Digital neural network implementation

Abstract

The authors provide a comparison of implementation approaches for digital neural networks. Digital neural networks of large size are feasible if the inputs and outputs are single-bit binary signals. A key component for this application is the parallel counter, which counts the number of inputs that are ONEs. Progress is reported toward the implementation of parallel counters with up to 1022 inputs, as required to realize multilayer neural networks with up to 1000 neurons per layer.<>