The parallelism of optics and dramatic miniaturization of optical components using nanophotonic structures, such as meta-optics, present a compelling alternative to electronic implementations of convolutional neural networks. However, lack of low power reconfigurability and optical nonlinearity presents serious challenges for creating an optical neural architecture, that can outperform their electronic counterpart.
In this webinar hosted by the Nanophotonics Technical Group, Arka Majumdar will present his work on phase-change material based ultra-low power nonvolatile switches, which can provide the reconfigurability needed for optical neural networks. Additionally, Prof. Majumdar will discuss a new architecture that utilizes a single electrical to optical conversion by designing a free-space optical frontend unit that implements the linear operations of the first layer with the subsequent layers realized electronically. They found that by intelligently designing the meta-optical frontend, they can improve the classification accuracy while minimizing power consumption and latency over a fully electronic neural network. Such hybrid optical-digital systems can not only implement optical neural networks but can also improve the quality of images captured via meta-optics. Finally, Prof. Majumdar will discuss ways the nonlinear activation can be implemented in the optical domain.
Subject Matter Level: Intermediate - Assumes basic knowledge of the topic
What You Will Learn in the Webinar:
• Optical neural network
• Phase-change material based ultra-low power nonvolatile switches
Who Should Attend the Webinar:
• Undergraduate students interested in practical applications of optics
• Graduate students and postdoctoral researchers in the field of nanophotonics
• Researchers who wish to keep up with the latest optical nanotechnology