Solutions to Fodor's Problem of Concept Acquisition

Solutions to Fodor’s Puzzle of Concept Acquisition Sourabh Niyogi (MIT, Co-Moderator), Jesse Snedeker (Harvard, Co-Moderator) Participants: Jerry Fodor (Rutgers), Dedre Gentner (Northwestern University) Alison Gopnik (University of California Berkeley), Frank Keil (Yale University) Stephen Laurence (University of Sheffield), Jean Mandler (University of Califonia San Diego) Eric Margolis (Rice University), James L. McClelland (Carnegie Mellon University) Timothy T. Rogers (University of Wisconsin) Viewpoint Abstracts Our goal is to engage workshop participants in a di- alectic on Fodor’s Puzzle of Concept Acquisition (Fodor 1975, Laurence and Margolis 2002), focusing on the fol- lowing interrelated questions: Baptism Jerry Fodor Despite various differences of formulation, all (nonbe- havioristic) accounts of concept learning, both in philos- ophy and in psychology, suppose that it exhibits the fa- miliar characteristics of nondemonstrative inferences. In particular, it requires a body of “data” specifying posi- tive and negative instances of the extension of the con- cept to be acquired; there must be a source of “hypothe- ses” specifying candidate identifications of that concept (i.e. candidate identifications of what the positive in- stances have in common), and some sort of “confirma- tion metric” that decides which of the candidates the data best support. The present issue concerns the vo- cabulary in which the hypotheses are formulated. On the one hand, on pain of circularity, it cannot itself be ac- quired by the process of concept acquisition in which it is exploited. On the other hand, it must be rich enough to specify a correct candidate hypothesis for each concept that the acquisition mechanism can learn. In traditional theorizing (e.g. in Hume) these conditions were recon- ciled by assuming that the primitive concepts in which candidate hypotheses are formulated are all innate, a conclusion that seems to be demanded if circularity is to be avoided. Traditional formulations generally assumed that most concepts can be identified in the vocabulary of a relatively small primitive basis. (Hume holds (i) that all the concepts that can be learned at all can be reduced to constructions out of sensory concepts and (ii) that the sensory concepts are themselves innate.) I take it that Hume was right about the logic of the situation; the assumption that the base concepts are innate is un- avoidable. But it now seems that he was wrong about the prospects for conceptual reductions; it now appears that very few concepts are reducible any others, sensory or otherwise. The required conclusion seems to be that our conceptual repertoire must be largely unlearned. This should be viewed not as a paradox but as a straightfor- ward inference from the principle that the mechanism of nondemonstrative inference is hypothesis formation and confirmation. • Can a person ever acquire a new conceptual primitive? If so, how? • How might a child expand the combinatorial expres- sive power of his or her representational system? • How might a child expand his or her hypothesis space of possible word meanings? This workshop will present and discuss admissibility of solutions from a diversity of perspectives: Development: Dedre Gentner Alison Gopnik Frank Keil Jean Mandler Jesse Snedeker Computation: James L. McClelland Sourabh Niyogi Timothy T. Rogers Philosophy: Jerry Fodor Stephen Laurence Eric Margolis We hope that the insights gained from studies of cogni- tive development (Mandler 1992, Gentner 2003, Mandler 2004, Keil 2003, Gopnik et al 2004) and computational models of learning (Rogers and McClelland 2003, Niyogi 2005) may be reconciled with the puzzle raised earlier and tested for admissibility (c.f. Piatelli-Palmarini 1980, Fodor 1998).