Page 261 - Deep Learning

244 Adaptation

this prediction, but it is consistent with the results of Bassok and others. In
general, skill acquisition theories that predict asymmetric transfer effects are
more likely to be psychologically accurate than those that predict symmetric
transfer effects.

The Primacy of Adaptation

It is unlikely that skills are ever constructed from scratch. A computer pro-
grammer can “teach” a computer a new procedure by typing code into a blank
file, but skill acquisition by a human brain is more likely akin to growth, differ-
entiation and specialization and to recycling existing structures to serve new
functions, one of Nature’s favorite tricks. Even in situations that we intuitively
conceptualize as learning a brand-new skill, many prior competencies provide
small pieces of the target competence. Learning to use SCUBA gear is a case
in point. The equipment and the procedures are all unfamiliar. Yet, fastening
the regulator on the air tank means fitting one metal piece over another and
tightening a screw with a rotating movement of the wrist – surely not entirely
unfamiliar; putting on a buoyancy compensator is somewhat similar to put-
ting on any other kind of vest; the dive mask is similar to a pair of ski goggles;
a diver moves underwater with almost the same kicking motion he uses to
swim on the surface; and so on. There are many precursor tasks, but each con-
tribution to the new skill set might be so slight and so implicit as to escape
conscious notice.
The theory of learning from error builds on this observation. It claims
that human beings can adapt to novel environments because we can both spe-
cialize general strategies to particular environments and fall back on slightly
more general strategies when the specialized ones become obsolete because
the environment shifts under our feet. The slightly more general strategies are
guaranteed to be available in memory, because, according to the principle of
specialization, they are intermediate products generated on the path to the
more specific strategies. Conflict resolution by specificity guarantees that
the most specific strategy available will be used as the initial strategy for the
changed circumstances, that is, that the cognitive architecture will make maxi-
mal use of past learning. In conjunction, constraint-based specialization, rule
genealogies and conflict resolution by specificity resolve the conflict between
the fact that highly general strategies, although widely applicable, provide
too little guidance for action, and the fact that highly specialized strategies,
although effective, are guaranteed to become maladaptive as the environment
changes.