Page 279 - Deep Learning

262 Adaptation

100
y=104.526x –1.036 2
r = 0.988

Steps 10

1
1 10 100
Trials
Figure 8.3. Learning curve for computer simulation model that learns from both
successes and failures. Adapted from Ohlsson and Jewett, 1997, Figure 10, p. 165.

learning mechanisms that are most important during the mastery phase
of practice; see Figure 6.2. In the mastery phase, the two most important
sources of information are positive and negative feedback generated by the
outcomes of the learner’s own actions. James J. Jewett and I performed a
series of simulation experiments with a computer model that could learn
from both successes and failures. We explored the effects of turning one
6
or the other learning mechanism on or off. We found that error correction
by itself generates learning curves that fit exponential equations. However,
with learning from successes and failures operating in parallel, the behavior
of the model was well described by a power law of the same type and shape
that is frequently observed in data from human learners; see Figure 8.3. The
power law of learning emerged in the interactions between learning from
successes and learning from errors. This result does not depend on the inter-
nal mechanics of the learning mechanisms. It does not matter at this level of
analysis exactly how learners capitalize on successes and correct their errors;
it only matters that they are capable of doing so. However, the rate of learn-
ing matters. If there are multiple modes of learning, each mode can occur
at a higher or lower rate, and the rate parameters can vary independently
of each other. The power law fit only appeared for certain settings of those