Flat optics for information processing

Abstract: Digital neural networks and the availability of large training datasets have allowed for rapid progress in the performance of machine-based tasks for a wide range of applications including image analysis, sound recognition, and natural language translation. The enhanced capability has, however, come at a computational cost ultimately resulting in an unsustainable growth in energy consumption and restrictions in real-time decision making when large computational systems are not available. In this talk, I will discuss the development of flat optic computational accelerators that can off-load computationally expensive operations into high-speed and low-power optics. The key to these architectures are the new freedoms afforded by nanostructured flat optics, or metasurfaces, including non-locality, polarization discrimination, and the ability to spatially multiplex, and demultiplex, information channels. I will discuss how these freedoms can be utilized for realizing optical metasurface front-ends that are co-designed with light weight digital back-ends for a variety of tasks including edge detection, segmentation, and objection classification. In all of these cases, a significant portion of the computational requirements can be off-loaded into passive front-end optics leading to increased processing speed while also reducing the energy consumption. This general architecture more efficiently bridges the gap between the natural world and digital machine vision systems and could be particularly advantageous in autonomous systems with limited onboard power or the necessity for high speed data processing.