Page 209 - Deep Learning

192 Adaptation

Computer simulation of skill acquisition resembles work on heavier-than-air
flight before the Wright Brothers: Would-be aeronautical engineers would
design an airplane, build it, try to fly it, sweep up the debris and build another
48
one. When confronted with the huge space of possible airplane designs, this
mode of operation made for slow progress. The Wright Brothers succeeded
because they worked separately on the three subproblems of lift, propulsion
and steering, and combined the solutions. The study of skill acquisition also
needs a principled way to guide the search through the space of possible rep-
ertoires of learning mechanisms.

The information Specificity principle

Hypotheses about skill acquisition can be organized in terms of their inputs.
improvements in a task strategy cannot arise out of nothing; structure has
to come from somewhere. if a learning mechanism produces rules that are
49
better adapted to the target task than the learner’s current rules, that mecha-
nism must have incorporated additional information about the target task
into those rules. A learning mechanism can therefore be conceptualized as
a process that takes one or more rules plus some additional information as
inputs and delivers an improved rule as output. The new rule is better adapted
to the task environment precisely because it incorporates the additional
information.
The question of the number and nature of the learning mechanisms
can therefore be approached by considering the number and nature of the
sources of information that are available to a learner in a practice scenario,
and the types of information they provide. The processes needed to make
use of one type of information typically differ from the processes needed to
make use of some other type. For example, learning from a solved example
requires different processes from learning from an error, and learning from
an analogous task requires different processes from learning from posi-
tive feedback. This Information Specificity Principle postulates a mapping
between types of information and types of learning mechanisms. it attri-
butes the multiplicity of learning mechanisms to the existence of multiple
types of information.
Multiple Mechanisms and Information Specificity are abstract principles
and they do not by themselves explain any one instance of skill acquisition or
any regularity therein. They are meta-principles in that they specify the kind
of explanation psychologists should seek. together they suggest that we can
make an informed guess about the repertoire of skill acquisition mechanisms