Symbol Grounding And Symbol Tethering
    By Samuel Kenyon | April 2nd 2013 11:33 PM | 7 comments | Print | E-mail | Track Comments
    About Samuel

    Software engineer, AI researcher, interaction designer (IxD), actor, writer, atheist transhumanist. My blog will attempt to synthesize concepts...

    View Samuel's Profile
    Philosopher Aaron Sloman claims that symbol grounding is impossible. I say it is possible, indeed necessary, for strong AI. Yet my own approach may be compatible with Sloman's.

    Sloman equates "symbol grounding" with concept empiricism, thus rendering it impossible. However, I don't see the need to equate all symbol grounding to concept empiricism. And what Sloman calls "symbol tethering" may be what I call "symbol grounding," or at least a type of symbol grounding.

    Firstly, as Sloman says about concept empiricism [1]:
    Kant refuted this in 1781, roughly by arguing that experience without prior concepts (e.g. of space, time, ordering, causation) is impossible.

    Well that's fine. My interpretation of symbol grounding didn't involve the baggage to bootstrap everything. Innate concepts, which are triggered by machinations in phylogenetic space, can contribute to grounding.

    Sloman also says that concept empiricism was [2]:
    finally buried by 20th century philosophers of science considering the role of theoretical terms in science (e.g. “electron”, “gene”, “valence”, etc.) that are primarily defined by their roles in explanatory theories.

    Here is the only inkling of why this would take down symbol grounding: abstract concepts might actually be defined in terms of each other. As Sloman explains here[3]:
    Because a concept can be (partially) defined implicitly by its role in a powerful theory, and therefore some symbols expressing such concepts get much of their meaning from their structural relations with other symbols in the theory (including relations of derivability between formulae including those symbols) it follows that not all meaning has to come from experience of instances, as implied by the theory of concept empiricism

    On the other hand, maybe theory concepts are grounded, but in a very tenuous way. Here is a metaphor, albeit not a great one: Imagine a family of hot air balloons with links between them, and this group is floating free. However, they aren't quite free because there is a single rope tying one of them, and indirectly all of them, to the ground. Sloman seems to be saying something like that, via mechanisms of how good a theory concept is at modeling something, hence the term "symbol tethering". Whatever the case, I don't see why all symbols have to be like theory concepts.

    If the goal is to understand how human minds create and use knowledge, then one is led down the road of grounding. Otherwise you're playing Wacky Mad Libs or tainting an experiment with an observer's human knowledge. Imagine if you could pause a human (or some other animal) and have access to the layer or point-of-view of internal mental symbols. You might then ask, what is the genealogy of a particular symbol--what symbols are its ancestors? The path to embodiment-derived symbols or innate symbols may be long and treacherous, yet there it is. And if the path stops short, then you have chosen a symbol which is somehow usable in a biological mind, yet is completely encapsulated in a self-referential subsystem.

    Sloman has hypothesized that theory concept networks don't need to be grounded in any normal sense of the word. But that doesn't mean we throw the baby out with the bathwater. As far as I can tell, we should add the theory tethering mechanism in as a method of grounding symbols. Or perhaps it is simply one of the other ways in which information structures can be handled in a mind. I think it is plausible to have ungrounded symbols generated by a mind which also has grounded symbols. The inherent structure of a ungrounded self-referential database could be useful in certain contexts. But ungrounded symbols are easy. That's the default for all existing computing systems. And that's what a dictionary is. The nature of those dictionary-like systems are at most a subset of the nature of human-like knowledge systems. We end up with humans injecting the meaning into the computer (or dictionary or whatever). The tricky problem is making systems that are grounded in the same way humans or other animals are. Those systems could have compatible notions of common sense and general (to humans) understanding. They would, in turn, be capable of doing the same kind of knowledge injection or anchoring that humans do with ungrounded systems.


    [1] A. Sloman, "Symbol Grounding is Not a Serious Problem. Theory Tethering Is," IEEE AMD Newsletter, April 2010.
    [2] A. Sloman, "Some Requirements for Human-like Robots: Why The Recent Over-emphasis on Embodiment has Held up Progress," in B. Sendhoff et al., Eds., Creating Brain-Like Intelligence, pp. 248-277, Springer-Verlag, 2009.
    [3] A. Sloman, “What’s information, for an organism or intelligent machine? How can a machine or organism mean?,” 2011.


    Gerhard Adam
    Perhaps this is a sidebar issue, but I was looking over the paper mentioned in note (3) and couldn't help but be struck by the confusion generated by the concept of "information".

    Maybe I'm way off here, but it seems that some of this confusion can be alleviated by being more precise in our definitions.  In my view, the "stuff" that is "out there" is simply data.  Once data is input and interpreted then it becomes information.  When information becomes integrated and can be applied then it becomes knowledge.

    In other words, "доброе утро" is simply data, until it can be interpreted properly to become information.  If I were to translate this into English ["Good morning"], then it becomes information to the appropriate interpreter.  Of course, this is then related to all manner of other concepts regarding good/bad and times of the day, etc., as well as whether it is merely convention.  Then concepts like "the sun comes up in the morning" take that information and convert it to knowledge.  From here it seems like we can construct exceptions that occur [i.e. earth's rotation, polar regions where the sun doesn't set, etc.].

    I guess I'm not really saying anything special here, but it seems that the discussion becomes more confused the less precise we are in what we mean by information.  So, in my view, a molecule is just a molecule until it binds to a cell site that can interpret it as information.  Even within a computer system, 1's and 0's are simply data unless they are viewed within the context in which they occur.  Are they instructions or data?  Can data be misinterpreted as instructions or vice versa?  Obviously yes, because if the context is wrong, then all manner of spurious errors occur.

    Anyway ... I don't know where I was going with this, so rather than continue rambling I'll stop :)

    Mundus vult decipi
    It's a sidebar, but semantics of information ties back to meaning. I tend to think of information as streams of arbitrary things (physical changes that == bits) which do not have any inherent meaning, but an interpreter can create a meaning from that stream. And the arbitrary bitstream must cause changes by the content of the bits (patterns), not by the physical causes alone.
    But it makes you wonder if we should call any bit-supporting stream "information" if there are no entities that use that bitstream at either end. It seems to be metaphysically the same type of stuff, so we just call it information anyway even if nothing can use it and nobody can create meaning out of it.
    Gerhard Adam
    Yes, it's the same type of stuff, but ....

    I guess my problem with it is that within such a definition, then everything becomes "information" because one can never know if it may be subject to interpretation by something, somewhere.  To me that doesn't seem particularly useful.

    It would seem that this is precisely the problem that biological organisms address, by having senses that focus on the relevant parts of the data to extract only that which can be useful as information. 
    Mundus vult decipi
    I have a hypothesis which really turns all of our current assumptions about information, physics, and consciousness upside down, but which I think is ultimately the only plausible solution. Samuel comes closest in seeing that without some sender or receiver on either end, there can't really be said to be information. The understanding that I have come to takes this one step further to say that there is no such thing as information which is independent of some sensory-motor phenomenon. Without some experience in which an intentional participant is informed (and in my hypothesis that includes non-human organisms, and even inorganic matter), there is no "in" for the "formation".

    I would disagree with Samuel in saying "Innate concepts, which are triggered by machinations in phylogenetic space" being a way around concept empiricism, as I think that it just sweeps it down into another level. After a lot of consideration over several years, my conclusion is that no form or function can exist independently of some context of pattern recognition and experiential presence, i.e. perception and participation. It doesn't matter if the forms are conceptual, symbolic, mechanical, digital, substantial, logical, etc, there has to be an underlying context of sanity: basic conditions of persistence of position, boundary and contrast, symmetry, repetition, significance, etc. If we look at a Turing machine to explain thought, we have already taken for granted the entire sub-mechanical support system symbolized by the machine's read/write head, it's controllable tape, the integrity of integers and arithmetic operations, equivalence and many fundamental conditions which themselves cannot be explained by more primitive phenomena must be taken for granted before any form of mechanism can exist. The universe has to make sense before anything can make sense of it. What this means, is that forms and functions are a category of sense and not the other way around.

    I would also take a deeper look at the assumptions being made behind Gerhard's statement "In other words, "доброе утро" is simply data, until it can be interpreted properly to become information." Rather than thinking of information in terms of being interpreted "properly" I suggest that we see interpretation as a multi-level continuum of sense. When I read "доброе утро", I can see that it looks Russian, but it could also look Greek. The characters already provide me with particular associations even though I don't know how to read it. This is important because it speaks to Searle's Chinese Room, etc, where a mechanical process can copy some phenomenon at some level of description from one form to another without being able to interpret the phenomenon at every level. Beyond simple copying from one dictionary to another by matching, a machine can take the next step to assemble complex strings of symbol forms using a larger dictionary, but still having acquired no semantic sense of what is intended to be communicated. The publicly extended form can be reproduced but the private intention cannot be simulated.

    The hard part is to realize that even the lines and curves in each character, and the pixels that make up those lines, and the molecular interactions which make up those pixels...all of those are levels of sense which supervene on optical form-functions and visual sense, then cognitive sense and motives. We have no access to phenomena which are beyond some kind of detection-participation, so we can't assume that such a thing could exist. Even if it could, without any aesthetic dimension at all, what would really be the difference between such a thing and nothingness?

    I would disagree with Samuel in saying "Innate concepts, which are triggered by machinations in phylogenetic space" being a way around concept empiricism, as I think that it just sweeps it down into another level. After a lot of consideration over several years, my conclusion is that no form or function can exist independently of some context of pattern recognition and experiential presence, i.e. perception and participation.

    The innateness of a concept or concept-precursor does not specify the mind in which the concept is instantiated and used. Genetic structures store instructions for generating the intentional beings which in turn end up running wetware with concepts.

    Certainly function is, as you say, dependent on a context of an entity that perceives recognizes. Form as we know it is constructed by our minds and in fact we have to assume overlaps between us (also we assume each other exists), and we also don't know how much our constructed forms correlate with objective reality. Indeed, we are phylogenetically-designed to construct our versions of reality. So innate concepts (or concept-precursors) make sense since phylogenetic space just specifies how to launch the creatures which in turn actually have the concepts during runtime (ontogeny).

    If, on the other hand, you are implying that I'm "sweeping it down another level" because something had to have a concept before phylogenetic space in some kind of chicken-egg paradox, then I don't buy that either since that seems to reject the whole concept of evolution.
    Thor Russell
    This whole question of "how are symbols grounded" seems a bit backwards to me. It seems that intelligent structures build up models/representations of something, say a part of the environment and then we try to explain this process. So rather than asking whether something really is a symbol or if a system is capable of having symbols, simply ask if you can explain the systems behavior without it. 
    Dennett in the Intentional Stance discussed similar issues. Take a simple example, the concept of temperature. Does temperature "really exist"? Well, you can calculate all the moving of molecules and say that you don't need temperature and can predict everything without it, however you cannot predict how I can make accurate predictions without such information (i.e. I can predict the water will boil without knowing the positions of all molecules) unless you then introduce the concept of temperature. You are forced to use the concept of temperature if you are to explain my skills. Saying it then "doesn't exist" doesn't make sense to me.

    You can apply similar reasoning to whether symbols exist or are "grounded". Say you have a baby and you want to know if there are any symbols in its brain. If you notice that it always smiles say when a face is close, then you could attempt to deny the symbol "face" by explaining its behavior by calculating molecules etc or characterizing faces in complete detail, however in order to explain how someone without this knowledge can also predict when the baby will smile, you need the concept of face. Asking "is the babies face symbol grounded" is not helpful. You need the symbol "face" to explain its behavior, "grounded" is not relevant. Patterns, relations, concepts etc are gradually built up in the brain(or AI system) until the point is reached that the only way to sensibly explain the behavior of a brain or similar complex system is with the concept of symbols. This also has to be Bayesian in nature, is a face detector that fires often for a round thing "really" a face detector? There is no clear dividing line. 

    Asking "how do symbols become ungrounded" is then the question. The answer is when you start taking only parts of such systems out of context and arranging them in different ways. A dictionary is of course such a thing, you don't have the neurons/synapses necessary for "face" along with the word, you just assume that the words correspond to the same neural patterns in anyone reading the word. 

    Relating to concept empiricism etc:
    Now of course you need some initial experience with the world to get the first thing that deserves to be called a symbol. Some data needs to be input that the system can learn. (Well it could born with it, but that's kind of a copy of an existing system) For example with any logical system like maths, you can't define the first concept as there is nothing to define it with. Then of course you can define other concepts as combinations of existing ones.

    It seems a lot of the problem is that asking "how are symbols grounded" is the wrong question and causes confusion before you even start. Figure out what is meant by symbol, then "ungrounded symbol" is a special case of it that shouldn't be the starting point.

    I don't see why there is anything else to it, if you want "symbols" that are grounded in the same way humans symbols are, then you need to build up (or perfectly copy the relevant part) the basic structure in your AI system in a similar way to that in humans (well I would start with something biologically less complex). Such a structure then deserves the same symbol only when it behaves in a similar way to the biological one and presumably has a similar internal representation.

    Now I don't mean necessarily using the same crude birth/death/genetic inheritance method, there is nothing special about this, in fact it seems quite inefficient. Copying and using the techniques that give rise to the neural plasticity that exists in a single individual or subsystem seems a more efficient, faster and powerful method to me.

    Going about things the other way and say having a "face" symbol and then trying to "ground" is then obviously seen to be misguided.
    Thor Russell
    Patterns, relations, concepts etc are gradually built up in the brain(or AI system) until the point is reached that the only way to sensibly explain the behavior of a brain or similar complex system is with the concept of symbols
    I would agree with that. The reason it may seem backwards to you is that you aren't thinking about the fact that most computing systems, even those with supposed commonsense reasoning do not link all the way back/down in a manner similar enough to humans. This shallowness and/or difference in low-level structures for grounding means computers and robots don't understand in the same way humans do. The other reason why symbol grounding seems to come up a lot is in the effort to explain the various interconnected ways in which a human baby-to-adult creates meaning. Your description is not invalid, but it hardly is detailed.

    Going about things the other way and say having a "face" symbol and then trying to "ground" is then obviously seen to be misguided.
    Only in terms of development. I'm not saying any biological creatures have a symbol out of nowhere and struggle to ground it. Going "backwards" is purely for discussion. If minds had diagnostics and introspection, one could look at a symbol at a particular paused moment of time and then trace the history and semantic build-up of that symbol in order to figure out where and how it came from.