Shadow targets: A novel algorithm for topographic projections by radial basis functions

The archetypal neural network topographic paradigm, Kohonen's self-organising map, has proven highly effective in many applications but nevertheless has significant disadvantages which can limit its utility. Alternative feed-forward neural network approaches, including a model called 'NeuroScale', have recently been developed based on explicit distance-preservation criteria. Excellent generalisation properties have been observed for such models, and recent analysis indicates that such behaviour is relatively insensitive to model complexity. As such, it is important that the training of such networks is performed efficiently, as computation of error and gradients scales in the order of the square of the number of patterns to be mapped. We therefore detail and demonstrate a novel training algorithm for NeuroScale which outperforms present approaches.