An analog-current mode local cluster neural net

The local cluster (LC) artificial neural net is a special kind of multilayer perceptron where the sigmoid functions combine in clusters that have a localised response in input space. The proponents of the LC architecture have shown that it is versatile and trains well. They also suggested that the LC nets could be suitable for realisation in analog VLSI. We investigated the feasibility of an analog realisation of LC nets by following through the complete cycle from design to fabrication. We found that the all the required mathematical functions call be realised in current mode bipolar and CMOS circuits. In this paper we discuss the main design issues paying special attention to the alternative training regimes for an LC chip.