Page 451 - Deep Learning

434 Notes to Pages 222–235

32. “A major lesson from AI [Artificial Intelligence] about generate-and-test situa-
tions is that it always pays to transfer knowledge from the test to the generator, so
that candidate solutions need never be created at all” (Newell, 1990, p. 100).
33. Langley (1983, 1985, 1987).
34. See Festinger (1957/1962), Piaget (1985) and Schank (1982). Popper’s falsification-
ist philosophy of science has been discussed as a theory of learning from error
by Berkson and Wettersten (1984). For action-conflict-change theories in motor
learning, see Hoffman et al. (2007).
35. The classic reference is Pauling (1935/1960). Niaz (2001) describes the historical
emergence of the co-valent bond.
36. Solomons (1988).
37. See Ohlsson (1993a, 1996b) for more details about the chemistry simulation.
38. The idea that change moves from concrete and specific knowledge structures
toward abstract ones has been expressed over and over again in a variety of
ways in different cognitive theories that otherwise differ in focus, formulation
and intent. Jean Piaget hypothesized that children’s cognitive skills advance from
sensorimotor skills through a stage of concrete thinking and on to a stage of
abstract thinking. (See Piaget, 1950, for an original statement of the stage theory,
and Flavell, 1963, for a comprehensive summary.) Contemporary theories of
skill acquisition envision a process of generalization that can apply, for exam-
ple, to production rules to generate more abstract rules (see, e.g., Sun, Merrill &
Peterson, 2001, for an example and Ohlsson, 2008a, for a review) and to descrip-
tions to generate more abstract mental representations often called schemas
(Gick & Holyoak, 1983; Marshall, 1995). This process apparently operates even
in implicit learning of such improbable learning targets as abstract grammatical
rules embodied in random-looking letter strings (Reber, 1996). The field of cat-
egory learning has contributed the theory of prototypes, which says that people’s
representations of categories like “bird” and “fruit” capture the central tendency
of the category instances that they have seen. That is, what is acquired in category
learning is a representation of the average bird or fruit, a very particular form of
generalization. See Rosch (1978) for an early statement of the prototype theory.
The literature on later developments is large (Osherson & Smith, 1981; Smith &
Minda, 1998; Smith, Osherson, Rips & Keane, 1988). Ashby and Maddox (2005)
review evidence, including neuroscience data, on performance of prototype-
related categorization tasks. The fundamental principle behind these and many
other cognitive theories is that knowledge moves from concrete and specific to
abstract and general in the course of learning.
39. Brain-imaging studies have shown that the Specialization Principle applies to
other types of learning as well, e.g., category learning (Little & Thulborn, 2006).
For the connection to the themes of differentiation and specialization in evolu-
tion and ontogenesis, see, e.g., Carroll (2005), Berenbaum (1996), Futuyma and
Moreno (1988), Raff (1996) and Wolpert (1992).
40. Like so many aspects of skill acquisition, research on transfer goes back to the
work by Edward L. Thorndike. Thorndike and Woodworth (1901) proposed that
the amount of transfer from task X to task Y is determined by the overlap or
similarity between the two tasks, measured in terms of the number of identical