Solving two-spiral problem through input data representation

This paper studies the effect of input data representation on the performance of backpropagation neural network in solving a highly nonlinear two-spiral problem. Several popularly used data encoding schemes and a proposed encoding scheme were examined. It was found that input data encoding affects a neural network's ability in extracting features from the raw data and therefore the network training time and generalisation property. Using a proper input encoding approach, the two-spiral problem can be solved with a standard backpropagation neural network.