Key-finding by artificial neural networks that learn about key profiles

We explore the ability of a very simple artificial neural network, a perceptron, to assert the musical key of novel stimuli. First, perceptrons are trained to associate standardized key profiles (taken from 1 of 3 different sources) to different musical keys. After training, we measured perceptron accuracy in asserting musical keys for 296 novel stimuli. Depending upon which key profiles were used during training, perceptrons can perform as well as established key-finding algorithms on this task. Further analyses indicate that perceptrons generate higher activity in a unit representing a selected key and much lower activities in the units representing the competing keys that are not selected than does a traditional algorithm. Finally, we examined the internal structure of trained perceptrons and discovered that they, unlike traditional algorithms, assign very different weights to different components of a key profile. Perceptrons learn that some profile components are more important for specifying musical key than are others. These differential weights could be incorporated into traditional algorithms that do not themselves employ artificial neural networks. (PsycINFO Database Record