Page 212 - Deep Learning

The Growth of Competence 195

get to the Far East from Europe by sailing west. The prevalence of reasoning
from first principles is likely to be limited by the complexity of the required
inferences.
A third source of information available to a novice learner is his or her
stock of prior practical knowledge. novel tasks are never completely novel;
there are always similarities to already mastered tasks. Edward Thorndike sug-
gested that already acquired “elements” of practical knowledge can be re-used
54
on a novel task. Thorndike did not define what he meant by “elements,” but
we can identify them with rules. if a rule was learned in the course of mas-
55
tering task X and that very rule also applies to another task Y, it need not be
r e-learned. no particular learning mechanism is needed to revise or edit the
rule. identity is a strong requirement. A weaker criterion is that the current
task is analogous to some already mastered task. to use analogical informa-
tion, the learner has to retrieve a useful analogue from memory, map the two
tasks to each other in a consistent way and then use the strategy for the already
mastered task as a template for the strategy-to-be-learned. Researchers have
proposed several different ways in which this type of learning mechanism
might be implemented in the human brain. 56
in real-world contexts, a person can learn by observing others perform-
ing the target task, as when a novice barista observes an experienced coffee
shop employee whip up a couple of complex espresso drinks (modeling). in
57
a training environment, the competent person might carry out a demonstra-
tion that is specifically designed to help a novice see how the target task is to
be done, as when a drill sergeant shows the recruits a rifle maneuver. in aca-
demic settings, lectures and textbooks often provide solved examples, canned
demonstrations that the novice can study in order to see how a particular
type of problem is supposed to be solved. A major difficulty with this type of
information is that models, demonstrations, and solved examples require the
learner to generalize the presented solution, and it may not be obvious what
or how to generalize.
if there are no helpful instructions, if the task situation is too opaque to
handle via practical reasoning, if the target task is not analogous to any prior
task, and if no model, demonstration or solved example is available, then the
learner can nevertheless act vis-à-vis the task. people possess general heuris-
tics that apply to almost any task (if you have no idea what to do, do something),
and those heuristics can at least generate task-relevant behavior. Acting gen-
erates new information in the form of observable outcomes. This source of
information becomes more extensive and important in the second phase of
skill acquisition.