Computational Templates, Neural Network Dynamics, and Symbolic Logic

This paper looks at the relationship between subsymbolic neural networks and symbolic logical systems from a philosophy of science perspective. More specifically, the point of view is that of Paul Humphreys' philosophical account of the organization of scientific knowledge [Humphreys, Paul. Extending Ourselves -Computational Science, Empiricism and the Scientific Method. Oxford: Oxford University Press, 2004.] Humphreys considers the units of analysis constituting scientific knowledge (in computational science) to be computational models, and computational templates. Computational templates are abstract syntactic schemata underlying domain specific models. Humphreys' own examples are mainly from computational research in physics, biology and statistics. Neural networks research in cognitive science can also be fruitfully considered from this computational models/templates perspective, illustrated here by two examples from recent neural networks research.