Page 452 - Deep Learning

Notes to Pages 235–236 435

elements. This formulation leaves open what is to count as an “element.” Kieras and
Bovair (1986) and Singley and Anderson (1989) have shown that if an element is
interpreted as a production rule, then Thorndike’s principle predicts the amount
of transfer to a high degree. However, rules have to have at least a minimal
level of abstraction to apply to multiple situations, so the identical rules inter-
pretation hides within it the second, and older, idea about transfer: Knowledge
transfers because it is abstract. This concept has been around since the begin-
ning of systematic thinking about cognition (e.g., James, 1890; see vol. 1,
pp. 505–508, and vol. 2, pp. 345–348) and it is still current (Bassok & Holyoak,
1989; Goldstone & Sakamoto, 2003; Ohlsson, 1993b; Ohlsson & Lehtinen, 1997;
Reed, 1993; Salomon & Perkins, 1989). A third idea about transfer is that it occurs
via analogy (Gentner, 1983; Gick & Holyoak, 1980, 1983; Hummel & Holyoak,
2003; Markman & Gentner, 2000). From a transfer point of view, the main dif-
ference between abstraction and analogy is that in the former case, the cognitive
work necessary to bridge from the training task to the transfer task – creating the
abstraction – is carried out in the context of mastering the training task, while in
the analogy case, that work – creating the analogical mapping – is carried out at
the time of encountering the transfer task. The transfer mechanism implied by
the constraint-based theory strikes a balance between processing for the pres-
ent and processing for the future, allowing for re-use of previously constructed
skill components while also recognizing that re-use typically requires a variable
amount of revision; see Ohlsson (2007a). However, it is highly unlikely that there
is a single mechanism behind transfer of training (Nokes, 2009). See Detterman
and Sternberg (1993), Haskell (2001) and Salomon and Perkins (1989) for useful
reviews of transfer research.
41. Detterman (1993) summarized studies that failed to find transfer and pointed
out that studies that produce measurable transfer effects tend to provide optimal
conditions, including great similarities between training and transfer tasks and
strong hints that the training task is relevant to the transfer task, factors that
might not be present outside the laboratory. An example is Gick and Holyoak’s
(1980) study of analogical transfer between two isomorphs of Duncker’s classic
radiation problem (or the convergence problem). Even in a situation in which the
two isomorphs follow each other in the course of a short experiment, 59% of the
subjects failed to solve the transfer problem without a hint to use the analogue;
24% did not solve it even with the hint (Exp. V, Table 12). However, there is no
standard metric and no widely accepted baseline or base rate against which to
compare transfer effects: Some experimental subjects will perform a transfer task
well even without transfer, so what level of performance is evidence for transfer?
Gagné, Foster and Crowley (1948) reviewed a variety of savings measures of trans-
fer – how much less training is required to master the transfer task after mastery
of a training task as compared to after no prior training – but recent experimen-
tal work on transfer does not typically use such measures (but see Nokes and
Ohlsson, 2004, and Singley and Anderson, 1989, for exceptions). In the absence
of a standard metric and an accepted baseline for the amount of transfer, claims
that transfer effects are great or small are moot. The amount of transfer is also a
concern in the design of training in industry and business (Baldwin & Ford, 1988;