High density VLSI implementation of neural networks and fuzzy systems

A new design flow for high density neuro-fuzzy systems is presented, based on the basic building blocks offered by the current-mode design style rather than on imitating one or other neuro-fuzzy mathematical model. A comparison between digital and analog approaches is carried out, concluding that the current-mode design technique is the most suitable for high density VLSI implementations of neural and fuzzy systems. Original implementation solutions for a nonmonotonic activation function and for a weighted summation module followed by a hard nonlinearity (CMOS perceptron) are presented.