Page 272 - Deep Learning

8

Error Correction in Context

The question is not so much how to prevent [errors] from occurring, as how to
ensure that their adverse consequences are speedily detected …
James Reason 1

… the real question is why only some societies proved fragile, and what distin­
guished those that collapsed from those that didn’t.
Jarred Diamond 2

The actions of hunter-gatherers have limited causal reach. A shot that misses
its target might mean sudden death for the hunter or warrior whose shot it
was, but the consequences of that particular error for the tribe or troop are
likely to be absorbed into the flow of events and make only a small difference
to the outcome of the hunt or the battle. With the advent of large-scale social
and technical systems like airlines, container shipping, global financial mar-
kets, irrigation agriculture, power grids and research hospitals humans came
to know man-made disasters. A single erroneous action can be the starting
point for a butterfly effect that propagates upward through the levels of the
relevant system until it makes the difference between cure and death, between
ubiquitous electricity and power blackout, between wealth and bankruptcy or
between victory and defeat; in general, between normality and disaster. The
unacceptable consequences of disasters make it imperative to understand
the dynamics of error correction in individuals, social systems and entire
societies.
The constraint-based theory of learning from error put forward in
Chapter 7 describes the inner mechanics of a single learning event, the correc-
tion of a single fault in a single person’s cognitive skill. Such an event might last
anywhere from a couple of seconds to a couple of minutes. Expertise in operat-
ing a complex social or technical system, on the other hand, consists of skills

255