Reconfigurable scattering potential for enhanced optical neuromorphic operations

Effective nonlinearity associated with multiple scattering in the linear optics regime has been demonstrated to be a promising approach to devise a hybrid optical-digital neural network with high energy efficiency as only low-power light sources are required, and the training or inference cost of simple digital neural networks is minimal. We show that the performance of such a system can be further enhanced if the scattering potential involved is tunable or reconfigurable, such as that provided by a liquid crystal-polymer composite under applied electrical voltages. Our study shows that the reconfigurable scattering potential enables one to explore and search for an optimal optical neuromorphic operation regime with higher classification accuracy. We also demonstrate what we believe to be, a new paradigm of optical ensemble learning, which improves the learning performance and inference accuracy by combining inference results from different applied voltages.