Optical multi-task learning using a multifunctional diffractive processor with "learned" structured illumination.

Area of Science:

• Optics
• Machine Learning
• Artificial Intelligence

Background:

• Diffractive deep neural networks (D²NNs) offer high inference speed and low power consumption for optical machine learning.
• Traditional D²NN designs lack flexibility for diverse computational tasks.

Purpose of the Study:

• To introduce a novel D²NN framework enabling multifunctional applications through learned structured illumination.
• To enhance the practicality and versatility of optical neural networks.

Main Methods:

• Incorporating illumination patterns as trainable parameters within the D²NN optimization process.
• Developing two training approaches: transfer learning and multi-input/multi-output (MIMO) networks.
• Testing the framework on MNIST, Fashion-MNIST classification, and computational imaging tasks.

Main Results:

• Demonstrated a D²NN framework capable of switching between different machine learning tasks using the same diffractive layers.
• Achieved successful classification and computational imaging results with the proposed multifunctional D²NN.
• Validated the effectiveness of both transfer learning and MIMO training strategies.

Conclusions:

• The proposed D²NN framework significantly improves the flexibility of optical neural networks.
• This work establishes a new pathway for creating versatile, multifunctional optical machine learning platforms.
• The learned structured illumination approach enhances the adaptability of diffractive optical computing.