How to modify Kohonen's self-organising feature maps for an efficient digital parallel implementation

Two new variants of Kohonen's self-organising feature maps based on batch processing are presented in this work. The motivation is related to the need of exploiting the hardware resources of neurocomputers based on systolic arrays. Ordering and convergence to asymptotic values for 1D maps and 1D continuous input and weight spaces are proved for both variants. Finally, simulations on uniform 2D data as well as simulations on speech 12D data using 2D maps are also presented to back the theoretical results.