Quantization-aware training of optical random neural networks

One of the important requirements for today’s machine learning algorithms running at the edge is real-time inference. In this regard, quantization-aware training (QAT) is incorporated into the training of conventional digital neural networks. Optical computing, offering versatility, inherent parallelism, and speed, is a suitable platform to test QAT. Here, we implement two strategies for QAT with optical random neural networks and show improvements in classification accuracy for a decision problem. Our results indicate potential applications of optical computing platforms in data-scarce settings where real-time performance is a concern.