High-Speed, Low-Power Computing of Convolutional Neural Networks using Meta-Optic Accelerators

Artificial intelligence algorithms have drastically improved the performance of persistent target recognition platforms for the United States Navy and Department of Defense as a whole. The key to these algorithms is deep learning, where new insights have allowed for the training of larger neural networks producing unprecedented results in machine-based tasks. However, these gains have come at the cost of reduced computational speed and increased power requirements. While electronic accelerators such as GPUs and TPUs have partially solved these issues, technologies that offer additional decreases in latency and power consumption could enable real-time, high-accuracy inference at the edge. In this work, we will develop a meta-optic frontend accelerator for a convolutional neural network object classification algorithm. Meta-optics are structured optical media that provide subwavelength control of phase, amplitude, and polarization in a flat form factor. We will design a near infrared or visible meta-optic accelerator using forward design methods and then fabricate and characterize a proof-of-concept device. Quantitative metrics such as latency reduction, decreased power consumption, and classification accuracy will be assessed. We will also develop a fully differentiable end-to-end pipeline for co-optimizing the meta-optic frontend and digital backend to improve channel density, bandwidth, field of view, and classification accuracy. By the end of Phase I, we expect to have a demonstrated concept, robust fabrication procedure, and an optimized design allowing us to quickly develop and demonstrate this technology in Phase II and beyond.