Page 469 - Deep Learning

452 Notes to Pages 364–366

see Jones, 2004, for a generally positive assessment of reductionism in science,
and Weinberg, 2001, for a persistent defense of reductionism in physics.) A non-
reductionist approach to unification is to conceptualize it as a gradual move-
ment toward substantive theories of wider and wider scope (Friedman, 1974). If
a theory of phenomenon X turns out to be sufficient to explain phenomenon Y
as well, then that theory can be said to unify X and Y. Kitcher (1988) calls this
explanatory unification. The programme of creating a unified theory of cogni-
tion by specifying the cognitive architecture (Newell, 1990) encompasses both
explanatory unification, in that one and the same set of principles are supposed
to explain a wide range, or even all, cognitive phenomena, and also reduction-
ism, in that higher-order cognitive phenomena are supposed to be explained in
terms of more basic cognitive processes.
8. See, e.g., Casper and Noer (1972) for an exposition of the synthesis of terrestrial and
celestial mechanics in the theory of gravitation. Kitcher (1988) presents a view of
unification as the use of a small number of explanation patterns to explain a large
number of diverse phenomena, using the Newtonian and Darwinian revolutions as
his primary examples. Evolutionary biologist Theodosius Dobzhansky emphasized
the unifying power of Darwinism: “Nothing in biology makes sense except in the
light of evolution” (Dobzhansky, 1973, p. 125). James Clerk Maxwell unified electric-
ity, radio waves and, eventually, light into a general field theory (Nersessian, 1992,
2002, 2008). The unification of medicine under the germ theory (Waller, 2002) is
only partial, given that there are other causes of illness than germs.
9. See Note 8 in Chapter 2.
10. For example, learning to categorize a set of stimuli might be analyzed into basic
processes like noticing a feature, strengthening or weakening a link between a
feature and a category, storing an instance in memory and so on, while decision
making might be analyzed into some combination of generating options, pre-
dicting outcomes and assessing and comparing expected outcomes, and problem
solving might be analyzed into goal setting, operator selection, operator execu-
tion, outcome evaluation and so on. Examples of this two-level style of theoriz-
ing are easy to find on the pages of psychological research journals such as the
Journal of Experimental Psychology: Learning, Memory, and Cognition, Cognitive
Psychology, Memory and Cognition and others.
11. Newell (1990, Figure 3–3, p. 122).
12. It is easy to forget exactly how ubiquitous flow diagrams were in cognitive
psychology in the 1950–1970 period. They were used to state models of atten-
tion (Broadbent, 1958, Fig. 7, p. 299), concept learning (Hunt, 1962, Figure 8–3,
p. 232), long-term memory retrieval (Shiffrin, 1970, Figure 2, p. 382), problem solv-
ing (Ernst & Newell, 1969, Figure 11, p. 42) and many other cognitive processes.
They are still with us (e.g., Alberdi, Sleeman & Korpi, 2000, Figure 1, p. 80).
13. See Newell (1972, 1973) for the basic concept. Anderson (1983) is the source for
the term “cognitive architecture.”
14. The best source for the Soar simulation system is the collection of articles edited
by Rosenbloom, Laird and Newell (1993, vol. 1 and 2). The ACT project has been
presented to the computing public in a series of books as well as numerous arti-
cles; see Anderson (1983, 1990, 1993, 2007) and Anderson and Lebiere (1998).