Remote training of a reservoir computer via digital twins

Area of Science:

• Optoelectronics
• Computational Science
• Artificial Intelligence

Background:

• Increasing energy demands of traditional information processing.
• Potential of optical and optoelectronic reservoir computing for low-energy applications.
• Challenges in training reservoir computing due to computational costs and matrix operations.

Purpose of the Study:

• To propose and validate a remote training approach for reservoir computing using digital twins.
• To enable continuous inference on physical reservoirs without task-specific interruptions.
• To reduce the computational complexity associated with training reservoir output weights.

Main Methods:

• Development of two digital twin models: differential equation-based and deep neural network (DNN).
• Implementation of a remote training strategy utilizing these digital twins.
• Validation using experimental data from an optoelectronic reservoir for a time-series prediction task (Santa Fe laser).

Main Results:

• Both digital twin models successfully replicated the optoelectronic reservoir's dynamics.
• Accurate predictions were achieved, and weights were transferable from digital twins to the physical reservoir.
• The equation-based model showed higher prediction accuracy; the DNN model exhibited better hyperparameter robustness.

Conclusions:

• Digital twins provide an effective method for remote training of reservoir computing systems.
• The proposed approach enables continuous operation of physical reservoirs for inference.
• This facilitates more efficient and less complex training of optical and optoelectronic computing systems.