Training Passive Photonic Reservoirs With Integrated Optical Readout

Area of Science:

• Photonics
• Neuromorphic Computing
• Optical Signal Processing

Background:

• Moore's Law is ending, driving demand for novel computing paradigms.
• Neuromorphic computing, particularly passive photonic reservoir computing, shows promise for high-bitrate and low-energy computation.
• Current limitations stem from electrical domain training and readout, hindering fully optical operation.

Purpose of the Study:

• To develop a training algorithm for integrated optical readouts in photonic reservoir computing.
• To enable all-optical domain operations, overcoming current electrical domain bottlenecks.
• To address the challenge of unobservable on-chip reservoir states during training.

Main Methods:

• Investigated several training algorithm options for photonic reservoir computing.
• Proposed a novel algorithm allowing observation of complex reservoir states via readout weight configuration.
• Utilized numerical simulations to compare the proposed method against an ideal observable baseline and CMA-ES.

Main Results:

• The proposed algorithm successfully enables observation of reservoir states.
• Demonstrated effective training for optical readouts without direct state observability.
• Numerical simulations validated the performance against established methods.

Conclusions:

• The developed training algorithm facilitates fully optical photonic reservoir computing.
• This approach overcomes key limitations, paving the way for advanced optical computing.
• Further research into optical readout design and training algorithms is crucial for realizing the full potential of photonic computing.