is not cheating) is that ungrounded computation (algorithmic symbol manipulation) can pass T2: "Stevan Said" only a grounded T3 could pass T2.
"How can Large Language Models (LLMs) be so good at manipulating language in a highly meaningful way without any grounding?"
Good question. But it's only meaningful to us, users, not to LLMs. They don't mean or understand a thing. They are ungrounded and cannot recognize or interact with the referents of their own words.
"Do LLMs gain an indirect sense of grounding through the syntactic formal structure?"
LLMs don't have any kind of sense, direct or indirect. Please see the replies to the other comments on direct (sensorimotor) grounding vs. indirect (verbal) grounding. The only way to gain the capacity to ground referents indirectly is bottom-up, from the sensorimotor ground (Chapter 8).
The symbol grounding problem will be solved when Cogsci successfully reverse-engineers T3 capacity. All it has produced so far is ungrounded toys.
Aya, Searle's argument was conditional: Even IF computation alone could pass T2, THEN it would not understand language.
ChatGPT (if it can pass T2 and the Big Gulp is not cheating) would show that ungrounded symbol manipulation can pass T2, hence that it can "manipulate language... in a way that is meaningful to a human reader", and can do that interactively, indistinguishably to and from a human. That's the essence of T2.
But that does not reverse-engineer and explain language capacity for any system that has not swallowed the Big Gulp. And we still don't understand how the Big Gulp enables ChatGPT to do what it can do. The computation alone (the algorithms) does not explain it.
Joann, about "vectorization" see 'Vector Grounding'.
ReplyDelete(skywriting 1) Chat GPT typed out that the difference between sensing and feeling is that “sensing is the objective process of perceiving external stimuli through the senses, while feeling is the subjective experience or emotional response that arises from the interpretation of sensory information.” You were right about the usual notion that feeling is just an emotional response. From what I gathered from the 2003 reading, feelings are needed in order to distinguish that we are in a state of meaning rather than just in a functional state; whereas, sense is utilized in terms of grounding and is more technologically focused, is more of an input output function that can merely be created, but feeling cannot be made through inputs and outputs.Additionally, I gathered sense has more to do with symbols in comparison to feeling.
ReplyDelete(skywriting 2) After reading and dissecting the Harnard (2003) paper, it seems to me that the concept of formal symbols is a concept created by our minds rather ta]han by the system itself. While the formal systems have rules and shapes that perpetuate symbols, without our minds, it does not hold any value or meaning. In the next section it refers to language as another example of a formal symbol system, and this reminded me of last week’s reading that mentioned language versus communication and how they relate to one another. Language is a tool to use in order to communicate and hold little to no value without our minds and knowledge used to understand and manipulate these symbols. The section “Natural Language and the Language of Thought” prompted a question. Does it explain that in order to be grounded, the symbol system must be able to hold meaning and relevance without the use of a human mind? Or that it has to be more fitting to the things the symbols are describing?
ReplyDeleteThere are direct (through sensorimotor capacities, trial-error, corrective feedback) and indirect (language) ways of grounding and learning about categories. I was wondering about the particular differences between these two ways, and do they operate in a completely distinct way? We can’t observe or interact (through sensorimotor capabilities) with more abstract concepts like “good” or “evil”, but we seem to be able to categorize them. Is this fully based on language, or verbal explanations from another person, or can there be a role of trial-and-error learning or corrective feedback in the linguistic learning for categories? Is it solely dependent on the other person teaching us the categories? I was thinking about Thorndike’s Law of Effect and was wondering if something similar (?) to that can play a role in the indirect way for grounding? Or does that only apply to observable things we can interact with via sensorimotor abilities? Can we learn linguistically through making errors and correcting them/repeating and strengthening correct responses?
ReplyDeleteAll grounding, hence all reference and meaning, whether indirect (verbal) or direct (sensorimotor) is ultimately based on direct grounding. About the grounding of "abstract" categories, see many of the replies above.
DeleteAfter reading a couple of the texts on the Symbol Grounding Problem written by Professor Harnad, I took a look at Steel’s chapter on representation and meaning. One section in particular, which reviewed a child’s image of a city bus, got me thinking. In the image, certain features of the bus are exaggerated or focused on, while others are completely discarded or misrepresented (e.g. the bus is very tall with many windows, however there are over twenty wheels). Steel states that this depiction emphasizes aspects of the bus which have been key characteristics based on the child’s experiences. The bus being large, having many more windows than a regular car, and featuring a conductor who takes tickets are likely all parts of the child’s experience riding a bus which stand out. I think this same idea translates to how people use symbols to represent the world, as the characteristics or details which tend to stand out to us are more prominent in our symbol system and how it is used. In the English language, for example, there are 3,000 words describing emotions, suggesting that the importance we place on human emotion is represented by the vastness of our emotional lexicon. Furthermore, when describing an experience of mine, I may use language which conveys certain details about things that matter to me and disregard others. If I go for dinner where I eat my favourite pasta and drink a glass of water, I may use different words to emphasize each depending on if I was very thirsty or very hungry beforehand. In this way, symbol grounding is not just grounded in our external world, but our internal response to it as well.
ReplyDeleteI am trying to think of any other way to ground “apple” to its referent. Drawing apples is not enough as this is just a simulation. As mentioned, explaining “apple” would be similar to the dictionary example, potentially causing infinite regress as other symbols and meanings are referred to. However, if every component of the explanation is grounded in the learner’s head, then they can build indirect grounding.
ReplyDeleteAn essential part of grounding is the ability to discriminate between different inputs and correctly categorize them (different apples can both be apples, fruits are not all apples). By categorize I mean assign the objects into groups of a higher level of abstraction (less specific) in order to do the right thing with the right kind of thing. Perhaps another way to ground would be to show/explain two objects/concepts, and have only one of them be familiar. For example, showing or explaining a pear (grounded) and apple (not grounded), and saying “one of these is a pear and the other is an apple”. In this scenario, the apple would become grounded indirectly through elimination.
These readings made me wonder, are there levels of grounding? And at which point do we understand a meaning? For example, surely someone who has tasted and seen an apple is more grounded for “apple” than someone who has only seen an apple. Does someone need the full experience to be grounded? And if not, how much understanding is enough for grounding? My guess for the minimum would be: one sense's input and the feeling of understanding.
The below text is compressed and modified a bit by chatgpt.
ReplyDeleteGrounding is spotlighting a key flaw in modern AI systems. Although neural networks (NN), especially ones demonstrating human-like capabilities like DeepMind's GATO, remain predominant, which may considered as T2 to some extend, they still tread a path fraught with scientific inaccuracies for cognitive science even Harnad raised the issue in 1987.(Harnad, S. (1987b) Category induction and representation. In S. Harnad (Ed.) Categorical perception: The groundwork of Cognition. New York: Cambridge University Press) Despite their ability to process inputs into numerical vectors, used in functions to derive decisions, these systems symbolize a core challenge in making meaningfully conscious AI. Harnad emphasized that symbol systems lack meaning without consciousness, presenting a significant obstacle in reverse engineering since they don’t physically correlate with reality, operating strictly in a statistical domain without addressing real-life complexities. The crux involves surpassing mere calculation and statistical functionality in NNs to genuinely bridge AI decision-making with physical robotics and their environment, thereby navigating beyond the existing confines of unified, yet meaninglessly vector-oriented, processing. This transition beyond pure statistical operation to fully proceed sen the Harder problem in evolving AI systems and T3 TT, and Harnad also makes his explanation and prediction in the state-of-art vector grounding paper.
Zoe, Categories are distinguished from each other by their -- different features of the same object may put it into different categories (which differ in what would be the "right thing to do with them"). And features themselves are categories too, so could be named. The vocabulary we lexicalize (give a name, and put in our dictionaries) depend on which ones are important to us -- to reduce uncertainty about what to do. I think that's what's behind what you mention.
DeleteJiajun, "Consciousness" is a weasel-word. Let's use "sentience", which means the capacity to have states that it feels like something to be it. The only thing that has been said about that so far is that Searle argued that if executed a (hypothetical) purely computational algorithm that could pass T2 in Chinese, he would not be understanding, and therefore neither would a computer executing the same algorithm.
Grounding is about connecting content-words to their referents so that the thinker can pick out and interact (T3-scale) with their referents in the world. This is not about "correlating with reality". GPT's words and sentences correlate with reality, but GPT is not grounded, and does not understand or mean anything. It's just manipulating words.
GPT has been hard-wired to keep reverting to unresolved to "complexities": a weasel world that says and means nothing (except in formal complexity theory). The last two sentences of your comment sound like GPT-speak and are incomprehensible to kid-sib. It sounds to me like it just says cogsci needs to find a way to reverse-engineer T3 capacity. Feeling ("consciousness") has nothing to do with that.
The statements in “The symbol Grounding problem” impresses me ：“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).” (Harnad)
ReplyDeleteIn my understanding, top-down processing and bottom-up processing are two types of processing.And these two kinds of processing cannot be carried out alternately.
Top-down processing starts with expectations & context to help sense/interpret incoming data stream; while bottom-up processing starts with distal stimulus or sensory data and build up representation. It is very strange that when two processing will meet together at some point and it doesn’t make sense. If they meet at one point, that might be only the stage of “perception” and “thought process”; however, it is not meaningful because they will be back to none since the end point on their journey is the opposite. End point of top-down processing is the starting point of bottom-up processing, vice-versa.
"Top-down" has many meanings and is somewhat weaselly, but symbol-grounding has only one meaning: connecting the referents of words to the words that refer to them. And there's only one way to do that: bottom up, building from direct sensorimotor interactions with the referents of words by learning the features that distinguish different referents (categories) from one another. Then language allows indirect grounding of further words' referents through verbal descriptions made up of propositions composed of already grounded names of the referent category's distinguishing-features.
DeleteI have some thoughts on this sentence from “Symbol grounding problem” :
ReplyDelete“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” (Harnad)
Modern technology can actually perform batch tasks in a short time. Whether it is classifying or sorting data, machines can complete the tasks. However, the symbols in the article require the machine to connect to the next point by itself to complete the task, but there is a problem, consciousness. Machines want to connect symbols, but current technology is limited to human control. Humans act as the consciousness of the robot to help the machine complete its tasks.
I think this is not a bad thing. Although the speed of completing the task will not be reached, making the machine conscious is a difficult problem according to current science and technology. Secondly, if machines are conscious, they will pose a certain threat to human life. They will slowly develop their own independence and try to escape from human control. Regarding consciousness, and whether machines should have it, harm is more than the good.
Except for Searle's use of the "Periscope" to show why executing a Chinese-T2-passing algorithm will not understand Chinese (because it feels like something to understand Chinese), the Hard Problem of reverse-engineering "consciousness", i.e. FEELING, will not be solved or treated in this course till Week 10.
DeleteThe rest of the course is about the "Easy Problem" which is reverse-engineering the capacity of thinkers to DO all the (cognitive) things they can do.
DeleteProfessor Harnad's notable contribution to artificial intelligence philosophy, notably the Symbol Grounding Problem (SGP), not only critiques the limitations of contemporary AI but also highlights the critical role sensors play within cognitive systems. The prevalent strategy in AI learning currently involves converting natural language into a machine-readable programming language, an approach inherently subject to entropy losses - with entropy referring to total informational content in this context. AI demonstrates that directly translating external systems into symbols without perceptual intermediaries results in informational loss, particularly when the input is in human language. Introducing conscious awareness to this process, the substantial informational loss during reverse engineering becomes catastrophic, a problem deeply rooted in grounding issues. Moreover, contrary to what is often assumed in numerous simulation environments, real-world agents do not receive neatly structured input vectors but are rather subjected to a ceaselessly variable stream of sensory stimuli, largely dependent on the agent’s current behavior. Additionally, research indicates that when GPT-4 is given a prompt, allowing two instances of GPT to converse, they begin to recycle their responses after a finite number of exchanges, unable to break free from the existing semantic space. Hence, Language Learning Models (LLM) intrinsically lack meaning, with humans being the ones assigning meaning to LLM’s language outputs.Therefore, an ungrounded LLM does not have any practical significance, and may even fall into a philosophical trap.
ReplyDeleteBrain in a Vat: On Missing Pieces Towards Artificial General Intelligence in Large Language Models(https://arxiv.org/abs/2307.03762)
DeleteThe problem with ungrounded words is not entropy; it's ungroundedness.
DeleteThe difference between grounding and meaning is essential to understand what the symbol grounding problem really is. Grounding is a bridge between the sensory input we get from the external objects we interact with, and the internal symbols (words, pictures) our brains have stored. In the context of language learning, let’s say you know that “piros” means “red”, and “zöld” means “green”. Those symbols become grounded. If I then point towards a red apple saying “piros alma”, and a green apple saying “zöld alma”, it is possible for you to infer and understand, using contextual cues, that “alma” means “apple”, which then becomes grounded. If neither words were grounded, I could tell you to point at the green apple saying the string of words “zöld alma” without you understanding what it means. Therefore, although groundedness is necessary for meaning, it is not sufficient for there to be meaning.
ReplyDeleteSee other replies about his hybrid grounding above. (What is the difference between reference and meaning? And between grounding and meaning?)
DeletePost #1: Harnad, 1990
ReplyDeleteI very much appreciated the clear explanation of what a symbol system is, I thought it made it much easier to understand how connectionist approaches differ from symbolic ones. We must answer the question: how does language gain meaning such that it is part of the language system itself and not external (only has meaning via some interpreter). Without connecting these symbols to the real world, they would be meaningless, Harnad points out it'd be like trying to learn the meaning of a Chinese word by looking it up in a Chinese-Chinese dictionary. Harnad argues that any understanding of a semantic interpretation of a formal symbol system comes from the reader, not from the symbol system itself and this point of view is supported by Searle's CRA. Harnad points to a common symbolist which is "just give the formal symbol system wheels and a camera" and in the footnote claims this argument is homuncular (I'd love an explanation why) and that it trivializes the symbol grounding problem. Instead he proposes a hybrid non-symbolic/symbolic system where the symbols are grounded either directly through sensorimotor experiences of by their relation to other symbols which have been grounded in sensorimotor experiences. If I understand correctly, these experiences give rise to categorical and iconic representations of words which ground their meaning in the real world. I'm really not sure I understand this part. What are these representations? Where are these representations? How do they give meaning to the symbol system that is not external? It seems to me that the forming of these categories and icons could be accomplished computationally (symbolically) so all you would need is access to the stimuli but Harnad claims that just adding wheel and a camera is a homuncular point view. I feel as though I've totally misunderstood something or missed something so any help that could be provided would be much appreciated!
See other replies in this thread: both symbolic AI and connectionist AI are computational.
Delete"Computer with wheels and camera" assumes the computer understands, and just needs to be able to look outside. That's the homuncularity. But see also the reply about the Sorites Paradox, about scaling up to T3.
Direct sensorimotor grounding is learning through supervised learning (mushroom-island-style trial-and-error) the features that distinguish members from nonmembers of the referent category. Indirect verbal grounding is to be told the distinguishing features in grounded definitions or descriptions.
I won't be able to reply to such long commentaries.
I found Prof. Harnad's 1990 reading very interesting. Particularly that the most promising answer to the symbol grounding problem involves a synthesis of both symbolic and connectionist systems. Connectionism helps us understand how exposure and feedback enable the association of objects with symbols through learning, relying on consistent patterns of sensory projections. On the other hand, the symbolic approach demonstrates that once we establish a foundation for elementary symbols in this manner, the remaining symbol combinations in a natural language inherit the grounding from these foundational symbols. When these apparently conflicting approaches are combined, they create a hybrid system that stands as the most viable solution in addressing the grounding problem.
ReplyDeleteYes, but the grounding is not "association." It's sensorimotor feature-learning. See other replies.
DeleteProf Harnad's 2003 article about the symbol grounding problem has deepened my understanding of its connection to many of the topics we've looking at. The crucial attribute enabling symbol grounding is the brain's ability to identify their corresponding real-world entities. This process requires our sensorimotor capabilities, enabling autonomous interaction with objects and properties in a manner that aligns with how humans interpret symbols as referring to these entities. For a symbol to be grounded, it must establish a connection to the entities it represents, ensuring that our sensorimotor interactions with the world align with the symbols' intended interpretations. A connection based on trial and error induction guided by the feedback from the reward of a correct categorization.
ReplyDeleteSee above reply, and other replies in this thread.
DeleteI am wondering whether looking at the symbol grounding problem from the field of developmental psychology would be helpful. Maybe looking at child development, and access how a child can learn how to relate language to tangible objects at first, and subsequently more complicated concept, would give us some clues about the underlying process.
ReplyDeleteIn Week 8 you'll see that the grounding words of dictionaries and more frequent in the language and learned at a younger age.
DeleteAll three articles seem to presuppose that all referring words have referents. For example, the scholarpedia article says "[s]urely all the (referring) words on this page, for example, have meanings, just as they have referents". But I don't think that we can take this for granted. I align myself more with Chomsky's view on this matter, namely that NO word has an objective referent (which I believe is the important criterion of "referring" here, since we are talking about the Symbol Grounding Problem, I will explain more later). The example of Aristotle's that I remember is the word "house": part of the word house is the physical components of the house, but the "form" of the house that we really just construct in our head is also necessary for it to be a house. The physical realization of the house does little to restrict what us humans think of as a house, the house could be anything as far as physical reality is concerned, even "a paperweight for a giant" (if giants were real) as Chomsky says. Or equally, think of the word "river": you can never cross the same river twice because its physical composition is constantly changing, but of course, we humans think of it as crossing the same river twice. The same goes for "Tony Blair", whose cells are constantly dying and being replaced, and for any living organism, and for really anything at all that a word in a natural language can pick out. Words aren't really grounded to referents at all, and just like meaning, referring also seems to be dependent on internally constructing what a thing is since anything that is out there in the world cannot be picked 1-to-1 by a word in human language.
ReplyDeleteIt will be helpful in this course to keep in mind that it is about cognitive science, not physics, metaphysics or even epistemology. Cogsci is trying to reverse-engineer how and why human beings on earth can DO what they can do.
DeletePeople can refer to apples, they can point them out, pick them and eat them, and they can describe them, their features, and what can be done with them. They can do the same with houses and paperweights, and giants, and Tony Blair).
Cogsci has to explain those capacities. They are real.
People can also refer to, recognize instances of, and describe apprehension, appraisal, and appositeness, which are not concrete objects you can pick up and eat. Cogsci has to explain this capacity too.
Many words are "polysemic," which means that the same word-form can have multiple senses, referents and meanings. Almost all words can be used metaphorically. Referents can be fictional. And distinguishing features may change, with a change in knowledge about them, or even a change in fashion. Cogsci has to explain these capacities too.
Once enough referents (and the right ones) have been grounded directly through sensorimotor category learning, indirect grounding of the referents of further words becomes possible, based on verbal descriptions of their features (as long as the words describing their features are themselves among the already grounded words).
The fact that Tony Blair's cells are dying and regenerating does not prevent us from recognizing, interacting with and talking about him -- not even in philosophy classes on the metaphysics of identity. It just takes a lot more words to talk about those nuances.
Moreover, the referents of most words (like "house") are categories, not individuals, and those categories have distinguishing features, whose names must be grounded categories too. They may be grounded indirectly by words too, but it can't be indirect grounding indefinitely. Features eventually have to be cashed in through direct sensorimotor grounding.
And both sensorimotor features and verbal definitions are approximate, not exhaustive or exact (except in maths or logic, where they are formal syntactic rules and their formal consequences rather than empirical observations about what, how, and why humans beings on earth can DO what they can DO.)
Noam Chomsky will make his appearance in this course in due course -- because linguistic capacity IS part of what Cogsci has to reverse-engineer (Chapters 8 and 9) -- but not his (or Aristotle's) speculations about how the brain grounds its words in its referents, or whether there exist any referents to ground.
in section 3.3 of the 1990 Princeton university reading the example of combining horse and stripes to form zebra is laid out. this example helped me to think that what splits names and symbols and how we identify a zebra is not just the stringing together of the concepts horse and stripes. if I'm being picky about it: the stripes of a zebra really aren't just flat out stripes (like the ones on a mime t shirt). while we can infer that if we see a horse the stripe-like pattern of a zebra it is probably a zebra (because horses don't typically have stripes) we cant just look at a horse painted in parallel horizontal stipes and say this is a zebra and still be right
ReplyDeleteYou're right, but see the preceding reply about the approximateness of features, both sensorimotor and verbally described.
Deletereading the scholarpedia article, made me think about referents which are less grounded than others. for example: we can point to the apple and say apple and mean "fruit with core that I can bite into", we can correctly call an apple an apple and proceed to do the right things with an apple and that's hard enough to work out as it is. but what about things which are more abstract like "courage" or "belief" and how do we categorize these referents such that we can also do the right things with them. not sure.
ReplyDeletePlease always read the other commentaries and especially the replies, before posting yours. This question has been raised many times. Just string-search for "justice".
DeleteNicole, excellent questions and answers. But the trouble with your hybrid grounding by elimination [saying or showing a grounded category + showing an ungrounded one] is that you can't show an ungrounded category! You can just show a membertell its distinguishing features, then you've made it almost all instruction (telling). Do you see that?
ReplyDeleteBut your ideas are good, and dictionaries (and instructors) sometimes do hybrid definitions, not just describing features verbally, but also showing an illustration.
Your question about levels is good too. First, not only all definitions but also all learned feature-detectors are approximate (just as theories are). The approximations can be tightened, but never exhaustive (except in formal maths and logic: necessary and sufficient conditions). We'll talk about this more in Week 6, on categories.
There is also a hierarchy in dictionary definition space (Week 8) like the one you try to describe. (J is defined in terms of features H and I, which are defined in terms of features D, E, F, G etc. These are definitional distances in indirect verbal grounding. But it can't be indirect grounding all the way down; some categories and their names have to be grounded directly (by trial and error, like on the mushroom island, to find the distinguishing features; and to tell those to someone too, they too need to become learned, named categories, like the features of the edible mushrooms).
And there is another hierarchy even in direct grounding, as you ground more and more general categories: apples, fruit, food. The higher-level, more general categories share fewer and fewer of the sensorimotor features (that's how they become more "abstract"). But that does not mmean they become less sensorimotor
And to make it (seem) more complicated still, generalization and feature-abstraction is not strictly hierarchical in one upward direction: You can have apple, fruit, food; or you can have seed, apple, tree.
Harnad, S. (2003) The Symbol Grounding Problem
ReplyDeleteThe scholarpedia summarizes everything we have learned so far section by section, in a way that is simple and easy to understand. I found this particularly interesting - “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” If everything in the world is pretty much meaningless until we humans gave meaning to it, then does it really matter if the robot understands the true meaning of the symbols and has grounded experiences if it is still able to connect/relate the information it knows?
We will discuss what "matters" in Week 11.
DeleteBut what you are asking about here is not whether "meaningfulness" (a weasel-word) matters: You are asking how and why FEELing matters. And answering that would require solving the HP.
But "matters" is a weasel-word too: Matters to whom? And how and why?
Stay tuned.
Harnad, S. (1990) The Symbol Grounding Problem.
ReplyDeleteCompared to the scholarpedia, this article explains the symbol grounding problem in a more detailed manner. Two theories about cognition are introduced: the symbol system - a set of symbols manipulated only based on its shape according to a set of rules, and the connectionist system - “dynamic patterns of activity in a multilayered network of nodes or units with weighted positive and negative interconnections.” To understand cognition, we can start with computation, then use connectionism to explain how the symbols reach its semantics.
Neural nets ("connectionism") that can connect words to their referents by learning to detect their distinguishing sensorimotor features can still be computational. Almost all the nets we read about are actually just computational simulations of neural nets. They are not really parallel and distributed. But unlike with computationally simulated ice-cubes, the difference does not matter. They could ground words to their referents anyway. But the sensorimotor actuator and effector components cannot be just computational simulations.
DeleteSo it's not (and never was) "computationalism" vs. "connectionism."
It seems that the text posits that for computers to exhibit the same kind of understanding that humans do, the symbols used by the computers would have to be grounded. They would need to be capable of identifying referents, and to achieve that, the symbol system would require the ability to interact with the external world. I would argue that, since knowledge is built upon previous knowledge, interaction with the external world would only be necessary for a set of fundamental units of the external world. The symbol system could then perform computations based on these fundamental units.
ReplyDeletePlease read the other replies. It's a robot, which cannot be just a computer. (Why not?) What is the symbol grounding problem?And what do you mean by "knowledge"? What kind of "fundamental units"? And what kind of "interaction"?
DeleteIf I am thinking correctly, a robot differs from a mere computer due to its capability to interact with and sense the external environment. By "knowledge built on previous knowledge," I’m not entirely clear on what you mean, like a database that’s updated? Though I remember in the discussion in class, we didn’t want to touch on acquiring knowledge but the concept of interpreting knowledge and also acting upon the environment. Are you imply accumulated understanding derived from sensory input?
DeleteThe "fundamental units" likely refer to basic sensory experiences that form the foundation for more complex knowledge. "Interaction" would mean the robot's direct engagement with the environment, allowing it to ground symbols in sensory data. Would that clarify the discussion, Liam and prof?
Please explain Searle's Periscope. What does it show, how?
ReplyDeleteWe have not been discussing how perception varies between individuals. Human words are grounded: How?
The HP is explaining "how and why we can feel at all", not "how and why some people feel this and others feel that."
What are Cogsci's EP and HP? And what is reverse-engineering?
While reading the article "Solving the Symbol Grounding Problem," I could not help but become increasingly confused by the Zero Semantical Commitment condition. The first element of the condition states that a solution to the SGP cannot possess any form of innatism. That is, no semantic resources should be presupposed in the artificial agent. To me this seems to be much too strong of a condition, and frankly, one that seems to ignore the role of Darwinian evolution. This is most evident when the paper refutes Sun (2000)'s Intentional Model. Sun states that the AA would first interact with the environment in a random, trial and error way, paying attention to the structure of the environment and the innate biases of the AA. The paper argues that these innate biases already violate the Z condition, and therefore cannot be a solution to the SGP. But this seems to ignore the innate biases that would have manifested through Darwinian evolution mechanisms. It does not seem like it would be possible to ground symbols without some degree of innate semantic resources guiding the interactions the AA is having with the world. It is most definitely the case with us that we possess a degree of innatism in regard to semantic resources.
ReplyDeleteYou are right. But what is the SGP? and what would be its solution?
DeleteThe SGP (I think) refers to the challenge of grounding the meanings of symbols in something other than more symbols, like in sensory experiences or real-world references that we are inquiring about. Its solution would require a mechanism which an artificial machine can derive meaning from symbols based on its interactions with the environment, rather than relying solely on pre-defined semantic relationships. While the Zero Semantical Commitment condition seems stringent, its intent is to ensure genuine symbol grounding. However, if we consider evolution, can we truly rule out any form of innatism in developing semantic understanding? Ishan, how do you think the balance between pure symbol grounding and some innate biases can be achieved?
Delete1. Cogsci is just trying to reverse-engineer and test T3 (or T3 & T4) for Turing indistinguishable DOing capacity (and hoping that it can also FEEL, but not being able to test that (because of OMP) or even able to explain, if it does feel, how and why (because of HP).
ReplyDelete2. Sensorimotor functions are physical, not just computational.
3. The brain does not "associate" words to their referents, it makes a direct sensorimotor connection, through robotic interaction, manipulation and feature detection. You can't get that, T3-scale, by just adding a camera and wheels to a computer. See again the Sorites reply.
4. The difference between computational and dynamical is the difference between a computer simulation of an ice-cube and an ice-cube.
5. I won't be able to reply to such long commentaries again. Please keep it under 150 words.
Yes, there are some categories that are unclear, or that we don't all agree on. All categories (hence their distinguishing features, whether sensorimotor or verbally described) are approximate and provisional, not exhaustive or definitive (except in maths). That's true of categorizing (on the mushroom island) as well as the categories of high-level abstract discourse.
ReplyDeleteAnd examples are not to be sneezed at. You can learn by trial and error what most people call "fair" or "just" from positive and negative examples plus social feedback. In that sense it's not just sensorimotor categories, like the mushrooms, whose distinguishing features can be learned either directly, by trial and error, or indirectly, by (grounded) verbal descriptions of their features: You could learn what most people consider "fair" or "just" and "unfair" and "unjust" from examples, by trial and error with social feedback (and perhaps some feedback from your empathy mirror-neurons). But in both cases (mushrooms and equity) a provisional (but grounded) verbal definition from someone who knows to someone who doesn't know is incomparably faster and more efficient (and less risky) than trial and error.
Of course disinformation can make hearsay categories really risky too...
See the other replies on this. You're not the only one who is struggling with the HP.
ReplyDeleteTuring was right that Cogsci cannot hope to do better than solving the EP by reverse-engineering DOing capacity and T-Testing (T2, T3, and T4) to test whether it has succeeded.
This is not just because of the OMP, which prevents testing for the presence of "consciousness" (feeling), because feeling is unobservable (except by the feeler).
Even if an omniscient deity could guarantee that your reverse-engineered TT-passer was not a zombie, we still could not explain how or why. So the HP would not be solved. (But don't ask me if an omniscient deity could solve it!)
We'll get to why the HP is so hard in Week 10. The short answer is that it's because of the solution to the EP.
I believe I understand the basic ideas in the symbol grounding problem, however, I am slightly perplexed by the idea that computers cannot ground symbols. I see how they just manipulate symbols and give output, but couldn’t there be a way to make connection networks and that are tied to a referent and store them, similarly to how we ground symbols to their referents. I may be missing something, but could someone clarify this for me.
ReplyDeleteSkywriting on SGP #1:
DeleteThis is of course from my understanding of the SGP, which might be wrong!
The issue in what you bring up is the "tying to the referent", which as out dear professor formulated, for that, there must be a connection between the "names" and "things-out-there", a connection between a symbol system, and a nonsymbolic, sensorimotor system.
The issue isn't only a software problem, it's as much a robotics (hardware) problem, and how to connect them. To be grounded, the "names" must be connected to sensorimotor notions, as names by themselves will lead to an endless loop of definitions.
Take the dictionary example, one looks up a definition, which brings one to another, and so on, it isn't grounded in what is "out-there". For grounding to happen at some point (or more likely in many) the names have to be connected to non symbolic. For example, let's deal with the word "ground", for "ground" to be grounded it isn't enough to have a definition, "that which structures and entities stand upon" must be connected to the feedback from muscles, the sight of that which usually lies beneath things, the intuitive understanding of gravity which we learn from our first steps (and first falls).
So a computer can't ground things by itself, what could possibly ground would be a system of which the computer is a part, which would be closest to what we think of as a "robot", as the computer (which only deals in symbols), would need to be connected to mechanical muscles and eyes, that can measure pressure and light, to which the symbols would then need to connect.
Johann, correct, but it's not just that the T3 robot has to have optical and acoustic (etc.) input and motor output but that it must be able to learn the features that distinguish and identify inputs from different kinds (or categories) of things, and aable also to learn what to DO with those things (besides naming them). That's why a grounded T3 robot is not just a computer with a camera on wheels.
DeleteI understand that a T3 passing robot must be able to go out in the world and correctly do things indistinguishable to a human, which entails more than just a camera and wheels. When it comes to the symbol grounding problem and T3, I wonder if a T3 passing robot would have to ground in exactly the same way as us, which is sensorimotor. I understand that it is necessary to ground via sensorimotor learning, as if it were only done by being told which category is which by whoever made the machine, it would be an infinite regress. Also, if we had a robotic chat-GPT that is able to do what we do in the real world, but it got all that information via the Big Gulp and neural nets, would this be a sufficient T3 passing robot that is properly grounded? I understand that it wouldn't, as the training data was already grounded by thinkers, and it simply makes connections based on those.
DeleteFor sensorimotor symbol grounding a T3 robot needs a body and head, not just a neural net. What the neural net is good at is learning to detect the features that distinguish the members of a category (e.g., apples) from others (e.g., billiard balls) so the T3 robot can do the right thing with the right kind of thing.
ReplyDeleteIf the category has a name, the features ground the symbols that are inside the robot's head in the robot's capacity to detect and act on the referent of the category-name. Algorithms manipulate symbols, but robots manipulate apples. (How much of what's going on inside a T3 robot that is just computation still remains to be discovered by Cogsci.)
Extra skywriting: In the 1990 and 2003 readings for “The Symbol Grounding Problem,” Professor Harnad proposes that the symbol grounding problem can be addressed through categorization. “Categorization” involves the appropriate handling of the correct type of thing (Harnad, 2003). More specifically, categorizing skills can be innate or learned, depending on what exactly is being categorized; this ability, done through sensorimotor interactions with the world, acts as a medium for which “grounding” can be achieved (Harnad, 1990). Regarding these parts of the reading, I have a few questions. In this discussion of innate vs. learned categorizing, what would be specific examples of this? Would language acquisition, as highlighted in Chomsky’s Universal Grammar Theory, be an example of an innate ability to categorize? Are there forms of categorizing that could be both innate AND learned?
ReplyDeleteCategorization can be complex and have many subjective layers. Given the subjectivity that can be involved, to what extent can grounding abilities vary across different people? How might these differences show on a neurophysiological-level? To add, as discussed, categorizing an abstract or intangible concept would involve “indirect grounding.” To my understanding, this means that “grounding” is executed not through sensorimotor experience; instead, abstract concepts would be “grounded” by having to connect to previously “grounded” information. With this in mind, are there limitations to grounding through categorization? In other words, if we know that the human brain is naturally able to create meaning, do these ideas of direct and indirect “grounding” through categorizing really capture the full complexity that may be involved in this process?
An example of learnable grammar is Ordinary Grammar (OG) (Week 8 and 9). We can learn that by imitation, trial and error and correction, or verbal instruction.
ReplyDeleteWith Universal Grammar (UG), we don't make any errors, and neither does anyone else. So we can't learn UG by imitation, trial and error and correction, or verbal instruction. But we don't make any UG errors, so it must be innate.
Maybe phonemes like ba/da/ga are partly innate and partly learned.
I don't know what you mean by "create meaning." Indirect grounding just means learning the referent of a word you don't know through words. This can be done from a dictionary definition -- as long as all the (content-) words in the definition have already been grounded for you, either by direct sensorimotor category learning or indirectly, by (grounded) definition.
"Concept" is a weasel-word for; usually it really just means a category, and sometimes something something as vague as an "idea."
There are limitations to grounding word referents' directly, but none on grounding them indirectly, through definition, description, explanation and instruction -- all through words.
Many categories are grounded both ways, by telling (words) and showing (as in illustrations with a dictionary definition.
But you can't learn a category by just showing: why not?
(Skywriting 1) The Symbol Grounding problem, 2003. I really liked this reading. The description of "a symbol system" was particularly interesting to me.
ReplyDeleteHarnad writes "The symbols are systematically interpretable as having meanings and referents, but their shape is arbitrary in relation to their meanings and the shape of their referents."
I was instantly reminded of the computationalist approach to cognition. The idea that these symbols only have meaning if they are within a system and that their shape is independent from their meaning is an exact parallel to: different parts of the brain have different functions but it does not matter exactly which one. It made me realize the reason we segregate different areas of the brain is in order to create functions.
Because the brain is a unit of sorts, we have to create the system ourselves... I thought it was an interesting thought.
In the realm of robotics, symbol grounding can lead to robots that interact more intuitively with their environment. Instead of relying solely on pre-programmed rules, robots could use their sensory input to ground symbols in real-world situations. This would enable them to adapt to dynamic environments, collaborate with humans, and even learn from their interactions. The practical applications of symbol grounding extend beyond AI and robotics, potentially revolutionizing fields like healthcare, education, and customer service. By deepening our understanding of how symbols are grounded in real-world experiences, we can enhance the capabilities of machines and systems, making them more effective and integrated components of our daily lives.
ReplyDeleteWhat are the ethical implications of symbol grounding (if any) would be present in the context of artificial intelligence and machine learning given it is even possible?
Harnad, S. (1990). [The symbol grounding problem](http://cogprints.org/615/1/The_Symbol_Grounding_Problem.html). *Physica D: Nonlinear Phenomena*, 42(1), 335-346.
ReplyDeleteHarnad tries to reconcile the purely symbolic and purely connectionist models of the mind by combining combining the two into a hybrid model. What stuck out to me was the link back to behaviourism and how this reconciliation echoes its sentiments. This argument is at its core what of the biggest obstacles in cognitive science; we can never definitively say “whether a semantic interpretation will bear the semantic weight placed on it,” for example. The tests we use to form judgements can at most be used to generate an educated guess. Also, who is to say whether a test is passed? In the case of the behavioural test, there is an infinite amount of objects that it is impossible to comprehensively assess whether a semantic interpretation can “discriminate, identify and describe all the objects and states of affairs to which its symbols refer”.
If referring to Searle's Chinese Room, it demonstrated that understanding isn't achieved merely by manipulating symbols, as a program would.
ReplyDeleteWhile human words are grounded, for example, the word "red" is grounded in our visual experience of red objects.
It’s because words are rooted in our sensory and motor experiences. This article helped put a term down that I couldn’t put an exact finger down for this concept we’ve been gleaming on since week 2.
The HP of cognitive science is about understanding consciousness—why and how sensations arise. EP (Easy Problem) deals with explaining cognitive functions, like discerning and responding to stimuli.
Reverse-engineering is deconstructing a system
to understand its design and functioning, which can provide insights into human cognition if we see robotics as a reflection of our neural processes.
Hey Stefan, I agree with you in that artificial intelligence's ability to symbol ground and utilize sensorimotor interaction with referents would absolutely take it to another level. In regards to how advanced we perceive it to be in its ability to present itself as what is typically deemed 'conscious'. However, I would refer you to Anaïs. T3 passing sensorimotor bot which is capable of symbol grounding using referents of the real world. I don't believe it is possible to prove Anais' consciousness, however, I also don't believe it matters much, given if it appears conscious to me, that is the only way I can judge it, in the same way it is the only way I can truly judge your consciousness.
ReplyDeleteI also had very similar thoughts about the consciousness and the zombie passing TT example. The example of a zombie passing the TT without consciousness, in the sense of lacking subjective experience, implies that a system could exhibit intelligent behaviour and even pass a TT without having consciousness or subjective experiences. This is a strong challenge to the assumption that consciousness is an absolute prerequisite for intelligent behaviour or meaning in my opinion. If subjective experience can be ruled out of consciousness, what is left that defines us as human? It seems as though the threshold to being truly conscious or human drops with every week of class hahaha.
ReplyDelete