A Fast Kohonen Net Implementation for Spert-II

We present an implementation of Kohonen Self-Organizing Feature Maps for the Spert-II vector microprocessor system. The implementation supports arbitrary neural map topologies and arbitrary neighborhood functions. For small networks, as used in real-world tasks, a single Spert-II board is measured to run Kohonen net classiication at up to 208 million connections per second (MCPS). On a speech coding benchmark task, Spert-II performs on-line Kohonen net training at over 100 million connection updates per second (MCUPS). This represents almost a factor of 10 improvement compared to previously reported implementations. The asymptotic peak speed of the system is 213 MCPS and 213 MCUPS.