simulates a large set of experimental data on perceived relations between tones, chords, and
keys. These simulations proposed activation as a unifying mechanism for different cognitive
tasks. A context activates the network and the activation state then influences the perception
(i.e. expectation, perceived similarity, memory). The HSOM network provided a model not
only for the learning of implicit knowledge about the tonal system, but also for a low-
dimensional and parsimonious representation of tonal knowledge. For example, the net-
work proposes a parsimonious account of the contextual dependency of musical functions
of an event and the changes associated with different keys. One single event has not mul-
tiple instances of representation, but its corresponding function in a given key emerges from
activation spreading in the system. As a consequence, events representing prototypes and
anchor points in a given key do not have to be stored separately, but emerge from the activa-
tion levels. A further emerging property of the network relates to the identification of a key;
the underlying key does not need to be inferred employing separate processing steps, but
emerges from the activation pattern in the key units.
The network also offers a framework for generating new predictions for further behavi-
oural studies on music perception. These studies and associated simulations could investi-
gate the perception of tones, chords, and keys. The experimental material will be presented
to the network and on the basis of the activation levels precise hypothesis can be derived
for the performance of human listeners. For example, the studies could be related to key
identification (e.g. number of notes necessary to establish a key, disturbing effect of an
unrelated event), to key modulation (e.g. how long the trace of a key remains) or to the
eventual link between activation decay and musical short-term memory span.
In the following, we point out two limitations of the HSOM network and discuss future
developments of the model. A first limitation is the restriction to pitch structures only, and
the second one is that the network is more abstract than brain structures and neural circuitry.
The simulations have focused on how regularities of the pitch dimension may be intern-
alized through passive learning processes. Although pitch is the most obvious form-
bearing dimension of Western tonal music, temporal regularities also contribute to listeners’
perceptual experience. Temporal regularities include the sensation of meter (sensation of a
regular succession of strong and weak beats superimposed over an isochronous pulse) and
patterns of intervals creating rhythms perceived against the metrical background. Two
frameworks have been proposed to account for the respective contributions of regularities
of pitch and time in the perception of musical events.^91 A single-component model^92 sug-
gests that temporal and harmonic accents are not processed independently, but are integ-
rated and together guide the attention of the listener during the unfolding musical
piece.51,54,93,94Based on neuropsychological cases (with double-dissociations of amelodia
without arhythmia and vice versa), and experimental data,91,95–98Peretz and Kolinsky^91 pro-
posed a two-component model suggesting independence between temporal and nontem-
poral information processing which are integrated only at a later stage. Two types of neural
net architectures can be used to mimic these two theoretical approaches and to account for
learning and perception of regularities in pitch and time. A single-component neural net
may learn the two regularities conjointly without a supplementary integrative step, by
adapting input codings of metrical information as proposed by Berger and Gang99,100or by
Stevens and Wiles.^101 A two-component network may learn separately regularities of pitch
117