Polarization-multiplexed diffractive neural networks for multi-task classification based on liquid crystals.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Materials Science

Background:

• All-optical diffractive neural networks (DNNs) offer high speed and low power but typically handle single tasks.
• Existing DNN frameworks lack the flexibility for multi-task operations in complex AI systems.

Purpose of the Study:

• To propose a novel polarization-multiplexed diffractive neural network (PMDNN) for flexible multi-task classification.
• To enhance the multi-task processing capacity of DNNs using liquid crystals and polarization multiplexing.

Main Methods:

• Incorporated liquid crystals into DNNs as Jones matrices for polarization modulation.
• Encoded multi-task inputs into multiple polarization channels.
• Designed and demonstrated a 3-task PMDNN for classifying MNIST, Fashion-MNIST, and KMNIST datasets.

Main Results:

• Consistent simulation and experimental results verified the effectiveness of the PMDNN framework.
• The proposed network demonstrated minimal inter-task crosstalk even with increasing task complexity.
• Successfully designed PMDNNs for classifying multiple datasets, including more complex tasks.

Conclusions:

• The PMDNN architecture significantly enhances the flexibility of diffractive neural networks.
• This approach paves the way for ultra-fast, low-power, multi-task integrated AI systems.
• Liquid crystal-based polarization multiplexing offers a viable route for advanced AI hardware.