Title
Bert Vaux (Cambridge): Disputation on the Power and Efficacy of Reductionism, ...

CASTL colloquium

Date:  30.11.2006

Time:  19:15

Venue:  E010 1

Description
"Disputation on the Power and Efficacy of Reductionism, with special reference to Language"

Gallistel 2006 observes that theories of cognition fall into two categories, those in which animals learn *to* do something and those in which animals learn *that* something is so. The former type of theory, in which animal brains are essentially recurrent switch networks, includes connectionist and related reductionist models such as Optimality Theory. The latter type of theory, in which the mind can be understood as a Turing Machine, includes rationalist models such as classical Chomskyan generative phonology.

Cognitive scientists disagree with respect to how to decide between these two sorts of theory:
i. Reductionists argue that we know that the brain contains a network of interconnected neurons, and that by Occam's Razor a "learning that" theory which mirrors what we know about the physical interactions between these neurons is preferable to theories that tack on additional cognitive mechanisms. (Rumelhart and McClelland, Prince and Smolensky, Bybee)
ii. Some computational scientists and linguists argue that the two sorts of theory are extensionally equivalent. (Karttunen)
iii. Rationalists identify predictive differences between the two theories and argue that a "learning to" model better captures what we know to be possible and impossible phenomena. (Gallistel, Idsardi, Vaux)

In this talk I support (iii) by identifying a number of phonological and morphological phenomena with respect to which, contra (ii), reductionist and rationalist models make distinct predictions, and for which, contra (i), a rationalist rule-based derivational model is both necessary and sufficient, whereas a reductionist constraint-based model is not.

Of particular interest in this investigation is the extraction by animals of generalisations about real-world contingencies in the form of rules that make predictions beyond the minimal scope required by
the data to which the animal has been exposed. More specifically I argue on the basis of priming and productivity studies, language games, morpheme structure constraints, free rides, and crazy rules
that, contrary to what is predicted by Optimality Theory and its attendant learning mechanisms Recursive Constraint Demotion, Richness of the Base, and Lexicon Optimisation, humans extract broad and
sometimes "crazy" language-specific generalisations from limited, static, and/or idiosyncratic data sets.

I then ground the form that rule-based learning appears to take in a larger information-theoretic model of animal cognition developed by Randy Gallistel, which I suggest has its roots in a particular evolutionary model of language and cognition.