Page 465 - Deep Learning

448 Notes to Pages 332–335

4. See Evans (2007), Chapter 4, for a review of the evidence regarding belief bias
from laboratory studies in cognitive psychology. Belief bias is real enough in
everyday contexts as well; see, e.g., Lynn and Williams (1990).
5. The concept of cognitive utility was discussed in Anderson (1990), especially
Chapter 5, and it plays a central role in the ACT-R theory (Anderson, 2007).
6. Cognitive load encompasses both the amount of information that needs to be
kept active at each moment in time and the amount of cognitive resources allo-
cated to processing that information. It corresponds to working memory load, as
measured by, for example, the Operation Span Test (see, e.g., Klein & Fiss, 1999;
Turner & Engle, 1989) or Working Memory Span (Baddeley, 2007). Sweller and
others have shown that high cognitive load can interfere with learning (Chandler
& Sweller, 1991; Sweller, 1988).
7. Evidence of this comes from a variety of empirical studies of students’ physics
reasoning (Clement, 1982; McCloskey, 1983; Robin & Ohlsson, 1989).
8. Miller (1996, pp. 136–138).
9. Anderson (1984), Goldstein (2008) and Markman (1999, Chapter 4).
10. Rokeach (1960, 1970).
11. The concept of subsumption has two historical roots, each relevant to this chap-
ter albeit in different ways. The earliest is Hempel and Oppenheim’s (1948) phil-
osophical proposal that a scientific explanation explains by subsuming particular
events under general scientific laws. This proposal has been much criticized and
is no longer considered a serious philosophical theory of explanation in the nat-
ural sciences (Salmon, 1989). Cummins (1983a, 1983b) has argued that it does
not apply to psychology. Although bad philosophy, the subsumption concept
might nevertheless be good psychology. Subsumption is central to the theory
of meaningful verbal learning proposed by David Ausubel (1963), especially pp.
24–26, and Chapter 4; see also Ausubel (1968). The key idea is that to learn is
to subsume new subject matter content under a prior conception, which sug-
gests that learning can be enhanced by providing that prior conception ahead
of time in the form of a so-called advanced organizer. This proposal obviously
suffers from a circularity problem: How is the learner to absorb the advanced
organizer? If he needs an organizer for the organizer, learning is in trouble.
Nevertheless, empirical studies tend to support the utility of advanced organizers
(Corkill, 1992; Dole, Valencia, Greer & Waldrop, 1991; Luiten, Ames & Ackerson,
1980; Mayer, 1979; Stone, 1983) although there are some exceptions (Barnes &
Clawson, 1975; Calandra & Barron, 2005; McDade, 1978). The subsumption the-
ory of learning was generalized beyond advanced organizers into the Elaboration
Theory of instructional design (Reigeluth & Stein, 1983; Van Patten, Chao &
Reigeluth, 1986), which says that learning is a process of progressive differentiat-
ing, elaborating and fleshing out initial concepts. The proposal in this chapter is
that subsumption is the relation that structures informal belief systems. This is a
hypothesis about representation, not about process. The Resubsumption Theory
is not committed to any specific claim about superior learning and retention in
the presence of an advanced organizer, because it is not committed to any partic-
ular theory of monotonic, routine belief formation.