Cognitive Science should be unified: comment on Griffiths et al. and McClelland et al.

The present paradigm debate [1xProbabilistic models of cognition: exploring representations and inductive biases. Griffiths, T.L. et al. Trends Cogn. Sci. 2010; 14: 357–364Abstract | Full Text | Full Text PDF | PubMed | Scopus (113)See all References, 2xLetting structure emerge: connectionist and dynamical systems approaches to cognition. McClelland, J.L. et al. Trends Cogn. Sci. 2010; 14: 348–356Abstract | Full Text | Full Text PDF | PubMed | Scopus (110)See all References] follows a long line of counter-productive clashes such as nature/nurture, symbolic/sub-symbolic, local/distributed. Such battles are particularly troubling for cognitive science, given that the motivation for the field was to combine disciplines and methodologies [3xUnified Theories of Cognition. Newell, A. See all References][3].My background includes over 40 years building Bayesian models and over 30 years on connectionist systems [4xEcological expected utility and the mythical neural code. Feldman, J.A. Cogn. Neurodyn. 2010; 4: 25–35Crossref | PubMed | Scopus (11)See all References][4]. Both fields encompass many approaches, but this debate focuses on pure Bayesian modeling versus Parallel Distributed Processing (PDP)-style layered networks trained with back-propagation [5xSee all References][5].The position papers and counter-arguments also exhibit several specific problems. For starters, both papers, despite being framed as addressing the very core questions of cognitive science, actually discuss only induction. Both groups derive from mathematical psychology where the tradition is fitting a mathematical model to data from one narrow range of experiments with no attempt to relate the specific model to other known constraints on cognition. From this perspective, the cultural clash is over whether the curve fitting should be largely based on analysis or allowed to ‘emerge’ from fitting data with a general model having an enormous numbers of parametersCultural purity is stronger in the back-propagation community [2xLetting structure emerge: connectionist and dynamical systems approaches to cognition. McClelland, J.L. et al. Trends Cogn. Sci. 2010; 14: 348–356Abstract | Full Text | Full Text PDF | PubMed | Scopus (110)See all References][2] than in the Bayesian group. Their emphasis on eliminative emergence is instructive. There are two, often conflated, notions of emergence. A bicycle is an emergent system, having properties unlike those of its constituents. But no one suggests that a box of bicycle parts will self-organize. Animal intelligence is the result of enormously complex evolutionary and developmental forces that bear little resemblance to back-propagation parameter fitting. The Bayesian group [1xProbabilistic models of cognition: exploring representations and inductive biases. Griffiths, T.L. et al. Trends Cogn. Sci. 2010; 14: 357–364Abstract | Full Text | Full Text PDF | PubMed | Scopus (113)See all References][1] errs in the opposite direction: co-opting all statistical approaches to language, etc., most of which explicitly forego any cognitive pretensions.The early back-propagation research [5xSee all References][5] performed an invaluable service to cognitive science by showing that even simple layered networks could be trained to model complex cognitive functions. This effectively eliminated a priori nativist arguments for anyone willing to observe the results.The greatest problem is that both sides fail to acknowledge any of the biologically based bottom-up research and integrative efforts. Within the vast enterprise of cognitive neuroscience, there are entire literatures on neuroeconomics [6xSee all References][6] and beautiful integrative work on biological Hebbian and reinforcement learning [7xDecision theory, reinforcement learning, and the brain. Dayan, P. and Daw, N.D. Cogn. Affect. Behav. Neurosci. 2008; 8: 429–453Crossref | PubMed | Scopus (142)See all References][7], both of which are much more biologically plausible than back-propagation.There is a great deal of integrative cognitive science that makes use of probabilistic and adaptive techniques as well as many others. Vision has always been more unified than other areas of cognitive science [8xVision Science: Photons to Phenomenology. Palmer, S.E. See all References][8]. As one example, Serre et al. [9xA feedforward architecture accounts for rapid categorization. Serre, T. et al. Proc. Natl. Acad. Sci. U. S. A. 2007; 104: 6424–6429Crossref | PubMed | Scopus (355)See all References][9] built a structured, neurally plausible model of rapid feed-forward vision, trained it, and compared it in detail with Thorpe group's remarkable findings on human performance [10xNeurons tune to the earliest spikes through STDP source. Guyonneau, R. et al. Neural Comput. 2005; 17: 859–879Crossref | PubMed | Scopus (65)See all References][10].The only sensible rationale for cognitive science is unified theory and it will come sooner if everyone keeps this goal in mind.