Two-layer all-optical deep neural network with photonic integrated weighted addition

Deep neural network architectures have shown their superior performance in visual classification, audio recognition, signal recovery, astronomy information processing, etc. Their intrinsic parallel computation scheme can boost computation speed and reduce power consumption. Electronic neuromorphic computing has already been developed to reduce processor power consumption. However, the computing speed is limited by the interconnection electronic bandwidth. We propose to use photonic computation as an alternative to pushing computing speed further while exploiting brain-inspired architectures. We employ an Indium Phosphide (InP) based photonic neural network to demonstrate an all-optical two-layer feed-forward neural network. A non-linear semiconductor optical amplifier (SOA) is used to generate the optical nonlinear activation function: it is possible to change the shape of the nonlinear function by tuning the driven current and the power ratio of the CW control laser to the optical input signal. The combination of the on-chip photonic cross-connect for weighted addition implementation and the off-chip optical nonlinear function allows to process 10Gbit/s data sequences with a normalized root mean square error of 0.21 at the very final output, which is comparable to using an E/O conversion approach after each layer. This result paves the way to implement a reconfigurable all-optical deep neural network with photonic integrated circuits with co-integrated optical amplifiers.