Abstract
Today, machine learning tools, particularly artificial neural networks, havebecome crucial for diverse applications. However, current digital computingtools to train and deploy artificial neural networks often struggle withmassive data sizes and high power consumptions. Optical computing providesinherent parallelism accommodating high-resolution input data and performsfundamental operations with passive optical components. However, most of theoptical computing platforms suffer from relatively low accuracies for machinelearning tasks due to fixed connections while avoiding complex and sensitivetechniques. Here, we demonstrate a genetically programmable yet simple opticalneural network to achieve high performances with optical random projection. Bygenetically programming the orientation of the scattering medium which acts asa random projection kernel and only using 1% of the search space, our noveltechnique finds an optimum kernel and improves initial test accuracies by 8-41%for various machine learning tasks. Through numerical simulations andexperiments on a number of datasets, we validate the programmability andhigh-resolution sample processing capabilities of our design. Our opticalcomputing method presents a promising approach to achieve high performance inoptical neural networks with a simple and scalable design.