Digital Modeling on Large Kernel Metamaterial Neural Network.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Materials Science

Background:

• Deep neural networks (DNNs) face computational burdens, latency, and power consumption issues, especially in edge computing and IoT.
• Metamaterial neural networks (MNNs) offer light-speed, energy-free computation but are limited by fabrication precision, noise, and bandwidth.
• Standard MNN designs do not fully leverage their potential due to limitations with conventional convolution kernels.

Purpose of the Study:

• To propose a novel large kernel metamaterial neural network (LMNN) that addresses the physical limitations of MNNs.
• To maximize the digital capacity and learning capabilities of MNNs by incorporating model re-parametrization and network compression.
• To explicitly consider and model the optical limitations of meta-optics within the digital learning scheme.

Main Methods:

• Developed a novel large kernel metamaterial neural network (LMNN) architecture.
• Implemented model re-parametrization and network compression techniques to enhance digital capacity.
• Designed a digital learning scheme that accounts for the physical constraints of meta-optic hardware.

Main Results:

• The proposed LMNN effectively offloads convolutional computations to fabricated optical hardware.
• Experimental results show improved classification accuracy on public datasets.
• Demonstrated a significant reduction in computational latency compared to existing methods.

Conclusions:

• The LMNN represents a significant advancement in optical neural networks, bridging digital learning with optical hardware constraints.
• This hybrid approach optimizes performance by maximizing MNN potential while respecting physical limitations.
• The LMNN is a promising step towards achieving energy-free, light-speed artificial intelligence for demanding applications.