In prior reading for this course, much of the focus has been on the capacity the human body for hearing and for ‘universals’ – or processes that are present in our biopsychological composition – in the perception of music. This chapter presents a significant departure from such a stance. Indeed, this article focuses on computer-simulated neural networks that ‘learn’ musical structure passively through mere exposure to Western music. This would represent the extreme of “nurture” in the seeming “nature versus nurture” dialogue that we have been exploring regarding the influence of human psychological composition versus acculturation in the processing of music by humans.
In this chapter, the authors present a summary of findings from the use of artificial neural networks on computers on the learning and perception of musical structures. Introductory material is given as to how such artificial neural networks function. The network consists of units (simulating a population of neurons) which ‘learn’ through repeated exposure. The patterns of learning in such a system, state the authors, can demonstrate how mere exposure (or acculturation) can function to influence musical perception.
The recent work of the authors involved what is called a Self-Organizing Map (SOM), in which units map themselves according to what is learned. The authors set up the neural network so that three levels (or “orders”) of musical syntax are recognized: tones (reduced to twelve through octave reduction), chords (three simultaneous tones) and keys (based on tone and chord recognition). By inputting tones and chords into the network, the SOM then creates the connections between the data to simulate passive learning of musical patterns (all of them from Western music in the research presented). One premise of such a system is that musical structures that are similar will emerge in close proximity in the SOM, while musical structures that are less similar will emerge in more distant proximity. This is what the researchers found; for example, key relationships emerged in a spatial pattern similar to the circle of fifths in music theory, where the keys with the most related tones are found next to each other in the circle. The authors note that the current data only relate to pitch and recognize the necessity for using a network where both pitch and rhythm are used in neural networking in future research; methods to do such are briefly discussed
The researchers summarize their findings: “Overall, the simulations showed that activation in the trained self-organizing network mirrored data of human participants in tonal perception experiments. This outcome suggests that the level of activation in tone, chord, and key units is a single unifying concept for human performance in different tasks” (p. 116).
Though the findings in the presented research are not surprising or beyond what would be expected, the promise of future developments in the area of artificial neural networks is promising, and this chapter demonstrates this potential. Though the current abilities of such networks seem limited, they nonetheless are not meaningless. As technology and neuropsychological science progresses, we can reasonably expect networks with greater power that could in turn be used for research with greater implications.
Artificial neural networks are a set of artificial neurons linked by different strengths of synaptic connections stimulating various levels of perception and cognition. Their primary purpose is to create a mathematical/computational model simulating the biological neural network. A mouthful, yes, but an integral part of every individual’s perceptive faculties, music being no exception.
ReplyDeleteIn layman’s terms, artificial neural networks are a way for scientists to create topographical representations of how the mind perceives and creates connections and sensitivity to regularities. It has been used for developments and research in language perception, visual perception, memory, and in certain aspects of music e.g. pitch regularities. Two approaches are commonly used viz. ART (adaptive resonance theory) and SOM (self-organizing maps). ART is used predominantly for tracking pattern recognition and prediction. SOM uses unsupervised learning to create a topological map of music cognition. Essentially and very simply, it creates a visual representation of data input.