Page 384 - Deep Learning

Elements of a Unified Theory 367

Table 11.1. The distinctions among routine processing, monotonic (additive) change
and non-monotonic change in three cognitive functions.

Routine processing Monotonic change non­monotonic change
Creativity:
Execute the current Follow an unexplored path in Revise problem space by acti­
task strategy. the current problem space. vating previously unheeded
options.
Adaptation:
Execute the current Extend strategy by add­ Alter strategy by constraining
task strategy. ing rules for previously (specializing) existing rules.
undecided situations.
Conversion:
Retrieve and Form new beliefs that are Revise and propagate the truth
articulate beliefs. consistent with prior beliefs. values of existing beliefs.

step toward a comprehensive treatment of this type of cognitive change was to
identify three subtypes of non­monotonic change – creativity, adaptation and
conversion – and to formulate a micro­theory for each. Each such theory is
based on a set of assumptions about the routine, steady­state processing that
forms the backdrop for cognitive change. other assumptions describe what
happens in monotonic learning. Finally, there are assumptions that explain
how non­monotonic changes can occur within a system that satisfies the first
two sets of assumptions; see Table 11.1 for an overview of these distinctions.
Each micro­theory conforms to the type of analysis that is standard in
cognitive psychology: The relevant behavioral patterns are explained by break­
ing down the high­level functions (creating, adapting, converting) into com­
ponent processes, parts and parts­of­parts, that interact to produce those
patterns. The three analyses are successful in two respects. First, the compo­
nent processes are not unique or exotic. on the contrary, the analysis in each
case breaks down non­monotonic change into well­known and familiar basic
processes that are empirically supported by prior cognitive research. second,
all homunculi have been discharged. in each case, the breakdown ends in basic
processes that are not themselves non­monotonic change processes but instead
are so simple that there is no doubt that we can specify them precisely; almost
all of them have, in fact, been implemented in running computer simulation
models.
These successes come at a price: The end points of the analyses, the set of
basic processes, differ from micro­theory to micro­theory. instead of break­
ing down creativity, adaptation and conversion into one and the same set of