Page 423 - Deep Learning

406 Notes to Page 41

known as the Baldwin effect after its originator James Mark Baldwin (1896a,
1896b) claims that the triggering condition for speciation is that members of
a species acquire a beneficial behavior or repertoire of behaviors via learning.
If the learned behaviors support survival and a high rate of reproduction, then
mutations that enable, facilitate, and support those behaviors will be more ben-
eficial than if those behaviors had not been acquired, and hence accumulate
over time via natural selection (Simpson, 1953). Consider the chimpanzee habit
of poking juicy insects out of holes with help of a straw or thin twig. If a troop
of chimpanzees ever became completely dependent on this type of diet, they
would obviously undergo selection for precise vision, eye-hand coordination
and fine-grained dexterity in hands and fingers. Once again, the change process
is the accumulation of beneficial mutations via natural selection, but the trig-
gering condition is completely different in Baldwin speciation than in allopatric
speciation, because it consists in the prior existence of a learned behavioral rep-
ertoire. Since a pioneering paper by Hinton and Nowlan (1987) demonstrated
an impact of learning on evolution in a simulation model, the Baldwin effect
has received attention in Artificial Intelligence, machine learning and related
fields. See Weber and Depew (2003) for appraisals of the Baldwin effect. The
main point for present purposes is that the triggering condition for a change
process is a separate piece of theory from the specification of the process itself.
It distinguishes ad hoc (merely descriptive) from explanatory accounts and to a
large extent determines which phenomena a particular change process can suc-
cessfully explain.
41. Modern philosophical work on scientific explanation began with Hempel and
Oppenheim’s (1948) article on the so-called covering law model, which holds that
to explain is to subsume a particular event under some scientific law that “covers”
it. Four decades later, this model is no longer accepted, but no other analysis
of what it means to explain or how scientific explanations differ, if at all, from
commonsense explanations has become authoritative (Cornwell, 2004; Keil &
Wilson, 2000; Pitt, 1988; Salmon 1989). The philosophical study of psychological
explanations, specifically, has expanded the early contributions by Boden (1972),
Cummins (1975) and Fodor (1968) in various ways; see, e.g., Cummins (2000).
One of the standard themes in this literature is that some explanations explain
by decomposing a system into its components and showing how the components
produce the system’s behavior. The prototypical example is explaining a piece of
machinery or a biological system with interacting parts (Bosse, Jonker & Treur,
2006). Cummins (1983b) contrasts componential with transition explanations,
the latter showing how a change of state came about by subsuming it under a
causal transition law. As I use the term, componential explanations combine ele-
ments of both. They are componential (or analytic) because they break down
an observed change into a chain of component transitions. In my case, both the
explanandum and the explanans are transitions in Cummins’s (1983b) sense. A
unit change has the same form as what Cummins calls a disposition (when so-
and-so is the case, unit process so-and-so happens, or tends to happen). Unlike
Cummins, I believe an explanation in terms of units of this sort can be genuinely