How Models of Creativity and Analogy Need to Answer the Tailorability Concern - Computational Creativity Research: Towards Creative Machines - page 96

semantic databases actually have to put these ideas to the test. If the state-of-the-art
research in the database field has significant difficulty with some style of repre-
sentation as required by an analogical matcher, then perhaps that matcher needs to
carefully consider its assumptions about the nature of knowledge representation, or
else be able to easily extract the necessary data. 3 The choice of database is therefore
a sensitive one.
It may help to go back to the criticisms whichmay havemotivated this requirement
in the first place. Hofstadter's group criticizes SME by saying that “the analogy is
already effectively given in the representations” [ 36 ]. The information provided to the
matcher is selected in such a way that it does not include a significant amount of extra
data that would lead to false analogies, and the choice of predicates and objects is done
in such a way that presupposes the desired matching. This is a criticism composed
of two points: the amount of information provided as input (which is not too little,
or too much, but just right ), and the nature of the information (the corresponding
predicates on both sides just happen to have the right amount of arguments in the
necessary order). The system is not robust enough to produce good analogies when
given the same input data in a variety of formats.
Two possible changes can begin to answer these critiques. One is to require that the
databases are large enough, and the input minimal enough, to introduce a significant
amount of false matches that would confuse less robust analogical matchers. Critics
claiming that the input is too carefully chosen could be pointed to the fact that the
analogical matcher must plumb through an answer space that is large (at least relative
to the input problem). The larger the search space, the more impressive the ability
of the analogical matcher to select only a few potentially relevant source analogs.
Furthermore, an inability to demonstrate scalability to large datasets weakens any
architecture's claim to psychological plausibility: if the architecture can't handle
a dataset large enough to produce non-trivial answers, how can it be an accurate
model of a process used by human-level reasoners? 4
Secondly, we could require robust and consistent performance on a variety of input
forms. For example, in the heat-flow problem (Fig. 5.1 ) Mitchell and Hofstadter [ 36 ]
note that there are many possible ways to structure the input: heat could be described
as an object, or as an attribute of coffee, or heat flow could be a relation with three
rather than four arguments [ 36 ]. Consistent performance across various input forms
puts more pressure on the analogical matcher's re-representation algorithm(s), rather
than relying on a separate NLP module. This also allows for a leveling of the playing
field across different systems: In order to show that a given example adheres to this
requirement, a localist, structured analogical system would have to demonstrate two
things with regard to that particular example:
3 We do not mean here to say that what works best for large artificial databases is the same as what
is employed by the human brain. But if a researcher declares that the structure of human knowledge
has certain properties, and large datasets cannot be created that do not have those properties for
reasons of scalability, then it should be at least a weak hint that perhaps the assumption of those
properties is not practicable.
4 This is a common criticism of Hummel and Holyoak's LISA system; see [ 24 ].