Page 447 - Deep Learning

430 Notes to Pages 192–197

49. The specificity principle was first proposed in Ohlsson (2008a).
50. Fitts (1964).
51. Fitts (1964, p. 261).
52. Neves and Anderson (1981). For a neural network view of the proceduralization
of instructions, see Schneider and Oliver (1991).
53. See, e.g., Ohlsson (1987a) for a simulation model that could execute a form of
practical reasoning in a very simple spatial task domain. See, e.g., Hiebert (1986)
for an analysis of this type of learning from the point of view of mathematics
education.
54. Thorndike and Woodworth (1901).
55. See Kieras and Bovair (1986) and Singley and Anderson (1989).
56. Carbonell (1983, 1986), Falkenhainer, Forbus and Gentner (1989), Forbus, Gentner
and Law (1994), Holyoak and Thagard (1989a, 1989b), Hummel and Holyoak
(1997, 2003), Keane, Ledgeway and Duff (1994) and Veloso and Carbonell (1993).
57. I know of no simulation model of learning from live demonstrations; but see
VanLehn (1991, 1999) and VanLehn and Jones (1993) for a model that learns from
solved examples.
58. Hilgard and Bower (1966, p. 3).
59. Barrow, Mitrovic, Ohlsson and Grimley (2008) and Ohlsson et al. (2007).
60. The quote is from Thorndike (1898, p. 45). The idea of gradual strengthening
was central to the behaviorist notion of learning as the building of stimulus-
response connections. After the cognitive revolution, the idea that repeated
traversal of a link or repeated application or execution of a knowledge structure
causes an increase in the strength associated with that knowledge structure has
remained one of the standard tools of cognitive models (Ohlsson, 2008a). The
consequences of an increase in strength are typically hypothesized to be that
the knowledge structure has a greater probability of being activated, retrieved
from memory, applied or executed. The conceptual difficulty inherent in this
idea is that once a knowledge structure S 1 has acquired a greater strength than
some other structure S 2 , it will always win over S 2 in any future situation, which
in turn will increase S 1 ’s advantage over S 2 , locking the cognitive system into
choosing S 1 in any future situation in which both knowledge structures apply.
This is not a promising hypothesis if the goal is to explain human flexibility.
Thorndike’s mentor, William James, saw the problem clearly: “So nothing is
easier than to imagine how, when a current [i.e., a neural signal] once has tra-
versed a path [inside the nervous system], it should traverse it more readily
still a second time. But what made it ever traverse it the first time?” (James,
1890, vol. 1, p. 109). The introduction of noise into the strength levels or the
decision-making process does not go very far toward alleviating this concep-
tual problem. In the end, this problem cannot be solved unless we also assume
that negative feedback has the power to reduce the strength of a knowledge
structure.
61. Sun, Merrill and Peterson (2001) and Sun, Slusarz and Terry (2005). In its
emphasis on interactions between implicit and explicit learning, the Clarion
system is a descendant of the instructable neural net model of Schneider and
Oliver (1991).