unit, a tone unit) is activated, the more stable the musical event is in the corresponding
context. For the experimental tasks, it was hypothesized that the level of stability affects
performance (e.g. a more strongly activated, stable event is more expected or judged to be
more similar to a preceding event).
For perceived relations between chords, the model succeeded in simulating data obtained
with similarity judgements,49,53recognition memory,^49 harmonic priming52,59,88,89and electro-
physiological measures.^69 Harmonic priming and ERP data provided evidence for the
development of harmonic expectations in a prime context that then influence the process-
ing of a target chord: a target is processed more slowly and less accurately and evokes a larger
late-positivity component LPC, when it is unrelated to the context than when it is related.
When the experimental material is presented to the network, activation levels of chord units
mirrored patterns of processing in the priming task (e.g. with higher activation for chord
units representing facilitated targets), and activation changes in chord units after the target
mirrored the amplitude of the LPC (e.g. with stronger activation changes for distant key tar-
gets). Both human listeners and network are sensitive to changes in chord stability caused
by key contexts. Further behavioural experiments have shown that human listeners perceive
keys underlying a given context, detect modulation (i.e. temporary changes in key) occur-
ring in a musical excerpt and have implicit knowledge of distances between keys.46,47,90A
further set of HSOM simulations showed that activation levels of key units and activation
changes over time mirror listeners’ behavioural data. For example, the changes of activation
patterns in key units were more important when the excerpt modulated to distantly related
keys.^46 As activation accumulates in the network over time, the model also tracks the key
changes in a sequence and simulates a dynamic aspect of tonality sensation.
The HSOM network also simulated behavioural data on the perceived relations between
tones, even if it was trained with chords only and not with melodies. Notably, the activa-
tion levels of tone units simulate the stability profiles obtained in probe-tone ratings,^90 the
patterns of similarity ratings^32 and memory for melodies.^56 For example, when human lis-
teners rate the similarity of a tone pair presented after a tonal context, the ratings reflect the
differences in tonal functions and show a perceived asymmetry with higher similarity for a
pair ending on a stable tone. The activation levels and activation changes of the tone units
mirrored these data sets.
Overall, the simulations showed that activation in the trained self-organizing network
mirrored data of human participants in tonal perception experiments. This outcome sug-
gests that the level of activation in tone, chord, and key units is a single unifying concept
for human performance in different tasks.
General discussion
The HSOM network was presented as an application of artificial neural networks to further
our understanding of learning and perceiving music. Based on self-organization (i.e. a
general unsupervised learning algorithm), the structure of the network adapted to tonal regu-
larities through repeated exposure to musical material. The network combines three levels of
music perception which can be placed on the top of networks simulating lower perceptual
processing steps (pitch extraction, octave equivalence). After training, the network
116