An optoacoustic field-programmable perceptron for recurrent neural networks

Area of Science:

• Photonics
• Optical Computing
• Artificial Intelligence

Background:

• Recurrent neural networks (RNNs) excel at processing sequential data but face limitations in tracking internal states.
• Photonic implementations of neural networks offer potential advantages, yet bidirectional optical RNNs require robust short-term memory solutions.

Purpose of the Study:

• To experimentally demonstrate an optoacoustic recurrent operator (OREO) for bidirectional optical RNNs.
• To address the challenges of programmable coherent computation, minimized noise, and scalability in optical RNNs.

Main Methods:

• Developed and experimentally validated an optoacoustic recurrent operator (OREO).
• Utilized acoustic waves to contextualize information from optical pulse sequences.
• Implemented all-optical control for pulse-by-pulse reconfigurability.

Main Results:

• OREO successfully contextualizes optical pulse sequence information via acoustic waves.
• Demonstrated OREO's capabilities in implementing recurrent drop-out and pattern recognition for 27 optical pulse patterns.
• Achieved programmable, coherent computation with minimized added noise and scalability.

Conclusions:

• OREO provides a viable solution for short-term memory in bidirectional optical RNNs.
• Introduced OREO as a bi-directional perceptron, paving the way for new optical neural network architectures.