Backpropagation decision region formation properties

Summary form only given, as follows. An experimental analysis was performed to determine decision region formation properties of the backpropagation (BP) paradigm for a 'winner-take-all' network using sparse training sets. Key objectives were to determine whether supervised training using the BP paradigm would result in nearest neighbor decision regions, and whether the resulting region boundaries were independent of the initial conditions of the network link weights. Using a three-layer network (i.e. an input layer, an output layer, and a single hidden layer), the experiments showed that the convergent decision regions were not necessarily nearest-neighbor solutions, nor were the regions necessarily independent of the starting condition. These results suggest precautions in designing adaptive networks, especially for high-dimensional networks (i.e. more than two input/output units) where the decision regions are not easily visualized.<
>