Training Photonic Mach Zehnder Meshes for Neural Network Acceleration

Photonic neural networks enable faster and energy-efficient inferences for deep neural network implementations when compared to electrical counterparts. Prior training methods for photonic neural networks modify the network or decompose the inputs or weights to simplify the training, however that could impact accuracy. Electrical training uses backpropagation that needs to partially differentiate each of the weights with respect to the loss function on a number of examples to achieve high accuracy. While the traditional backpropagation algorithm can be applied to electrical networks, it is difficult to directly apply to photonic neural network with Mach-Zehnder Interferometer (MZI) meshes that have different parameters (frequency and phase). In this paper, we implement the traditional backprop-agation algorithm to train the parameters of the MZI meshes with gradient descent. Parameters of the mesh are trained by implementing a modified backpropagation algorithm on the hardware, and those operations may be performed on device. The implementation does not modify the circuits or add any additional lasers and therefore, self-trains the MZI to achieve the desired accuracy and achieve O(n2) speedup over electrical methods. We test our training algorithms on MNIST datasets and show that our method improves the training when compared to state-of-the-art electrical backpropagation.