Page 414 - Deep Learning

Notes to Pages 15–16 397

39. Tool use, drawing and language emerge naturally and without obvious strain or
stress in normal infants. Although behavioral researchers disagree whether tool
use is best seen as an extension of perceptual-motor activity (Lockman, 2000)
or as a form of problem solving (McCarty, Clifton & Collard, 2001), there is now
neuroscience evidence that “behaviors associated with complex tool use arise
from functionally specialized [neural] networks” (Johnson-Frey, 2004, p. 71), i.e.,
that tool use is supported by special features of our brains. Although it might
take a team of dedicated cognitive scientists several years to teach 150 words to
a chimpanzee, all normal children learn thousands of words spontaneously and
seemingly effort-free. There is a long-standing argument, originally formulated
by Noam Chomsky and known as the argument from the poverty of the stimulus,
to the effect that language is supported by special innate capabilities. Crain and
Pietroski (2002) show how this argument explains why “language acquisition is a
snap.” I know of no similar argument with respect to drawing, but the tendency
to draw emerges spontaneously; see, e.g., Gardner (1980) and Lambert (2005).
40. For the idea that biological evolution and cognitive change are mechanisms
for tracking change in the environment at two different time scales, see Plotkin
(1994), especially Chapter 5.
41. Semon introduced the term “engram” in 1904 in a book called Die Mneme, writ-
ten in his native German. Schacter (2001) describes the history of this contribu-
tion; see Chapter 7.
42. Describing experiments in which rats first acquired the skill (“habit”) of finding
their way through a maze, then had a portion of their brain removed and finally
were observed as they tried to navigate the maze again, Lashley (1929) wrote, “It
is certain that the maze habit, when formed, is not localized in any single area of
the cerebrum and that its performance is somehow conditioned by the quantity of
tissue which is intact. It is less certain, though probable, that all parts of the cor-
tex participate equally in the performance of the habit …” (p. 107). Neuroscience
evidence gathered since then has not confirmed this holographic view of memory
at the neural level. The evidence points rather to multiple, separate brain systems
(Rolls, 2000). For example, localized lesions sometimes give rise to very content-
specific impairment in a person’s knowledge (Caramazza & Shelton, 1998).
43. There are many textbooks in cognitive psychology that cover what is known
about memory, e.g., Anderson (2004), Goldstein (2008) and Reisberg (2006).
44. The history of the supposed process that cognitive scientists variously refer to as
abstraction, generalization and induction stretches across two millennia. Aristotle
wrote that multiple perceptions form universal categories (e.g., in the Posterior
Analytics, Book II, 19: 100a, 1–15). The British empiricist philosopher David
Hume (1777/1910) is responsible for posing the problem of justifying inductive
inferences in a way that was so intractable as to provide centuries of work for
philosophers: To say that inductive inferences have been shown to be accurate in
the past and therefore likely to be accurate in the future seems circular, so how
can such inferences be justified instead? This formulation of the problem turned
out to be so intractable that philosophers eventually decided to abandon it; see
Goodman (1954/1983) for an attempt to reformulate the problem. A closely related