Implementation of the selectron on transputers
Frans Moyson · 2003
When a system consists of a few hundred or more processors, it becomes virtually impossible to program each and every processor by hand with current software techniques. A possible way to avoid this problem is to let the machines program themselves, as advocated in the connectionist approach. The author describes the implementation of an attempt at doing just that. Learning in connectionist networks is a consequence of self-organizational behavior. A problem is that the learning procedures take an extremely long time. An alternative called the selectron is implemented and discussed at length. A selectron is an abstract device consisting of a collection of units, organized in layers, all having the same number of units. It has one input layer and one output layer, with hidden layers in between. The communication goes between units in different layers. It is a formal automation, induced to implement a method for the self-organization of massively parallel networks based on the idea of selectionism.>