Page 217 - Deep Learning

200 Adaptation

% Phase 1 Phase 2 Phase 3
100

50

0

Instruction Feedback
Solved Quantitative
examples regularities
Figure 6.4. A hypothetical case of how the relative importance of four learning
mechanisms might shift across the three phases of skill acquisition. The y-axis repre-
sents the percentage of overall improvement within each phase that is accounted for by
each mechanism. to avoid clutter, the figure assumes that there are only four mecha-
nisms, so their contributions sum to 100% within each phase.

The relative prevalence of each information type is likely to vary across the
course of learning. in the beginning stage, instructions, examples, prior strate-
gies, practical inferences and outcomes of tentative steps are likely to be the
most available sources of information and the mechanisms that make use of
them hence must account for much of the initial improvements. in the middle
stage of learning, those types of information will typically fade because their
usefulness has been exhausted. instead, the information generated by acting
vis-à-vis the task – feedback – and the mechanisms that utilize feedback domi-
nate. Eventually, the task environment is so well understood and explored that
there is little that can be learned from yet more feedback. in this stage, addi-
tional speed-ups and error rate reductions are possible only because execution
histories enable the learner to discover shortcuts, and statistical regularities in
the environment enable the learner to optimize his decision making. The nature
of the changes that take place is itself gradually changing as practice progresses;
see Figure 6.4 for a simplified illustration with four types of learning.
in principle, there could be other types of information that are not yet
included in the preceding list, but to date no such addition has come to mind.
The possibility that the nine constitute the complete list raises the question of
how we should interpret the learning mechanisms proposed to date.