Integrated high-neuron-density diffractive neural networks  performing near-infrared inference         

Dr. Elena Goi, University of Shanghai for Science and Technology, Shanghai, China

Abstract

Optical machine learning has emerged as an important research area that, by leveraging the advantages inherent to optical  signals such as parallelism and high speed, paves the way for a future where optical hardware can process data at the speed  of light. In this work, such optical devices for data processing in the form of multi-layer nanoscale diffractive neural networks  trained to perform optical inference tasks are presented. We show the functionality of these passive optical devices on the  example of decryptors trained to perform optical inference through symmetric and asymmetric decryption and multi-layer  diffractive neural networks for direct phase retrieval. The perceptrons, designed for operation in the near-infrared region,  are nanoprinted on complementary metal-oxide–semiconductor chips by galvo-dithered two-photon nanolithography with  axial nanostepping of 10 nm, achieving a neuron density of 108 neurons / mm3. The compact form factor of the resulting  optical neural networks and the lithographic fabrication technology that allow for directly integration on opto-electronic  sensors, enable the co-integration of the optical perceptrons with additional layers of electronic neural networks, or the  use of the sensor’s nonlinear response as nonlinear activation function, in this way forming deep neural networks. This  power-efficient commixture of machine learning and on-chip integration may have a transformative impact on optical  decryption, sensing, microscopy, high-precision laser nanolithography, medical diagnostics, and computing.