Information Technology Reference
In-Depth Information
As we shall see below, MusEng builds a musical memory in terms of small
segments of music. Ideally, the system would segment the music based on per-
ceptual criteria. The original iMe system sported such a method, inspired by Gestalt
psychology (Eysenck and Keane
2005
). However, for this project, we programmed
MusEng to segment the music according to a user-speci
ed number of measures,
for instance, every measure, or every two measures, or every three and so on. The
rationale for this decision is that we wanted to synchronise the fMRI analysis to the
input score by handling the fMRI data on a measure-by-measure basis, as it was
shown schematically in Figs.
12.3
and
12.4
. Therefore, it made more sense to
establish the measure as a reference value to segment the music.
MusEng
'
s memory consists of a series of Feature Tables (FTs), which comprise
vectors of musicodes for material that the system has been exposed to. As the
musicodes are extracted from incoming measures, the system may or may not create
new FTs, depending on whether the respective musicodes have already been seen
by the system. If a certain vector of musicodes is identical to one that has been
previously seen by the system, then the system updates the relevant FT by
increasing a weighting factor, represented by the variable
(Eq.
12.1
). This var-
iable is generated by summing the total number of FTs and then dividing the
number of instances of each individual FT by the total. In essence, this becomes a
simple moving average. In Eq. (
12.1
), the value of
ω
indicates the weighting factor
associated with a given FT. The variable
x
represents the number of instances of a
given FT in the series, and
n
the total number of FTs in the series.
ω
P
FT
ð
x
Þ
P
FT
ð
n
Þ
x
ð
x
Þ
¼
ð
12
:
1
Þ
for vectors of
musicodes that do not appear as often as more frequent ones, in the same way that it
raises the value of
ω
for more commonly used vectors, to a maximum value of 1.0.
The value of
This moving average has the effect of lowering the value of
ω
informs the probability of a given musical segment being generated
later on by the system. Typically, a decrease in the value of
ω
causes the system to
ω
'
forget
to utilise the corresponding FT entry in the subsequent generative phase.
In order to illustrate how MusEng
'
is memory is built, let us examine a hypo-
thetical run through the sequence previously shown in Fig.
12.10
, commencing
with an empty memory. The
'
first measure (Fig.
12.11
) is analysed, and the
respective musicodes are generated. For the sake of clarity, this example will focus
on three of the
five features: melody direction (dir), melody interval (int) and event
duration (dur).
MusEng creates in its memory the
first feature table, FT1, with musicodes
derived from the
first measure of the training sequence (Fig.
12.11
) as follows:
Fig. 12.11
The first measure for the example analysis
