Digital chip architecture for the emulation of a biology-oriented neural network

Abstract

The circuit design and chip architecture for an emulator for an exemplary biology-oriented neural network are presented. This emulator is able to emulate 16 fully interconnected neurons of the Marburg type. It is cascadable to larger fully interconnected networks and multilayer networks with or without feedback. It is shown that chip architectures not suffering from a connection problem can be found even for complicated neural network architectures. The full interconnection and the multilayer architecture can be transformed to matrices. Neighborhood networks such as those used in lateral inhibition networks can also be transformed to a regular structure.<>