Musical Notation and Fine Motor Skills when Playing the Soprano Recorder: Making A Neural Network Lift a Finger

Most studies on notation-fingering mapping used the piano where one finger covers one key to produce one tone. This study used the recorder as a model, where learning of finger combinations is needed to produce one tone. In three simulations, a 13 x 5 x 5 x 10 three-layer feedforward neural network was required to transform a binary spatial representation of a musical notation of the chromatic C ’ scale on the stave into the appropriate fingering output on the ten holes of the Soprano recorder. The network could play random sequences of tones and ‘Amazing Grace ’ successfully. It was not the case that musical notes were gradually grouped into large chunks, as assumed, but instead the network just grouped pairs of tones and halftones, and worked harder on adapting the motor output to the properties of the instrument. An initial network with few internal nodes learned to keep all fingers down (activation), but when the number of internal layer nodes matched output nodes, the network could develop an activation/inhibition representation in the connections between internal layer and output nodes to keep fingers down, but also lift them. A robust finding in all three network simulations was that the activation and inhibition pattern indicating motor flexibility showed particularly in the lower part of the Soprano Recorder. This was explained with statistical learning, as fingers in the lower part of the instrument could stay in place for low pitch tones, but needed to be lifted for high pitch tones, while in the upper part of the instrument, fingers could stay in place most of the time. Hence the network systematically developed motor flexibility on a localized part of the spatial scale of an instrument determined by statistical learning