Online research in philosophy

What language allows us to do is to "steal" categories quickly and effortlessly through hearsay instead of having to earn them the hard way, through risky and time-consuming sensorimotor "toil" (trial-and-error learning, guided by corrective feedback from the consequences of miscategorisation). To make such linguistic "theft" possible, however, some, at least, of the denoting symbols of language must first be grounded in categories that have been earned through sensorimotor toil (or else in categories that have already been "prepared" for us through Darwinian theft by the genes of our ancestors); it cannot be linguistic theft all the way down. The symbols that denote categories must be grounded in the capacity to sort, label and interact with the proximal sensorimotor projections of their distal category-members in a way that coheres systematically with their semantic interpretations, both for individual symbols, and for symbols strung together to express truth-value-bearing propositions.

1.1 The predominant approach to cognitive modeling is still what has come to be called "computationalism" (Dietrich 1990, Harnad 1990b), the hypothesis that cognition is computation. The more recent rival approach is "connectionism" (Hanson & Burr 1990, McClelland & Rumelhart 1986), the hypothesis that cognition is a dynamic pattern of connections and activations in a "neural net." Are computationalism and connectionism really deeply different from one another, and if so, should they compete for cognitive hegemony, or should they collaborate? These questions will be addressed here, in the context of an obstacle that is faced by computationalism (as well as by connectionism if it is either computational or seeks cognitive hegemony on its own): The symbol grounding problem (Harnad 1990).

Cognitive science is a form of "reverse engineering" (as Dennett has dubbed it). We are trying to explain the mind by building (or explaining the functional principles of) systems that have minds. A "Turing" hierarchy of empirical constraints can be applied to this task, from t1, toy models that capture only an arbitrary fragment of our performance capacity, to T2, the standard "pen-pal" Turing Test (total symbolic capacity), to T3, the Total Turing Test (total symbolic plus robotic capacity), to T4 (T3 plus internal [neuromolecular] indistinguishability). All scientific theories are underdetermined by data. What is the right level of empirical constraint for cognitive theory? I will argue that T2 is underconstrained (because of the Symbol Grounding Problem and Searle's Chinese Room Argument) and that T4 is overconstrained (because we don't know what neural data, if any, are relevant). T3 is the level at which we solve the "other minds" problem in everyday life, the one at which evolution operates (the Blind Watchmaker is no mind-reader either) and the one at which symbol systems can be grounded in the robotic capacity to name and manipulate the objects their symbols are about. I will illustrate this with a toy model for an important component of T3 -- categorization -- using neural nets that learn category invariance by "warping" similarity space the way it is warped in human categorical perception: within-category similarities are amplified and between-category similarities are attenuated. This analog "shape" constraint is the grounding inherited by the arbitrarily shaped symbol that names the category and by all the symbol combinations it enters into. No matter how tightly one constrains any such model, however, it will always be more underdetermined than normal scientific and engineering theory. This will remain the ineliminable legacy of the mind/body problem.

After people learn to sort objects into categories they see them differently. Members of the same category look more alike and members of different categories look more different. This phenomenon of within-category compression and between-category separation in similarity space is called categorical perception (CP). It is exhibited by human subjects, animals and neural net models. In backpropagation nets trained first to auto-associate 12 stimuli varying along a onedimensional continuum and then to sort them into 3 categories, CP arises as a natural side-effect because of four factors: (1) Maximal interstimulus separation in hidden-unit space during autoassociation learning, (2) movement toward linear separability during categorization learning, (3) inverse-distance repulsive force exerted by the between-category boundary, and (4) the modulating effects of input iconicity, especially in interpolating CP to untrained regions of the continuum. Once similarity space has been "warped" in this way, the compressed and separated "chunks" have symbolic labels which could then be combined into symbol strings that constitute propositions about objects. The meanings of such symbolic representations would be "grounded" in the system's capacity to pick out from their sensory projections the object categories that the propositions were about.

Harnad's main argument can be roughly summarised as follows: due to Searle's Chinese Room argument, symbol systems by themselves are insufficient to exhibit cognition, because the symbols are not grounded in the real world, hence without meaning. However, a symbol system that is connected to the real world through transducers receiving sensory data, with neural nets translating these data into sensory categories, would not be subject to the Chinese Room argument. Harnad's article is not only the starting point for the present debate, but is also a contribution to a longlasting discussion about such questions as: Can a computer think? If yes, would this be solely by virtue of its program? Is the Turing Test appropriate for deciding whether a computer thinks?

Connectionism and computationalism are currently vying for hegemony in cognitive modeling. At first glance the opposition seems incoherent, because connectionism is itself computational, but the form of computationalism that has been the prime candidate for encoding the "language of thought" has been symbolic computationalism (Dietrich 1990, Fodor 1975, Harnad 1990c; Newell 1980; Pylyshyn 1984), whereas connectionism is nonsymbolic (Fodor & Pylyshyn 1988, or, as some have hopefully dubbed it, "subsymbolic" Smolensky 1988). This paper will examine what is and is not a symbol system. A hybrid nonsymbolic/symbolic system will be sketched in which the meanings of the symbols are grounded bottom-up in the system's capacity to discriminate and identify the objects they refer to. Neural nets are one possible mechanism for learning the invariants in the analog sensory projection on which successful categorization is based. "Categorical perception" (Harnad 1987a), in which similarity space is "warped" in the service of categorization, turns out to be exhibited by both people and nets, and may mediate the constraints exerted by the analog world of objects on the formal world of symbols.

In this commentary on Harnad's "Grounding Symbols in the Analog World with Neural Nets: A Hybrid Model," the issues of symbol grounding and analog (continuous) computation are separated, it is argued that symbol graounding is as important an issue for analog cognitive models as for digital (discrete) models. The similarities and differences between continuous and discrete computation are discussed, as well as the grounding of continuous representations. A continuous analog of the Chinese Room is presented.

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 (arbitrary) 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: (1) iconic representations, which are analogs of the proximal sensory projections of distal objects and events, and (2) 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 (nonsymbolic) categorical representations. Higher-order (3) symbolic representations, grounded in these elementary symbols, consist of symbol strings describing category membership relations (e.g., An X is a Y that is Z). 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.

Computation is interpretable symbol manipulation. Symbols are objects that are manipulated on the basis of rules operating only on theirshapes, which are arbitrary in relation to what they can be interpreted as meaning. Even if one accepts the Church/Turing Thesis that computation is unique, universal and very near omnipotent, not everything is a computer, because not everything can be given a systematic interpretation; and certainly everything can''t be givenevery systematic interpretation. But even after computers and computation have been successfully distinguished from other kinds of things, mental states will not just be the implementations of the right symbol systems, because of the symbol grounding problem: The interpretation of a symbol system is not intrinsic to the system; it is projected onto it by the interpreter. This is not true of our thoughts. We must accordingly be more than just computers. My guess is that the meanings of our symbols are grounded in the substrate of our robotic capacity to interact with that real world of objects, events and states of affairs that our symbols are systematically interpretable as being about.

It is unlikely that the systematic, compositional properties of formal symbol systems -- i.e., of computation -- play no role at all in cognition. However, it is equally unlikely that cognition is just computation, because of the symbol grounding problem (Harnad 1990): The symbols in a symbol system are systematically interpretable, by external interpreters, as meaning something, and that is a remarkable and powerful property of symbol systems. Cognition (i.e., thinking), has this property too: Our thoughts are systematically interpretable by external interpreters as meaning something. However, unlike symbols in symbol systems, thoughts mean what they mean autonomously: Their meaning does not consist of or depend on anyone making or being able to make any external interpretations of them at all. When I think "the cat is on the mat," the meaning of that thought is autonomous; it does not depend on YOUR being able to interpret it as meaning that (even though you could interpret it that way, and you would be right).