Dynamic scheduling for feed-forward neural nets using transputers

The modeling of neural networks on conventional digital computers can be a very time consuming operation. The authors evaluate one way to ease this time problem by mapping the processes involved onto an array of parallel processors. The neural approach to computing is inherently parallel with a fine level of granularity. This is to some extent incompatible with commercially available parallel processing systems, and in particular transputer-based systems. However, by exploiting the parallelism in the training or classification data, multi-transputer-based systems can efficiently model neural processing for a wide range of real-world problems. The paper describes a dynamic load balancing arrangement, based on a division of the training data, that produces near-linear improvement against the number of processors in use