LINGUIST List 3.127

Sun 09 Feb 1992

Confs: Conference on Cognition and Representation

Editor for this issue: <>




From: William J. Rapaport <rapaportcs.Buffalo.EDU>


			 April 3-5, 1992
	Center for Tomorrow, State University of New York at Buffalo


 "Connectionism, Compositionality, and the Explanation of Productivity"

 The most fundamental fact that any theory of human cognition must
 explain is the incredible productivity of human knowledge. Most
 obviously in language, but also in other cognitive domains, human
 knowledge supports a range of appropriate behaviors that is
 appropriately idealized as infinite. The strategy for explaining this
 fact that has virtually defined cognitive science since its inception
 has been to (1) analyze the behaviors in question as possessing
 combinatorial structure, and (2) to identify human knowledge as part of
 an internal causal mechanism that physically realizes this
 combinatorial system. Possessing a finite repertoire of basic
 constituents and the means to (recursively) compose them, this
 mechanism explains productivity. Connectionism can respond to this in
 at least the following three ways:
	 (1) Deny the problem.
	 (2) Implement the standard explanation.
	 (3) Offering a new explanation, possibly while revising the problem.
 Radical eliminativist connectionism takes option (1)--by default if
 not by explicit intent. Connectionist critics such as Fodor & Pylyshyn
 have advocated (2), and some connectionist research has also followed
 this option. In this talk, I will present results of my program of technical
 connectionist research based on option (3), and discuss its implications
 for the foundations of cognitive science.

 "Distributed Represenation--An Outline"

 What is distributed representation? This question is central to
 many practical and philosophical concerns, both in connectionism and in
 cognitive science more generally, yet it has never been given an answer
 that is both comprehensive and preceise. In this talk I propose a way
 of defining distribution that, on one hand, reveals the fundamental
 similarity between (for example) the gross functional neuroanatomy of
 the various brain areas and connectionist hidden unit activity patterns,
 while on the other is strict enough to yield mathematically precise
 descriptions in real modeling contexts. The key concept is that of
 semantic superimposition; I elaborate this suprisingly tricky concept,
 offer a formal framework rendering it precise, and explain how
 superimposition can be incorporated into a general definition of
 distributed representation.

 "Representing, Similarity, and the Storage of Information"

 Representing is an activity, a process through which a subject cognizes
 the world. Most theories of representation take one element or
 component of this activity and identify it as the representation by
 attributing to the element, in isolation, properties it has only in the
 context of the act of representing. In particular, I argue that the
 similarity of an icon to an object is neither necessary nor sufficient
 for representation and that seeing why this is so reveals fundamental
 defects with views that see representations as (1) stored information,
 which represents in virtue of an isomorphism effected by an information
 storage and retrieval algorithm; or (2) as a distributed pattern of
 activity over a set of units or phase space, which represents in virtue
 of a topological isomorphism to the represented object. All of these
 models identify the representation with some element of the cognitive
 process that cannot intrinsically represent apart from its situation in
 a wider context. I argue that such models of representation will fail
 to account for the cognitive role of representation as long as they
 mistake part of the representing process for the representating itself.

 "Primary and Mature Conceptual Structures--Evidence from Child Language"

 Currently, in semanic enquiry some researchers represent
 lexical-conceptual structure as an architecture of sets and truth
 conditions. Others borrow from psychology such terms as 'conceptual
 primitives', 'canonical' and 'marginal' structures, and 'prototypes'
 (e.g. Jackendoff, 1990).

 The documentation of usage of verb argument structures, may shed
 considerable light on this division. Differentials in production of
 argument structures by children over time may shed light on the
 construction of underlying representations. These differences over time
 appear to also have an effect on complexity of context sentences.
 Production of tense, mood and negation markers appears to be contingent
 on the stability of conceptual structures underlying production of
 arguments. The current study supports the view that childrens'
 conceptual structures can be characterized as prototypes, and that
 gradual extension of structures is contingent on transition to new

 Early on (1968), Fillmore advocated the need for distinct treatment of
 propositional and modal information. Our results show that the presence
 of modality in a sentence is contingent on high-frequency argument
 structures. If the development of modality (in Fillmore's sense) and
 propositional content are distinguishable but show this relation, how is
 it to be characterized?

 Jackendoff (1990) distinguishes between I-languages (internal,
 based on innate predispositions) and E-languages (input-dependent).
 He holds that truth-conditional semantics requires a theory of language
 as an abstract artifact extrinsic to speakers. A possible interpretation of
 our findings is that modality overlayed on propositional content may be the
 means whereby this abstract artifact is reconstructed as part of a speaker's
 internal representation. Truth-conditional semantics, then, involves
 modelling mature inferential processes, or mature representation. This
 possibility will be examined in light of actual developmental sequences, where
 modal forms can be shown to emerge gradually to condition verb meaning
 in ways at least reminiscent of model theoretic semantics.


 "Representing Grammatical Knowledge as a Repertory of Constructions"
 (abstract not available)


 "Inferring Semantic-Memory Structure from Behavioral and Electrophysiological

 Researchers have investigated the structure of semantic-memory
 representations by examining subjects' performance in tasks in which
 they must judge the truth of sentences relating familar categories
 Differences in time to verify various classes of sentences were initally
 interpreted in terms of characteristics of the semantic representations
 retrieved from memory. Subsequent investigators have reinterpreted
 these findings in terms of characteristics of the verification
 _processes_ operating on these representations, rather than in terms of
 the representations themselves. We have taken a different approach.
 Instead of inferring the nature of knowledge representations based on
 how people _use_ them (i.e., "behavioral" response date), we have been
 investigating electrical protentials in the brain during sentence
 verification. A certain component of these potentials seems to reflect
 the _access_ or _retrieval_ of the stored knowledge, and not the
 processes that use this information to judge truth. These
 access/retrieval mechanisms better reflect structural characteristics of
 semantic memory than do behavioral measures, yielding a different
 picture of semantic-memory structure.

 "Specifying Classifying Representational Systems: A Critique and Proposal for
 Cognitive Science"

 Much of the work in cognitive science presupposes a theory of
 representation complete with a classification scheme; a scheme which
 allow us to say that two representations are interestingly similar or
 interestingly different for particular purposes.
 It is argued that such a scheme needs to meet at least the following
 eight constraints:
 (i) It must be grounded in some intuitions or a discipline-specific theory;
 (ii) It must not beg the crucial questions;
 (iii) It must result in an interesting number of categories (i.e.,
 something other than a unit or infinite number);
 (iv) It must individuate on the basis of relevant/constitutive
 properties of symbol systems;
 (v) It must be readily applicable;
 (vi) It must be widely applicable;
 (vii) The distinctions must be detectable by our behavioral data and
 (viii) It must be compatible with the computational story of mind.

 The most widely used apparatus for classifying symbol systems is that of
 informational and computational equivalence. This is critiqued and
 found wanting on most accounts. A diagnosis of the problem is offered.
 Time permitting, some prescriptive suggestions will also be made.

 "Notions of Representation and the Diverging Interests of Philosophy and
 Empirical Science"

 Contemporary discussions of mental representation often seem to assume that
 there is a single sense of the word `representation' that (a) is applied
 univocally to such disparate objects as pictures, maps and symbols, (b) is
 utilized by empirical researchers in cognitive science, and (c) can readily be
 used to provide a philosophical account of intentionality. In fact, however,
 the notion of "representation" is paradigm-driven, and all of the familiar
 paradigms (symbols, etc.) are convention- or interpretation-dependent. This
 undercuts one philosophical strategy for explaining the content of mental
 states in representational terms. However, a non-conventional notion of
 "representation" as a theoretical term can be developed which seems to capture
 the empirical scientist's needs even if it does not explain the intentionality
 of mental states. This accords well with the following view of the importance
 of the computer paradigm: that what it provides is (i) a formalism for the
 mathematization of psychology and (ii) suggestive strategies for

 "Representation, Compositionality, and Analyticity"

 We claim that three principles, all of which have been widely espoused in the
 philosophy of language and in cognitive science, cannot be simultaneously
 satisfied. These are:

 1. That meaning is inferential role.
 2. That there is no analytic/synthetic distinction.
 3. That meaning is compositional.

 If 1-3 are not simultaneously satisfiable, at least one of them must be
 abandoned. We consider the question of which to give up, and of what the
 implications of doing so are likely to be.


 "Matching Logical Form to Linguistic Form"
 (abstract not available)

 "Disentangling Conceptual and Linguistic Knowledge"

 Language is the principal mediator of thought and one of the few vehicles with
 which to explore abstract conceptual structures. Cognitive psychologists use
 linguistic evidence to study psychological functions such as memory and
 categorization and to construct models of knowledge representation. However,
 researchers in cognitive psychology seldom address how to disentangle
 conceptual and linguistic knowledge. In fact, the field seems to assume that
 the two systems are isomorphic. The related field of neurology provides
 insights into the relationship between cognition and language. Findings from
 both split-brain and aphasic populations encourage a separation of linguistic
 and conceptual structures. Specifically, Antonio Damasio's neurological theory
 of convergence zones is presented and used to explain the differential
 cognitive and linguistic abilities of neurologically impaired individuals
 such as split-brain populations and aphasic populations. The paper analyuzes
 both simple concepts and complex conceptual structures known as scripts.

 "Natural Language and Thought"

 Hauser defends the proposition that our languages of thought are public
 languages. One group of arguments points to the coincidence of clearly
 productive thought with overt possession of recursive symbol systems. Another
 group relies on phenomenological experiences of mental discourse and making
 thoughts physical. A third group cites practical considerations, e.g. Occam's
 razor and the `streetlight principle' (look under the lamp) motivating looking
 for instantiations of outer languages in thought first.

 Abbott points to the literature and adduces a number of specific replies to
 Hauser. Examples of productive behavior showing thatnatural language is not
 necessary for productive thought includeproblem solving by chimpanzees, dreams,
 and feral human cases (Genie). On phenomenological and practical grounds,
 Abbott argues that communication of thoughts should be trivial if the inner
 language is the outer language, but it is not; the decryption analogy Hauser
 uses to apply the `streetlight principle' is flawed; and Occam's razor doesn't
 cut any ice with Mother Nature.


 "Behavioral and Computational Constraints in Human Shape Representation"

 Do visual representations use an object-centered or viewer-centered
 reference frame? Studies suggest that recognition is
 orientation-dependent under many circumstances. The resulting theory,
 Multiple-Views-Plus-Transformations, hypothesizes that recognition is
 achieved by using a mental transformation to match input shapes to
 object representations in a viewer-centered reference frame. Moreover,
 these representations are orientation-specific, e.g. "views", and are
 stored according to the frequency of occurrence of an object in a
 particular orientation. However, familiarity is not the only factor
 that determines represented views. First, there is evidence that views
 are contingent upon the frequency with which other objects appear at
 particular orientations. Representations of familiar objects in novel
 views may arise as a result of the frequent appearance of an object's
 visually similar cohorts. Second, there is evidence that views are
 contingent upon the geometry of an object. The likelihood of a
 representation arising increases with the distinctiveness of visible
 surfaces at each orientation -- novel orientations are likely to be
 represented to the extent that their geometry is unique, while
 orientations in which the geometry differs only slightly from that
 depicted in preexisting views are unlikely to be represented.

 "Brain Mechanisms for Perceptual Representation"

 The brain is designed on a pattern of converging and diverging fiber
 tracts with their associated broadly tuned receptive and responsive
 fields. This puts certain constraints on the processing, transmission
 and representation of information. The properties of perceptive fields
 and target cells are taken as neural analogs of cognitive and logical
 operations. Analysis and synthesis can then be carried out in
 parallel; and the neural representations of elementary constituents
 and categorical constructs can be grounded in associational programs
 of activity in groups of cells.

 "Is Perception for Real?"

 What is the relation between the "external" world and our
 conceptualization of this world? At one extreme an independent
 external reality is denied, whereas at the other, an external reality
 is a requirement for any conceptualization. Perception lies at the
 heart of this controversy: can our percepts really reflect (or
 approximate) the true structure of the world independent of our
 observations or not? To address this question we need a clear
 understanding of just what a percept is and what it entails. I offer
 one definition and provide support for this choice using examples from
 vision (Jepson & Richards 1991). For our percepts to be useful,
 enabling us to predict the consequences of events and actions, certain
 conditions must be met. Two I will discuss are (1) the ability to
 manipulate representations or internal models, and (2) criteria for
 data (observations) which generate reliable interpretations. This
 second condition imposes limitations on the scope of useful percepts,
 and shows that percepts (perhaps like scientific theories?) are
 critically dependent upon a matching of cognitive concepts to modal
 regularities in the world.
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue