Saturday 2 January 2016

5. Harnad, S. (2003) The Symbol Grounding Problem

Harnad, S. (2003) The Symbol Grounding ProblemEncylopedia of Cognitive Science. Nature Publishing Group. Macmillan.   

or: Harnad, S. (1990). The symbol grounding problemPhysica D: Nonlinear Phenomena, 42(1), 335-346.

or: https://en.wikipedia.org/wiki/Symbol_grounding

The Symbol Grounding Problem is related to the problem of how words get their meanings, and of what meanings are. The problem of meaning is in turn related to the problem of consciousness, or how it is that mental states are meaningful.


If you can't think of anything to skywrite, this might give you some ideas:
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?
In M. de Vega (Ed.), Symbols and Embodiment: Debates on Meaning and Cognition. Oxford University Press.
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)

93 comments:

  1. - I’m curious how symbols that can have multiple referents work (e.g.: bow as a verb or as a noun, # as number or as hashtag, etc.)? Is it because we’re viewing each context as being part of a different symbol system? Is it cause there can be multiple iconic representations associated with a symbol?
    - How are different symbols with the same referent connected with each other (e.g.: ‘plus’ vs ‘+’)? Or is this part of the SGP?
    - With the zebra example, how does one form a superordinate category if it seems that symbols are categorized on the basic level, and combine to form subordinate categories?
    - Also, from Kripke’s rule-following paradox, how is it that we know what interpretation/rule to apply to a symbol when interpretations themselves do not determine meaning?

    ReplyDelete
    Replies
    1. Multiple meanings is called "polysemy." Automatically disambiguating polysemous words in computational linguistics is a problem, but not an insurmountable one, with the help of context, along with knowing the multiple meanings. Where the uncertainty among multiple meanings of the same word becomes a problem, we invent new words to distinguish them. (Often the multiple meanings of the same word are related, as in: bear a child, bear weight, bear discomfort, bear tidings -- but not bear in the woods, or bare naked...)

      Meaning is not the same as reference, and multiple words all having the same referent is almost the opposite of polysemy. As Frege noted, the referent of the "Morning Star," "Hesperus, the "Evening Star," "Phosphorus," and "Venus" (the planet) is exactly the same. All five refer to the planet venus. But they don't have the same meaning. That shows that meaning (also called "sense") and reference, though related, are not the same. The sense (or meaning) of a word or phrase is the way you pick out its referent, by calling it by its proper name (Hesperus, Phosporus, Venus) or by s describing it (the morning star, the evening star) using other words.

      You can use words to create higher-or-lower order categories: mammals are animals that nurse their young; bats are mammals that fly; all mammals are warm-blooded, and so on.

      And there's no such thing as "basic level" categories. What we refer to with that descriptor is the default level at which we usually talk about things, but the level is arbitrary and varies with context and background. The category hierarchy goes on and on above and below it, and it isn't even a hierarchy, it's a network, like the web. Connections can be up, down or criss-cross.

      In a 3-D (XxYxZ) vector space, any point (e.g., (2, 4, -7) can be expressed as a linear combination of the basis vectors 2x(1,0,0) + 4x(0,1,0) - 7x(0,0,1) or it can be a (complicated) linear combination of any 3 linearly independent vectors -- but you have to start somewhere, and in 3-D space you need at least 3 linearly independent vectors.

      It's similar with meaning. You need a certain number of "grounded" words to start. Then the rest can be described or defined using combinations of those grounded words whose meaning you already know.

      But how many do you need?

      And which ones?

      (We'll be discussing this later in the course.)

      Delete
    2. - I find it confusing why, then is there a need to assign multiple meanings to a symbol? If the multiple meanings is just dependent on context, then there is no need to create new words or have a vast vocabulary of morphemes; everything could just be depending on syntax to give meaning to the symbol in the context.

      "The sense (or meaning) of a word or phrase is the way you pick out its referent, by calling it by its proper name"
      - Can you elaborate on this part a little bit more? I'm a little puzzled by how a meaning of a word is dependent on its name if the name is made up of meaningless symbols.

      - Since we're going to talk about these "grounded" symbols more later on in the course, I just want to know if the grounded that you suggest we need to start with is innate or not, because there are individuals that are born with disability in an area of sensory perception, thus I would assume to be "grounded" would be of different interpretation for those unless these grounded words were innate? Or am I completely off the mark?

      Delete
    3. 1. Polysemy. I'm not an expert on polysemy, but my guess is that the reason there are multiple meanings to the same word (symbol) in natural language is that many of the later meanings of a word evolved from earlier meanings. The multiple meanings are not "assigned"; they evolve with usage. It's no problem for human speakers, who understand the words and context, but it's a big headache for computer processing of text. (And context is not syntax; it depends on word meaning, not just word shape.) The symbols in formal computer programs should have only one meaning (interpretation), but the meanings are not in the symbol system; they are in the head of the user (interpreter).

      (The only exceptions to single meanings in maths and computation are what are called mathematical "duals." Relatedly, sometimes the very same symbol system can have multiple interpretations.

      2. Sense and Reference. "Chair" refers to those things you sit on that usually have four legs. What connects the word to them is your senses and your brain's (learned) sensorimotor feature-detectors: You have learned to recognize the members of the category "chair." You can also describe chairs and define "chair" in words. There's a second sense of "chair": the head (!) of a department or a meeting. You can recognize, describe and define the chair in this second sense too. (It derived from the first.)

      3. Categories, Symbols, Grounding and Learning. Before a category name can be grounded, you have to be able to distinguish the members from the non-members. (Not necessarily by naming them: by doing the right thing with them, which might be sitting on them, as with chairs.) Once the category is known, its name can be easily grounded. Most categories are learned. Some are inborn (but not their names, obviously), e.g., colors, facial expressions, perhaps phonemes and, for musicians with perfect pitch: pitch.

      Delete
  2. First I’m going to take a stab at a kid sib version of the symbol grounding problem.

    The symbol grounding problem says that only words that exist in our head have the capacity to be grounded or connected to referents and therefore have meaning, while words on paper or in computers cannot be. This is partially because words on paper lack the dynamic sensorimotor capability of humans, and also because of consciousness which is (Steven Says) impossible, and “there is no way we can hope to be any the wiser”

    The symbol grounding problem seems to be focused on a) referents in the “outside” world and b) the mechanism that links the referent to the word to create meaning. The article proposes one property of the brain that allows words (symbols) to be given meaning to understand language. The proposed property is that it is dynamical and therefore implementation-dependent. This intuitively makes sense; in order to connect reference to word we must have sensorimotor experiences.

    “[The symbol system] would have to be able to pick out the referents of its symbols, and its sensorimotor interactions with the world would have to fit coherently with the symbols’ interpretations”.

    Lets set aside the possibility that a robot would have sensorimotor capabilities and just consider us (humans). How in the world do humans know or feel that whatever it is they are sensing is the referent to connect to the word? How do we know that what we see, feel, taste and hold is an apple rather than a pear? Are these pairings of referent + word always taught? My intuitive answer is no, but the symbol grounding problem doesn’t seem to explain how. How are we not completely swamped by all of the information coming into our brain at once?

    ReplyDelete
    Replies
    1. Problems, Problems

      1. Grounding does not require consciousness (unless you can solve the hard problem and explain how and why!). Grounding requires connecting symbols to their referents using T3 (robotic) capacity. What we "can't be any wiser" about is whether a T3 grounded robot (or anyone) is conscious (i.e. feels), but that's because of the other-minds problem, not the symbol grounding problem.

      2. The sensorimotor connection between words in the head (human or T3) and their referents in the world is necessarily dynamic (as all sensorimotor activity is dynamic). It cannot be just simulation (i.e., symbol/symbol connections): The symbols "in" a simulated "T3" robot in a simulated world are not grounded in their referents. They are just squiggles connected to squoggles. The symbol-referent grounding has to be real (sensorimotor) -- but explaining whether, how, or why it needs to be felt (conscious) would again have required a solution to the hard problem. Turing's "solution" is simply to assume it's so, because we can never tell the difference.

      3. We learn what symbol connects to what referent through trial-and-error sensorimotor category learning (week 6). The world "corrects" us when we do the wrong thing with the wrong kind of thing and it "rewards" us when we do the right thing with the right kind of thing. (That's categorization.) The feedback from the trial-and-error learning allows our brains to figure out how to tell apart the members from the non-members of the category, and so to make the connection between symbols (names of categories) and their referents. The pairings of symbol and referent can only be taught (i.e., verbally) once you have a language, plus enough grounded symbols to explain (describe, define) the meaning of any new symbols through verbal instruction. But the grounding of those first symbols cannot come from verbal instruction: it has to come from direct sensorimotor induction (trial-and-error category learning) guided by feedback from doing the right or wrong thing with the right or wrong kind of thing (referent). (There is an extra-credit experiment you can do -- ask Riona or Maddy about it -- to see what trial-and-error category learning is like.)

      (Now the question is: how many symbols need to be grounded directly through sensorimotor category learning -- and which ones -- so that the rest can be learned from being recombined into verbal definitions/descriptions/explanations? This too will be discussed next week, along with another experiment you will be able to do (a computer game) to get an idea of the symbol grounding problem and why some words first have to be grounded in some other way than just verbally (computationally).

      Delete
    2. You've highlighted a few granny objections, (which i will post again here http://users.ecs.soton.ac.uk/harnad/CM302/Granny/sld002.htm ), especially with regards to granny objection number 5 (computers are mechanical, we are flexible), number 2 (computers can't do anything new) and number 7 (computers can't choose) where replies point to the malleability of computer programs I would like to understand what you mean when you say that we are not all computation. I am not disputing the necessity of T3 over T2, I definitely agree that a fixed computer with determined inputs will not be able to pass a lifelong TT and a properly implemented T3 that is mobile and able to interact directly with the world has a much better chance but, if computer programs are indeed so flexible and computation is indeed incredibly powerful because of it, why could we not conserve the purely algorithmic model implemented in the robot? What is the 'extra something' that mediates the bridge between the symbols and the referents and that is not computational? Or is that not what you mean by we are not all computation? In which case, which parts are, which parts aren't ? I think this will allow me clarify quite a few jumbled points in my own head!

      Also, time and again you've answered the question "i don't want to know how robots do it, i wanna know how we do it" with the issue of underdetermination and an argument of the sufficiency of the T3 over the T4, where essentially to be cognizing like us a man-made-non-human dynamic system does not require the same kind of neural tissue and as long as an explanation at hand is infallible, it doesn't matter if it is the 'right' one (as nature intended) ! Again, i am unable to convince myself that this would be sufficiently satisfying, unless to marvel at the grandeur of a man made creation. Perhaps I am resisting the idea because i am unable to wrap my head around what a T3 would look like in the underlying mechanics that would produce a 'cognizing mind' (would it be a 'mind' if it is unable to know what 'it feels like' to understand red apple despite knowing what someone is referring to when they say red apple) ! What would this composite system of algorithm and dynamism look like? Would it not be a supra-algorithm style model that allows for sensorimotor grounding through 'independent' (to what extent would it be independent?) interaction in the world?

      Delete
    3. 1. T2 is not a "fixed computer with determined inputs." Like T3, it gets new inputs every day. It's just that they're all just symbols. The difference between T2 and T3 is that T3 has sensorimotor contact with the objects in the world that it's symbols refer to. T2 does not. Therefore T2's symbols can't really mean to anything to T2, any more than the words in a book mean anything to the book. The meaning of the symbols in a book or a computer comes from our minds, which are grounded in the world, like the T3 robot. A T3 robot is not just a computer on wheels. The connection between T3's symbols and their referents is dynamic: it's a sensorimotor connection. Computation, with all its formal power, cannot do anything dynamic.

      (A computer can, because it is a dynamical system, but what its hardware is made of or what it is doing -- its dynamics -- are irrelevant to the computation (program) that it is implementing -- other than the fact that the computation has to be dynamically implemented somehow. But, depending on which algorithm (software, program, computation) the computer is running, its symbols could be interpretable as payroll salaries, mathematical proofs, a virtual vaccum-cleaner, a virtual solar-system, a virtual airplane, or Chinese T2 messages.)

      The power of computation is formal, not dynamic. It can do anything a mathematician can do (Weak C/T Thesis) and it can symbolically simulate -- but not be -- just about any dynamical system (Strong C/T Thesis).

      Anything going on inside T3 could be dynamic rather than computational, but at the very least, its sensorimotor system would have to be dynamical. (That too can be simulated computationally. But you can no more make a T3 using simulated sensorimotor function than you can fly to Chicago in a simulated airplane.)

      2. No one knows how to make a T3 yet, we're not even close, and we're not going to find out in this course! But if Renuka and Riona were T3 but not T4, for what would that not be "sufficiently satisfying" for you, Naima?

      There might be other ways to pass T3. (It might even turn out that only a T4 could pass T3!) But however it turns out, the point is that if the robot is indistinguishable from us the way R & R are, then that solution to the "easy problem" has been reduced to "normal underdetermination," which doesn't just apply to cognitive science or biology but also to physics and astronomy and the Grand Unified Theory of Everything: Once a model has been shown to be capable of accounting for all the available data, it is "Turing-indistinguishable" from any other model that can also account for all the available data. There's no way to know which (if any) is the right one.

      T4 does account for more data than T3 -- but how do we know whether the extra T4 properties are relevant -- unless they turn out to be essential for passing T3! (Which still leaves T3 as the decisive test -- or T2, if only a T3 could pass T2, as "Stevan Says" -- because of the symbol grounding problem.)

      3. What would the hybrid computational/dynamic system look like? Like Renuka and Riona, at least in all they can say and do in the world.

      4. Would T3 have a mind (i.e., feel)? Only T3 would know. And that extra bit of underdetermination (the other-minds problem) is unique to cognitive science. GUTEs don't have it. But Turing points out there's no way we can do better. -- And even if a god told us that T3s feel, that still would not solve the "hard problem" of explaining how and why they feel -- rather than just do.

      Delete
    4. Regarding your first reply above:
      “Grounding does not require consciousness (unless you can solve the hard problem and explain how and why!). Grounding requires connecting symbols to their referents using T3 (robotic) capacity.”

      I was hoping you could clarify this in contrast to the following point from your paper:
      “Consciousness. Here is where the problem of consciousness rears its head. For there would be no connection at all between scratches on paper and any intended referents if there were no minds mediating those intentions, via their internal means of picking out those referents.
      So the meaning of a word in a page is "ungrounded," whereas the meaning of a word in a head is "grounded" (by the means that cognitive neuroscience will eventually reveal to us), and thereby mediates between the word on the page and its referent.”

      Was it simply a little misleading to bring up consciousness here? Because it seems from your Skywriting comment that, in fact, it isn’t consciousness that “mediates between the word on the page and its referent,” rather just the internal mechanism of the mind (or the T3 robot).

      Regardless of whether or not my above point hits the mark, can you elaborate more on what you mean by “the problem of consciousness rear[ing] its head”? If the problem isn’t that grounding requires consciousness, as you point out in your comment, then I’ve missed the point.

      Delete
  3. I agree with a lot of points in Harnad’s 2003 paper including that consciousness is necessary for symbols to become grounded in the brain. However I was a little confused that it was asserted the cognitive neuroscience would eventually tell us how symbols are grounded in the human brain. Is this wrapped up into the “easy problem” ?

    “AI researchers had independently already come to the conclusion that autonomous grounding was necessary”

    Steels accepts that one critique of AI’s abilities is valid: that the robot does not make its own symbol system and instead has a semiotic web based on whoever programmed it. To me, this is the only criticism necessary because it brings all sorts of probably distinctly living things into play: fear, pain, joy etc. Humans (and animals) have motivation to avoid negative things and experience positive things and this motivates us to do things, which consequently expands our semiotic network, which therefore changes dynamically with new encounters.

    I guess the issue I’m stuck most on with Steels is motivation. A robot may be able to sort red marbles from yellow marbles but this action has hollow motivation. There is nothing I can see that would suggest that a robot wants to perform this action to experience anything whereas a human might find some sort of joy in sorting marbles and think to expand their sorting repertoire to include blue marbles and pink marbles because they have linked “sorting” to “makes me happy.” Why would a robot want to ground any symbols? What use does it have for them? It does what it needs to do and need not expand on its duties for its own agenda. Without human input, doesn’t it have about as much motivation as a piece of paper?

    “Symbols co-occur with other symbols in texts and speech, and this statistical structure
    can be picked up using statistical methods (as in the latent semantic analysis proposal
    put forward by Landauer and Dumais”

    I’m having a hard time with these points. Isn’t finding ways to model a human’s cognitive abilities kind of irrelevant to solving this problem? If it’s not THE way we do it, it’s just a copy and doesn’t give us any insight into how the process really works. Don’t we want to reverse engineer?

    “They seem to believe that the (bio-)physics of the brain must have unique characteristics with unique causal powers.”

    Well…this argument and counter argument seems unsolvable. How can we ever know about these casual powers without solving the other minds problem?


    ReplyDelete
    Replies
    1. ***it brings all sorts of mechanisms that are probably distinct properties of living things into play (Not really sure what happened to that sentence)

      Delete
    2. Grounding ≠ Feeling: I

      1. If you can explain how and why "consciousness is necessary for symbols to become grounded in the brain" you will have solved the "hard problem." Otherwise grounding just seems to be correlated with feeling, but we have no idea how or why -- or even whether (except in our own private case, introspecting -- the rest is just Turing indistinguishability!) grounding is felt. Cognitive neuroscience (T4) can just (maybe) help solve the "easy problem" of explaining how the brain can do all the things it can do. But feeling is not doing! It's not moving. It's not something we can observe (except in our own private case, introspecting: the rest is again just Turing T3/T4 indistinguishability!)

      2. (I assume you are talking about is Steels, L. (2008). The symbol grounding problem has been solved. so what’s next? Symbols and embodiment: Debates on meaning and cognition, 223-244.) The symbol grounding problem has definitely not been solved till we have a robot that can pass T3. Steels's robots are just toy robots in toy worlds, learning "socially" to use the same symbol to "refer" to the same category amongst themselves. Toy problems are are not T3; they are far too underdetermined to make us believe that the way they do it is the way T3 does it: to show that you would have to scale up to T3 itself. (No, language is not just social agreement on what to call what!)

      3. If you can explain how and why "motives" have to be felt, rather than just executed (i.e. done) you will have solved the hard problem. Toy robots can easily be designed to have (unfelt) "motives," i.e., goals, to do this or that. Those are just behavioral capacities and dispositions, as in machine-learning and machine problem-solving. And the goals can be dynamic (such as in a thermostat, a simple dynamical system that turns on the furnace whenever the temperature gets too low.) And dynamic does not mean felt.

      4. By "motivation," you mean felt wanting: it feels like something to want something. A robot that has been designed (or has learned) to do certain things in order to reach a goal state (like the thermostat) can also be described as having internal "motivation" to achieve a certain state or outcome, but not a felt motivation. Explaining how or why motivation needs to be felt would again need a solution to the... [fill in the blanks...]. It feels like motivation needs to be felt in order to really be motivation, but explaining how and why that's true is another matter...) But, again, grounding ≠ feeling!

      5. Latent semantic analysis is a statistical technique based on the correlations (co-occurrences) of symbols (words) in text. Some patterns of word co-occurrence are correlated with word meaning. ("Chair" and "table" tend to occur together more than "hair" and "table.") But this is all just symbols and symbol manipulation, as in computationalism. The symbol grounding problem is that symbol-symbol connections are not the same thing as symbol-referent connections, so they cannot generate, or constitute, meaning. You are right that Luc Steels is dreaming if he imagines that toy robots agreeing on what to call what, plus word-word correlations, are enough to generate meaning. You need T3 plus the power of language (T2) -- which is not just agreement on what to call what, plus word-word correlations. It is the power of the proposition (week 8 ).

      Delete
    3. Grounding ≠ Feeling: II

      6. Explaining how to pass T3 (or T4) is not the hard problem but the easy problem. It's just that Steels' robots don't even solve the easy problem. They are just toys. Because of the other-minds problem, there is no way to know whether grounded T3 robots feel. Explaining how and why grounding generates feeling would require solving the hard problem.

      Delete
  4. The Symbol Grounding Problem is related to the problem of how words (symbols) get their meanings, and hence to the problem of what meaning itself really is. I’d like to take a stab at the first issue “how words (symbols) get their meaning”. If we take a step back and look at how we learn a language it is in our childhood. Through socialization we start to understand, learn and acquire new words and an understanding for a language. Only listening to spoken words does not do this; associations form. The environment is filled with objects so that when a word is spoken a child can pair this word to that object. For example, when a parent says to their child “do you want to read a book?” or “yum! Applesauce!”, this sentence would go hand in hand with pointing at the book or showing a spoon filled with applesauce, so they can learn the association. Then, when a child goes to school and they begin to learn to read, a similar thing happens… They are not just shown the word “book” on a page and are expected to know the meaning. Think back to kindergarten when you are taught to read. Your workbooks and class lessons were filled with pictures to go with labels and words. So the symbols that form the word “book” on a page would be shown with a picture of a book, which the child learns to associate with the object that their parent read to them. All these connections form a meaning of the word/symbol for “book”. Maybe this isn’t the full answer to how words get their meanings but I think this could be a part of the process (forming associations with spoken words, objects, and the symbols).

    ReplyDelete
    Replies
    1. I agree – the symbol grounding problem made me think about the problem of how symbols get their meanings as well.

      If we consider how attaching meaning to symbols could work with learning two languages concurrently (i.e. being bilingual) or learning a second language later in life, then how would these two processes be the same or differ? When learning a language later in life, such as in elementary school, we always learn the French/Spanish/German word in relation to the English word. So, would we be attaching one meaning representation to both words in each language? Or would we be creating new meaning representations for each word separately? (I admit I do not know much about the mechanisms of bilingualism).

      So, going off what you said, would you think that we form an association with the already learned word (as well as the meaning of the word) and the new word in another language?

      Delete
    2. You guys are touching upon some questions I was wondering myself!

      I myself wondered whether bilinguals have only one or two representation(s) for both languages. I also wondered if one learns the second language after the critical period (5-6 years old), if that also has an impact. Relying on my own experience (and introspection is by no means a reliable source, as we’ve seen), when learning another language, at first I tend to use the meaning of the word in the new language to mean the same thing as the word in my mother tongue. But then as I acquired more experience, my new language became much better – as if it had become it’s own representation?
      But yeah, super interesting question, Melissa!

      Delete
    3. Melissa, Yes I think we form an association with the already learned word and the new word. So "le livre" will come to represent or be associated with the english word book which is in turn associated with paper binded object that you say. Then with enough exposure, those symbols forming the word livre on a page eventually will have the meaning of that object that is a book.
      But then this brings me back to something I brought up in Searle's Argument Skywriting which is whether when we learn a second language, we ever fully UNDERSTAND it or is it constant symbol manipulation and constant translation through your mother tongue.

      Delete
    4. I think you guys brought up some really interesting points (and lot of questions!) concerning language learning and the association of meaning. I thought I could bring some neuroscience background to those questions regarding the brain mapping of bilingual individual.
      I remember from one of my neuroscience class that we talked about the fact that learning a language before or after the critical period (as you mentioned) has an important effect on the brain mapping. From what I remember, the fMRI studies showed that adult who has learned simultaneously two languages before the critical period, showed similar brain activation during the testing of either language. The subject that learned the second language after its mother tongue had brain activation for the two languages that didn’t overlap as much as the early bilingual. What we could attempt to conclude is that an early bilingual would understand both languages in a more consolidated fashion.
      Unfortunately, I didn’t find the particular article. However, in my attempt to retrieve it, I came across plenty of article about bilingualism and its implication on the brain. I shall try to make a connection from those readings with the symbol grounding problem. As Jordana first mentioned, children first learn world when they see contiguous objects reference. So their brain must first shape this association. If they learn both languages at the same time, this association could be made following the same brain pathway or connection. Whereas a monolingual person who then learn a second language must first recognize a new word, and then try to associated with the former language. I think the brain has its own built up mechanism to built a semantic world. It clearly has the capacities to relate any referent to the rightful meaning. I also believe that it is the complexity of the brain that allows it to build a complex network of associated meaning. And if there is consciousness arising, it might always be unreachable to our understanding.

      Delete
    5. "Association" vs. Explanation

      JS: The trouble with association (or correlation) as an explanation of how symbols get connected to their referents is that it does not explain a thing! Try telling a roboticist who is trying to design a T3 that the way to connect the robot's symbols to their referents is to form an "association" between them. But how can you even connect the symbol "apple" with apples in the world? There are trillions of them, and they all look different in color, form, space and time, every time. Association is fine for connecting individual nonsense syllables with other individual nonsense syllables. But to connect words with their referents?

      Most referents are not individuals but categories (not individual thins but kinds of things). So you have to learn which things are and are not in that category. "Association" would mean memorizing every single example you ever encounter. And even that would not cover new examples you had not yet encountered.

      No, what is needed is not associations but feature-detection: What features are shared by apples that are not shared by other things. Finding shared features requires a mechanism. And all of that has to be done at Turing-scale, for all the words (category names) that we know. And grounding does not even begin with referent -symbol "associations," but with referent-action "associations," which means learning to do the right thing with the right kind of thing. The eventual outcome is indeed an association, a correlation, between a kind of thing and an action, but the name of the game is explaining the mechanism that generates that "association."

      (Word-picture associations are no better than word-referent associations unless you already know the kind (category) of thing the picture is a picture of! (Otherwise it's just an association between a word and that particular picture.) And you can't have learned what kind of thing that picture is a picture of by association either. [Notice how we take the "association" between words and their meanings for granted exactly the same way we take it for granted that our brains will deliver to us the names of our 3rd-grade school-teacher when asked. Cognitive science has the burden of actually explaining how all those "associations" are really made.)

      Delete
    6. MW & CA: Once you ground a symbol (in a first language) directly through sensorimotor learning, that grounded symbol can then be used to define other symbols purely verbally, whether within the same language or, via translation, in a second language. (If you think about it, every definition of a new word is a translation in terms of words you already know.) The difference between "compound" and "coordinate" bilinguals in their internal representation of word-meaning, if it's real, might be that compound bilinguals simply use the translation (and hence the original sensorimotor grounding) of their first-language meaning whereas co-ordinate bilinguals have a different sensorimotor grounding for their two languages, and also a different network of within-language definitions grounded in it.

      JS: If the first language is grounded, then the second language inherits the grounding via translation (just as a new word within the same language inherits the grounding via definition). But of course even after enough words have been grounded to do all the rest by verbal definition alone (in principle), I'm sure that in practice the learning of new word-meanings is always hybrid, with the new verbal definitions and explanations supplemented by sensorimotor examples.

      RPB: First, again, "association" explains nothing. It describes an outcome, not how it is achieved. And how word-meaning is "represented" in the brain either in monolinguals or in bilinguals -- early, late, balanced, unbalanced, "compound," "coordinate" -- is not yet understood (and the neuro-imagery correlates certainly do not explain it to us). The brain encoding of word meaning will differ for words that have direct sensorimotor grounding and words whose meaning was learned indirectly from verbal definition. Direct sensorimotor groundings can differ from one another, and so can indirect verbal groundings, both within and between languages. And probably many word meanings are hybrid, the moreso the more concrete the words are. (But concrete/abstract is not a clearcut description, and every category, hence every (content) word, is abstract; categorization is always an act of abstraction. It's just a question of degree.)

      Delete
  5. After reading The Symbol Grounding Problem, I was mulling over the concepts of “grounded” versus “ungrounded” symbols and the process to which we establish the connections from our presently “grounded” symbols to those “ungrounded” symbols that might eventually become “grounded” through learning. I was wondering if there has been any thought or discussion given to how symbols become “connected directly to (i.e. grounded in) their referents”? The process of “grounding” a symbol is discussed as being how it is “augmented with nonsymbolic, sensorimotor capacities” of which the connections that are established “must not be dependent only on the connections made by the brains of external interpreters like us.” However, it leaves me questioning whether the process of establishing a symbol or idea with previously recognized connections to forge meaning within the new symbol would ever be genuinely possible in a system like robotics – where we define as the “robotic capacity to detect, identify, and act upon the things that words and sentences refer to.” What about establishing unique reactions, or crafting new meaning to a word or sentence that previously required ‘x’ as s response versus ‘y’ – where ‘y’ is a reaction dependent on new connections that might override the previous symbolic meaning? These thoughts often feel like I am describing how to provide meaning through learning as an overall concept, but I feel at a loss as to how to differentiate it from “grounding” – or why I feel like robots won’t be able to achieve it.

    ReplyDelete
    Replies
    1. Yes, grounding is related to learning, since word meanings are not inborn: First you have to learn what's the right thing to do with what kind of thing. That's sensorimotor category learning (week 6). Then one of the things you can do with your directly grounded sensorimotor categories is learn their names. That grounds their names (symbols), again directly. Then, when we get to language (in weeks 6, 8 and 9), the grounded category names can be combined to form subject/predicate propositions that describe or define new categories (including ones you have not yet seen, or that cannot be seen).

      Direct sensorimotor grounding is learning categories by induction, from exposure, trial-and-error, and corrective feedback from doing the right or wrong thing with the right or wrong kind of thing. Verbal learning of word meanings from verbal definitions, descriptions and explanations, made up of already grounded words, is learning categories by instruction. All mammals, at least, can learn new categories directly by sensorimotor induction, but only our species can learn new categories indirectly by symbolic instruction, i.e., language.

      Delete
  6. The more we go along in this class, the more I am convinced that cognition cannot all be computation. I agree, for the most part, with Harnad (2003) and I think the symbol-grounding problem is a serious problem that computationalists have to face.

    After Searle’s CRA, it became pretty clear to me that even though T2 could pass the TT through computation alone, it would do so solely on the basis on input output as it doesn’t understand anything –it’s symbols are ungrounded! T2 is unable to connect the symbols to its referents!
    Now, I think with the addition of sensorimotor capacities, a T3 robot’s symbols would be grounded, however we still cannot know for sure that it would understand anything because of the other-minds problem (Searle’s periscope doesn’t work for T3).
    So this led me to think, is grounding a necessary and sufficient condition for understanding? After Searles CRA, it certainly seems like it is necessary, however is it sufficient?

    I also had a question in regards to the paragraph on “words and meaning” – you say, “some have suggested that the meaning of a (referring) word is the rule or features one must use in order to pick out its referent” (p.1). I was wondering more specifically what was meant by features? I came to understand that let’s say we’re trying to pick out the referent for the concept CAT (the symbol in my brain that stands for the cat out there in the world), are the features things like has a tail, meows, has whiskers, etc? If this is indeed what is meant, I am a little confused, as I don’t see how that would make sense. Those features (has a tail, has whiskers) hold meaning in themselves, no? So how do we come to understand the meaning of those features themselves? We keep reducing it, in the sense that we can ask what kinds of features are encompassed in “has a tail”—but this would lead to an infinite regress? And if its about the shape of the feature, how is that different than computation? Or is it about the sensorimotor properties attached to the features? Perhaps, that is not what is meant by features.

    Then I get confused when Harnad (1990) starts talking about “iconic & categorical representations” (p.6). So iconic representations are sensory projections & all their internal analogs (whatever that may be)? If one sees a HORSE, it causes a certain sensory projection on my retina and that is somehow encoded and stored in the brain? Then Harnad says “for identification, icons must be selectively reduced to those “invariant features” of the sensory projection that will reliably distinguish a member of a category from any nomembers with which it could be confused”. What are these invariant features? Are those the unique sensory “features” for HORSE?
    What if, as I go about in the world and look at objects that are out there in the world (let’s say a DOG), the dog causes a certain sensory projection on my retina, that is encoded in my mind and stored in a box, call it box X. Let’s say dogs have a certain unique sensory feature on my retina, and all of those are stored in box X. Now let’s say you encounter a CAT in the world, the cat causes a certain sensory projection on my retina, and since it has a unique sensory feature, it’s stored in box Y. And the way that we relate the two, are through inferential domains. Let’s say X has a set of inferences A {a,b,c} that are caused by X and Y has set of inferences B {c,d,e} that are caused by Y, we related the two when one set entails the other {c}.

    Sorry for getting carried away. I particularly like the topic of concepts and categories and I thought this paper was particularly interesting.

    ReplyDelete
    Replies
    1. Hi Cait,

      In response to your first point re: grounding as necessary and/or sufficient. I spent some time mulling over the same issue, and I think that you've actually answered your own question (at least if my conclusions about it are valid). I think that precisely because of the other-minds problem, it's impossible to know whether grounding is sufficient for understanding. While grounding certainly seems necessary, since to understand really means to feel understanding, we can never really know if it alone is sufficient. i.e. we'll never know if Riona and Renuka actually understand anything when we talk to them, or have any feeling of 'meaning', despite their clear capacities for grounding.

      Delete
    2. Is Grounding ≣ Meaning (Understanding)?

      CA: "Is grounding a necessary and sufficient condition for understanding? After Searles CRA, it certainly seems like it is necessary, however is it sufficient? "

      That is exactly the right question to ask. No one knows the answer (except the grounded T3 robot itself, hence Renuka or Riona) because of the other-minds problem.

      Picking out referents by features only leads to an infinite regress if the features are just verbal (symbols). If they are detected by sensorimotor feature-detectors the circularity is broken -- and that's exactly what grounding means. Your brain learns to detect ("abstract") the features in the sensory projection (icon) from and sensorimotor interactions with cats ("affordances") that allow you to identify cats or horses). That grounds the word "cat" or "horse" for further use in talking and thinking (and understanding).

      "Invariant" features are the ones that distinguish the members from the non-members of a category.

      Not sure where you were going with the inferences. Being able to categorize cats and dogs (knowing what kind to do what with, including naming them) is what category learning is about and for. And it's based on learning to abstract the invariant features of the category

      AB: Good response. Not only does the other-minds problem prevent us from knowing whether grounding is sufficient for meaning, but the hard problem prevents at from explaining how or why it's necessary or sufficient, even if it is. (Btw, Meaning = Grounding + Feeling.)

      Delete
    3. Prof. Harnad: “Invariant features are the ones that distinguish the members from non-members of a category”.

      I had a question regarding this. I understand the idea that the features allow us to distinguish between members and non-members of a category. However, what threw me off is the notion that these features somehow do not undergo change. There’s a tremendous amount of ambiguity in categorization and I don’t think there are features that do not vary.

      Categorizing would be very simple if each object had a feature that was truly invariant and unique to it. But this doesn’t seem to be the case. For example, we would still categorize a completely rotten apple (to the point where it no longer resembles one) as an apple. We would still call an a fruit that looks, tastes and feels like an apple an apple even if it was twice the size of other apples we had seen. The same can be said if we came across a purple apple. At what point would we no longer call it an apple? None of these features that give meaning to the idea of “apple” are invariant (unless I’m missing something).

      So when categorizing, would it make sense to rely on features that are strictly invariant? Instead, I feel like it makes more sense to look at it in terms of objects needing to meet a threshold of features in order to be categorized. For example, if we come across an object that is a) shaped like an apple, b) coloured like an apple (many potential options here), c) tastes like an apple, d) crunches when bitten into, etc, we may categorize it as an apple. If one of these criteria fails to be met, we may still categorize the object as an apple. But if none of them are met, we wouldn’t do so.

      Delete
    4. Features can be either/or too. In fact they could be based on a long, complicated Boolean rule, like a google search (X or if not P then Zand...), but if you can distinguish all members from non-members, there better exist some rule that describes the difference, and that's the invariant.

      But we're not talking about metaphysics here, just categorization. Categorizing is doing the right thing with the right kind of thing, based on feedback from the consequences of getting it right or wrong. If something that looks like an apple but is purple is an apple for botanists, then it's an apple. But they too must have their invariant Boolean rule.

      Category membership is all-or-none. You can't half-belong. So something is not an apple if it's "appley" enough. (And a threshold too can be all-or-none, like a melting or freezing point. And of course where it doesn't matter or we don't know whether something's a member, that's not categorization, it's categorization failure.)

      Delete
  7. This article cleared up some concerns I had over the Chinese Room Argument. More specifically, that Searle is using consciousness to demonstrate that understanding cannot just be computation. So, since he is not conscious of the meaning of the Chinese symbols, he concludes that cognition is not equivalent to computation, as the computationalists propose. However, because Searle is working at the T2 level, he does not mention the symbol grounding problem that arises. If I have this correct, that means that the symbols in the CRA are neither understood nor grounded. In making this connection between consciousness and symbol grounding, is one required for the other? My initial thought would be that we would need symbol grounding for consciousness (so that we could connect the symbols in the system to the real world referents), but we would never know if consciousness is needed for symbol grounding because of the other minds problem. So, even if the T3 robots are unconscious beings, they will still require symbol grounding to be able to do everything we as humans can do (thus, be able to pass T3). That being said, if consciousness is required for cognizing, then we could say that the T3 robots are not cognizing beings. Now, as Harnad states, I am beginning to understand how it’s very unlikely we will ever adequately be able to explain the how and why of consciousness, and thus it's connection to cognition.

    ReplyDelete
    Replies
    1. My understanding was that CRA showed that cognition cannot be computation alone, using T2. That would mean that a simply symbolic T2 is not grounded, hence there is no understanding.

      Harnad proposes that we build a hybrid model that processes perceptual information from sensorimotor signals in a bottom-up approach (1990). But how will we ever know if this hybrid model contains an "intrinsic interpretation" of the symbols? Will the 3 proposed stages by Harnad 1) iconization 2) discrimination 3) identification (1990) still just facilitate an "extrinsic interpretation"? How would we ever get over the hurdle of making it a true intrinsic interpretation? In that sense Melissa, I'm also wondering how will we know if a T3 (who successfully passes the TT) is actually cognizing? For this T3, would its interpretation not just be "a symbol is a symbol is a symbol" and so forth?

      Delete
    2. Or hold on.. T3 would be grounded because it can interact with its environment and the people in it? I've got myself confused here.

      Delete
    3. Hello,
      From what I understand is that T3 is grounded because it is a robot and no longer just a computer (in other words that it has dynamic properties). And in referring back to solving the grounding problem is that to have a robot or system that is partly symbolic and partly sensor-motor. So that T3 is capable to be computation and symbolic but, like you say, to also have the capacity to interact with the environment and in attempt to connect symbols to the things that they refer to. What do you think?
      But the discussion above about "extrinsic interpretations" vs "intrinsic interpretation" to couple with consciousness is very interesting. I wasn't sure if it ties back the other mind's problem, but it would seem very circular. Do you think you can expand more on the link you were making between consciousness and its relation to symbol grounding problem? Thank you!

      Delete
    4. Grounding -- connecting symbols to their reference -- is necessary for passing T3 and probably necessary for meaning (which is grounding + feeling). Because of the other-minds problem we cannot know whether people or T3s feel, but we can be pretty sure people do, so since we can't tell people apart from T3s (based on what they can say ans do), T3s probably also feel. But even if they do, we can't know how or why they feel unless we can solve the "hard problem" (that "Stevan Says" we can't...)

      Delete
  8. Can we separate grounding and understanding?

    When Searle is in the Chinese room, he does not understand the symbols he is manipulating - there's no feeling associated with the words, and the words are ungrounded since he cannot connect them to referents like a T3 sensorimotor machine could. This had me thinking about two conditions: meaningless + grounded vs. meaningful + ungrounded.

    Meaningless and grounded words:
    Last week I came across a complicated word in English - Vociferous - and had clue what it means. After a quick dictionary.com search, I was able to "ground" my understanding of the word. But even after connecting the word with its referent in the real world, I still don't get an intuitive feeling (i.e. meaning) upon seeing the word a week later. Is it possible that some words, no matter how well connected they are to real life objects, simply don't evoke a meaningful experience in us? I mean who's to say what's is and what isn't meaningful (other minds problem)?

    Meaningful and ungrounded words:
    I'm bilingual, but my hebrew is not as perfect as my english. It is not uncommon for me to come across a word I have never seen or heard before, thus cannot associate with a referent in the outside world, but get a certain feeling/intuition/inkling into the meaning of the word upon hearing it. If meaning is really a subjective internal feeling, can a word be meaningful even though it is not grounded?

    ReplyDelete
    Replies
    1. Correction: "Last week I came across a complicated word in English - Vociferous - and had NO clue what it means"

      Delete
    2. Hi Maya,

      I think that as for your “meaningless + grounded” category, it is more a question of memory retrieval and not of symbol grounding. The feeling of understanding comes both from the initial process of grounding a symbol and conferring a meaning to it, but also from subsequent retrievals of the meaning whenever the symbol is encountered again. The example you give, in my opinion, is a great example of the limitation of theories. A theory usually can’t describe all the possible situations in an imperfect world. We have to assume the symbol grounding is accomplished in perfect conditions, i.e. retrieval is not an issue because it is always a success.

      As for the “meaningful +ungrounded” category, if my understanding of this paper is correct, I don’t think you could really encounter any real examples of words that are meaningful to you, without having been grounded first. I think that the feeling we sometimes have that something is meaningful to us, yet seems completely novel yet meaningful is a result of other factors not linked to the symbol grounding itself. It is either a case of something we forgot we once knew (happens to me all the time in Spanish, I have brief flashes of insight about the meaning of words because I once knew them but they feel new when I hear them again) or it is a case of transfer (i.e., the meaning of another word being assigned to something new, but the meaning had been grounded during the initial word learning process).

      Delete
    3. Hi Maya ,

      Not sure if this was what Hernán already meant by 'transfer' at the end of his response, but I think that the feeling of meaning/inkling of understanding you experience upon viewing or hearing a new Hebrew word could also be attributable to the fact that context and existing knowledge of other words in a language give very powerful clues to a novel word's definition. Could it not just be that the intuition you have that gives you the feeling of meaning could be due to other words in the sentence, or words that are similar/related to the novel one, that you have grounded in the past?

      Delete
    4. Hi Maya, Hernán and Adrienne,

      I think that is a good point you bring up. For instance, I have a mild hearing loss and as a result, I often mishear words. For example, if someone says ‘That movie was great!’ I can easily mishear the sentence as 'That movie was grape.’. I am able to quickly figure out that I did, in fact mishear the last word and replace the word with something that makes more sense (‘great’). How do I know that the word that I think I heard (‘grape’) was not actually the correct word? Is it because I can use the context to quickly identify the inappropriate word ‘grape’ and replace it with ‘great’? I think I heavily rely on the context of the sentence, rather than the ‘feeling’ of the word. (I don’t know if that is actually what is happening though, I could actually be relying on the feeling of the word).

      Delete
    5. Some word meanings can be vague, but they're not meaningless (like "laylek" unless you haven't a clue, and then they are just nonsense syllables -- for you.)

      If you figure out a word from verbal context rather than by direct grounding through experience or indirect grounding through definition, its meaning can remain vague and approximate -- or even wrong -- until you shore it up with more context, definition or direct experience. Btw even direct grounding can be vague, approximate or wrong for a while, because you are still just learning the category and your brain has not yet abstracted its relevant features. I would say the category's not fully grounded till you are no longer making mistakes (of over- or underextension).

      (I've always liked the word "vociferous": it means someone who is very vocal about something, arguing vigorously for it.)

      As for meaningful but ungrounded: Think of the meaningfulness of "the smell of petroleum pervades throughout" (while you are under the influence of nitrous oxide.) Feelings can be deceiving...

      Delete
  9. A symbol system alone without this capacity for direct grounding, is not a viable candidate for being whatever it is that is really going on in our brains.


    The symbol grounding problem makes sense to me I think – I understand how the grounding/ungrounding of symbols is related not just to the referents of a word but also the relation a subject has to the word. So words are grounded when I have use for them and can use them but random “squiggles” are not grounded.

    I think I understand how this relates to the CRA, in that symbol manioulation rules are all based on shapes and so the Chinese language is meaningless to Searle – like any other squiggle.

    But is the whole symbol system (the rules that go along with the manipulation of the symbols) not being used in the CRA? It seems to me that it is and that whatever Searle has going on in his head is a conscious process – even if it is a very basic one it is still a symbol manipulation in the same way (though a much more basic order) that me typing this skywriting is. So now I am a little bit more confused than before about the role of consciousness in “cognitive” vs. “vegetative” processes.

    So the symbol grounding problem is claiming that unless computationalism is true, a word in a mind is grounded differently than a word in a computer (which is not grounded). To me this seems to complicate things further – I know what it feels like to understand but I don’t know what my understanding actually is, so how can we be totally sure that a symbol in my head is grounded while one in a computer is ungrounded? I do not at ALL believe in computationalism but to me it just seems like we are trying to extrapolate from what we feel like understanding a symbol is to things beyond that…

    ReplyDelete
    Replies
    1. Hi Julia,

      I am going to try and clear up what you’re confused about, but it’s a little long (sorry ... I hope this will help).

      “So words are grounded when I have use for them and can use them but random “squiggles” are not grounded”

      I think this is a bit of a misunderstanding. A symbol (word) is grounded when there is a concrete, tangible referent. It has nothing to do with whether or not you ‘have use’ for it per say (also not quite sure what you mean by ‘have use’). Grounding just indicates that there is a definitive end to a circle of symbols.

      I imagine it like a game of Charades, where you have to act out the word ‘ladder’ with someone who has never seen or heard of a ladder before. You can act out ‘ladder’ in 5000000 different ways. You could even give up on the rules of charades and start verbally explaining a ladder. The person will grasp that all these things are equivalent, but if the person has never had the sensory experience with the ladder, they will still have no idea what on earth you are talking about. The only case in which they could eventually learn what you mean by ‘ladder’, is either by using a ladder, or by you describing it with enough other words that they have grounded, that the concept of a ladder begins to make sense. But this eventuality is dependent on having previously grounded something, somewhere. Grounding is still eventually necessary for meaning. The issue is that with a computer, this grounding can never happen. It would instead be an infinite loop of explanations and equivalent symbols that never end at a grounded word, because they have no capacity for sensory experience. Any explanation that could possibly be offered will still contain symbols that are not grounded to an actual thing.

      [continued below]

      Delete
    2. [cont]

      For your confusion about Searle’s CRA:

      “is the whole symbol system .. not being used in the CRA? It seems to me that it is and that whatever Searle has going on in his head is a conscious process ... it is still a symbol manipulation in the same way .. that me typing this skywriting is.”

      I think there’s two things you’re saying, 1) is the system’s argument (that if he has the entire system within him, then he would “understand”), 2) is that the act of manipulating Chinese symbols is a conscious process, so therefor he would “understand” what he is doing by the same token that you understand what you are writing.

      To address the system’s argument:

      The CRA says Searle has Chinese symbols that are separated according to ‘a script’, ‘a story’ and ‘questions’, and English rules instructing how to change the input ‘questions’ to symbols of the ‘script’ and ‘story’ in an ‘answer’. Searle says that only the English rules are the actual program, but the systems argument says everything is a part of the program and Searle is the entire system.

      The reason this doesn’t make sense is that it assumes the people who speak Chinese and who are writing the questions would actually be within Searle – suggesting that Searle has multiple personalities – the Chinese one writing the inputs for the English Seale, and the English Searle. The ‘program’ would somehow result from the simultaneous existence of these two (i.e. the rules from the Chinese Searle are somehow translated to English within Searle for the English Searle to follow). But humans don’t have multiple personalities, and the argument is further discredited while addressing your second point:

      That he would “understand” in the same way you do in your skywriting:

      First, I think we need to differentiate between ‘understanding what he is doing’ (i.e. consciously understanding that he is manipulating symbols), and ‘understanding Chinese’. Searle uses himself, as the ‘hardware’ in his example. Because he is human, then yes, he may understand that he is translating something Chinese into something else that is Chinese. But this capacity of understanding is not a result of the manipulation of Chinese symbols, it is because it so happens that the ‘hardware’ is human – it is a capacity of the hardware. Since the computationalist argument says that the ‘hardware’ is insignificant, then him understanding that he is manipulating symbols is irrelevant.

      If instead your statement is based off of Searle being the entire system, then if this were true, it means that Searle’s ‘understanding’ of Chinese would have to come as a result of switching between Chinese and English symbols, while additionally relying on an English connection to the outside world. In Searle’s Chinese Room example, the only reliance he has on his understanding of English is in the form of the program: if squiggle then squabble. This is the same capacity a computer would have to read a program. But for your argument and the systems argument, there is still an assumption that English has meaning, and that through its meaning, meaning is automatically created for Chinese, and thus understanding. The problem is that this comes directly back to the symbol grounding problem: he can only give meaning to Chinese via English, and he can only give meaning to English via sensory experiences. Or worded another way: the ‘Chinese Searle’ could only extract meaning from the Chinese output with some sort of sensory experience to ground the Chinese symbols. No matter what, some of either the English or Chinese symbols have to be grounded for there to be any meaning anywhere. So then even the systems argument, meant to disprove the CRA and prove computationalism, still proves that sensory experience is necessary to create meaning. We don't know if this meaning necessarily produces understanding, but we do know that there can absolutely be no understanding if there is no meaning to the symbols.

      I hope this helps ... let me know if you have questions / if I've said something wrong!

      Delete
    3. JP: The question is not whether Searle understands that he is manipulating symbols but whether he understands Chinese.

      EJ: Not all word meanings need to be directly grounded. A definition or description is enough, as long as all of its words are grounded (directly, or via grounded definitions, etc.)

      The only other understander in Searle is the one speaking the Chinese, not all the outside pen-pals that are sending the letters. But Searle's point is that there is not other understander in his head, just Searle himself, doing exactly what he says he's doing: manipulating meaningless symbols based on their shapes according to rules he memorized.

      The symbol manipulation rules just are not translations of Chinese into English or vice versa. They are just rules for manipulating symbols based on their shapes: if squiggle then sqoggle.

      We don't know if grounding is enough for meaning, which is also something felt, not just "done."

      Delete
  10. “To be grounded, the symbol system would have to be augmented with nonsymbolic, sensorimotor capacities -- the capacity to interact autonomously with that world of objects, events, properties and states that its symbols are systematically interpretable (by us) as referring to. It would have to be able to pick out the referents of its symbols, and its sensorimotor interactions with the world would have to fit coherently with the symbols' interpretations.”

    I understand the argument here that in order for a symbol to be related to a referent then there should be some sort of physical input that represents that referent. But I get lost at why this has to be any sort of sensorimotor input such as the kinds we use to get info about real world objects. My thinking is that regardless of how we get information from the real world, it all gets pushed through neurons that for the sake of argument can be represented by binary symbols that are firing or not, 0s and 1s. So if all this input just comes into the brain as patterns of symbols that some higher cognitive function will connect to another set of symbols (words, thoughts, etc.) then why do we need these sensorimotor capacities at all? To generate the new symbol patterns for input? A computer can do that on its own.

    What strikes me as the issue this is trying to solve is how enough information can be generated to properly ground a symbol by showing it enough real world instances of what is to be grounded (whether that just be many views of a single scene). I don’t understand why the same sort of pattern matching to create referents cannot occur solely through text.

    The argument for consciousness needing to be involved was that something would have to implemented to make the connections between symbols and their referents but certainly there are programs that can do this. By pure statistical analysis and brute force entering of sentences and stories could not the word “table” be paired and match to various other words the same way we could describe what makes a table a table to another person? The returning argument I suppose would then be “But how do the words describing “table” get their meaning? And merely connecting symbols to other symbols allow for real “understanding”. The first answer is to use the same process, finding the connections between words the same way that we would make connections between visual representations (which are just patterns) to other visual representations or to other modes of patterns such as the word or the sound it makes. As for “understanding” I cannot see the difference in our pattern matching to computers and would thus have to assume that this results in “understanding” if such a task can be done without creating some sort of consciousness… But I will let others prove me wrong here.

    My points is it’s all just patterns connecting to patterns and computers are pretty good with patterns.

    ReplyDelete
    Replies
    1. I'm not sure if that was what the argument for consciousness meant. I don't think consciousness is simply just connecting the symbols and referents. If it were, I think there are many programs we could categorize as "conscious." What "conscious" refers to is that distinct feeling where you understand a statement. If I don't know any Spanish the word "cuchara" doesn't give me a feeling of understanding. However, for the word "spoon" I have a distinct feeling that I know exactly what it means. But consciousness is irrelevant to the visibly observable ability of a computer being able to do what we do since we can't even deduce consciousness is within other humans.

      I think sensorimotor capabilities are seen as a pre-req to the symbol grounding problem. Evolutionarily speaking, for the level of intelligence exhibited by humans, they needed to be able to efficiently act on objects and observe objects, so there had to be a way to systematically store knowledge about these things. I think the symbol systems we have today are a byproduct of this need. Though that's not necessarily an argument for why computers need to have sensorimotor capabilities, I think it is hard to say the shapes that we store in computers by themselves will have this level of understanding. If understanding was simple pattern matching, then Searle should have displayed understanding in the Chinese Room.

      Lastly, I think just simulating neurons would not constitute a form of understanding. I could simulate neurons firing with a network of dominoes as well, but that doesn't have any form of understanding. It seems that over the millions of years of evolution, there is a distinct hardware dependent property that has allowed neurons to construct the range of feelings and abilities we are capable of doing.

      Delete
    2. Hey friends, this is quite a complicated topic. I think William is right that the argument for consciousness comes into play because of feeling. What separates basic symbol grounding from true meaning is that “it feels like something to be in the meaning state, whereas it does not feel like anything to be in the merely grounded functional state” (from the Wikipedia page). So Jordan is correct that there are programs that can perform basic symbol grounding, but these programs are not creating meaning because meaning requires feeling (and therefore consciousness). These programs are simply connecting squiggles to squaggles.

      Also, while I agree with William that “just simulating neurons would not constitute a form of understanding,” I'm not sure his final point is correct:

      It seems that over the millions of years of evolution, there is a distinct hardware dependent property that has allowed neurons to construct the range of feelings and abilities we are capable of doing.

      This idea seems to suggest that for a robot to pass the Turing Test (i.e. be totally indistinguishable from us for its entire life), it must be a T4 robot. This robot must have our physiology to be able to feel. Our T3 robots Riona and Renuka would not even appear to feel, as they aren't made of neurons and tissue, so we can kick them all we want. However, I'm not convinced that T4 is necessary over T3 yet. It still seems feasible that Riona and Renuka could appear to feel the same way I do, even if neurons facilitate my feelings and circuit boards facilitate theirs. Furthermore, the symbol grounding problem only takes us from T2 to T3; it does not require we go all the way to T4, as Dr. Harnad points out in his article:

      The necessity of groundedness, in other words, takes us from the level of the pen-pal Turing Test, which is purely symbolic (computational), to the robotic Turing Test, which is hybrid symbolic/sensorimotor. (page 4)

      Delete
    3. JP: It's not that the referent has to be "represented." The referent is an outside object, and the symbol has to be connected to it by the T3 capacity to recognize it, manipulate it, categorize it, name it.

      Neurons do whatever they do. The fact that they can be "represented" by 0/1 is just the fact that they (like almost any object and process) can be simulated by computation. That does not mean that real neurons are just implementing a computation. They could be part of a dynamical system.

      It is not clear that the projections of objects on our sensory surfaces are digitized and then processed computationally. The projection can be part of a dynamical system. (And the sensory surfaces and motor effectors have to be dynamic.)

      What's needed to ground a symbol that refers to an object or a category of objects is the sensorimotor feature detectors that can recognize, manipulate, and categorize the symbol's referent through sensorimotor (robotic, T3) interactions with it. These interactions are dynamic, not computational.

      It's not consciousness (feeling) that's needed for grounding, it's sensorimotor interactions and categorization. No one knows what feeling is for, just what doing is for.

      AH: No, computation alone cannot do symbol grounding at all. Symbol grounding is necessarily hybrid: sensorimotor/symbolic.

      Delete
  11. The reasons why it is thought that robots or computers will not be able to pass the Turing test with the techniques that are used to build them now are often listed as the issues of the symbol grounding problem and the problem of intentionality.
    The symbol grounding problem, as described by Harnad is that computers or robots are able to recognize symbols, however, they are unable to associate these symbols with their referents. They are therefore unable to pick out the intended referent of a certain set of symbols.
    The problem of intentionality, as described by Steels, is what all humans and living organisms have and what is makes them willing to do certain actions. It ‘’that feature of certain mental states by which they are directed at or about objects and states of affairs in the world’’.
    To pass the Turing test, depending on its level, the computer or the robot needs to be able to communicate with a human for a life-time and that human should not be able to understand that he or she is communicating with a non-human being. The striking feature about this kind of testing is that the test lasts a life-time but little time is given to the computer or the robot to ‘’prepare’’. Adult humans and children have large amounts of time where they can learn things about interacting with other human beings and other facts about the world. It is also during this period that they have time to establish links between symbols (words) and referents. Robots that are considered for the Turing test do not have that much time to ‘’observe’’ things about the world. Several years are not available for them to just observe things in the world and learn (or compute how to simulate human behaviour). Observation is crucial in order to be able to establish links between the symbol and its referent. In the examples given by Steel, a person that was never exposed to French wines would not be able to figure out what another person asking them to pass them ‘’the Bordeaux’’ meant.
    As mentioned previously, the lack of intentionality is another reason why computers will not be able to pass the Turing test with the techniques that are used to build them. However, intentionality also
    arguably comes from observation and interaction with others. In terms of the intentionality argument, the robot or the computer can be given the task to pass the Turing test (trick a human into believing that it is interacting with a human). This could be the feature giving intentionality to the computer or the robot. On the other hand, if the computer is performant enough, it can gain intentionality by observing how humans behave and what are their intentions and simulate similar intentions in order to accomplish its task of behaving like a human being.
    The arguments given highlighting the difficulties of making a computer or a robot that would be able to pass the Turing test are reasonable, however, it is important to take into account the amount of experience that is provided to the computer or the robot before attempting to pass the Turing test. Even humans, that pass the Turing test by definition, need some experience to create links between words and their referents. Furthermore, the intentions that humans have, such as buying a house, getting a diploma or marrying somebody are not present at birth but emerge through experience and from the pressure of social norms. A computer or a robot that just got assembled might only care about getting electricity (or other means to keep it working), like an infant that worries about getting food.

    Steel’s article:
    http://homepage.univie.ac.at/nicole.rossmanith/concepts/papers/steels2008symbol.pdf

    ReplyDelete
    Replies
    1. Computers don't "recognize" symbols, they manipulate them (by implementing a computer program). Computers are not robots or vice versa. Grounded T3 robots do recognize the referents of their symbols.

      "Intentionality" is a weasel-word (a mixture of intending, meaning, and feeling: but intending and meaning are already felt states). It explains absolutely nothing.

      It is part of T3 to have the capacity to learn. Renuka & Riona can learn and "prepare." It is true, however, that if learning something changes the state of a learner (whether human or T3) and imparts new information across a period of time, then, in principle, the end result could also have been implanted directly, rather than in real time. That's why I say that R & R were built in MIT 2 years ago. They have a "history," but it is a virtual or fictional history. In practice, of course real experience and real time is the way brains are changed and information is acquired, but, in principle, there is nothing sacred about real time -- the past, that is. T3 does require real-time learning capacity in the future, however. (Turing also talked about making a child T3 and letting it "grow up" through real-time development and learning.) If and when a real T3 is developed, a fictional past will probably be completely unnecessary and it can be Turing-tested on its own terms: as an MIT-built robot whom we are interacting with for a lifetime to see whether it really does have the full capacity to behave and speak indistinguishably from a real person. To the extent that real-time experience is needed for certain capacities, the T3 must have it, and can use and develop it in real time, just as people do.

      The Turing Test is not a trick. It is a test of whether the real capacity to do what thinking people can do has been successfully reverse-engineered.

      I don't think Steels has a very good grasp of the symbol grounding problem. ("Stevan Says.")

      Delete
  12. I agree with the conclusion that sensorimotor grounding is required to give meaning to the symbols in a formal system, but I wonder just how many and which symbols would need to be grounded. Say we were to map out all of the words in an English dictionary by assigning each word to a node and connecting that node to the nodes corresponding to all of the words used to define that word: how much would the intrinsic structure of the graph tell us (who already know the meaning of the words, and hence have our own grounded dictionary somehow in our heads) about the referents of those words? Could we somehow process the resultant graph to yield a condensed dictionary where the fewest possible words require grounding, and all of the rest are defined in terms of those words? I really don’t know how such efforts would proceed, but it strikes me that the symbol grounding problem suggests such questions as an interesting area of investigation for computer science, linguistics and, hence, cognitive science.

    ReplyDelete
    Replies
    1. Hi Timothy,
      I think the point of the symbol grounding problem is not to say that every symbol needs to be grounded. In fact, I think the example you gave of consolidating definitions is exactly what the human brain does by categorization. The point that symbol grounding is trying to make, is that no matter what, all you need is at least one symbol, somewhere, to be grounded. Even if you could consolidate every word in the English language down to one concrete representation, that symbol would still have to be grounded in the real world so that every other symbol could stem from it. The point of symbol grounding is to show that whether it is all symbols or one symbol or 10 symbols that need to be grounded, they still have to be grounded, and grounding only comes by connecting the symbolic representation to the real world referent. This capacity to ground 1, 10, or 10^10 symbols is simply not a capacity that a computer can have, because they have no capacity for sensory input. The “area of investigation” that you refer to that would allow for this is perhaps the idea of a T3 – a robot that has the capacity for sensory experience to create ground symbols and thus allow for meaning. (Note: meaning does not necessarily mean understanding – while the words may have meaning because they can be grounded in reality, there is no way to know if the robot has understanding).

      Delete
    2. TA: This is how we have reduced dictionaries to their minimal grounding sets:
      https://www.newscientist.com/article/mg21929322-700-why-your-brain-may-work-like-a-dictionary/
      (Is this what you were referring to, or did you ask the question without knowing anything about this?)

      In a week or so, everyone in the class will get a chance to play the dictionary game at:
      http://lexis.uqam.ca:8080/dictGame/
      (Not this week: it's being updated.)

      EJ: You couldn't ground T3 with just one word (category). You'd need to ground enough words directly so that all the rest can be grounded indirectly out of definitions made up of already grounded words. The question of how many words and which ones is a question for research. (We're working on it. We can already say that the minimal grounding set is under 2000 words but it is not unique: there are many possible grounding sets consisting of different combinations of words. It's also almost certain that we don't actually ground one minimal grounding set directly through sensorimotor experience and then do everything else verbally using those words. We probably keep grounding some categories directly throughout life; but the more verbal and bookish or textish ones among us might do more of it the indirect, "virtual" way...).

      Probably the right way to put it is not that "meaning does not necessarily mean understanding" but rather that grounding does not necessarily mean meaning. (Apart from that, "meaning," "understanding," "thinking," "intelligence," "intentionality" etc. all just refer to the fact that cognition is not just "done" [the easy problem] but also felt [the hard problem].)

      Delete
  13. ''The means of picking out referents...whatever ''processing' is required to connect the inner world to the outer object''
    Reading the paper made me start to wonder whether things from our imaginations can be said to be ''grounded'' in the real world. Take for example, artists. What many artists do seems to be the reverse process. By using a fantastical image they have in their minds, they can 'translate' this onto paper and into the real world. In this case it comes from an internal mental process and is then made into a tangible referent in the exterior world. In this case, does meaning precede the referent and does it initially exist solely within the entity? How is it possible for people to conjure up meaningful images of things they have never actually seen or experienced? Additionally, people can create meaningful images of things in other people's minds using solely language and descriptions. Is it possible to program a robot to imagine concepts they have never experienced? How would a connectionist model, which is so reliant on past experience, explain this?

    ReplyDelete
    Replies
    1. Hey Rose, you raise some great questions here.

      Is it possible to program a robot to imagine concepts they have never experienced? How would a connectionist model, which is so reliant on past experience, explain this?

      These last two questions reminded me of a recent article (http://www.ncbi.nlm.nih.gov/pubmed/23888038), where scientists at MIT claim to have created a “false memory” in a mouse's brain. They did this by activating the neurons encoding the memory of a room (via optogenetics) at the same time they presented a fear stimulus (an electric shock). In this way, the mouse learned to fear the room, despite never actually being shocked in it. Essentially, they “programmed” the mouse to fear the room, even though there was nothing in the room to scare it and it had never been scared in the room before.

      Similarly, if we could create false memories in robots, then it would be possible to program a robot to imagine concepts they have never experienced. And the connectionist model explains this because it is so reliant on past experience. Past experience connects the circuits encoding separate “events” to link the two events together (e.g. an electric shock and being in a room), just like these optogenetic techniques do. These optogenetic techniques simply offer another way of implementing this process. So if there is a way to implement this process in robots, then past experience becomes irrelevant and we should be able to program a robot to imagine concepts they have never experienced (or at the very least, remember things they have never experienced). We are the ones creating the connections instead of past experience.

      However, this does not even begin to answer the main question:

      How is it possible for people to conjure up meaningful images of things they have never actually seen or experienced?

      This question has stumped me as well. Maybe it has to do with recursion, and people simply combine images of things they have seen/experienced to create something entirely new. Which brings me to a related question of my own: can images/concepts be “meaningful” if there is not a symbol assigned to them? For example, let's consider the concept of “two” (..), something that is more than one (.) and less than three (…). On the one hand, the concept (..) has always existed, but it wasn't until the symbol 2 was introduced that “meaning” became necessary at all. (..) is always (..), but “meaning” is totally irrelevant until a symbol is connected directly to (and therefore grounded in) the concept, at which point 2 begins to “mean” (..). But on the other hand, there is a feeling associated with having (..) of something that is different than the feeling of having (.) or (…) of it, regardless of what symbols are used to describe them. And this feeling encodes the concept of (..), so it must mean (..).

      Or maybe the answer is a combination of these two approaches, and meaning only comes into play here because the feeling encoding the concept of (..) is simply a physiological symbol meaning (..). I guess this all boils down to another question you asked:

      In this case, does meaning precede the referent and does it initially exist solely within the entity?

      Can meaning ever precede a referent? Or does a symbol need to be grounded to its referent for meaning to come into play? Ultimately, is feeling enough to provide meaning? Can concepts be grounded in feeling alone?

      Delete
    2. RW: Before you can imagine shapes, you have to ground real shapes.

      (Robots are not "programmed." And computers execute algorithms, which, if they are algorithms for learning, can change as a result of practice or input. It is misleading to imagine that they are "programmed" in the movie sense, which means compelled to do this or that in advance. Learning changes them.)

      Neural nets may help in learning and feature abstraction, hence in grounding.

      AH: In the cruel experiment you describe, there is no "program" in the computational sense involved. They manipulated the little victim's brain dynamically -- and got a published article out of it.

      Yes, in principle, T3 could have "false memories" (a non-existent virtual past -- see reply to Anastasia above) but in practice there's no need for it in real reverse-engineering aimed at T3: Real learning capacity in real time is enough. (But neural nets, though they may help in some functions, like feature learning, certainly can't do the whole job.)

      You are also using the verb "to program" too loosely. It just refers to implementing an algorithm, not to dynamic processes, nor to "mind control."

      Kid-sib couldn't follow what you were saying about concepts...

      Meaning = T3 grounding + feeling.

      Delete
  14. The idea of sensorimotor capabilities constituting a crucial component of symbol grounding is an interesting one. I believe it is a necessary condition, but not a sufficient condition. Part of being a sentient organism, with the cognitive capabilities that humans have, requires a way for knowledge about experiences in the world, objects, and actions to be communicated and stored for future use. I think the logical order is we sense/do things, we have knowledge of this experience and form a precept of it, and then we systematically make symbols to refer to it again.

    This brings a new question of what level of sensory and motor capabilities is required to do symbol grounding. A person who may be blind, deaf, insensitive to pain etc clearly demonstrates symbol grounding for cross-modal experiences. However, they may not display the conscious feeling of understanding or be able to attach a word to a referent say for the word “red” if a person is blind. This does indeed illustrate that aspects of sensorimotor capabilities are necessary to form the connection necessary to connect a word or symbol to its meaning.

    The symbol grounding problem seems to be heavily connected to semantic memory. How do we store our abstract knowledge of the world is one of the first questions that must be answered in order to see how we connect words to their referents. It seems that the brain has a property where it stores abstract knowledge without any specific information as to its particular sensations in the temporal poles, but then connects it with modality specific information of certain lobes. Perhaps words activate the amodal representation we have of the concept, which then connects to the modality specific information we have of the concept to allow us to precisely connect words to their referents. Therefore, it seems the symbol combinations we use to denote the amodal representation are arbitrary as long as they are systematically interpretable from other symbol combinations that represent other amodal representations.

    ReplyDelete
    Replies
    1. Hi William,

      I very much agree with you that the SGP seems to be heavily connected to semantic memory. There is a lot of evidence to suggest that perceptual systems are what semantic memory depends upon.

      1) In a study done by Martin (1995), we can see that there is a network of brain regions that represent semantic memory. For example, the generation of colour or action words seemed to activate regions of the brain involved with those respective perceptual processes (ie. Colour retrieval activated the ventral temporal cortex bilaterally) (Martin 1995).

      2) You mentioned that “modal-experiences” will lead to different symbol grounding (ie. A person who has been blind their entire life trying to make sense of the word “red”). This leads me to wonder how there can be a multitude of different meanings per symbol since there can be different representations of it in the brain. Some people encode the word “seaweed” for example, for its taste (memory stored in gustatory cortex), for its touch (in somatosensory), for its smell perhaps (in olfactory), other identifiable characteristics as an living organism (it is part of the algae group, can find it in oceans, a eukaryote, etc). Others may have never eaten seaweed before so they may just represent their meaning of it by thinking of it being a living organism out in the ocean. So is there a boundless space for the possible different meanings that there could be and in that case, is a symbol grounded in the interaction (or perhaps network) of sensory areas?

      3) A study on subjects with semantic dementia revealed that most of their brain damage and atrophy was restricted to their anterior temporal lobes (ATL). This suggests that the ATL could be serving as an amodal hub. Points 2 and 3 support what you said (William), that “it seems that the brain has a property where it stores abstract knowledge without any specific information as to its particular sensations in the temporal poles, but then connects it with modality specific information of certain lobes”.

      4) Another way understanding the meaning of a symbol can follow from logic, rules and reasoning. If we examine retrieval characteristics of semantic memory, we can see that we are able to retrieve information that has never been “encoded” so to speak. For example, if we know that mammals have mammary glands and then find out that bats have mammary glands (surprise to some who never knew), it will follow that bats are mammals. By using this kind of logical reasoning, we have been able to manipulate symbols to give new meaning.

      In my mind, this all sounds like connectionism (correct me if I’m wrong) - the idea that there is a dynamic system at play. This hybrid system is able to learn from experience (sensorimotor) through a giant network of interconnected nodes.

      Delete
    2. WB: I"m not sure what you mean by "knowledge" (data? skill? felt data/skill?), "experience" (felt? unfelt?), "percept" (?). And we don't make symbols, but we do agree on words to name categories we have learned to identify.

      Blind people don't see things, but they sense them with other senses and with the help of analogy. Blind people can't see color but they can talk about it, just as we can talk about echolocation in bats.

      "Representation" is a weasel word. It refers to something-or-other in the brain, but we have no idea yet what, or how...

      LK: Yes, symbol grounding is related to what we call "semantic memory" -- but to "episodic memory" and "procedural memory" too..

      Yes, the same category can be identified using different features (where that is possible). It is another form of "underdetermination."

      Yes, neural nets can play a role in feature extraction and category learning but, no, connectionism does not yet "explain" anywhere near as much through concrete causal models as it seems to in vague metaphors.

      Delete
  15. I'm interested in the causal mechanisms that allow us to ground "words" in their "external referents". In this paper it is described as the "'processing' required to connect the inner word to the outer object." I am curious as to how much of this grounding is mediated by a learning process (un/supervised) vs. how much is given by a genetically programmed framework.

    When people talk about machine learning they often describe supervised and unsupervised learning.
    Supervised learning is a type of learning in which a machine is given a set of labeled data, and its programming is such that it must 'reverse-engineer' a categorisation system which could have generated the labeled data. This is the sort of learning that allows for check reading and voice recognition. If you provide the machine with a sufficiently large sample of checks, each with the signatures on the checks pre-labelled, a machine capable of supervised learning will generate a system capable of reading other checks, not part of the original sample.
    Unsupervised learning is learning in which a machine takes unlabelled data and attempts to find patterns or consistent structures for the purpose of developing categories that could be used to label a data set. This sort of learning has been used in the programming of robot vision, which can be used to detect statistical regularities in images, clumping together similar objects.

    When we talk about human cognition, it's easy to see how these frameworks for learning might play a role in symbol grounding. Unsupervised learning is thought to play a role in the general structuring of the world. The brain is naturally good at finding patterns, and an excellent example of this is visual perception. Our vision becomes attuned to detecting objects, forms, and abstract features of those objects -- underlying structures in the world. This seems to me to be a plausible way that we learn to categorize sensory information.
    How then do we give those categories names? How do we connect referents to words? It appears that supervised learning is a good way to finely tune our category learning. In childhood we are given vast sample sizes of labeled objects. For example, we are given many different examples of 'cars', which we use to adjust our schema of what 'cars' are, in order to create an accurate system for categorizing objects as 'cars'. Somewhere in here is where the meaning for 'car' is created. Since we describe the "meaning" of words as "the means of picking them out" then supervised learning seems to be a good guess as to how words get their meanings.

    I feel like I'm grasping at something here, but I'm left with a number of questions. To what extent do these learning frameworks really play a role in human cognition? How do we transition from unlabeled categories (unsupervised learning) to labeled categories (supervised learning) and how related is this to the learning of language? To what extent are these categories genetically programmed, and how does THAT relate to language? How feasible is it to create robots that are better at unsupervised learning (which seems to be the current difficulty)? When we do create robots that are good at unsupervised learning, how much will the gap to Turing indistinguishability be shortened?

    ReplyDelete
    Replies
    1. CL: From the look of it, sensorimotor category learning for symbol grounding looks to be largely trial-and-error learning with corrective feedback (reinforcement/supervised learning) though there is also some learning from passive exposure and correlations (unsupervised learning). These are discussed next week.

      Giving a neural net a huge set of "labelled" data is the computational way of doing supervised learning, but in real-life it is trial-and-error learning with corrective feedback.

      Grounding requires being able to identify the members of the category. Words are the names of the categories. We have to learn which kinds of things are called what. So we need to learn the kinds (categories) and their names. Categories are acuired by learning to abstract the features that distinguish the members from the non-members. Unless the features are very obvious, this can only be done via supervised (trial and error) learning, guided by the consequences of doing the right or wrong thing with the right or wrong kind of thing. Unsupervised learning is unlikely to be enough, not only when the relevant features are hard to find, but also when the same things can be sorted in many different ways, hence different categorizations.

      The transition from unsupervised to supervised learning comes from the consequences of mis-categorizing.

      A dictionary is mostly the names of categories (nouns, verbs, adjectives, adverbs). Take a random sample of words from a dictionary, say 100 of them.Count how many of them are likely to have been innate. (Very few.) Now how many of the learned ones are likely to have been learnable from passive exposure alone, from their shapes and correlations, without any feedback as to what's what and what to do with what. (Let me know the percentage of each -- innate, unsupervised, supervised -- that you come up with. But count also what percentage of the learned categories is likely to have been learned indirectly from grounded verbal definitions rather than from direct sensorimotor experience. That's the portion contributed by language.)

      T3 is a long, long way off as yet...

      Delete
    2. CL: From the look of it, sensorimotor category learning for symbol grounding looks to be largely trial-and-error learning with corrective feedback (reinforcement/supervised learning) though there is also some learning from passive exposure and correlations (unsupervised learning). These are discussed next week.

      Giving a neural net a huge set of "labelled" data is the computational way of doing supervised learning, but in real-life it is trial-and-error learning with corrective feedback.

      Grounding requires being able to identify the members of the category. Words are the names of the categories. We have to learn which kinds of things are called what. So we need to learn the kinds (categories) and their names. Categories are acuired by learning to abstract the features that distinguish the members from the non-members. Unless the features are very obvious, this can only be done via supervised (trial and error) learning, guided by the consequences of doing the right or wrong thing with the right or wrong kind of thing. Unsupervised learning is unlikely to be enough, not only when the relevant features are hard to find, but also when the same things can be sorted in many different ways, hence different categorizations.

      The transition from unsupervised to supervised learning comes from the consequences of mis-categorizing.

      A dictionary is mostly the names of categories (nouns, verbs, adjectives, adverbs). Take a random sample of words from a dictionary, say 100 of them.Count how many of them are likely to have been innate. (Very few.) Now how many of the learned ones are likely to have been learnable from passive exposure alone, from their shapes and correlations, without any feedback as to what's what and what to do with what. (Let me know the percentage of each -- innate, unsupervised, supervised -- that you come up with. But count also what percentage of the learned categories is likely to have been learned indirectly from grounded verbal definitions rather than from direct sensorimotor experience. That's the portion contributed by language.)

      T3 is a long, long way off as yet...

      Delete
  16. This article left me a little confused about the relationship between meaning, understanding, and consciousness, especially in relation to something Dr. Harnad mentioned in class on Monday: that “intelligence,” “cognition,” and Searle's “understanding” all refer to the same concept. With this in mind, I'm going to attempt to breakdown the relationship of these terms with respect to one another, at least from my point of view. Please point out where you disagree or think I'm simply just wrong.

    How does Searle know that there is no meaning going on when he is executing the TT-passing program? Exactly the same way he knows whether there is or is not meaning going on inside his head under any other conditions: He understands the words of English, whereas the Chinese symbols that he is manipulating according to the program's rules mean nothing to him. (page 2)

    So according to Searle, symbols need to be understood to have meaning. In other words, meaning requires understanding (or intelligence/cognition). Then meaning must be a mental process, which brings up the issue of consciousness, as Dr. Harnad points out in his next paragraph:

    Note that in pointing out that the Chinese words would be meaningless to him under those conditions, Searle has appealed to consciousness. Otherwise one could argue that there would be meaning going on in his head under those conditions, but he would simply not be aware of it. (page 2)

    So the difference between understanding and mere symbol manipulation is that it feels like something to understand. And feeling requires consciousness, so understanding must require consciousness as well. In summary, we have a top-down hierarchal relationship between the three terms, where meaning requires understanding, which in turn requires consciousness. So for there to be meaning, there must consciousness, and understanding (or intelligence/cognition), whatever that process may be, is the link that joins the two together. And we can relate all this back to the symbol grounding problem because symbol grounding is the process that gives words their meanings, and meaning requires consciousness, so symbol grounding must require consciousness as well.

    And to take this train of logic one step further, if symbol grounding requires consciousness and feeling, then the symbol grounding problem – how and why words get their meanings – is directly related to the hard problem of cognitive science – how and why we feel. Does any of this make sense? What do you think?

    ReplyDelete
    Replies
    1. Hi Alex,

      I'm also looking to clear up how these ideas all relate to one another, so I thought I'd take a stab at giving you a response.

      As a disclaimer I had some trouble working out the exact meaning of the word "understanding" in the context of the class and the readings, because in some cases understanding="capacity to interpret symbols", and other times understanding="feelings of understanding". For my response, I took it to mean the former.

      "So according to Searle, symbols need to be understood to have meaning. In other words, meaning requires understanding (or intelligence/cognition)."

      I think we agree on this point. If symbols are not understood, or interpreted by some external observer then they are simply scratches on paper. In that sense, for symbols to have meaning, there must be someone around to interpret (understand) them.


      "So the difference between understanding and mere symbol manipulation is that it feels like something to understand."

      Here I think we disagree. I think that in order to distinguish between symbol manipulation and "understanding", we don't yet have to rely on consciousness, or "the feeling of understanding". It seems to me that the main difference between symbol manipulation (computation) and "understanding" is the addition of sensorimotor capacities, which doesn't necessarily relate to "feelings of understanding".

      When Searle appeals to consciousness, I think he's just using a simple, dirty metric of true cognition. When he does the Turing Test in Chinese using just symbol manipulations, he's asking himself, "Am I cognizing in Chinese by virtue of computation alone?" And since cognition and consciousness are a packaged deal (feelings, and feelings of understanding are a capacity generated by our cognizing brains) something must be missing. Computation is not sufficient for Searle to be cognizing since he lacks an obvious aspect of his own cognition.

      So what happens if we add sensorimotor capacities (dynamics) to a symbol manipulating machine (compuation)? What if we create a robot that can pass T3? A robot that can ground formal symbols in real world referents? Is it necessarily conscious and having "feelings of understanding" in the way that Searle never was? I don't think it's necessarily conscious, and I don't think that's a question we can answer because of the other-minds problem.

      Searle's thought experiment was handy specifically because it could penetrate the other-minds problem by making Searle the T2 robot. In this way Searle could be aware of the consciousness (or lack of consciousness) in the mind of T2 robot. As soon as we upgrade to a T3 robot, and start grounding symbols by way of sensorimotor capacities, the whole thought experiment falls apart because Searle can't "be" a T3 robot, nor can he penetrate the mind of a T3 robot to observe its "feelings of understanding". (Or maybe I'm just not creative enough to think of a cooler thought experiment).

      My point being that I don't think that "understanding must require consciousness". I think that "understanding" requires symbol grounding: a system (dynamical & computational) capable of interpreting symbols based on form and linking them to real-world referents. I don't believe that the "understanding" is contingent upon the actual conscious "feeling of understanding", nor do I think that symbol grounding requires "consciousness" itself. A T3 robot capable of doing what I've described, and passing the Turing Test would surely "understand" or be able to interpret symbols as they are related to their referents, but its "conscious understanding" would remain impenetrable to us because of the other-minds problem.

      Delete
    2. So to speak to the hierarchy you described, I think that meaning requires understanding (in the form of sensorimotor interpretation and symbol grounding) but that these two things are not necessarily contingent upon consciousness or the conscious feeling of understanding since it's unclear where consciousness actually fits into this whole mess.

      God I hope that wasn't an appalling trainwreck. It is altogether possible that I'm completely wrong on some (or all) counts, or that I misinterpreted some (or all) of your post, so feel free to tear into my reply. I'd love to generate some discussion!

      Delete
    3. Hey Chris, thanks for the response. I get confused over the exact meaning (ba-dum, tss) of certain words as well, especially since different authors tend to use them in different contexts. I'm glad someone is down in the trenches with me trying to hash it out though. Anyways, on the two different meanings of understanding, I think the first one you mention, understanding = “capacity to interpret symbols,” is simply symbol grounding. The two terms describe the same concept, which is best illustrated by another point you make:

      It seems to me that the main difference between symbol manipulation (computation) and "understanding" is the addition of sensorimotor capacities,

      Sensorimotor capacities are what takes us from basic symbol manipulation to symbol grounding. Now this “understanding” (symbol grounding) is necessary for meaning, but not sufficient. What else is required then? Consciousness, which is where the second “understanding” (feeling of understanding) comes into play. So you're right about the T3 robot you mention at the end of your comment:

      A T3 robot capable of doing what I've described, and passing the Turing Test would surely "understand" or be able to interpret symbols as they are related to their referents, but its "conscious understanding" would remain impenetrable to us because of the other-minds problem.

      A T3-passing robot is definitely “understanding” when using the first definition (symbol grounding), and we can't know if it is “understanding” when using the second definition (feeling of understanding). Because of this, we can't know if there is actually meaning going on there. As Dr. Harnad points out at the end of his article:

      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. (page 4)

      Piecing this all together, I'd say meaning requires both forms of “understanding.” There needs to be the “capacity to interpret symbols,” as well as the “feelings of understanding.” And because of this second requirement, meaning is, in fact, “necessarily contingent upon consciousness.”

      Delete
  17. Reading Dr. Harnad's paper immediately reminded me of my recent attempts at running word2vec on Twitter data. word2vec is a group of shallow neural nets that are trained to reconstruct linguistic contexts https://en.wikipedia.org/wiki/Word2vec. The reason why these two-layer neural nets actually work is not well understood, but they appear to support, and be supported by the "distributional hypothesis”: linguistic items with similar distributions have similar meanings. Neural nets therefore, seem to have supported some sense of semantic understanding.

    I read the other linked papers above (re the solution/solvability of the symbol-grounding problem). I understand the plea to T3. At the same time, however, I would still ascribe neural nets/deep learnng a greater capacity for semantic understanding than I would a ‘straight’ symbol-manipulation program. Text summarization fairly reliably brings out the ‘meaning’ of a given document, as validated by human coders. Even a ‘simple’ clustering algorithm can result categorizations and classifications that approach what we hypothesize must be part of our meaning-making (I’m thinking of icons). Event detection is a large field of research, where researchers attempt to discern (using various (neural) network-based techniques) when an important event has happened in some medium (including social networks and video), and classify this event. This again, seems similar to our sensory processing. My question then becomes: given the above, can there be varying degrees of solvability of the symbol grounding problem?

    (Side note, somewhat arguing against myself: it appears that Twitter data is too sparse for word2vec to reliably pull anything out (i.e. differentiate between polysemous words))

    ReplyDelete
  18. "Intentionality" has been called the "mark of the mental" because of some observations by the philosopher Brentano to the effect that mental states always have an inherent, intended (mental) object or content toward which they are "directed": One sees something, wants something, believes something, desires something, understands something, means something etc., and that object is always something that one has in mind. Having a mental object is part of having anything in mind.”

    I disagree with this statement. When we meditate there is consciousness and intention there, but focused on nothingness. The absence of thought in this case requires more consciousness than thoughts about objects or content. Brentano says “Even feeling depressed feels like something” clarifying that intentionality is not only the notion of thinking of something, but feeling something too. However, back to my meditation argument, there are times during meditation where one feels absolutely nothing, but this does not mean that consciousness was lost during that time. This may be just one instance where Brentano’s definition does not work. There are still cognitive processes going on during meditation, so how could we now define intentionality? Does it require a more scientific and specific explanation?

    “Meaning, in contrast, is something mental. But to try to put a halt to the name-game of proliferating nonexplanatory synonyms for the mind/body problem without solving it (or, worse, implying that there is more than one mind/body problem), let us cite just one more thing that requires no further explication: feeling. The only thing that distinguishes an internal state that merely has grounding from one that has meaning is that it feels like something to be in the meaning state, whereas it does not feel like anything to be in the merely grounded functional state”

    The article loses me here. I understand the differentiation between feelings which cannot be fully put into words, and grounding of symbols into a meaning. However, when I look at a chair and think of it as a chair, there is a meaning there, yes, but there is also a feeling associated with that chair specifically and all chairs in general. I do not understand where the line is drawn between grounding and feeling. The article says “Grounding is a functional matter; feeling is a felt matter” but don’t they come hand in hand? When you ground an object you feel something too.

    ReplyDelete
  19. Having skimmed Luc Steels’ article, I noticed that he brushed quickly over a claim that he would only be discussing “groundable” symbols, those being symbols that are grounded in perceptual experience. As a result, he suggested that words such as “serendipity” or “holy water” or other words with cultural meaning will be rejected from the argument. I was wondering if these words are also rejected in the description of the symbol grounding problem in the Harnad reading for this week?

    In my opinion they should not because the concept of being grounded is defined by a symbol having meaning. A symbol is grounded if it has meaning or is understood, and with this logic, abstract words like serendipity or culturally relevant words like holy water, do contain meaning and are understood. In light of this discussion, I was wondering how words that describe emotion or “feeling” fall into the symbol grounding problem. It seems that there is the attempt to separate consciousness from the grounding of symbols or in other words that grounding does not require consciousness. With this in mind, how can we separate consciousness when discussing words that describe emotion? When consciousness is defined as feeling, it is directly implicated in the grounding of words describing emotion. For example, we know the meaning of the word “sad” because we know what it feels like to be sad, there is no direct sensorimotor interaction in the physical world that we can ground this word with, but rather the word is directly entwined with how we feel. As a result, are emotion words ungrounded? Or in this case is consciousness needed to ground emotion symbols (words)?

    ReplyDelete
  20. Here is my attempt at the kid sib definition of the symbol grounding problem:
    The symbol grounding problem addresses how specific symbols (words) have specific meanings. The symbol grounding problem explains that words that are in our mental state (in our minds) are meaningful words. Words or phrases which exist in our minds have the ability to refer to unintuitive definitions. Words on a piece of paper or displayed on a computer screen do not have consciousness or sensorimotor capacity, so theses words lack meaning.

    This makes me wonder about words or phrases that have unintuitive definitions (For example, ‘to kick the bucket’ doesn’t mean to actually kick a bucket, but rather to die). How do these words/phrases/expressions get their meaning?

    ReplyDelete
    Replies
    1. Hey Lucy!

      I'm not sure if my description will make sense but I think we can think of the phrase "to kick the bucket" as a series of symbols/shapes that, when placed together in that very order, our mind picks out the referent ‘to die’ and that it its meaning. As in, we know that when someone says "to kick the bucket" or when we read that phrase, that group of symbols in our heads has the referent of death just as the symbols that make up the phrase 'red apple' leads our mind to pick out the referent of a red apple (and I guess it would go along that this rule for picking out that specific referent in our mind is the meaning of the word or phrase in that context). In this way, your question seems not to refer to unintuitive definitions per se (because technically speaking is any definition 'intuitive' without having learned it?) but that there are multiple referents that one symbol (or set of symbols put together) can lead to. So "to die" is just one referent of the symbol-string "to kick the bucket" but also someone physically kicking a bucket is another referent of that word. It clearly necessitates some sort of context in order for the rule to be used correctly by our minds. Think of the way that 'ocean' has multiple referents because there are multiple oceans or there may even be a cafe named ocean, or a human named ocean, for that matter, but we know the correct referent to pick out because of the context in which we encounter the symbols for ocean. So it seems even more, symbols and referent are context dependent. As Dr. Harnad puts it "a symbol system alone, whether static or dynamic, cannot have this capacity [to pick out referents], because picking out referents is not just a computational property; it is a dynamical (implementation-DEpendent) property". The important point that Dr. Harnad makes is that symbols have referents because the symbol system must be dynamical, it cannot just exist alone on a computational basis. So when you ask how phrases like the ones above get their meaning, I think that's your answer – “it is a dynamical (implementation-dependent) property”. Hope that helped and wasn't too confusing… I’m having a hard time trying to keep it straight for myself as well.

      Delete
    2. Hey guys,
      I really like your explanation Alba of dynamical in order to pick out the correct referent of a string of symbols that may have several referents. This was a point that bugged me quite a bit while reading the article, and I feel that what you wrote makes a lot of sense.
      But this brings me to the follow up observation: though it’s very helpful to know that meaning is dynamical (/context dependent), there is no explanation provided as to how the meaning arrises. That is, how does the environment yield the correct interpretation of whatever symbol string is provided? Maybe this is another Fodor case where “it doesn’t matter how meaning and context interact, it just matters that they do”. But this always ends up seeming like a cop out for questions that are difficult to answer. I feel like it isn’t sufficient to know that a symbol system is dynamic, because that statement doesn’t offer any explanation as to what dynamic is or how it functions.
      Or maybe I’m just missing a key point and not understanding the paper correctly. So if anyone has any input, please share!!

      Delete
  21. If I am understanding this right, the Symbol Grounding Problem is just getting back to the easy problem? How do we do what we do in terms of language and understanding languages referents. We are looking at one part of the brains abilities and examining if it can be explained by computationalism, before applying it to all the brains abilities. But what we cannot understand is whether there is the involvement of consciousness. When thinking about this I cannot separate it in my mind from the hard problem, why/how we feel what we feel. In the brain a word is not just grounded in sensorimotor function but also in memories and feelings. The word "mother" is grounded in one's experience with their mother and also with what society has portrayed as what the word "mother" should symbolize. Feelings can be evoked in the friction between one's actual experience and societal expectations, or in just one or the other. This is all to say I cannot understand how consciousness could possibly not be involved in words referents and groundedness when the root can be feelings. But now we are just back to the unsolvable hard problem...

    ReplyDelete
    Replies
    1. I think the first part of your analysis is spot on: that the symbol grounding problem is a counter-argument to computationalism that demonstrates that we couldn't possibly do all the things that we could do -- ie speech -- if our minds were just computer programs. This objection is on the grounds of the easy problem.

      I think I disagree with your second point, though it is interesting to think about philosophically/aesthetically/existentially. Despite the emotional baggage that the word "mother" bears, I can't help but think that Renuka or Riona could build this category and behave appropriately with respect to mothers because they passed T3. This illustrates, in my mind, that feelings and the hard problem can be sufficiently divorced from the hard problem in order to make some headway into the subset of the easy problem that is how we produce speech.

      Delete
  22. "So if Searle is right, that (1) both the words on a page and those in any running computer-program (including a TT-passing computer program) are meaningless in and of themselves, and hence that (2) whatever it is that the brain is doing to generate meaning, it can't be just implementation-independent computation, then what is the brain doing to generate meaning?"

    This kind of reminds me of the whole, "if a tree in a forest falls down and there's no one there to hear it, does it even make a sound?" question. Words on a page are symbols; inherently meaningless in and of themselves, yes, and arbitrary scratches which are really not linked at all to that to which they refer. However, as soon as someone with a brain sees them, they have meaning - just like how as soon as a computer gets an input of various symbols, it can categorize them and process them. I think the word "meaning" is being thrown around here perhaps a little too generously. "Meaning", in my opinion, is the prerequisite for the next set of steps in interpretation/response. If a computer sees a set of symbols and then, from that set, can use the rules to respond (let's say in Chinese to a set of Chinese symbols) with an appropriate response, then that is enough to say that the computer understands them. And that understanding, in my opinion, is a type of meaning.

    I think that what people are getting at here when they talk about "meaning" hearkens back to the hard problem. A huge part of the meaning we ascribe to arbitrary symbols like "bread" or "2" lies in how we feel about the referents. My past experiences with and feelings about that thing over there on my counter which we call "bread" allows me the capacity to pick out referents and have more of a relationship with those symbols. Therefore, I think that what we're really talking about here is just another look at the hard problem, and I think that that's what this boils down to. If we determine why we feel what we feel, or how we feel, I think that that would help elucidate the middle step that computers seem to lack when we're talking about "meaning". Things become grounded, in my opinion, because we feel.

    ReplyDelete
    Replies
    1. Hi Hillary,
      I liked that you linked to symbol and tree falling questions together, because they both follow the same logic:
      When a tree falls down and there’s no one there to hear it, it does NOT make a sound. A sound is defined as a traveling air pressure fluctuation received by an eardrum. The tree falling does physically create a pressure fluctuation, but if there’s no one there to receive the result of the fluctuation then there’s no sound.
      The same thing applies to symbols on a piece of paper: they need to not only contain potential for meaning, but that meaning is only realized once there is a mind present to receive/interpret it. So I would say that squiggles on a piece of paper have no meaning. Instead, meaning arise from a mind’s interpretation of those squiggles.
      And this all ties in to consciousness once again. A digital recorder can store the “sound” of the tree falling, it still won’t be a sound until an eardrum perceives it. That does not mean a computer can’t perform measures on it. Likewise, a computer can analyze squiggles and output the correct squaggles, but there’s no meaning until a human mind receives said squiggles. Therefore I don’t think I would agree that there’s any understanding going on in the computer in the first place… (that is, I’m not quite sure I agree with your definition of understanding/meaning)

      Delete
  23. In terms of whether or not consciousness is needed for grounding, we would say that sensory input is, correct? For example, someone who is blind, deaf, and cannot feel (as unfortunate as that would be) would not be able to learn any type of language. There would be absolutely no way for them to ground any symbols because they couldn’t interact with the world in order to sense or feel the referents of the symbols’ meanings. So if we say then that sensory input is necessary for grounding, would we not say that consciousness is necessary for sensory input to be felt? Or would perhaps a computer be able to devise a way around this (ie using a camera for “sight”)?

    ReplyDelete
    Replies
    1. I think the unsatisfying answer to this is that you could probably program a computer to react to thinks like what we would think of as sensory input. I'm imagining a robot that can pass T3 because it can be programmed to react to something kicking it and then saying "Ouch" or "seeing" a stimulus and having a programmed reaction based on that stimulus-- I mean how else could a robot realistically pass T3 and convince humans it is another human without this? Though, it definitely seems like there's a limit to the programability of this so I understand your frustration.

      Delete
    2. Yes I believe sensory input is necessary to ground the symbol with it's referent (which is what allows the robot to pass T3). I think you are confusing 'doing' (the easy problem) and 'feeling' (the hard problem). If I have this correct, symbol grounding problem deals with the easy problem, arguing that cognition cannot be just computation (therefore it cannot be solved with T2 since the symbols are not grounded in real world referents). However, our Riona and Renuka robots could have synthetic sensors to process sensory inputs (just like we 'see') and produce the appropriate actions in response (which they would have to do in order to do everything that we can do). However, as you were mentioning, I am unsure how many 'sensory systems' the T3 robot would need be able to sufficiently ground the symbols with their referents. That being said, I do not think that whether T3 robots can feel the sensory inputs or whether the robot would be conscious are being questioned here.

      Delete
    3. I agree... I think this might be confusing consciousness with the ability to detect sensory input, which clearly is not the same thing

      Delete
  24. "Another symbol system is natural language. On paper, or in a computer, it too is just a formal symbol system, manipulable by rules based on the arbitrary shapes of words. In the brain, meaningless strings of squiggles become meaningful thoughts"

    From this reading, it seemed that language only exists if there is a conscious being to which its symbol system is grounded.

    What about languages of ancient civilizations, like Egyptian hieroglyphs, where meanings are lost to the modern man? Does that language only exist as a formal symbol system, until a mind is capable of mediating the language and its intended referents?

    If someone is only able to decipher a language system, say Egyptian hieroglyphs, with the help of rules like the Rosetta Stone, which is an ancient artifact that was essential for modern understanding of the hieroglyphs, then is he not comparable to Searle in the Chinese room, wherein he does not ascribe meaning to the language but instead follows rules to decipher its output?

    ReplyDelete
    Replies
    1. "Discrimination is independent of identification. I could be discriminating things without knowing what they were. " (Harnad, 1990)

      Hi Freddy!

      I think you might have the Chinese room experiment and the Rosetta stone a bit confused!

      In Searle's thought experiment, he is receiving inputs of Chinese characters and has some rules in english that say "if you receive THIS string of characters, then send out THAT string of characters." You can think of it as being equivalent to him sitting in a room and having inputs and outputs of strings of shapes with rules like: "if you receive the string: square, triangle, circle, then send out star, octagon, circle" and other arbitrary rules for every type of string input you could possibly receive. Whether Searle is receiving strings of Chinese characters or strings of shapes, it remains that he cannot determine meaning from these arbitrary symbol manipulations without some reference to their real world meaning (thus why T2 is not adequate enough to pass the Turing Test, since interaction with the world (T3) is the only way that one can ground such symbols).

      The Rosetta Stone on the other hand was a decree written in 3 languages: Ancient Egyptian hieroglyphs, Demotic script, and Ancient Greek. While there had been no knowledge of how to read hieroglyphs, upon discovery of this stone, the Greek text, which was widely known by scholars, provided a source of reference for hieroglyphic text so that meaning could be understood from the hieroglyphs.

      To contrast this with Searle's Chinese room:

      If Searle's thought experiment followed a Rosetta Stone-esque template, then the rules he would for deciphering, say a made up shape language, would be something more like this:
      "If you receive the string: 'square, triangle, circle' the person is asking 'how are you?', you should respond with, 'star, octagon, circle'."
      After a while, of back and forth, you would in fact come to ascribe meaning from these words - however, as I mentioned before, this is not how Searle's actual thought experiment (nor how a T2 computer program with no outside references) functions.

      I think the quote I put at the top sums it up nicely:
      In Searle's case he is just discriminating between symbols provided, on the other hand, your Rosetta Stone example is actually a case of identification.

      Hope this makes sense!

      Delete
  25. These are just some observations I had while reading the article.
    How would we account for when the wrong referent is picked out for a symbol? I understand the system is dynamic and context dependent, but I feel there would still be cases where that would not help elucidate which referent the symbol is pointing at at a given time. For example with “Tony”, I’m quite positive there are many people named Tony in the world, so making a simple statement such as “Tony said his pants were blue” does not give any clue as to which Tony that may be.
    Another point about the mind “grounding” meaning. To make sure I’ve understood correctly, when the mind is mediating it’s grounding a symbol? So tying this is with what Searle says, this seems to mean that grounding is a conscious process; it’s not enough for it to just happen in the mind, it also has to be in the conscious awareness of whoever is interpreting the symbol at a given time. So Searle set up his experiment premises so that any kind of understanding/grounding that may occur has to be conscious in order to be defined as such. He doesn’t exclude the fact that there might be unconscious understanding going on, but because it is a mentioned, unconscious, it cannot be reported or accounted for, just like we cannot know about those same processes in computers, therefore it is irrelevant. But what about if there truly is some sort of unconscious understanding going on? Does that mean anything at all with regard to symbol grounding? Would there be any influencing going on? How to be sure?

    ReplyDelete
    Replies
    1. Yes, sometimes we don't know what a word or name refers to. So what? To find out, you ask some more questions, or do some more looking...

      When you read a sentence in a book, you know what it means, but the book doesn't. It doesn't mean anything to the book. Ditto for the words spoken by a T2 computer. It doesn't mean anything to the computer, So what's actually going on in your head can't just be T2 computation, like what's going on in the T2 computer, otherwise the words wouldn't mean anything to you either (the way the Chinese symbols don't mean anything to Searle).

      To ground the words in your head in their referents, you need to be a T3 robot able to detect and interact with those referents. But unlike for T2 computation, for a grounded T3 hybrid robot, there is no "Searle's Periscope" to check whether T3 really understands (i.e., whether the words mean anything to T3). So there's no way to know whether T3 has meaning, or just grounding.

      And even if there were a divine periscope, which a god used to check whether T3 really understands, and the god assured you that, yes, T3 really does understand and feel, that still wouldn't solve the "hard problem" of explaining how and why it feels, rather than just acts T3-indistinguishably from someone that feels. So although T3 (or T4) solves the "easy problem" of explaining how and why we can do what we can do, it cannot explain how and why we feel (hence why grounding either isn't enough for understanding or meaning, or it is enough, but we just can't explain how or why).

      Unconscious "understanding," until further notice, is just T3 grounding, whereas the question is whether (and how and why) grounding is understanding. Calling it "unconscious (unfelt) understanding" just begs the question (and, if you think about it, it actually says nothing at all: it just renames T3 grounding as "unconscious understanding."

      Delete
    2. Yes, sometimes we don't know what a word or name refers to. So what? To find out, you ask some more questions, or do some more looking...

      When you read a sentence in a book, you know what it means, but the book doesn't. It doesn't mean anything to the book. Ditto for the words spoken by a T2 computer. It doesn't mean anything to the computer, So what's actually going on in your head can't just be T2 computation, like what's going on in the T2 computer, otherwise the words wouldn't mean anything to you either (the way the Chinese symbols don't mean anything to Searle).

      To ground the words in your head in their referents, you need to be a T3 robot able to detect and interact with those referents. But unlike for T2 computation, for a grounded T3 hybrid robot, there is no "Searle's Periscope" to check whether T3 really understands (i.e., whether the words mean anything to T3). So there's no way to know whether T3 has meaning, or just grounding.

      And even if there were a divine periscope, which a god used to check whether T3 really understands, and the god assured you that, yes, T3 really does understand and feel, that still wouldn't solve the "hard problem" of explaining how and why it feels, rather than just acts T3-indistinguishably from someone that feels. So although T3 (or T4) solves the "easy problem" of explaining how and why we can do what we can do, it cannot explain how and why we feel (hence why grounding either isn't enough for understanding or meaning, or it is enough, but we just can't explain how or why).

      Unconscious "understanding," until further notice, is just T3 grounding, whereas the question is whether (and how and why) grounding is understanding. Calling it "unconscious (unfelt) understanding" just begs the question (and, if you think about it, it actually says nothing at all: it just renames T3 grounding as "unconscious understanding."

      Delete
  26. I found this article to be very insightful and informative. There are just a few things I had trouble grasping that I wanted to clarify. Firstly, if the meaning of a word is the rules/features that one must use to pick out of its referent and there is no explicit way to uniformly pick out rule, would we obtain some kind of uniform result if we were to work backwards? If we were given the meanings would we be able to identify the rules; and thus identify the referent. Would this help establish some kind of uniformity for the rules for referents and therefore meaning? Due to the large ambiguity; if we are unable to do so, then how can we even come up with the meaning of words?

    In addition, I just wanted to clarify the following statement. "So the meaning of a word in a page is "ungrounded," whereas the meaning of a word in a head is "grounded" (by the means that cognitive neuroscience will eventually reveal to us), and thereby mediates between the word on the page and its referent.”

    I don’t think there is enough evidence to support this statement and it seems like it is slightly flawed. It seems like they are both intertwined with the meaning on a page and in one’s head. In order to even have a word appear on a page; wouldn’t it have to go through a writer and therefore through one’s head? How can it be separated then if it goes through both the writers head and paper; how would the meaning change.

    ReplyDelete
  27. I find the effect Harnad (2003; 1990) terms the “hermeneutic hall of mirrors” very interesting. Harnad (1990) defines the hermeneutic hall of mirrors as “[the] illusion one creates by first projecting an interpretation onto something (say, a cup of tea
    leaves or a dream) and then, when challenged to justify that interpretation,
    merely reading off more and more of it, as if it were answerable only to itself” (Harnad, 1990). I wonder how this may relate to the earliest interpretations of cryptologists, who are only able to decipher “ancient languages or secret codes” in modern day because they can ground their exploration in their first language and in their knowledge of the real world. In the footnotes it is explained that “cryptologists also use statistical information about word frequencies, inferences about what an ancient culture or an enemy government are likely to be writing about, decryption algorithms, etc.” (Harnad, 2003). Having access to this sort of relevant contextual information seems de facto to the task of interpreting any code or symbol system that is entirely foreign and initially appears very arbitrary. But what did cryptologist do before this type of information was available? Clearly it is possible to eventually “crack the code”, even without statistical information, algorithms, or the like. How would this be accomplished? In the same vein, without the help of modern technology/contemporary knowledge how did cryptologists not fall victim to the hermeneutic hall of mirrors? How did one recognize the presence or absence of a pattern/syntax in a series of symbols that may or may not be grounded?

    ReplyDelete
  28. From reading some of the other comments here, it seems that some individuals are confusing the hard problem and the easy problem with regards to the symbol grounding problem, and I can certainly see why. I myself was a little confused as to what 'solving' this problem would mean at first as well, since we talked about how it feels like something to 'understand,' and yet solving the symbol grounding problem apparently does not require such feelings.

    I think I better understand now the separability of the easy and hard problem now.

    "The present grounding scheme is still in the spirit of behaviorism in that the only tests proposed for whether a semantic interpretation will bear the semantic weight placed on it consist of one formal test (does it meet the eight criteria for being a symbol system?) and one behavioral test (can it discriminate, identify and describe all the objects and states of affairs to which its symbols refer?). If both tests are passed, then the semantic interpretation of its symbols is "fixed" by the behavioral capacity of the dedicated symbol system, as exercised on the objects and states of affairs in the world to which its symbols refer; the symbol meanings are accordingly not just parasitic on the meanings in the head of the interpreter, but intrinsic to the dedicated symbol system itself." (Harnad, 1990)

    These lines in the closing paragraph I found helpful in teasing apart just what sort of solution we are after for the symbol grounding problem. While feeling certainly accompanies the vast majority of our our conscious mental states, the easy problem is searching for why and how we have the capacity to behave and cognize the way we do, NOT why and how it feels like something to when we do. In order to have a T3 robot pass the Turing Test, it must be behaviourally equivalent, and that requires it possessing a grounded symbol system in the sense mentioned above. Without this capacity to reference the outside world as a link to the symbols it manipulates, it can never be considered behaviourally (or cognitively?) equivalent.

    In short, the easy problem here is understanding why and how we have the capacity to understand (language), which is what the symbol grounding problem posits, and the hard problem is why and how it feels like something to understand, which is is inessential to the question at hand.

    ReplyDelete
  29. "If both tests are passed, then the semantic interpretation of its symbols is "fixed" by the behavioural capacity of the dedicated symbol system, as exercised on the objects and states of affairs in the world to which its symbols refer; the symbol meanings are accordingly not just parasitic on the meanings in the head of the interpreter, but intrinsic to the dedicated symbol system itself."

    I feel like this passage captures the essence of symbol grounding. I feel as if a lot of people have failed to grasp what the symbol grounding problem really tells us about cognition, so I would like to refer back to the opening statement (where Harnad refers to the Chinese room), to clarify what exactly the intrinsicity of the system implies:

    Talking about the Searle's Chinese room: "The symbols and the symbol manipulation, being all based on shape rather than meaning, are systematically interpretable as having meaning -- that, after all, is what it is to be a symbol system, according to our definition. But the interpretation will not be intrinsic to the symbol system itself: It will be parasitic on the fact that the symbols have meaning for us, in exactly the same way that the meanings of the symbols in a book are not intrinsic, but derive from the meanings in our heads. Hence, if the meanings of symbols in a symbol system are extrinsic, rather than intrinsic like the meanings in our heads, then they are not a viable model for the meanings in our heads: Cognition cannot be just symbol manipulation."

    So here Harnad is saying that in Searle's scenario, interpretation of the symbol system is parasitic on the meanings that symbols have in our heads and not intrinsic to the system itself (i.e. extrinsic). If meaning of symbols is extrinsic to the "system" then cognition cannot be just symbol manipulation because symbol manipulation alone would not produce intrinsic meaning. Going back to the passage at the end, Harnad suggests that symbol meanings are not only parasitic on the meaning inside our heads but are also intrinsic to the symbol system itself. Hopefully this makes sense, although with all this "intrinsic" and "extrinsic" business it isn't exactly kid sib.

    ReplyDelete
  30. I think the most enjoyable aspect that I pulled out this article is how Stevan creates a distinction between grounding and meaning. Well, pointing out that meaning might actually be more than just grounding. All of meaning might not be captured by understanding what grounding is. This, I think, is where the paper becomes especially fascinating. This is because Stevan really can’t prove the distinction between grounding and meaning in the same way the rest of his paper focused on the distinction between computation (symbol-manipulation) and cognition (everything humans can do). I would like to ask what Stevan means when he is talking about some fascinating extra property that humans have which computers do not necessarily have. Stevan says that a T3-passing robot might not have what Searle has in his head: “It could be a zombie, with no one home, feeling feelings, meaning meanings.” What does that mean, exactly? What is this functional capacity?

    Anyways, sensorimotor grounding is super interesting to me. According to Stevan, words are not grounded in other words, but they can be grounded through sensory capacities and motor capacities. So when one smart-ass-undergrad asks another undergrad to explain the meaning of coffee, the explainer could explain by funneling hot coffee directly down the throat of the questioner. It is undoubtedly a powerful definition of coffee, which is grounded in the sensorimotor capacities of the meaning-making subjects that are involved.

    ReplyDelete
  31. The symbol grounding problem deals with how words get their meanings and then raises questions related to what meaning actually is. From the examples of the Chinese Room and the Chinese Dictionary, it can be seen that the meanings of words are grounded in other words. Essentially, words are symbols and we see that these symbols have meanings grounded in other symbols. This is the symbol grounding problem.

    Obviously, the meaning of a word has a sense and reference. But this problem helps us conclude that there is something more than that; something more than just sensorimotor experience and computation grounded in meaning. Can we say that there is a sense of feeling grounded in words and meanings?

    Since the symbol grounding problem is essentially a symbol system, it is closely related to computation and this really interests me. The words we process, are they the dynamic in our heads or are they static as we see them on paper? The symbol grounding problem raises more questions against the view of computationalism.

    ———

    Another perspective: since word definitions lead to an infinite regress, there is an external factor that we are not accounting for that links words to meanings. In this case we assume it is “feeling”. But what if it’s not feeling and it’s senses? We could be using our senses/perception to link words we hear/read to referents in the physical world.

    Maybe AI will figure it out once we have reached a phase where AI perception is strong enough.

    ReplyDelete
  32. I just had one question that I wanted to address. I like how the author explained the derivation of meaning of a word by linking it to the various representations and subsequent experiences. However, what would then be the common feature that would help unify all of these variations? Would it be a singular feeling? Is the author saying that ambiguous word’s meaning comes from sensorimotor capacities? Do we need to experience this common feeling to help understand the meaning of a word?

    ReplyDelete
    Replies
    1. This is more fully explained in the readings on category learning and language. The members of the same category differ but share some common features (invariants), although the invariants may be described by a complex rule ("A or not B or if C then D..."

      Words with multiple meanings name different categories, with different invariants.

      None of this explains anything about feelings, because of the "hard problem."

      Delete
  33. After re-reading this article and the comments on this thread; I had another question that I think deserves further research. After doing all the readings from the course, I am aware that most words do not have to be grounded directly; we can learn their meaning indirectly from things like explanations, as long as the word in the defenition is grounded (directly or indirectly). Since the symbol grounding problem is that there cannot be fully indirect grounding, my question is then how many words have to be grounded directly?

    ReplyDelete
    Replies
    1. A tentative first estimate of the "minimal grounding set" -- the smallest number of words that can define all the other words -- is made in 8b. We now think it's about 1500 words, but it's not unique. There can be many variations. And it's almost certain that it's not just matter of learning exactly 1500 directly and then spending the rest of one's life in a room doing T2! Categories surely continue to be learned both ways -- by sensorimotor induction as well as verbal instruction -- throughout life.

      Delete