Machine Learning to Identify Neural Correlates of Music and Emotions - Guide to Brain-Computer Music Interfacing - page 90

5.5 Machine Learning

Machine learning refers to the science of getting a computer to learn a rule
describing the regularity of patterns embedded in data without having to explicitly
program it. Instead, machine learning systems attempt to identify generalisable
rules directly from the data. The identi ed rules should be applicable to previously

unseen data and may be used to partition them via appropriate decision boundaries
or translate it into some more appropriate space (Alpaydin 2004 ;M

ü ller et al.

2004 ).
For example, given a set of neurological measurements, it may be desirable to
identify a rule which relates them to measures of musical stimuli from a piece of
music played to the participant. Alternatively, one may uncover neural correlates of
emotion by identifying a rule which relates neurological activity to participants

'

self-reported emotions.
Rule identi cation of decision boundaries which

may be applied to the data. For example, given EEG recorded during two tasks (e.g.
listening to music vs. listening to noise), rule identi

cation often amounts to identi

cation may amount to iden-
tifying a rule for finding whether a new EEG segment corresponds to piece of music
or a noisy auditory stimulus.
More formally, a two-class problem classi

cation learning may be expressed as

f : R N !f 1 ; þ 1 g the process of identifying a function from a function class

F using a set of training data such that f will classify unseen examples with min-
imum error. For problems with more than two classes, f is modi

ed appropriately.
Machine learning methods can be broadly described as either supervised or
unsupervised. Supervised machine learning methods use labelled data as a part of
the learning process, whereas unsupervised methods do not.

5.5.1 Unsupervised Machine Learning Methods

Unsupervised machine learning does not use labelled data and hence concentrates
on removing redundancy in the dataset or on emphasising components of the data
which may be of interest, for example, components of high variance. This means
unsupervised machine learning methods are often used for dimensionality reduc-
tion, for example, principal component analysis (PCA) (Smith 2002 ; Lee and
Seungjin 2003 ), for identifying advantageous translations that may be applied to the
data, for example, independent component analysis (ICA) (Comon 1994 ; Qin et al.
2005 ), or for identifying hidden structure in the data, for example, clustering
algorithms (Guyon and Elisseeff 2003 ;Dy 2004 ) or Markov modelling (which may
be used for either clustering or classi

cation) (Obermaier et al. 2001 ). They may use
translations or transformations to reduce the dimensionality of the dataset and hence
select subsets of the data that may better illustrate or highlight features or structures
of interest.