Analog circuits for self-organizing neural networks based on mutual information
Janusz A. Starzyk, Jing Liang · 2002
Comparing to conventional neural network structures, this organization greatly reduces the interconnections in neural network by using local interconnection based on statistical analysis, and eliminates the need to store large number of synaptic weights. The network is characterized by evolvable hardware structure and adjustable threshold values based on selection criteria, which use mutual information. Next, a mix-signal implementation scheme is proposed for this organization in order to achieve the best performance. The digital implementation is used for the evolvable structure of the network for the better ability to be reconfigured. Analog implementation is used for the entropy-based evaluator (EBE), which is used for statistical analysis and mutual information evaluation, in order to achieve smaller area and faster on-chip learning process. Either on-chip analog memory or off-chip digital memory can be used to store the threshold values of the neurons and organization of resulting interconnection of neurons. Finally, circuits used for the analog implementation of the EBE are presented, the simulation results of the circuits are shown and discussed.