Pulse stream current mode CMOS CNN chip

Abstract

A new idea of analog cellular neural network (CNN) VLSI implementation is described. The main problem in neural networks realization is the size of adjustable connections between neurons in the net. The weight circuits may occupy more than 90% of the single neuron space. The main advantage of current mode pulse stream neurons is small size of the weight circuit. The simplicity of the building blocks used allows one to operate them with relatively high speed. As compared with the other pulse stream circuits, this approach is based on current pulses. Therefore, for both basic neural operations, summation and multiplication are easy to implement, if only the phase is irrelevant.