Page 66 - Deep Learning

The Nature of the Enterprise 49

5. An articulation of the mechanisms and the triggering conditions vis-à-vis
the explanatory target. An explanation is a demonstration that (a) the rele-
vant triggering conditions held in the situation in which the target change
is supposed to have happened, and (b) the specified learning mechanisms,
if triggered under those conditions, would in fact produce the observed
change. If the explanatory target is a pattern of change, then the articula-
tion needs to show why that type of change tends to recur.
6. An explanation is the more satisfactory if it comes with an argument to
the effect that the postulated change mechanisms scale up, that is, produce
observed or plausible outcomes over long periods of time and across sys-
tem levels.
7. Last, but not least, an explanation is more satisfactory if it comes with a
demonstration that the postulated learning mechanisms can support suc-
cessful practice.

In the terminology of the philosophy of science, these seven points are
criteria of adequacy. Their satisfaction does not guarantee the truth of a theory.
They constitute a test that a purported explanation has to pass in order to be a
viable candidate. Bluntly put: If a theory or hypothesis lacks these features, it
is not worth considering.

THE PATH AHEAD
The research strategy behind the investigations reported in this book is to study
specific types of non-monotonic change and propose micro-theories to explain
them. Once the micro-theories have been clearly formulated, they can be mined
for deeper principles, if any. In this approach, theory construction does not pro-
ceed in an inductive, bottom-up or top-down fashion. The choice of phenomena
to be studied is guided by the prior decision to focus on non-monotonic change,
itself a theoretical concept. On the other hand, theory construction does not pro-
ceed by pushing a single concept or principle into every corner and crevasse of
the cognitive landscape. Instead, the principles of each micro-theory are designed
to provide understanding of the case that inspires them without regard for their
applicability elsewhere. The deeper theory, if any, is to emerge from the concep-
tual analysis of the micro-theories. In this layered approach, the degree of unifica-
tion to be sought is itself an outcome of the investigation rather than something
given at the outset. Figure 2.3 shows the overall structure of the enterprise.
Following this strategy, Parts II–IV investigate three cases of non-
monotonic change: the creation of novelty, adaptation to an unfamiliar or
changing task environment, and conversion from one belief system to another.