Optical hardware backpropagation neural network

Optical error backpropagation is experimentally demonstrated in a feedforward optical neural network developed by the authors. To our knowledge, this is the first report of hardware backpropagation training in an optical system. The network uses the trainable light steering behavior of an optical nonlinear material to implement both neural processing and connectivity. The nonlinear material steers (by modulation of the phase front) a forward propagating information beam by dynamically altering the index of the refraction profile of the material via a stronger weighting beam. Effectively, the weight beam creates spatially varying lensing effects in the nonlinear material. These trainable effects steer the information beam to produce the correct output value at an optical detector. A photorefractive crystal is used as a phase conjugate mirror to generate the backpropagated optical error. This backpropagated optical error is detected and used in a gradient descent algorithm to update the weighting beam profile in real time. Both computer simulation and experimental results are presented.