Optical pulse subtractor for fiber neural networks

Negative weight modulation has always been a challenge in constructing fiber neural networks. This problem can be alleviated by separately modulating neighboring pulses with positive weights and negative weights, respectively, and performing subtractions after PD detection electronically. Therefore, both inserting a nonlinear activation module and cascading layers optically become hard for fiber neural networks since all these two operations are based on subtractions of four neighboring pulses so as to obtain the final results after each layer. In order to address this problem, a novel scheme for an optical pulse subtractor, to the best of our knowledge, is proposed and demonstrated in this manuscript. By adopting devices such as a tunable attenuator, a beam splitter, and a photo-diode, this optical pulse subtractor can do both optical pulse subtraction operation and the nonlinear activation function optically, which will later become a powerful tool for deep fiber neural network construction.