Page 421 - Deep Learning

404 Notes to Pages 34–37

statistical analysis of large numbers of test scores. J. B. Carroll (1993) has summa-
rized the psychometric analysis of mind and its empirical support. H. Gardner’s
theory of multiple intelligences is a later development of the same concept
(Gardner, 1983). For Fodor (1983), a cognitive module is a processing unit with
a specialized purpose and a certain independence from the rest of the cognitive
system. Modules receive inputs from and deliver their results to other compo-
nents of the cognitive system, but their internal processing is isolated from the
latter. Abilities and modules differ in several respects, but they share the assump-
tion that cognition is accomplished through a set of components that operate
relatively independently from each other.
28. Analysis by decomposition into functions needs to be distinguished from what is
often termed functional explanation or functional analysis (Hempel, 1959; Nagel,
1961, Chap. 12). The latter is said to be of special importance in the biological as
opposed to the physical sciences (but see Hanke, 2004, for a biologist’s point of
view). A prototypical target of analysis is the statement that the function of the
heart is to pump blood. The goal of philosophical analysis as usually stated is to
grasp why functions are explanatory even though they are not causal, and a typ-
ical answer is that they are explanatory precisely to the extent that they can be
reduced to some type of causal account. Accepting that explanations that refer
to a function or a purpose are irreducible and nevertheless rigorous and scien-
tific was an important step in the cognitive revolution (Boden, 1972; Cummins,
1975, 1983a). Philosophers continue to struggle with the concept of functional
explanation (Davies, 2003; Griffiths, 1993; McLaughlin, 2001) and there is now
even a psychology of functional explanation (Lambrozo & Carey, 2006). For pre-
sent purposes, the meaning of “function” as used in the discussion of functional
explanation is not the same as its meaning in discussions of functional decom-
position. The latter ultimately derives from the mathematical rather than the bio-
logical concept of a function: A function maps inputs to outputs and is defined
by a set of computations that produce the correct output, given any one input.
In the functional decomposition of the cognitive system, functions like perceiv-
ing, remembering and learning are not explanatory but targets for explanation; in
philosophical parlance, they are explananda rather than explanans. The explan-
atory task is to demonstrate how a cognitive function is or can be implemented,
i.e., to specify the processes that compute that function down to some desired
level of detail (Ohlsson, 2007c). This is the sense in which I use the term “func-
tion” in this book. It is more closely related to the psychological movement called
functionalism in the first half of the 20th century (Schultz, 1969, Chaps. 6–8) than
to philosophical analyses of functional explanations in biology.
29. For example, a skilled driver might appear to be able to drive and talk on his cell
phone at the same time, but Strayer and Johnson (2001) showed that this is not
so: The phone conversation affects reaction time when braking and the probabil-
ity of overlooking traffic signs, indicating that this dual task situation is handled
by switching attention back and forth between the two tasks.
30. See, e.g., Alvarez and Emory (2006), Baddeley (1996) and Shanahan and Baars
(2005). The term “cognitive architecture” originated with Anderson (1983).