Trainable dynamical masking for readout-free optical computing.

Area of Science:

• Leveraging principles of nonlinear systems and optical communication technologies for advanced computation.

Background:

• Nonlinear systems transform input signals into high-dimensional feature spaces, enabling non-conventional computing.
• Traditional training methods require modifying software coefficients, unlike the parameter changes needed in physical systems.

Purpose of the Study:

• To introduce a trainable dynamical mask using off-the-shelf optical communication devices.
• To implement this mask within extreme learning machine (ELM) frameworks for enhanced computational capabilities.

Main Methods:

• Utilizing high-speed optical communication devices and technologies to construct a dynamical mask.
• Integrating the mask as an alternative or supplement to traditional readout layers in ELM.
• Conducting numerical simulations to validate the approach.

Main Results:

• Demonstrated computational potential through successful regression tasks.
• Validated effectiveness in time series prediction simulations.
• Showcased the feasibility of using optical hardware for trainable computational components.

Conclusions:

• The proposed optical dynamical mask offers a viable alternative for ELM-based non-conventional computing.
• This approach effectively addresses the need for parameter modification in physical systems for training.
• The method shows promise for advancing machine learning hardware and computational tasks.