Implementation of Free-Space Optical Networks based on Resonant Semiconductor Saturable Absorber and Phase Light Modulator

Optical Neural Networks (ONNs) have gained traction as an alternative to the conventional computing architectures used in modern CPUs and GPUs, largely because light enables massive parallelism, ultrafast inference, and minimal power consumption. 

As with conventional deep neural networks (DNNs), free-space ONNs require two main layers: (1) a nonlinear activation function which exists to separate adjacent linear layers, and (2) weighting layers that applies a linear transformation given an input.

Firstly, a Resonant Semiconductor Saturable Absorption Mirror (RSAM) was investigated as a viable nonlinear activation function. Several mechanisms have been used to create nonlinear activation functions, such as cold atoms, vapor absorption cells, and polaritons, but these implementations are bulky and must operate under tightly controlled environments while RSAMs is a passive device. Compared to typical SESAMs, the resonance structure of RSAM also reduces the saturation fluence compared to non-resonant SAMs, allowing low power laser sources to be used. A fiber-based optical testbed demonstrated notable improvement of 8.1% in classification accuracy compared to a linear only network trained with the MNIST dataset.

Secondly, Micro-electromechanical-system-based phase light modulators (PLMs) were evaluated as an alternative to LC-SLMs for in-situ reinforcement learning. PLMs can operate at kilohertz-scale frame rates at a substantially lower cost compared to LC-SLMs but have lower phase resolution and non-uniform quantization which impacts fidelity. Despite these disadvantages, the high-speed nature of PLMs allows for significant decrease in optimization time, which not only allows for reduction in training time, but also allows for larger datasets and more complex models with more learnable parameters. A single layer optical network was implemented using policy-based learning with discrete action-space to minimize impact of quantization. The testbed achieves 90.1%, 79.7%, and 76.9% training, validation, and test accuracy, respectively, on 3,000 images from the MNIST dataset. Additionally, we achieved 79.9%, 72.1%, and 71.7% accuracy on 3,000 images from the Fashion MNIST dataset. At 14 minutes per epoch during training, it is at least a magnitude lower in training time compared to LC-SLMs based models.