Stevan Harnad

There has been much discussion recently about the scope and limits of purely symbolic models of the mind and about the proper role of connectionism in cognitive modeling. This paper describes the symbol grounding problem : How can the semantic interpretation of a formal symbol system be made intrinsic to the system, rather than just parasitic on the meanings in our heads? How can the meanings of the meaningless symbol tokens, manipulated solely on the basis of their shapes, be grounded in anything but other meaningless symbols? The problem is analogous to trying to learn Chinese from a Chinese/Chinese dictionary alone. A candidate solution is sketched: Symbolic representations must be grounded bottom-up in nonsymbolic representations of two kinds: iconic representations, which are analogs of the proximal sensory projections of distal objects and events, and categorical representations, which are learned and innate feature-detectors that pick out the invariant features of object and event categories from their sensory projections. Elementary symbols are the names of these object and event categories, assigned on the basis of their categorical representations. Higher-order symbolic representations, grounded in these elementary symbols, consist of symbol strings describing category membership relations. Connectionism is one natural candidate for the mechanism that learns the invariant features underlying categorical representations, thereby connecting names to the proximal projections of the distal objects they stand for. In this way connectionism can be seen as a complementary component in a hybrid nonsymbolic/symbolic model of the mind, rather than a rival to purely symbolic modeling. Such a hybrid model would not have an autonomous symbolic module, however; the symbolic functions would emerge as an intrinsically dedicated symbol system as a consequence of the bottom-up grounding of categories ' names in their sensory representations. Symbol manipulation would be governed not just by the arbitrary shapes of the symbol tokens, but by the nonarbitrary shapes of the icons and category invariants in which they are grounded

We are accustomed to thinking that a primrose is "concrete" and a prime number is "abstract," that "roundness" is more abstract than "round," and that "property" is more abstract than "roundness." In reality, the relation between "abstract" and "concrete" is more like the (non)relation between "abstract" and "concave," "concrete" being a sensory term [about what something feels like] and "abstract" being a functional term (about what the sensorimotor system is doing with its input in order to produce its output): Feelings and things are correlated, but otherwise incommensurable. Everything that any sensorimotor system such as ourselves manages to categorize successfully is based on abstracting sensorimotor "affordances" (invariant features). The rest is merely a question of what inputs we can and do categorize, and what we must abstract from the particulars of each sensorimotor interaction in order to be able to categorize them correctly. To categorize, in other words, is to abstract. And not to categorize is merely to experience. Borges's Funes the Memorious, with his infinite, infallible rote memory, is a fictional hint at what it would be like not to be able to categorize, not to be able to selectively forget and ignore most of our input by abstracting only its reliably recurrent invariants. But a sensorimotor system like Funes would not really be viable, for if something along those lines did exist, it could not categorize recurrent objects, events or states, hence it could have no language, private or public, and could at most only feel, not function adaptively (hence survive). Luria's "S" in "The Mind of a Mnemonist" is a real-life approximation whose difficulties in conceptualizing were directly proportional to his difficulties in selectively forgetting and ignoring. Watanabe's "Ugly Duckling Theorem" shows how, if we did not selectively weight some properties more heavily than others, everything would be equally (and infinitely and indifferently) similar to everything else. Miller's "Magical Number Seven Plus or Minus Two" shows that there are (and must be) limitations on our capacity to process and remember information, both in our capacity to discriminate relatively (detect sameness/difference, degree-of-similarity) and in our capacity to discriminate absolutely (identify, categorize, name), The phenomenon of categorical perception shows how selective feature-detection puts a Whorfian "warp" on our feelings of similarity in the service of categorization, compressing within-category similarities and expanding between-category differences by abstracting and selectively filtering inputs through their invariant features, thereby allowing us to sort and name things reliably. Language does allow us to acquire categories indirectly through symbolic description....

William Gardner's proposal to establish a searchable, retrievable electronic archive is fine, as far as it goes. The potential role of electronic networks in scientific publication, however, goes far beyond providing searchable electronic archives for electronic journals. The whole process of scholarly communication is currently undergoing a revolution comparable to the one occasioned by the invention of printing. On the brink of intellectual perestroika is that vast PREPUBLICATION phase of scientific inquiry in which ideas and findings are discussed informally with colleagues, presented more formally in seminars, conferences and symposia, and distributed still more widely in the form of preprints and tech reports that have undergone various degrees of peer review. It has now become possible to do all of this in a remarkable new way that is not only incomparably more thorough and systematic in its distribution, potentially global in scale, and almost instantaneous in speed, but so unprecedentedly interactive that it will substantially restructure the pursuit of knowledge

Libet, Gleason, Wright, & Pearl (1983) asked participants to report the moment at which they freely decided to initiate a pre-specified movement, based on the position of a red marker on a clock. Using event-related potentials (ERPs), Libet found that the subjective feeling of deciding to perform a voluntary action came after the onset of the motor “readiness potential,” RP). This counterintuitive conclusion poses a challenge for the philosophical notion of free will. Faced with these findings, Libet (1985) proposed that conscious volitional control might operate as a selector and a controller of volitional processes rather than as an initiator of them.

Many special problems crop up when evolutionary theory turns, quite naturally, to the question of the adaptive value and causal role of consciousness in human and nonhuman organisms. One problem is that -- unless we are to be dualists, treating it as an independent nonphysical force -- consciousness could not have had an independent adaptive function of its own, over and above whatever behavioral and physiological functions it "supervenes" on, because evolution is completely blind to the difference between a conscious organism and a functionally equivalent (Turing Indistinguishable) nonconscious "Zombie" organism: In other words, the Blind Watchmaker, a functionalist if ever there was one, is no more a mind reader than we are. Hence Turing-Indistinguishability = Darwin-Indistinguishability. It by no means follows from this, however, that human behavior is therefore to be explained only by the push-pull dynamics of Zombie determinism, as dictated by calculations of "inclusive fitness" and "evolutionarily stable strategies." We are conscious, and, more important, that consciousness is piggy-backing somehow on the vast complex of unobservable internal activity -- call it "cognition" -- that is really responsible for generating all of our behavioral capacities. Hence, except in the palpable presence of the irrational (e.g., our sexual urges) where distal Darwinian factors still have some proximal sway, it is as sensible to seek a Darwinian rather than a cognitive explanation for most of our current behavior as it is to seek a cosmological rather than an engineering explanation of an automobile's behavior. Let evolutionary theory explain what shaped our cognitive capacity (Steklis & Harnad 1976; Harnad 1996, but let cognitive theory explain our resulting behavior

The causal structure of cognition can be simulated but not implemented computationally, just as the causal structure of a furnace can be simulated but not implemented computationally. Heating is a dynamical property, not a computational one. A computational simulation of a furnace cannot heat a real house (only a simulated house). It lacks the essential causal property of a furnace. This is obvious with computational furnaces. The only thing that allows us even to imagine that it is otherwise in the case of computational cognition is the fact that cognizing, unlike heating, is invisible (to eveyrone except the cognizer). Chalmers’s “Dancing Qualia” Argument is hence invalid: Even if there could be a computational model of cognition that was behaviorally indistinguishable from a real, feeling cognizer, it would still be true that if, like heat, feeling is a dynamical property of the brain, a flip-flop from the presence to the absence of feeling would be undetectable anywhere along Chalmers’s hypothetical component-swapping continuum from a human cognizer to a computational cognizer -- undetectable to everyone except the cognizer. But that would only be because the cognizer was locked into being incapable of doing anything to settle the matter simply because of Chalmers’s premise of input/output indistinguishability. That is not a demonstration that cognition is computation; it is just the demonstation that you get out of a premise what you put into it. But even if the causal topography of feeling, hence of cognizing, is dynamic rather than just computational, the problem of explaining the causal role played by feeling itself – how and why we feel – in the generation of our behavioral capacity – how and why we can do what we can do – will remain a “hard” (and perhaps insoluble) problem

It is hypothesized that words originated as the names of perceptual categories and that two forms of representation underlying perceptual categorization -- iconic and categorical representations -- served to ground a third, symbolic, form of representation. The third form of representation made it possible to name and describe our environment, chiefly in terms of categories, their memberships, and their invariant features. Symbolic representations can be shared because they are intertranslatable. Both categorization and translation are approximate rather than exact, but the approximation can be made as close as we wish. This is the central property of that universal mechanism for sharing descriptions that we call natural language.

Both Artificial Life and Artificial Mind are branches of what Dennett has called "reverse engineering": Ordinary engineering attempts to build systems to meet certain functional specifications, reverse bioengineering attempts to understand how systems that have already been built by the Blind Watchmaker work. Computational modelling (virtual life) can capture the formal principles of life, perhaps predict and explain it completely, but it can no more be alive than a virtual forest fire can be hot. In itself, a computational model is just an ungrounded symbol system; no matter how closely it matches the properties of what is being modelled, it matches them only formally, with the mediation of an interpretation. Synthetic life is not open to this objection, but it is still an open question how close a functional equivalence is needed in order to capture life. Close enough to fool the Blind Watchmaker is probably close enough, but would that require molecular indistinguishability, and if so, do we really need to go that far?

Europe is losing almost 50% of the potential return on its research investment until research funders and institutions mandate that all research findings must be made freely accessible to all would be users, webwide. It is not the number of articles published that reflects the return on Europe's research investment: A piece of research, if it is worth funding and doing at all, must not only be published, but used, applied and built upon by other researchers, worldwide. This is called 'research impact' and a measure of it is the number of times an article is cited by other articles ('citation impact').

Almost all words are the names of categories. We can learn most of our words (and hence our categories) from dictionary definitions, but not all of them. Some have to be learned from direct experience. To understand a word from its definition we need to already understand the words used in the definition. This is the “Symbol Grounding Problem” [1]. How many words (and which ones) do we need to ground directly in sensorimotor experience in order to be able to learn all other words via definition alone? The answer may shed some light both on the developmental origin of word meanings and on the evolutionary origin and adaptive value of language. We used an algorithm to reduce each of our dictionaries (Longmans LDOCE, Cambridge CIDE and WordNet) to its “grounding kernel” (“Kernel”) (which turned out to be about 10% of the dictionary) by systematically eliminating..

This is a paperback reissue of a 1988 special issue of Cognition - dated but still of interest. The book consists of three chapters, each making one major negative point about connectionism. Fodor & Pylyshyn (F&P) argue that connectionist networks (henceforth 'nets') are not good models for cognition because they lack 'systematicity', Pinker & Price (P&P) argue that nets are not good substitutes for rule-based models of linguistic ability, and Lachter & Bever (L&B) argue that nets can only model the associative relations between cognitive structures, not the structures themselves.

Behavioral scientists studied behavior; cognitive scientists study what generates behavior. Cognitive science is hence theoretical behaviorism (or behaviorism is experimental cognitivism). Behavior is data for a cognitive theorist. What counts as a theory of behavior? In this paper, a methodological constraint on theory construction -- "neoconstructivism" -- will be proposed (by analogy with constructivism in mathematics): Cognitive theory must be computable; given an encoding of the input to a behaving system, a theory must be able to compute (an encoding of) its outputs. It is a mistake to conclude, however, that this constraint requires cognitive theory to be computational, or that it follows from this that cognition is computation

I am going to attempt to argue, given certain premises, there are reasons, not only empirical, but also logical, for expecting a certain division of labor in the processing of information by the human brain. This division of labor consists specifically of a functional bifurcation into what may be called, to a first approximation, "verbal" and "nonverbal" modes of information- processing. That this dichotomy is not quite satisfactory, however, will be one of the principal conclusions of this chapter, for I shall attempt to show that metaphor, which in its most common guise is a literary, and hence a fortiori a "verbal" phenomenon, may in fact be more a function of the "nonverbal" than the "verbal" mode. (For alternative attempts to account for cognitive lateralization, see e.g. Bever, 1975; Wickelgren, 1975; Pendse, 1978.).

"Symbol Grounding" is beginning to mean too many things to too many people. My own construal has always been simple: Cognition cannot be just computation, because computation is just the systematically interpretable manipulation of meaningless symbols, whereas the meanings of my thoughts don't depend on their interpretability or interpretation by someone else. On pain of infinite regress, then, symbol meanings must be grounded in something other than just their interpretability if they are to be candidates for what is going on in our heads. Neural nets may be one way to ground the names of concrete objects and events in the capacity to categorize them (by learning the invariants in their sensorimotor projections). These grounded elementary symbols could then be combined into symbol strings expressing propositions about more abstract categories. Grounding does not equal meaning, however, and does not solve any philosophical problems

Explaining the mind by building machines with minds runs into the other-minds problem: How can we tell whether any body other than our own has a mind when the only way to know is by being the other body? In practice we all use some form of Turing Test: If it can do everything a body with a mind can do such that we can't tell them apart, we have no basis for doubting it has a mind. But what is "everything" a body with a mind can do? Turing's original "pen-pal" version (the TT) only tested linguistic capacity, but Searle has shown that a mindless symbol-manipulator could pass the TT undetected. The Total Turing Test (TTT) calls for all of our linguistic and robotic capacities; immune to Searle's argument, it suggests how to ground a symbol manipulating system in the capacity to pick out the objects its symbols refer to. No Turing Test, however, can guarantee that a body has a mind. Worse, nothing in the explanation of its successful performance requires a model to have a mind at all. Minds are hence very different from the unobservables of physics (e.g., superstrings); and Turing Testing, though essential for machine-modeling the mind, can really only yield an explanation of the body.

Turing's celebrated 1950 paper proposes a very general methodological criterion for modelling mental function: total functional equivalence and indistinguishability. His criterion gives rise to a hierarchy of Turing Tests, from subtotal ("toy") fragments of our functions (t1), to total symbolic (pen-pal) function (T2 -- the standard Turing Test), to total external sensorimotor (robotic) function (T3), to total internal microfunction (T4), to total indistinguishability in every empirically discernible respect (T5). This is a "reverse-engineering" hierarchy of (decreasing) empirical underdetermination of the theory by the data. Level t1 is clearly too underdetermined, T2 is vulnerable to a counterexample (Searle's Chinese Room Argument), and T4 and T5 are arbitrarily overdetermined. Hence T3 is the appropriate target level for cognitive science. When it is reached, however, there will still remain more unanswerable questions than when Physics reaches its Grand Unified Theory of Everything (GUTE), because of the mind/body problem and the other-minds problem, both of which are inherent in this empirical domain, even though Turing hardly mentions them.

Turing set the agenda for (what would eventually be called) the cognitive sciences. He said, essentially, that cognition is as cognition does (or, more accurately, as cognition is capable of doing): Explain the causal basis of cognitive capacity and you’ve explained cognition. Test your explanation by designing a machine that can do everything a normal human cognizer can do – and do it so veridically that human cognizers cannot tell its performance apart from a real human cognizer’s – and you really cannot ask for anything more. Or can you? Neither Turing modelling nor any other kind of computational r dynamical modelling will explain how or why cognizers feel

The ethical case for Open Access (OA) (free online access) to research findings is especially salient when it is public health that is being compromised by needless access restrictions. But the ethical imperative for OA is far more general: It applies to all scientific and scholarly research findings published in peer-reviewed journals. And peer-to-peer access is far more important than direct public access. Most research is funded so as to be conducted and published, by researchers, in order to be taken up, used, and built upon in further research and applications, again by researchers (pure and applied, including practitioners), for the benefit of the public that funded it – not in order to generate revenue for the peer-reviewed journal publishing industry (nor even because there is a burning public desire to read much of it). Hence OA needs to be mandated, by researchers' institutions and funders, for all research