Page 301 - Deep Learning

284 Adaptation

we can characterize the ongoing change as consisting of both the elimination
of steps – the discovery of shortcuts – and the speedup of the remaining steps.
The fact that change of this dual sort generates negatively accelerated improve-
ment curves is a level-invariant characteristic.

Summary
A collective – group, team, population or organization – can be regarded as a
single entity that learns from error. The abstract principles of constraint-based
learning appear to be level-invariant. That is, they apply at the collective level
as well, even though the material implementation of the relevant processes is
different. A collective fails, perhaps among other ways, by applying underspeci-
fied decision rules and operating procedures, thereby performing actions in
situations in which they should not have been performed. Errors of this sort are
discovered on the basis of error signals that violate constraints on the appro-
priate, correct or useful system states. When there is an opportunity to learn
from such errors, the response is typically to specialize – make more explicit,
extensive and precise – the applicability conditions for the relevant decision
rules and operating procedures. If the principles of constraint-based specializa-
tion are level-invariant, the patterns that characterize collective change over
time should be similar to the patterns in individual change. Available data
indicate that this is indeed the case. Specifically, learning in collectives follow
the same negatively accelerated learning curve that characterizes learning in
individuals.

Safety Implications
Although individual error rate curves never fall all the way to zero, it is pos-
sible to design collective systems for higher levels of safety than such curves
suggest. Safety scientists characterize commercial airlines, the U.S. Navy’s
nuclear submarines and the European railroads as ultrasafe systems, because
the error rates for those systems have fallen so far that the probability of a
disastrous accident is 0.0000005 per unit, where a unit is, for example, a unit
28
of time or distance or a passenger, depending on the system. This level of
safety requires that errors are caught before they develop into disasters, which
in turn requires that the participating individuals are provided with multiple
sources of information about the current system state. In technical systems,
instruments provide such information. Recommendations in other areas of
experience follow the same principle of providing the operators with more
dense information and status indicators along the way, so that pre-failure