Digital implementation of programmable neural networks in synchronous pulse mode

In this paper, we propose a new digital implementation of neural networks which operates in synchronous pulse-mode and allows massively-parallel processing. The basic signaling scheme is devised based on the Manchester encoding of pulse-stream signals. This pulse mode coding allows signed addition with a single OR gate and signed multiplication with a single AND gate and a two-to-one multiplexer. Each neuronal element keeps a set of registers which hold the information of weights. The activation function is implemented as a look-up-table (LUT). This structure provides programmability in weights, the number of inputs, and activation function of a neuron.