Improved artificial digital neuron, neural network and network traning algorithm

摘要

The invention relates to an improved artificial digital neuron, an enhanced artificial neural network architecture together with a reduced training neural network training algorithm. The digital neuron comprises an n bit input one bit output device utilising a random access memory having 2" locations of one bit each. The neuron is programmed by the network training algorithm with the neuron fire state (on or off) in accordance with the synaptic weights allocated to the input code combination pattern for that neuron within the neural network. The digital neuron operates as a look-up table device deriving the neuron firing state, for each n bit digital code pattern applied to its input paths, from the random access memory location identified by the digital code pattern. The synaptic weights allocated to each digital code pattern for a particular neuron in the network cause the neuron to provide as an input correlation function or an interconnect function. The network architecture is based upon digital neurons providing either correlation or interconnection functions allowing a reduced interconnection arrangement where the output of each neuron in the input and hidden layers is connected only to an input path of a corresponding or immediately above or below neuron in the next layer. The training algorithm maps each pattern (as a function or interconnect) into neurons that lie on the path between a positive input (1) and the network output.