creating artificial lesions or suggesting biologically plausible networks (i.e. neurobiological
models with autonomous and adaptive behaviour).
The present chapter reviews some artificial neural networks that have been proposed in
auditory and music perception. Neural nets are helpful in understanding how we learn
musical patterns by mere exposure, how these patterns might be represented and how this
knowledge arising from acculturation influences perception. To date, applications of net-
works for music cognition have been restricted to behavioural data. However, future develop-
ments in combination with neuropsychological cases and brain imaging methods
should make it possible to establish links to brain structures and neural circuitry for the
investigation of music perception. These developments are in progress in other domains
(e.g. language perception, visual perception, memory). In these domains, artificial neural
networks are used not only to simulate cognitive processes and behavioural data, but to
postulate correspondences to cerebral structures, their interactions and the effects of brain
damage. For behavioural data, a variety of networks have been proposed in connectionist
psycholinguistics for language acquisition, sentence processing, reading, and production
(cf. Ref. 1 for a recent review). Well-known networks developed by McClelland and
colleagues on visual word and speech perception2–5simulate a variety of behavioural data
and generate numerous hypotheses for further behavioural experiments. The creation of
‘artificial lesions’ in neural net models simulating behaviour of healthy participants have
enabled the simulation of neuropsychological cases, and have provided further insights
into the normal functioning of the brain, possible representations of knowledge and the
modularity of subprocesses (see Ref. 6 for a review). Simulations in computational
neuropsychology have been proposed by eliminating network units or by weakening
connections in order to model clinical cases of aphasia,7–9dyslexia10,11or semantic impair-
ments.12,13For example, comparing the behaviour of damaged connectionist models to
double-dissociations observed in the behaviour of brain-injured patients suggests that
behavioural deficits arise from impairments of two types of lexical information (semantic
and phonological) and explains past tense formation of English verbs without rules or
symbolic mechanisms.14,15
In what follows, we briefly review some basics of neural net modelling and summarize
the regularities underlying Western tonal music. Next, we present our recent work with
Self-Organizing Maps (SOMs) as an example of neural networks providing further insights
to music cognition. We then suggest some possible, though still speculative, future direc-
tions in which neural nets can help us further to investigate the processing of music,
notably by simulating neuropsychological cases and establishing links to neural structures
and circuitry.
Some basics of artificial neural networks
A principal advantage of artificial neural networks is their capacity to adapt in such a way
that representations, categorizations or associations between events can be learned.
Connectionist models have the characteristic that rules governing the materials are not
explicit, but emerge from the simultaneous satisfaction of multiple constraints represented
110