Abstract:
We present an overview of our design for a fully digital hardware implementationof the Self Organising Map (SOM) [1]. Our approach has resulted in a modular system (Modular Maps) which utilises fine grain parallelism with each neuron being a separate entity implemented as a small RISC processor. The essence of the SOM has been maintained by this design although minor modifications have been made to the original algorithm to facilitate implementation. Modules can be used as either stand alone systems or combined to enable large networks to becreated and large input vectors to be catered for. A simulator system was developed to facilitate investigation into the high level behaviour of Modular Map systems, and as Modular Maps are computationally intensive and parallel in nature, it was implemented on a parallel computer system. A series of simulations was carried out using encoded images of human faces where it was found that the classification accuracy of a Modular Map system offered an improvement over that of the traditional SOM.