Free-space real-time training of a single-layer optical neural network on a phase light modulator with quantized...

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Materials Science

Background:

• Optical neural networks (ONNs) offer potential for faster computation but face challenges with training speed and hardware limitations.
• Existing ONNs often use liquid-crystal spatial light modulators (LC-SLMs), which are slower and more expensive.
• Micro electromechanical system (MEMS)-based phase light modulators (PLMs) present a cost-effective and high-speed alternative.

Purpose of the Study:

• To demonstrate a single-layer, feed-forward free-space ONN for image classification using MEMS-PLMs.
• To address the challenges of low phase resolution and non-uniform quantization in PLMs for ONN optimization.
• To significantly reduce the in-situ model-free training time of ONNs.

Main Methods:

• Utilized a mass-producible MEMS-based PLM for the ONN architecture.
• Employed an evolutionary strategy with a discrete categorical distribution for ONN optimization, contrasting with continuous distributions used for LC-SLMs.
• Systematically investigated the impact of phase quantization levels on ONN performance.

Main Results:

• Achieved comparable image classification performance to ONNs based on LC-SLMs.
• Demonstrated an approximately 10-fold reduction in in-situ training time.
• Showed that discrete categorical policies are robust to low phase quantization levels, unlike continuous policies.

Conclusions:

• MEMS-PLMs are a viable and efficient alternative for building practical ONNs.
• Discrete optimization strategies are crucial for overcoming PLM phase quantization challenges.
• The developed ONN architecture offers a significant speed advantage for real-world applications, especially in dynamic environments.