rik%roland@SDCSVAX.UCSD.EDU.UUCP (05/21/87)
I now see why you consider my use of ``subsymbolic'' sloppy. It is because you have a well thought out, concrete proposal for three distinct representational levels that captures extremely well the distinctions I was trying to make. In the main, I think I accept and even like your ``psycho-physically grounded'' symbol theory. I do have a few questions, however. \section{Icon/Pointer/Symbol != Icon/Category/Symbol} First, what evidence causes you to postulate iconic and categorical representations as being distinct? Your distinction appears to rest on differences between the types of performance at task these two representations each ``subserve.'' Apart from a relatively few cognitive phenomena (short-term sensory storage, perhaps mental imagery) I am aware of little evidence of ``... continuous, isomorphic analogues of the sensory surfaces'' that is the basis of your iconic representations. In any case, I see great difficulty in distinguishing between such representations and ``... constructive A/D filters which preserve the invariant sensory features'' based simply on performance at any particular task. More generally, could you motivate your ``subserve'' basis for classifying cognitive representations. I use ``icon'' to mean much the same as your ``categorical representations'' (which I'm sure will cause us no end of problems as we discuss these issues!). These representations --- whatever they are called --- are characterized by their direct, albeit statistical, relationship with sensory features. This distinguishes icons from ``symbols'' which are representations without structural correspondence with the environment. Your, more restricted, notion of ``symbol'' seems to differ in two major respects: its emphasis on the systematicity of symbols; and its use of LABELS (of categories) as the atomic elements. I accept the systematicity requirement, but I believe your labeling notion confounds several important factors. First, I believe you are using labels to mean POINTERS: computationally efficient references to more elaborate and complete representations. Such pointers correspond closely to Peirce's notion of INDICES, and are valuable not only for pointing from symbols to icons (the role you intend for labels) but also from one place in the symbolic representation to another. Consider Peirce's view on the primacy of pronouns. However, I have come to use the term ``pointer'' instead of ``index'' because I also mean to refer to the vast economy of representation afforded by such representational devices, as recognized by computer science. Pointers have obviously been an integral part of traditional data structures in computer science since the beginning. Quillian's use of TOKEN --> TYPE pointers is still a classic example of their benefit to AI knowledge structures. More recently, many connectionists have taken this pointer quality to be what they mean by ``symbol.'' For example, Touretzky and Derthick say: \begin{quotation} Intelligence seems to require the ability to build complex structures and to refer to them with simpler objects that may be passed among processes easily. In this paper we use ``symbol'' to denote such objects... Symbols in Lisp are mobile [one of five properties Touretzky ascribes to symbols] because their addresses are easily copied and passed around. In connectionist models where symbols are identified with activity in particular units, symbols are not mobile. [Touretzky \& Derthick, ``Symbol structures in connectionist networks'' IEEE COMPCON 1987] \end{quotation} A more sophisticated form of pointer has been discussed by Hinton as what he calls a ``reduced description.'' The idea here is to allow the pointer to contain some reduced version of the description to which it is pointing. (For example, consider the use of tag bits in some computer architectures that indicate whether the pointer address refers to an integer, a real, a string, etc.) If the reduced description is appropriately constructed, the pointer itself may contain sufficient information and so the computational overhead of following it to the full description can be avoided. In general, however, it might seem impossible to construct such appropriate reduced descriptions. But if a PROCESS view of cognition is adopted, rather than relying on a STATIC structure to encode all information , such generalized pointers become more conceivable: reduced descriptions correspond to PARTIALLY ACTIVE representations which, when more FULLY ACTIVE, lead to more completely specified descriptions. The other feature of your labeling notion that intrigues me is the naming activity it implies. This is where I see the issues of language as becoming critical. I would go so far as to propose that truly symbolic representations and language are co-dependent. I believe we agree on this point. It is important to point out that by claiming true symbol manipulation arose only as a response to language, I do not mean to belittle the cognitive abilities of pre-lingual hominids. Current connectionist research is showing just how powerful iconic (and perhaps categorical) representations can be. By the same token I use the term language broadly, to include the behavior of other animals for example. In summary, it seems to me that the aspect of symbols connectionism needs most is something resembling pointers. More elaborate notions of symbol introduce difficult semantic issues of language that can be separated and addressed indepently (see below). Without pointers, connectionist systems will be restricted to ``iconic'' representations whose close correspondence with the literal world severly limits them from ``subserving'' most higher (non-lingual) cognitive functioning. \section{Total Turing Test} While I agree with the aims of your Total Turing Test (TTT), viz. capturing the rich interrelated complexity characteristic of human cognition, I have never found this direct comparison to human performance helpful. A criterion of cognitive adequacy that relies so heavily on comparison with humans raises many tangential issues. I can imagine many questions (e.g., regarding sex, drugs, rock and roll) that would easily discriminate between human and machine. Yet I do not see such questions illuminating issues in cognition. On the other hand, I also want to avoid the ``... Searlian mysteries about `intrinsic' vs. `derived' intentionality....'' Believe it or not, it is exactly these considerations that has led me to the information retrieval task domain. I did not motivate this well in my last message and would like to give it another try. \section{Semantics in information retrieval} First, let's do our best to imagine providing an artificial cognitive system (a robot) with the sort of grounding experience you and I both believe necessary to full cognition. Let's give it video eyes, microphone ears, feedback from its affectors, etc. And let's even give it something approaching the same amount of time in this environment that the developing child requires. I want to make two comments on this Gedanken experiment. First, the corpus of experience acquired by such a robot is orders of magnitude more complex than any system today. Second, there is no doubt that even such a complete system as this would have a radically different experience of the world than our own. In short, I simply mean to highlight the huge distance between the psycho-physical experience of any artificial system and any human. The communication barrier between the symbols of man and the symbols of machine to which I referred in my last message is a consequence of this distance. When we say ``apple'' I would expect the symbol in our heads to have almost no correspondence to the symbol ``apple'' in any computers. Since I see such a correspondence as a necessary precondition to the development of language, I am not hopeful that language between man and machine can develop in the same fashion as language develops within a species. So the question for me becomes: how might we give a machine the same rich corpus of experience (hence satisfying the total part of your TTT) without relying on such direct experiential contact with the world? The answer for me (at the moment) is to begin at the level of WORDS. I view the enormous textual databases of information retrieval (IR) systems as merely so many words. I want to take this huge set of ``labels,'' attached by humans to their world, as my primitive experiential database. The task facing my system, then, is to look at and learn from this world. This experience actually has two components. The textbase itself provides the first source of information, viz., how authors use and juxtapose words. The second, ongoing source of experience are the interactions with IR users, in which people use these same words and then react positively or negatively to my systems interpretation of those words. The system then adapts its (connectionist) representation of the words and documents so as to reflect what the consensus of its users indicate by these words. In short, I am using the original authors and the browsing users as the systems ``eyes'' into the human world. I am curious to see what structural relationship arise among these words, via low level connectionist learning procedures, to facilitate access to the IR database.