Optical neuromorphic computing via temporal up-sampling and trainable encoding on a telecom device platform

Area of Science:

• Neuromorphic computing
• Optical signal processing
• Photonic devices

Background:

• Mapping input signals to high-dimensional spaces is crucial for neuromorphic computing models like reservoir computing (RC) and extreme learning machines (ELM).
• Existing methods often require complex or specialized hardware.
• High-speed optical data transmission technologies offer potential for efficient signal manipulation.

Purpose of the Study:

• To propose and demonstrate a novel approach for implementing RC and ELM using commercially available telecom optical devices.
• To leverage nonlinear optical mapping for efficient signal processing in high-dimensional spaces.
• To investigate the impact of temporal up-sampling and wave-division multiplexing (WDM) on feature space manipulation.

Main Methods:

• Utilizing telecom devices like semiconductor optical amplifiers and nonlinear Mach-Zehnder interferometers (MZI) for nonlinear optical signal mapping.
• Implementing temporal up-sampling with a trainable encoding mask to increase the feature space's representational capacity.
• Applying wave-division multiplexing (WDM) to manipulate the output feature dimension.

Main Results:

• Demonstrated the feasibility of using readily available photonic devices for RC and ELM.
• Showcased the effectiveness of dynamical phase masking for flexible input signal manipulation.
• Characterized the nonlinear mapping in terms of enhanced controlled separability and predictability of the output state.

Conclusions:

• The proposed method offers a flexible and efficient way to implement neuromorphic computing models using existing telecom infrastructure.
• Photonic devices provide a powerful platform for creating high-dimensional feature spaces essential for advanced AI.
• This approach paves the way for scalable and cost-effective neuromorphic hardware.