4.4.1. The Multi-Layer Perceptrons
MLP software should be implemented with the capacity to vary the number of input nodes (equal to the number of quantised frequency bands), hidden layers, hidden nodes and output nodes. Recurrency should be available (Jordan or Elman types). The backpropagation algorithm should incorporate a momentum term and should record the total output error per epoch to a file.
4.4.2. The Kohonen Self Organising Map
The SOM should be of variable input and grid size. A method of displaying the ordering of the map should be available. Learning Vector Quantisation should also be optional.
4.5. Post classification
Where the MLP makes an unlikely multiple classification, a decision structure should be used to determine probable type. The amplitude of each drum should be used to determine whether it is a hard hit or light hit. The frequency of maximum magnitude for each drum identified should be used to group these drums into different pitch types (eg low or high conga; high, medium, floor tom).