Page 435 - Deep Learning

418 Notes to Pages 95–101

21. “… a fundamental feature of the anatomy of cortical pathways is a feedback of
information from higher order cortical areas to areas that are closer to the input
from the periphery” (Gilbert, 1998, p. 479). See Felleman and Van Essen (1991)
and Gilbert and Sigman (2007) for reviews of the relevant evidence. Researchers
are pinpointing the exact brain areas and mechanisms involved in such top-down
influences (see, e.g., Grent-’t-Jong & Woldorff, 2007; Ullman, 1995).
22. See Behrens (1987) for an overview of Albert Ames Jr.’s work on visual illusions,
and Levine (2000, pp. 262–264) for an explanation. For a recent study of the
Ames Room illusion, see Dorward and Day (1996).
23. For the description and origin of these figures, see Note 55 to Chapter 3.
24. Early evidence for parallel activation of multiple meanings was presented by
Swinney (1979). The study of lexical ambiguity has since swelled to a research
area of its own. There is now neuroscience evidence for the activation of multiple
meanings in parallel; see, e.g., Mason and Just (2007).
25. The concept of a situation model was introduced into the psychology of reading
by van Dijk and Kintsch (1983, Chap. 10). It has since become a standard con-
ceptual tool for the study of language comprehension (Graesser, Millis & Zwaan,
1997; Zwaan & Radvansky, 1998), with some unexpected applications (Radvansky
& Dijkstra, 2007).
26. The average educated adult knows approximately 50,000 words (Miller, 1996),
and hence approximately as many concepts. Each concept is likely to enter
into more than one knowledge element. The space of possible knowledge ele-
ments built out of two concepts (e.g., dogs are mammals) is thus approximately
2,500,000,000. We do not know how many of those combinations make sense
or are typically acquired by a person in the course of a lifetime, but even if the
proportion is one tenth of a percent, the number of two-concept knowledge ele-
ments is of the order of a million. Similar calculations would yield even higher
numbers for knowledge elements built out of three or more concepts.
27. Anderson (1984), Anderson and Pirolli (1984) and Collins and Loftus (1975).
28. The study of word associations has a long history in psychology (Deese, 1965),
and, unlike many other techniques, it is used in both clinical psychology and in
basic cognitive research on memory and language. The two basic measures are
the probability and the speed with which one word elicits another; see Miller
(1996, Chap. 8).
29. This view of long-term memory is due to Anderson (1989, 1990). It is an advance
on the more commonly held view that the task of memory is to retain informa-
tion accurately.
30. Gibson (1977).
31. Newell and Simon (1972a).
32. There is a long line of developmental research on how children make transitive
inferences (Donaldson, 1963, Chap. 3; Glick & Wapner, 1968). Transitive inferences
are closely related to Piaget’s (1952) notion of seriation (Murray & Youniss, 1968).
Piaget’s main work on seriation appears not to have been translated into English
[Piaget, J., & Inhelder, B. (1959). La Genèse des structures logiques élémentaires: clas-
sifications et sériations. Neuchâtel: Delachaux et Niestlé], but see Inhelder and
Piaget (1964) for a briefer treatment. Studies with adults, where problems of this