Binary Pattern Classification with a Photonic Neuromorphic Device Based on Optical Receptrons

Photonic systems for data processing are receiving increasing interest because of their potential to realize fast and massively parallel computing at low power consumption. Many challenges still must be faced and overcome due to the complexity of the fabrication and integration of optical computing devices replicating the architectures of Complementary Metal-Oxide-Semiconductor (CMOS) systems, difficulties in implementing optical nonlinearities, and lower performance compared with their digital counterparts. Recently, the use of an elemental building block, called receptron is proposed, for the realization of optical neuromorphic systems for classification. The receptron model is used for the implementation of an all-optical device that exploits the non-linearity of optical speckle fields produced by a solid scatterer. Here the practical implementation of an array of optical receptrons for the classification of digital patterns (handwritten digits) is reported with training based on a random search protocol substantially different from standard machine learning approaches. The optimization of a very simple hardware set-up and of binary data manipulation is obtained by a general matrix formalism that allows the a priori determination of the distinguishability and recognizability of different classes by the network.