Implementation of Kohonen's self-organising maps on MANTRA I

This paper presents a computer dedicated to neural algorithms. The computational core of the system is a systolic array of dedicated processing elements (PEs). This accelerator provides the primitives to implement several widely used neural models, including Kohonen's Self-Organizing Maps (SOMs). This paper concentrates on the problems related to the mapping of the SOM algorithm on the machine. Some modifications to the algorithm, required to parallelize it efficiently, are also proposed. Some aspects of the complexity of programming a systolic system are discussed and the chosen software approach is presented.