Optical Computing for Machine Learning with Integrated Waveguides

Computational resource demand for large neural networks (NN) has increased exponentially. Therefore, both electrical and optical research communities are looking for new means of computing as an alternative to von Neumann architecture. Optics' inherent parallelism promises efficient computing substrate using light [1]. There are two main approaches in the field: replacing digital operations with an analog optical counterpart, or direct implementation of neurons internally by an optical system. A common approach is performing matrix-vector multiplication operations, frequently used in NN calculations and heavy for digital computers, using interference with photons. Although improved energy efficiency, the high cost of optoelectronic conversion and the lack of optical nonlinearities on such systems become a burden to maintain scaling [2]. Another approach is to use the propagation of light in a system, such in optical waveguides to effectively mimick a complex analog neural network [3]. Here, we demonstrate a technique that relies on ultrafast pulse propagation in a ridge waveguide based on LiNbO 3 (LN) on an insulator.