IDyOT: A Computational Theory of Creativity as Everyday Reasoning from Learned Information - Computational Creativity Research: Towards Creative Machines - page 134

another sensory modality, e.g., describing in language a scene representation derived
from a vision input) in the statistical model, in the style of Eshghi et al. [ 11 ].
In summary, IDyOT's memory consists of multiple structures like those in
Fig. 7.2 a, in parallel, tied together by observed co-occurrences of feature values
expressed in multidimensional perceptual input sequences. The whole constitutes
a Bayesian Network, stratified as explained above, and constrained to predict only
to the subsequent symbol; note, though, that the subsequent symbol may represent
something arbitrarily far in the perceptual future, because higher-level, more abstract
models predict in parallel with, and conditioned by, more concrete ones. From this
model, IDyOT's generators make predictions and their outputs are selected on the
basis of probabilistic matching with input—and the cycle begins again.
7.4.3 Matching, Similarity and Representation
One important factor in the above description is the notion of similarity, and—
implicitly—the provenance of the symbols used in the sequences. We address both
of these issues by appeal to Gärdenfors' theory of Conceptual Spaces [ 17 ], in which
a low-dimensional “conceptual level” mediates between a “symbolic” or “linguistic”
level, and the high-dimensional continuous representation that hypothetically cap-
tures thewetware. Using this approach, each of our symbols correspondswith a region
in a conceptual space, and thus can its relationship with other symbols be defined.
In some perceptual domains (such as musical pitch) there are well-defined theories
of the geometry of the specific conceptual spaces required, and these can support
the measurement of similarity. More excitingly, Gärdenfors' theory has an account
of the development of such spaces in a way that may be construed as information-
theoretic, and a future avenue for development of IDyOT will be the inclusion of
learned representations for the learned structures in its memory.
In the current formulation, perceptual inputs are matched for similarity in a met-
ric space distorted by the predicted distribution, in such a way that more expected
symbols are more tolerant of a poor match with the symbols encountered. Without
such tolerant but predictive matching, it would be difficult to understand one's own
language when spoken with an accent different from one's own, and the phonemic
restoration effect would not take place. This effect can also be modelled by lowering
the tolerance of a simple sequential match, so that unmatched intermediate symbols
do not cause failure. This approach, however, leaves no room to account for the abil-
ity to learn an unfamiliar accent over time, nor indeed, does it account for learning
perceptual spaces at all.