A Mechanistic Account of Computational Explanation in Cognitive Science
Marcin Miłkowski · Cognitive Science · 2013
A Mechanistic Account of Computational Explanation in Cognitive Science Marcin Milkowski ([email protected]) Institute of Philosophy and Sociology, Polish Academy of Sciences ul. Nowy Świat 72, 00-330 Warsaw, Poland Abstract Explanations in cognitive science rely predominantly on computational modeling. Though the scientific practice is systematic, and there is little doubt about the empirical value of numerous models, the methodological account of computational explanation is not up-to-date. The current paper offers a systematic account of computational explanation in cognitive science in a largely mechanistic framework. The account is illustrated with a short case study of modeling of the mirror neuron system in terms of predictive coding. Keywords: computation; computational modeling; explanation; mechanism; levels; information-processing. Importance of Computational Modeling Computational modeling plays a special role in contemporary cognitive science; over 80 percent of articles in theoretical journals focus on computational 1 models (Busemeyer & Diederich, 2010). The now dominating methodology forcefully defended by (Marr, 1982) has turned out to be fruitful. At the same time, the three-level account of Marr is not without problems. In particular, the relationship among the levels is interpreted in various ways, wherein the change of level is both the shift of grain and the shift of the boundary of the system under explanation (McClamrock, 1991); it is not at all clear what is the proper relation between competence and its realization or whether bottom-up modeling is entirely mistaken; and, last but least, whether one model should answer how, what and why questions related to the explanandum. My goal in this paper is to offer a descriptive account, which is close in spirit to the recent developments in the theory of mechanistic explanation (Bechtel, 2008; Craver, 2007; Glennan, 2002; Machamer, Darden, & Craver, 2000). According to mechanism, to explain a phenomenon is to explain the underlying mechanism. Mechanistic explanation is a species of causal explanation, and explaining a mechanism involves the discovery of its causal structure. While mechanisms are defined variously, the core idea is that they are organized systems, comprising causally relevant component parts and operations (or activities) thereof. Parts of the mechanism interact and their orchestrated operation contributes to the capacity of the mechanism. Mechanistic explanations abound in special sciences and it is hoped that the adequate description of the principles implied in explanations generally accepted as sound will furnish researchers also with normative guidance. I am not using the word ‘computational’ here in the sense used by Marr to define one of the levels in his account. The claim that computational explanation is best understood as mechanistic gains popularity (Piccinini, 2007), and I have defended it at length against skeptical doubt elsewhere (Milkowski, 2013). Here, I wish to succinctly summarize the account and, more importantly, add some crucial detail to the overall mechanistic framework proposed earlier. I cannot discuss Marr’s theory in detail here (but see (Milkowski, 2013, pp. 114–121)) and it is used only for illustration purposes. My remarks below are not meant to imply a wholesale rejection of his largely successful methodology. Marr’s account did not involve any theory of how computation is physically realized, and it is compatible with a number of different accounts. I will assume a structural account of computational realization here, defended also by Piccinini (2008) and Chalmers (2011). For an extended argument, see also (Milkowski, 2011, 2013). One particular claim that is usually connected with the computational theory of mind is that the psychologically relevant computation is over mental representation, which leads to the language of thought hypothesis (Fodor, 1975). Here, no theory of mental representation is presupposed in the account of computation, one of the reasons being that representation is one of the most contentious issues in contemporary cognitive science. As the present account is intended to be descriptively adequate, assuming one particular theory of representation as implied by computation would make other accounts immediately non- computational, which is absurd. Another reason is that mechanistic accounts of computation do not need to presuppose representation (Fresco, 2010; Piccinini, 2006), though they do not exclude the representational character of some of the information being processed. In other words, it is claimed that only the notion of information (in the information-theoretic sense, not in the semantic sense, which is controversial) is implied by the notion of computation (or information-processing). Explanandum phenomenon Marr stressed the importance of specifying exactly what the model was supposed to explain. Specifying the explanandum phenomenon is critical also for the mechanistic framework, as several general norms of mechanistic explanation are related to the specification of the capacity of the mechanism. All mechanisms posited in explanations have an explanatory purpose, and for this reason, their specification is related to our epistemic interest. For the same reason, the boundaries of the mechanism, though not entirely arbitrary, can be carved in different ways depending on what one wishes to explain.