IDyOT: A Computational Theory of Creativity as Everyday Reasoning from Learned Information - Computational Creativity Research: Towards Creative Machines - page 125

with a multidimensional encoding of speech based on feature pairs,

was capable of segmentation into syllables with an F 1 scoreof0.67[ 54 ]. Further,
an electrophysiological study suggests that there are neural signals that correspond
with reported experience of expectation, which in turn corresponds with IDyOM's
predictions [ 35 ]; thus, IDyOM is demonstrably capable of predicting human behav-
iour.
Following Meyer's original proposal, that expectation (and its denial or fulfill-
ment) is a contributor to the emotional experience of musical listening, Egermann
et al. [ 10 ] compared IDyOM's information-theoretic predictions with physiological
and consciously reported measures of emotional response, and found significant
correlations. This lends general empirical support to Huron's suggestion [ 23 ] that
expectation (or “sweet anticipation”) is fundamental to humans and other organisms
and exapted 2 into music appreciation because it uses a more general mecha-
nism: IDyOM's mechanism is completely domain-general, and uses only temporal
sequence of the observed data to predict them (thence, it can be used for speech seg-
mentation, as above). This argument is expounded in full by Huron [ 23 ] and Wiggins
[ 55 ].
The abovework supplies a substantial body of evidence, supplementing that which
exists in the statistical linguistics literature, to hypothesise that the learning of lan-
guage and music are essentially the same statistical process, running over different
data representations. Expectations are represented as statistical distributions across
known symbol-sets (which might represent musical pitches, chords, phonemes, or
words). That representation is in turn subject to analysis in terms of Shannon's infor-
mation theory [ 46 ], and it is that analysis that will drive the cognitive architecture
presented below. The work is summarised by Pearce and Wiggins [ 39 ].

phoneme,stress

7.3.2 Conceptual Representations of Information Structure

While the IDyOM model is primarily concerned with the prediction of expecta-
tion, and therefore with the establishment of a model of the statistics of perceived
sequences, the work also makes useful predictions about the representations over
which those sequences are formed. In particular, given a selection of features that
might be used to represent the data over which it works, IDyOM uses a hill-climbing
search technique to choose the subset that allows the data to be represented in the
most compact way, with respect to average information content (also called cross-
entropy [ 8 ]). This embodies a hypothesis that brains seek to find representations that
store what they have learned efficiently (perhaps, we speculate, as part of memory
consolidation). However, this model is probably only descriptive, in that the features
are given as program functions, and (expensive) brute force search and re-evaluation

2 Exaptation is the appropriation of a biological capacity driven by given evolutionary pressures
into a different function. An alternative view is that these capacities form spandrels , supporting
other behaviours, but not becoming part of them.