Bidirectional Optical Neural Networks Based on Free-Space Optics Using Lens Arrays and Spatial Light Modulator

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Neural Network Architectures

Background:

• Traditional artificial neural networks (ANNs) primarily utilize forward connections, limiting their efficiency in supervised learning algorithms.
• Implementing backward connections, essential for algorithms like backpropagation, often requires complex electronic circuitry.
• Optical neural networks offer potential for high-speed computation but typically lack inherent bidirectional capabilities.

Purpose of the Study:

• To introduce a novel bidirectional optical neural network (BONN) architecture.
• To enable efficient implementation of algorithms requiring backward connections, such as backpropagation.
• To explore BONN's potential for enhanced supervised learning and hardware compactness.

Main Methods:

• Development of a novel BONN architecture incorporating laser diodes and photodiodes.
• Utilizing Köhler illumination to establish optical channels for backward signal propagation.
• Investigating BONN's scaling limits, throughput, and potential for multilayer emulation using cascaded extensions and clustering techniques.

Main Results:

• BONN architecture successfully provides bidirectional functionality for ANNs.
• Achieved a scaling limit of 96x96 for input/output arrays and a throughput of 8.5 x 10^15 MAC/s.
• Demonstrated potential for overcoming scaling limits and achieving full interconnections via clustering techniques.

Conclusions:

• BONN architecture offers a promising solution for implementing bidirectional ANNs.
• The bidirectional nature of BONN enhances supervised learning and increases hardware compactness.
• Further development, including fast spatial light modulators, is needed for optimal backpropagation algorithm throughput scaling.